Simulating Ulcerative Colitis Clinical Trials Using Knowledge Graph–Enhanced Real-World Data Modeling: Validation Across 21 Studies

Simulating Ulcerative Colitis Clinical Trials Using Knowledge Graph–Enhanced Real-World Data Modeling: Validation Across 21 Studies | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Simulating Ulcerative Colitis Clinical Trials Using Knowledge Graph–Enhanced Real-World Data Modeling: Validation Across 21 Studies Flavio Dormont, Michael Shapiro, Amichai Perlman, Amina Alaskarov, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9272163/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Clinical trials are costly, time-consuming, and often yield uncertain outcomes. Predictive modeling using machine learning (ML), real-world data (RWD), and biomedical knowledge graphs offers new opportunities to improve translational efficiency and trial outcomes. Methods This study validated an ML-based modeling framework, ClinBoost , which integrates de-identified RWD from insurance claims and electronic health records with a drug-centered biomedical knowledge graph. Patient severity was modeled using an electronic Mayo Score, and longitudinal data inputs were transformed into patient-drug journey embeddings. Model performance was evaluated against 21 randomized controlled trials using a robust validation framework with multiple performance metrics. Results ClinBoost a achieved high concordance with outcomes for 16 of the 21 trials, achieving an F1 score of 80%. The model demonstrated minimal bias in estimating treatment effect sizes. Additionally, the framework was able to refine trial design by identifying sub-cohorts with better response and increased statistical power. Conclusions These findings demonstrate the potential of the ClinBoost modeling approach, which combines RWD and knowledge graph-informed ML, to improve clinical trial outcomes, optimize study design, and accelerate drug development. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Machine learning real-world data knowledge graphs graph neural networks predictive modeling Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction The development of new therapeutic agents is resource-intensive and often hindered by clinical trial challenges. Traditional randomized controlled trials, despite being considered the gold standard for evaluating drug efficacy and safety, are limited by high costs, lengthy durations, and recruitment difficulties [1-5]. Predictive modeling has sought to address these challenges by shifting from traditional in vitro and in vivo studies to advanced in silico simulations, improving translational efficiency and accelerating drug discovery [6-8]. Computational in silico models[9, 10] can simulate biological processes and help to predict drug responses, providing a cost-effective and efficient exploration of new therapeutic mechanisms of action (MoA). These models help identify promising compounds based on their structure and anticipated interactions with other target biomarkers, simulate pharmacokinetic and pharmacodynamic behaviors, optimize dosing, and detect potential adverse effects early [11-13]. Such approaches can markedly reduce reliance on extensive in vivo studies and expedite drug candidate identification [7, 8] yet their application to clinical-trial modeling and protocol optimization remains limited. Recent legislative and technological advances, such as the European Health Data Space (EHDS) and the US HITECH and Cures Acts, have facilitated the integration of clinical real-world data (RWD) in drug development. Coupled with advanced machine learning (ML) methods [14], RWD-based models offer dynamic time-dependent predictions of patients’ trajectories, crucial for management and decision-making in progressive or fluctuating diseases [15]. These models leverage sophisticated ML techniques, enhancing predictions by parsing complex patient data and the causal frameworks influencing treatment response [16, 17]. Integration of knowledge graphs, encoding biological relationships further enhances ML-driven predictive modeling. Knowledge graphs enable the representation of complex and interconnected biomedical processes in a way amenable to machine learning frameworks [18, 19]. This facilitates discovering novel drug-target interactions, repositioning of existing drug, and enhancing the causal inference of clinical outcomes [12, 20]. The integration of real-world data and knowledge graph–driven modeling holds particular promise for predicting clinical outcomes and optimizing trial protocols. Building on these advances, we developed ClinBoost, a modeling framework that combines RWD and knowledge graph drug embeddings to simulate clinical trial outcomes and guide trial design. The unique capabilities of ClinBoost enable scalable trial designs by modeling clinical responses using endpoints aligned with those used in prospective trials. It supports comparative simulations between investigational therapies, standard-of-care, and competitors, including scenarios absent from historical data such as first-in-class or repurposed drugs. The framework can simulate diverse treatment settings, including placebo, monotherapies, and combination therapies (regimens), while capturing drug–patient interactions to enable subgroup-level predictions. It also supports protocol optimization through flexible evaluation of cohorts, treatments, and endpoints. Although predictive models of drug efficacy based on RWD have advanced considerably, few studies subject them to rigorous validation, and those that do typically include only one or two comparisons, limiting generalizability. To address this gap, we applied a robust validation framework to compare model predictions with outcomes from 21 clinical trials assessing 13 different drugs, through rigorous statistical and Bland–Altman analyses to ensure technical reliability and clinical relevance in line with contemporary regulatory standards [21]. Our approach assessed not only binary trial outcomes (e.g., success vs. failure), but also continuous treatment effect estimates by precisely simulating trial conditions. This included matching clinical endpoints, eligibility criteria, time points (e.g., 12 weeks, 24 weeks) and sample sizes to replicate trial designs using real-world patient cohorts. In this study, we demonstrated the practical applicability of our ML process by developing and validating a RWD predictive drug response model using ulcerative colitis (UC) as a use case. UC is a chronic inflammatory disease with significant unmet medical needs and where patient heterogeneity hampers drug development efforts [22, 23]. By integrating ML, RWD, and knowledge graph methodologies, we demonstrate accurate predictions of drug efficacy against past clinical trials and demonstrate how such modeling frameworks can be applied to optimize future trial design, thereby optimizing recruited patient populations, increasing the probability of technical success, and accelerating therapeutic development. Methods Data Sources Real-world Clinical Data For modeling patient outcomes, we used deidentified patient-event level data from the PurpleLab® insurance open claims database and electronic health records (EHR) from Eversana’s Electronic Health Record database (EVERSANA Life Sciences Inc.), which were linked via a third-party tokenization service (Datavant Inc., San Francisco, California). The PurpleLab® open claims database covers approximately 350 million patients representing 98% of healthcare claims filed in the United States between 2014 and 2023. The database contains anonymized information on surgeries, procedures, medications, diagnoses, and various healthcare encounters, and includes comprehensive demographic and death information. The Eversana EHR dataset covers 120 million patients who were treated primarily in the community setting in the United States and includes data on medical procedures, diagnoses, healthcare encounters, medications, laboratory results, vitals, and socio-demographic factors. The contents of the linked dataset used in this study were standardized to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) prior to model development. Bio-pharmacological Data for Knowledge Graph Construction A comprehensive biomedical dataset compiled from a combination of open-source and proprietary data sources was used to build a biomedical knowledge graph. This dataset contained information on a wide range of small molecules and biologics and their pharmacological profiles, namely pharmacokinetics/-dynamics (PK/PD), absorption, metabolism, excretion, toxicity, mechanism of action, cellular targets and known drug-drug interactions. We also included data for relevant biological processes, diseases, anatomical structures, genes, biological pathways, and molecular functions affected by these agents. This knowledge graph consists of approximately 100,000 nodes representing biomedical entities and over 5 million edges representing the relationship between these entities. To represent novel drugs lacking mechanistic information in the bio-pharmacological dataset, we manually enriched the knowledge graph by adding connections to newly created compound nodes, based on published preclinical studies. This included, for example, their mechanism of action, additional indications for the drug under investigation, and pharmacodynamics components such as metabolism and excretion. For this project, we ensured that all medications indicated for UC patients were thoroughly represented in the graph, and we enriched the graph further whenever necessary, as described below. Framework for Model Training and Prediction Overall Cohort Definition The study cohort included patients with a confirmed diagnosis of UC. To emulate this criterion using RWD, patients eligible for the cohort were required to have at least three documented diagnosis records of ulcerative colitis and at least one prescription for a UC-related medication. The disease start date was defined as the earliest recorded date of a diagnosis of UC. For the training set, patients were permitted to enter the training process multiple times. Entry into the cohort was based on the first occurrence of the disease being classified as mild, moderate, or severe after the disease start date defined above, thus allowing for up to 3 separate cohort entries per patient. The classification of severity was done using the electronic Mayo score depicted in the next section. This overall cohort was used to train the machine learning model (described in the following section) to be able to predict the joint effect of patient features and treatments on patient outcomes. However, when predicting the outcomes of the clinical trials in the validation set, we used a distinct cohort for each trial, consisting only of patients who met the eligibility criteria for that specific trial. Importantly, during the prediction process, we used the patient characteristics and assign to them the intervention or control arm drugs, rather than using the actual drugs they received in the RWD (described further in the section Prediction of RCT outcomes, below). Engineering of Patient Outcomes – EHR-based Mayo Score (eMS) The Mayo Score (MS) is a clinical measure of UC activity that is extensively used to evaluate treatment response. The MS has four components: stool frequency, rectal bleeding, endoscopic findings of mucosal inflammation, and physician’s global assessment. The partial MS (pMS) composed of stool frequency, bleeding components and physician assessment on a 9-point score has been shown to perform on par with MS in measuring response to UC therapy [24, 25]. It has also been used in several clinical trials to monitor responses [23, 26-29]. To model UC disease severity outcomes based on the MS, we constructed a retrospective EHR-based Mayo score (eMS) composed of elements that can be found in Table 1 . The weights contributing to the eMS were determined through a combination of expert consultation and empirical calibration. A gastroenterologist specializing in inflammatory bowel disease initialized the weights to reflect clinical assessment of severity. To further calibrate the weights, monthly eMS scores were generated by summing the weighted records and compared to Mayo score distributions from both the literature and curated datasets containing actual scores. Table 1: Retrospective EHR-based Mayo Score Element Record Weights Abdominal pain 1 Oral steroid usage 1 Frequent endoscopy procedures (>1 procedure within 6 months) 1 Anemia due to blood loss 2 Hospitalization due to UC 2 Diarrhea 2 Rectal bleeding 3 The total eMS score ranges from 0 to 12. A 0-12 scale was first employed and subsequently harmonized to a 0-9 range for consistency with the pMS scoring framework ( Supplementary Figure S1 ). Our validation results using curated data showed a high concordance between the electronic Mayo Score (eMS) and pMS ( Supplementary Figure S1 ). For this study, the harmonized eMS scoring system was categorized into three disease severity levels: mild (eMS 0-2), moderate (3-6), and severe (7-9). The eMS for each patient were calculated on a monthly basis. Disease progression was defined as an increase in eMS by one point in a given month relative to the previous 3-month window. This definition of progression was validated against published retrospective cohort analyses of disease relapses among patients with moderate or severe UC. The eMS score successfully captured UC disease course over time and was internally consistent and concordant with previously published pMS estimates. The eMS’s capacity to monitor disease progression was validated through the analysis of the following elements: absolute progression rates relative to literature, relapse rates 1 year after diagnosis, and progression by neutrophil/lymphocyte ratio (NLR) and C-reactive protein (CRP) strata.[22] Results from these validations of the eMS can be seen on Supplementary Figure S1 . The predictive model, described further below, was trained to predict progression in patients with UC. Engineering of Patient Outcomes – Progression and Remission As clinical visits in RWD often reflect episodes of heightened disease activity, they provide a granular view of disease progression and exacerbation, while episodes of low activity in the data can result from missing data rather than “real” remission. Thus, disease progression events are more directly measurable than remission in real-world settings. We developed a two-stage model to predict remission in line with clinical trial benchmarks, despite the lack of directly observed remission in RWD. First, as described above, we trained our machine learning model to predict UC progression over time using real-world clinical data. Second, we developed an additional model to infer expected patient remission rates in clinical trials based on the predicted treatment effect on UC progression. To do this, we constructed a reference sigmoidal remission curve over time using reported placebo remission rates from clinical trials involving patients with moderate to severe UC, extracted at time points from 5 to 52 weeks post-randomization. We then created a transformation function that maps this reference curve to predicted placebo progression rates, capturing the inverse relationship between progression and remission in UC. Finally, we applied this function to the model’s predicted treatment effect on progression (as described below) to estimate corresponding remission rates. Drug-centered Knowledge Graph Embeddings The knowledge graph, constructed from a combination of bio-pharmacological data and manual enrichment, was converted to graph embeddings for model training. We tested several embedding methods, including TransE, TransR, MURE, metapath2vec and graph neural networks, and evaluated the quality of embeddings using a scoring metric that assessed the relative cosine similarity of drugs with similar versus dissimilar mechanisms of action. The embeddings that achieved the highest evaluation score were then used for model training. Knowledge Graph Enrichment for Novel Drug Representation To enable modeling of treatment effects for novel drugs not represented in the knowledge graph (KG), a process termed KG enrichment is applied. In this process, relevant information is extracted from the literature, including mechanisms of action, drug-drug interactions, molecular targets, side effects, and approved or investigated indications. This information is structured to match the KG format and used to construct a corresponding subgraph. The subgraph is then integrated into the main KG, and updated drug embeddings are generated to include the novel drug, relevant standards of care (SoC), and comparators. These embeddings are subsequently used to train models that predict the effect of the novel drug in comparison to SoC and alternative treatments, leveraging both the newly integrated data and the existing KG context. Throughout the study, such novel drugs whose information was added using this mechanism are referred to as enriched drugs. Patient-Drug Journey Representation The index date for representing patient-drug journey within the predictive model was defined as the first instance when the patient’s disease was classified as mild, moderate, or severe during the course of the patient’s condition, occurring after the UC start date specified in the cohort definition paragraph above. “Patient journey” was defined as the time series consisting of clinical events recorded in their real-world clinical data, starting six months prior to their index date, the end date was determined according to the corresponding clinical trial matching date – therefore allowing to predict remission over different period durations. Data for patient demographics, current and prior treatments, evidence of hospitalization within the last year, high levels of inflammatory markers (namely, C-reactive protein, erythrocyte sedimentation rate and fecal calprotectin levels) were also included as predictors. This data for patient-drug journey was converted into a vector embedding that was then used in the prediction model. Temporal Prediction Framework – Model Training and Hyperparameter Optimization We developed the ClinBoost algorithm that combines patient-drug journey embeddings and the knowledge graph embedding into the XGBoost algorithm, followed by per-time-window logistic regression. This algorithm includes trainable weights to combine patient clinical data, demographics, and drug embeddings into a single model to predict outcome probabilities for each patient for each time window. This model was trained using real-world clinical data and graph embeddings, with a train/validation/test split of 80:10:10 used for tuning model hyperparameters. For hyperparameter optimization, we considered a three-dimensional grid of hyperparameters covering tunable components of (1) drug-centered knowledge graph embedding, (2) clinical data definitions and filters for selection of the overall cohort used for model training, and (3) the prediction model hyperparameters. For hyperparameter optimization, we evaluated the area under the receiver operating characteristics curve (ROC-AUC) or c -statistics by comparing the predicted patient outcomes to those observed in the validation set of the RWD. After selecting and validating the best hyperparameters using the RWD we tested the best model on a set of validation clinical trials. Model Validation using Randomized Controlled Trial (RCT) Data Selection of RCTs for Validation Information relating to study design and results of the clinical trials used in this study for validation was sourced from Clinicaltrials.gov and PubMed (pubmed.ncbi.nlm.nih.gov). Clinical trials were included if they were phase II or phase III RCTs evaluating a pharmacological treatment for moderate and/or severe ulcerative colitis compared to placebo or standard of care. Trials had to be registered at clinicaltrials.gov, with well-documented inclusion and exclusion criteria that could be modeled using RWD; and to report results on disease remission. Trials without published results (accessible at clinicaltrials.gov or pubmed.ncbi.nlm.nih.gov) were excluded. The 21 trials ultimately included for validation spanned 2007 to 2022. The following elements were extracted for each trial for the purposes of modeling and validation: inclusion and exclusion criteria, investigational and comparator treatments, endpoint definitions, sample sizes per treatment arm, baseline patient characteristics, and trial results including point estimates and confidence intervals for the relevant endpoints and treatment effect estimates, as well as reported p-values. Prediction of RCT Outcomes For prediction, eligibility criteria for each validation RCT were first applied to select an eligible population of patients from the study’s real-world dataset. For each trial, all clinical parameters that could be replicated in RWD were applied; a full list of these parameters is supplied in the supplementary materials ( Appendix 1 ). Patient cohorts with age and sex distributions, and sample sizes matching those from the corresponding RCT were then sampled to create the prediction cohort. The trained prediction model was used to predict patient-level outcomes for the prediction cohort separately under the treatment and control regimen from the respective trial, regardless of what treatment the patients may have received previously in the real-world clinical data. The prediction of remission rates was performed at the same time point as the corresponding RCT to enable comparability. For subgroup-level simulations, the set of patients sampled were based on additional eligibility criteria for a variety of sample sizes, as required. Performance Metrics for Model Validation To measure the trial-level validation performance of our model, we evaluated the following criteria: Trial success agreement: A binary analysis examining the accuracy of predicting trial success or failure based on a direction and statistical significance of effect, which is estimated using confidence intervals and p -values. In other words, the predicted trial should reach the same conclusion as the original RCT. We report four levels of agreement: Concordant Alignment: The prediction aligns with the original trial in both direction (treatment better than control) and statistical significance (p < 0.05). Directionally Aligned with Threshold Misalignment: Both the original trial and the predicted trial outcome indicate the same direction of effect (treatment better than control for instance), but the prediction fails to reach statistical significance while the original trial was statistically significant. Paradoxical Agreement: Both the original and emulated trials are statistically non-significant (p > 0.05), but the direction of the effect differs (treatment better in one, control better in the other). Complete Divergence: The original trial shows a statistically significant positive outcome for the treatment (treatment better than control), while the prediction indicates a statistically significant negative outcome (control better than treatment). Absolute outcome agreement: This was achieved when the estimate from the predicted RWE study fell within the 95% confidence interval of the corresponding observed RCT estimate. The results were presented visually as forest plots and a Bland-Altman plot. Cohen’s d statistic [30]: The standardized difference between observed and predicted effect sizes was calculated using Cohen’s d, which expresses the magnitude of difference in standard deviation units. Values of 0.2, 0.5, and 0.8 were interpreted as small, moderate, and large differences, respectively. This metric enabled comparison of predictive accuracy across trials with varying populations and measurement methodologies. The average Cohen’s d across all validation trials was calculated to summarize overall model performance. Trial success and failure across 21 trials was evaluated using a confusion matrix and quantified by standard metrics including F1 score, sensitivity, and specificity. Additionally, model performance was also evaluated on trial strata by drug mechanism of action (e.g., biological versus non-biological). Model-guided Trial Design Optimization. To evaluate the utility of the multivariate prediction engine in informing clinical trial design, a series of simulation analyses were conducted. These analyses aimed to identify patient sub-cohorts with differential treatment responses, optimize eligibility criteria, and assess power and sample size requirements for achieving statistically significant results. Cohort Variant Generation and Effect Prediction: The analysis began with a base cohort derived from an existing clinical trial population. Using a multivariate predictive modeling framework, multiple cohort variants were generated by systematically modifying inclusion and exclusion criteria. These criteria included prior exposure to biological therapies (e.g., anti-TNFs), baseline inflammatory marker levels, and systemic steroid use. For each variant, the expected treatment effect, defined as the difference in remission rates between treatment and control arms at a pre-specified time point, was predicted, and the total cohort size was estimated by extrapolating to a representative real-world population. Predicted treatment effects and corresponding cohort sizes were visualized in a scatter plot. Gray markers indicated cohort variants unlikely to achieve statistical significance given the original trial’s sample size. From the set of predictions, two cohorts of interest were selected for in-depth comparison: an “Optimized” cohort with enhanced predicted response, and a “Low Effect” cohort with diminished efficacy. Longitudinal Outcome Projection: Weekly remission rates over a 52-week period were predicted for each selected cohort using the model’s time-to-event outputs. These projections enabled longitudinal comparison of remission dynamics across treatment arms and cohorts. Power and Sample Size Estimation: An analysis was performed to evaluate the total sample size needed to reach statistical significance, while maintaining the original treatment-to-control allocation ratio. For each cohort, p-values corresponding to the predicted treatment effect were calculated, and the minimum sample size required to achieve statistical significance (α = 0.05) was determined. Monte Carlo Simulation for Probability of Success: To assess variability and robustness in trial outcomes, Monte Carlo simulations were conducted. Patient profiles meeting the inclusion/exclusion criteria for each cohort were repeatedly sampled. For each simulation, the predicted treatment effect was calculated and aggregated into a distribution. The probability of technical success (PTS) was defined as the proportion of simulations in which the treatment effect exceeded a predefined threshold (e.g., 10% remission difference between treatment and control). Resulting effect-size distributions and PTS estimates were visualized. Results Modeling Framework Overview Figure 1 provides an overview of the predictive model architecture and workflow used in the study. The model was trained using data from a bio-pharmacological database in the form of a knowledge graph and longitudinal real-world clinical data from a linked claims/electronic health records database of patients with UC. Embeddings of the knowledge graph and patient-drug profiles from these two data sources were combined into a gradient-boosted tree model to predict patient-level outcomes on therapy over time. The study cohort included 482,870 unique patients, which could potentially be recruited up to 3 times in relation to a change in their clinical status. The median age of patients in this cohort was 56 years (ranging 18 to 85 years), 59.5% of whom were women. For the model training and predictions, patients were permitted to enter the simulation multiple times. The demographic, clinical history, and treatment characteristics of this overall study cohort, including the multiple entries (n = 723,263) used for model development, including outcome rates, are provided in Supplementary Table S1 . The drug-centered knowledge graph and derived drug embeddings are shown in Figure 2 . The graph included both approved therapies represented in the real-world clinical data, as well as novel therapies with known mechanisms of action which were manually added to the graph ( Figure 2, a ). The quality of the knowledge graph embeddings was assessed by comparing the cosine similarity of drugs sharing similar mechanisms of action and therapeutic indications. Figure 2b illustrates a two-dimensional t-SNE visualization of the compound embeddings, color-coded by drug groups. We found that the learned embeddings preserve the underlying graph topology (Figure 2c), and that agents with similar mechanisms of action within the graph embedding clustered together preferentially and demonstrated high cosine similarity, as expected ( Figure 2, b and d ). Model Validation Results Table 2 provides an overview of the 21 randomized controlled trials that were included for the validation of our predictive model. Of these 21, 7 (33.3%) were unsuccessful trials, 5 (23.8%) were phase 2 trials and 6 (29%) included a novel agent that was not commonly administered in the real-world at the time of this study. The trials also encompassed a range of sample sizes (from 28 to 476 patients per arm) and varied timings of outcome measurements (6 to 52 weeks post-randomization). The treatments included diverse mechanisms of action such as biologics (e.g., tumor necrosis factor (TNF) inhibitors) and small molecule drugs ( Table 2 ). Therefore, we expected this set of 21 trials to provide a sufficiently diverse cohort for evaluating our machine learning model’s ability to predict remission rate per study arm, and the treatment effects (i.e., difference in outcomes between treatment arms) . Trial Number Trial Short Name NCT ID Success/ Failure (1/0) Time of outcome Measurement (weeks) Treatment Arm Treatment Arm Remission (% Patients) Control Arm Control Arm Remission (% Patients) Treatment Arm Target / MoA Treatment /Control Arm Size (N) Age (mean±std) Study phase Novel dug? 1 ELEVATE UC 12 NCT03996369 1 12 Etrasimod 24.8 Placebo 15.2 S1PR1 agonist 222/112 40 (± 13) 3 1 2 NCT03934216 NCT03934216 0 12 Deucravacitinib 14.8 Placebo 16.3 selective TYK2 inhibitor 88/43 42 (± 15) 2 1 3 U-Accomplish NCT03653026 1 8 Upadacitinib 33.5 Placebo 4.1 JAK inhibitor 341/174 42 (± 15) 3 1 4 NCT03482635 NCT03482635 0 12 Spesolimab 7.1 Placebo 0 Anti IL-36 28/23 44 (± 15) 2 1 5 SELECTION NCT02914522 1 10 Filgotinib 26.1 Placebo 15.3 JAK inhibitor 245/137 NA 3 0 6 NCT02819635 NCT02819635 1 8 Upadacitinib 26.1 Placebo 4.8 JAK inhibitor 319/154 44 (± 14) 3 1 7 VARSITY NCT02497469 1 52 Vedolizumab 31.3 Adalimumab 22.5 Integrin blocker/ Anti TNF-alpha 386/383 41 (± 14) 3 0 8 NCT02435992 NCT02435992 1 10 Ozanimod 18.4 Placebo 6 S1PR1 agonist 429/216 NA 3 1 9 NCT02289417 NCT02289417 0 12 Apremilast 21.8 Placebo 12.1 PDE 4 inhibitor 55/58 43 (± 15) 2 0 10 HIBISCUS 2 NCT02171429 0 10 Etrolizumab 18.2 Placebo 11.1 Integrin blocker 143/72 41 (± 14) 3 0 11 HIBISCUS 1(A) NCT02163759 0 10 Etrolizumab 19.4 Adalimumab 22.5 Integrin blocker 144/142 40 (± 13) 3 0 12 HIBISCUS 1(B) NCT02163759 1 10 Etrolizumab 19.4 Placebo 6.9 Integrin blocker 144/72 40 (± 13) 3 0 13 GARDENIA NCT02136069 0 10 Etrolizumab 20.6 Infliximab 32.8 Integrin blocker 199/198 40 (± 15) 3 0 14 HICKORY NCT02100696 1 14 Etrolizumab 18.5 Placebo 6.3 Integrin blocker 384/95 41 (± 13) 3 0 15 NCT01694485 NCT01694485 1 8 Abrilumab 13.4 Placebo 4.4 Integrin blocker 79/116 40 (± 12) 2 0 16 OCTAVE(A) NCT01465763 1 8 Tofacitinib 18.5 Placebo 8.2 JAK inhibitor 476/112 NA 3 0 17 OCTAVE(B) NCT01458951 1 8 Tofacitinib 16.6 Placebo 3.6 JAK inhibitor 429/112 NA 3 0 18 GEMINI 1 NCT00783718 1 6 Vedolizumab 16.9 Placebo 5.4 Integrin blocker 225/149 40 (± 13) 3 0 19 NCT00488774 NCT00488774 0 6 Golimumab 13 Placebo 11 Anti TNF-alpha 77/73 40 (± 14) 2 0 20 PURSUIT-SC NCT00487539 1 6 Golimumab 17.9 Placebo 6.4 Anti TNF-alpha 257/251 41 (± 14) 3 0 21 ACT 1 NCT00096655 1 8 Infliximab 38.8 Placebo 14.9 Anti TNF-alpha 121/121 42 (± 14) 3 0 Table 2: UC Validation Trials and Experimental Drug MoA The 21 trials were chosen to assess model performance in predicting trial outcome. A diverse validation set was composed, including trials with different drug modalities, phase 2 or 3, novel vs. approved drugs, and different study sizes. Prediction cohorts include age and sex distributions similar to original trial, and the prediction of remission rates were done at the actual time point of the trials’ outcome measure and with similar arm size. Model validation results are presented in Table 3 and Figure 3 . Our model correctly predicted the effect size – defined as the difference in remission rates between treatment and control groups – for 16 of the 21 trials (76.2%), correctly matching the direction and statistical significance of the corresponding RCT. These concordant predictions would have led to the same conclusions regarding trial outcome ( Table 3 ). Among the five trials where predictions did not fully align with the observed trial outcomes, four (80%) showed effect sizes in the same direction as the RCTs, but differed in statistical significance, which affected the conclusion. Only one trial, VARSITY (evaluating adalimumab vs vedolizumab), showed a misclassification in the direction of the effect size. This trial had the longest follow-up duration (52 weeks), and the discrepancy stemmed primarily from inaccurate predictions in the control arm. Notably, in another adalimumab trial HIBISCUS-1 (NCT02163759), the model accurately predicted the effect size at week 10, fully consistent with the corresponding RCT ( Figure 3a ). Importantly, for all six trials involving novel agents, the model’s predicted effect sizes aligned with those from the RCTs, supporting the same regulatory decisions. Overall, the model exhibits an F1 score of 0.8 when compared to RCT data on the binary level of success/failure agreement ( Figure 3, b and c ). Furthermore, the trained model performance was evaluated on the patient level real world data, and a high performing model was selected using the ROC-AUC ( Figure 3d ). The model performed slightly better when predicting outcomes at shorter follow-up lengths of time. Nonetheless, the AUC declined by only 7.2% when choosing a simulated follow-up length of 10 days and up to 400 days (AUC 0.83 and 0.77, respectively ( Figure 3d ) . Table 3: Effect Size Estimates and Agreement Metrics Trial N Trial name Treatment Arm Target / MoA Observed Treatment Effect* Predicted Treatment Effect** Difference Agreement 1 ELEVATE UC 12 S1PR1 agonist 0.1 (0.04, 0.15) 0.15 (0.09, 0.2) 0.05 (-0.03, 0.13) 1 2 NCT03934216 selective TYK2 inhibitor -0.01 (-0.1, 0.07) 0.06 (-0.03, 0.14) 0.07 (-0.05, 0.19) 1 3 U-Accomplish JAK inhibitor 0.29 (0.25, 0.34) 0.11 (0.07, 0.15) -0.18 (-0.24, -0.12) 1 4 NCT03482635 Anti IL-36 0.07 (-0.02, 0.15) -0.04 (-0.12, 0.04) -0.11 (-0.21, -0.01) 1 5 SELECTION JAK inhibitor 0.11 (0.05, 0.16) 0.19 (0.13, 0.24) 0.08 (0.00, 0.16) 1 6 NCT02819635 JAK inhibitor 0.21 (0.17, 0.26) 0.12 (0.08, 0.16) -0.09 (-0.15, -0.03) 1 7 VARSITY Integrin blocker/ Anti TNF-alpha 0.09 (0.04, 0.13) -0.15 (-0.19, -0.1) -0.23 (-0.30, -0.17) 0 8 NCT02435992 S1PR1 agonist 0.12 (0.09, 0.16) 0.1 (0.06, 0.14) -0.02 (-0.07, 0.03) 1 9 NCT02289417 PDE 4 inhibitor 0.1 (-0.0, 0.19) 0.25 (0.13, 0.35) 0.15 (0.00, 0.30) 0 10 HIBISCUS 2 Integrin blocker 0.07 (0.0, 0.14) 0.05 (-0.02, 0.12) -0.02 (-0.11, 0.07) 1 11 HIBISCUS 1(A) Integrin blocker -0.03 (-0.1, 0.04) -0.13 (-0.2, -0.07) -0.10 (-0.20, -0.01) 1 12 HIBISCUS 1(B) Integrin blocker 0.12 (0.06, 0.19) 0.05 (-0.02, 0.11) -0.08 (-0.17, 0.01) 0 13 GARDENIA Integrin blocker -0.12 (-0.18, -0.06) -0.12 (-0.18, -0.06) 0.00 (-0.08, 0.09) 1 14 HICKORY Integrin blocker 0.12 (0.08, 0.16) 0.01 (-0.03, 0.04) -0.11 (-0.17, -0.06) 0 15 NCT01694485 Integrin blocker 0.09 (0.03, 0.15) 0.04 (-0.02, 0.1) -0.05 (-0.13, 0.03) 0 16 OCTAVE(A) JAK inhibitor 0.1 (0.06, 0.14) 0.08 (0.04, 0.11) -0.03 (-0.08, 0.02) 1 17 OCTAVE(B) JAK inhibitor 0.13 (0.09, 0.16) 0.08 (0.04, 0.11) -0.05 (-0.10, -0.00) 1 18 GEMINI 1 Integrin blocker 0.11 (0.07, 0.16) 0.04 (0.0, 0.08) -0.07 (-0.13, -0.01) 1 19 NCT00488774 Anti TNF-alpha 0.02 (-0.05, 0.08) 0.06 (-0.0, 0.12) 0.04 (-0.05, 0.13) 1 20 PURSUIT-SC Anti TNF-alpha 0.12 (0.08, 0.15) 0.06 (0.03, 0.09) -0.06 (-0.11, -0.00) 1 21 ACT 1 Anti TNF-alpha 0.24 (0.16, 0.31) 0.11 (0.05, 0.17) -0.13 (-0.22, -0.03) 1 *Observe Treatment Effect is the difference in remission rates between the investigational and control groups of the actual clinical trials; 95% CI is provided around this estimate. **Predicted Treatment Effect is the difference in remission rates between the investigational and control groups in the predicted results; 95% CI is provided around this estimate. Column “Difference” shows the difference between the clinical trial Observed and the Predicted Treatment Effect sizes; 95% CI is provided around this estimate. Column “Agreement” corresponds to the evaluation metrics “Concordant alignment” described in section Performance Metrics for Model Validation of the method section. We also showed visually the difference in point estimates between predicted versus observed treatment effects ( Figure 4 a and b ). The forest plots present the absolute difference between observed and predicted treatment effect, expressed in the original units of measurement (percentage points), with the related 95% CI (Figure 4 a). These differences were relatively small on average, ranging from 0 for the GARDENIA trial to a maximum of –0.23 for the VARSITY trial, with a median difference of -0.05 . The standardized mean difference ( Figure 4b) shows a narrower dispersion To further evaluate model performance, we applied Bland–Altman analysis to quantify agreement between predicted and observed remission outcomes. Figure 4c presents the Bland–Altman plot comparing predicted and observed remission rates at the arm level (treatment and control). The mean difference across arms was minimal (–0.009), indicating no substantial systematic bias, with most trial arms falling within the 95% limits of agreement (LoA: –0.148 to 0.129), suggesting good concordance. However, a trend of increasing prediction error was observed for arms with higher average remission rates (>0.2), indicating that model accuracy may decline slightly in high-response scenarios. Figure 4d shows a Bland–Altman plot comparing predicted versus observed effect sizes. The average difference was –0.045, indicating a slight negative bias, i.e., modest underestimation of treatment effect. Despite this bias, prediction accuracy was consistent across the range of effect sizes, with the majority of values well within the LoA bounds (–0.219 to 0.129), demonstrating stable model behavior across trials with both small and large treatment effects. These analyses confirm that ClinBoost provides accurate and unbiased estimates of both individual arm outcomes and relative treatment effects, with persistent performance across a diverse range of trial conditions Sub-analyses on trial strata were conducted to compare the prediction accuracy according to the type of the administered drugs, (biological products vs. small molecules) and the drug development phase (phase II or III clinical trials). We found that the predictive model performed slightly better on average for trials involving non-biologic agents (F1 score 0.94 versus 0.80 for trials with biologics) and those predicting the results of phase III trials compared to phase II trials (F1 score of 0.90 versus 0.75; Figure 5 ). Furthermore, analysis comparing the prediction accuracy according to the drug’s mechanism of actions showed the best results for JAK inhibitors (5 trials) compared to anti-integrins (5 trials) showing F1 values of 1 and 0.75, respectively; a third group that included other mechanism of actions revealed an F1 score of 0.84 ( Figure 5 ). Finally, the model’s flexibility was evaluated by processing trials involving pharmacological agents that required manual graph integration, and that were withheld from the training RWD dataset. To enable this, the KG was updated to map the mechanistic attributes of these agents, followed by the generation of corresponding embeddings. The subset of trials utilizing this manual enrichment process (6 trials) showed robust alignment with observed results (F1 = 1.0) compared to the standard automated pipeline (15 trials; F1 = 0.80). These findings validate the consistency of the graph embedding technique, demonstrating that mechanistic signal can be effectively preserved even when integrating pharmacological profiles that were not part of the primary training data. Figure 5: Stratified Performance of the Predictive Model by Clinical Trial Design Model-guided Trial Design Optimization A key application of drug response modelling is to inform clinical trial design by predicting responder populations, optimizing eligibility criteria and improving sample size estimation. To demonstrate this, we first evaluated predicted treatment outcomes for the ELEVATE UC-12 trial base cohort, as well as various sub-cohorts after applying additional inclusion and exclusion criteria. ELEVATE UC-12 was selected as a recent clinical trial with modestly positive treatment efficacy findings of a novel MoA (not observed in RWD) and for which enrichment for specific responder subpopulations could have afforded improved clinical outcomes. Treatment effect predictions were plotted alongside cohort size to assess treatment effect heterogeneity within these sub-cohorts, which revealed distinct response subgroups of patients with higher and lower effect sizes than those in the base cohort ( Figure 6 ). Specifically, Figure 6a shows how adding or omitting different combinations of inclusion/exclusion criteria change the results of the trial in terms of treatment effect and sample size (i.e., the available number of patients in the US). The predicted remission rates over time for selected cohorts, including the base cohort, a responding optimized cohort, and a low treatment effect cohort, are presented with their associated p-values over a range of sample sizes ( Figures 6b and 6c) . Comparing the original validation cohort to the optimized sub-cohort in a phase 3 trial setting, we observed a significant improvement in remission rates at 24 weeks post-randomization in our simulations ( Figure 6b ). Additionally, it was possible to estimate the minimum sample size that would be required to achieve statistically significant results using the model ( Figure 6c ). Patient-level simulations across varying arm sizes demonstrated that this optimized sub-cohort required fewer participants to achieve statistical significance, as determined by a standard p-value threshold of 0.05 ( Figure 6c ). Finally, we simulated the distribution of possible trial results for each cohort and computed the probability of technical success (PTS) for a specific threshold of remission difference (in this case 10% more remission in the treatment vs. the control arm). Figure 6d shows in how many simulations the treatment effect exceeded the threshold with PTS variations according to the type of cohort. These findings emphasize the potential of model-guided patient selection in enhancing clinical trial efficiency and success rates. Discussion Our study demonstrates that a predictive model leveraging a combined knowledge graph and RWD-based machine learning framework can effectively generate out-of-sample drug response predictions that align well with clinical trial findings in UC. This is evidenced by the strong convergence between observed and model-predicted trial results, with 16 of 21 comparisons showing an agreement regarding the direction, the effect sizes, and the statistical significance of the results. This robust performance extended across a diverse set of trials encompassing various drug mechanisms (11 small molecules and 10 biologics), phases (Phase II and III), and endpoint duration (6 to 52 weeks), showcasing the model's adaptability. Notably, the model showed high concordance with outcomes for all seven trials involving manually enriched agents, highlighting the utility of the knowledge graph framework in estimating effects based on pharmacological attributes, even when historical utilization data is unavailable. While the alignment was broadly strong, some discrepancies were observed. In four cases out of the five where full concordance was not achieved, the effect sizes remained directionally consistent with the RCTs, but conclusions diverged due to differences in statistical significance. Only one comparison -the VARSITY trial—showed a contradictory result between the original RCT that showed a positive effect for vedolizumab [0.09 (0.03-0.15)] in contrast to the model-generated effect size that indicated a negative effect compared to adalimumab [-0.15 (-0.21, -0.08)], both statistically significant but in opposite directions. A plausible explanation for this outlier is the extended follow-up duration (52 weeks) in VARSITY, which is substantially longer than most other trials (8 to 14 weeks, with most spanning 8–10 weeks) and can introduce increased uncertainty from informative censoring in RWD-derived estimates and potential for uncontrolled concomitant interventions. The differences in effect size between observed and predicted results were generally small, ranging from 0 to -0.23 with a median of -0.05 and a mean of -0.045 (SD: 0.09; Table 3 ). Figure 4 a & b visually confirm the alignment of predictive model results with actual RCTs through forest plots of treatment versus control. The Bland-Altman plots ( Figure 4 c & d ) further demonstrate minimally significant differences and little systematic bias, particularly at lower remission rates (<20%) or for novel agents (in red). Our efforts build upon and expand previous work that uses real-world evidence to emulate clinical trial findings, including target trial emulation frameworks and initiatives like RCT DUPLICATE [31-33]. Unlike replication-focused studies, our approach incorporates knowledge graphs to enable outcome estimation for mechanisms not yet represented in the RWD – a significant advantage in scenarios where historical utilization data is unavailable.. The successful validation across 21 trials represents a more extensive comparison than typically reported in studies validating results in cohort-level predicted trial results [11]. Importantly, our validation went beyond binary success/failure predictions to assess the accuracy of precise treatment effect estimates, matching clinical endpoints, patient eligibility criteria, evaluation time points, and sample sizes with the original trials. The use of multiple performance metrics, including trial success agreement, absolute outcome agreement, Cohen’s d, and F1 scores, provided a comprehensive evaluation of the model’s performance against contemporary regulatory standards. This rigorous, multi-faceted validation provides a strong basis for confidence in the model's outputs. Considering the substantial heterogeneity among patients with UC, characterized by varying clinical expressions and disease trajectories including frequent relapses versus more stable courses, achieving a strong concordance between predicted and RCT observed results is noteworthy. This heterogeneity is driven by multiple factors, including environmental factors, diet, life stress, psychological influences, and genetics. Additionally, the diversity of studied drugs further contributes to the variability [34, 35]. Our analysis encompassed seven drug types, each with distinct mechanisms of action, spanning two major categories: small molecules (JAK inhibitors, S1P receptor modulators, TYK2 inhibitors, PDE4 inhibitors) and biologics (monoclonal antibodies targeting TNF-α, β7 integrin, and IL-36). The high consistency of results across this heterogeneity underscores the model’s robustness, predictive reliability, and validity in the context of UC treatment. Several factors likely contribute to the model’s robustness: the comprehensive knowledge graph capturing diverse mechanisms of action, rigorous validation of individual outcome model components and, critically, the large RWD sample size (N=723,263 patient events) facilitating well-matched cohort creation. A critical application of a validated drug response model lies in its capacity to inform clinical trial design. Focusing predictions on specific patient subgroups allow to preemptively identify responder subgroups and estimate treatment effects across diverse patient profiles, paving the way for personalized therapeutic strategies. Supplementary Figure S2 and Table S2 show factors associated with modified treatment effects in response to S1P modulator treatment. For instance, men under 40 without prior biologic exposure exhibited enhanced response. Subgroup identification also directly informs cohort optimization ( Figure 6 a & b ) and increases the probability of technical success in clinical trials ( Figure 6 c & d ). Notably, our drug response model predictions for optimized patient cohorts in ELEVATE-UC-12 aligned closely with post-hoc analyses of the trial - analyses that were not available at the time of the study. These retrospective analyses confirmed that patients with low inflammatory markers who were naïve to biologic therapy consistently showed higher response rates to etrasimod (an S1P modulator) [36]. Additionally, patients not receiving concomitant steroid therapy demonstrated increased response to S1P modulator treatment in the post-hoc analyses [37]. Future research should elucidate whether the identified predictive markers of response are specific to S1P modulator therapy or represent generally predictive factors for better response to induction therapy in ulcerative colitis. Overall, these convergent findings validate the use of real-world data-informed predictive drug response modeling to guide future clinical trial design and advance personalized therapy development. It is important to emphasize that predictive models should be applied within the context in which they were validated. In our case, induction therapy trials may be more amenable to implementing these insights compared to maintenance therapy trials, given the specific patient populations and treatment paradigms used in model development. Beyond trial optimization, the validation and modelling frameworks presented here aim to reinforce the evolving role of RWE in drug development and evaluation. While RCTs remain the gold standard, our findings suggest that well-designed RWE studies, enhanced by appropriate ML approaches, can confirm and extend RCT findings to broader, more heterogeneous real-world populations where adherence and concomitant care differ, and provide out-of-sample estimations for novel drug classes. This offers a robust framework for comparative research in a context where head-to-head trials are scarce, and where assessing efficacy in smaller subpopulations is often impractical. Here, validated computational approaches can provide key signals to warrant investment in larger studies involving primary collection of healthcare data. The strong concordance between our real-world predictions and RCT findings, established through a comprehensive validation process, reinforces the reliability of such model-based RWE approaches and has significant implications for regulatory and clinical decision making. As healthcare practitioners, drug developers and regulatory agencies increasingly consider the role of AI and RWD in their processes, robust, patient level, validation frameworks offer tangible pathways to evaluate and integrate these data sources effectively. Demonstrating alignment between predicted outcomes and trial data builds confidence in using predictive models for regulatory and clinical decision-making. Gradual implementations of such models could support prioritizing resource allocation, optimizing designs, and assessing the feasibility of new indications or stratifications before undertaking expensive RCTs. As regulatory agencies increasingly integrate RWE into decision-making processes, RWD models have the potential to support adaptive approvals, label expansions, and health technology assessments where traditional trials are impractical. Given its demonstrated validity across multiple drug classes, this approach could be extended to other disease areas, broadening its applicability in real-world drug evaluation and accelerating the translation of biomedical knowledge into tangible patient benefit. Conclusion This study demonstrates the value of integrating RWD, biomedical knowledge graphs, and machine learning to improve the prediction of clinical trial outcomes in UC. The ClinBoost framework showed strong concordance with observed results across 21 RCTs, accurately reproducing treatment effects in diverse settings, including those involving novel agents not represented in existing datasets. By leveraging enriched knowledge graph embeddings and longitudinal patient journeys, the model provided unbiased estimates of remission outcomes and treatment effects while identifying responder subgroups that could enhance trial efficiency. Importantly, ClinBoost not only emulated past trials but also offered actionable insights to optimize trial design, such as refining eligibility criteria, guiding sample size estimation, and improving probability of technical success. These findings highlight the potential of data-driven modeling to complement traditional randomized trials by informing trial strategies, reducing uncertainty, and accelerating drug development. Broader application of this approach may extend to other therapeutic areas, supporting more efficient, personalized, and cost-effective pathways for evaluating novel interventions. Abbreviations Abbreviation Definition CDM Common Data Model CI Confidence Interval CRP C-Reactive Protein EHDS European Health Data Space EHR Electronic Health Records eMS Electronic Mayo Score JAK Janus Kinase KG Knowledge Graph LoA Limits of Agreement ML Machine Learning MoA Mechanism of Action MS Mayo Score NLR Neutrophil/Lymphocyte Ratio OMOP Observational Medical Outcomes Partnership PDE4 Phosphodiesterase 4 pMS Partial Mayo Score PTS Probability of Technical Success RCT Randomized Controlled Trial RWD Real-World Data RWE Real-World Evidence SD Standard Deviation SoC Standard of Care Sphingosine-1-Phosphate Sphingosine-1-Phosphate TNF Tumor Necrosis Factors t-SNE t-distributed Stochastic Neighbor Embedding TYK2 Tyrosine Kinase 2 UC Ulcerative Colitis Declarations Funding This study was funded by Sanofi. Acknowledgements The authors would like to thank Jean-Paul Collet and Maninder Anand of Evidinno Outcomes Research Inc. (Vancouver, BC, Canada) for medical writing assistance. Author Contributions F.D., O.M., and R.H. conceived and supervised the study, F.D. and O.M. led the study design and the interpretation of results. M.S., O.M., A.A., and A.P. led the data curation, preprocessing, cohort definition, and development of the modeling framework and predictive models, including the integration of real-world data and knowledge graph components. F.D., O.M., M.S., and A.P. led the design of the statistical validation and simulation analyses. A.K. and B.S. contributed domain expertise in ulcerative colitis, assessed clinical relevance, and supported interpretation of the study findings. F.D., O.M., M.S., and A.P. jointly drafted, wrote, and revised the manuscript. All authors critically reviewed the manuscript, contributed to its final version, and approved it for submission. Competing Interests Flavio Dormont, Annie Kruger, and Ramon Hernandez are employees and shareholders of Sanofi (Morristown, NJ, USA). Amina Alaskarov, Amichai Perlman, Omri Matalon, and Michael Shapiro are employees of QuantHealth (Tel Aviv, Israel), which was contracted by Sanofi to conduct this study. Dr. Sands reports personal fees and non-financial support from Abbvie, personal fees from Alimentiv, personal fees from Adiso Therapeutics, personal fees from Agomab Therapeutics, personal fees from Amgen, personal fees from AnaptysBio, personal fees and non-financial support from AstraZeneca, personal fees from Biolojic Design, personal fees from Biora Therapeutics, personal fees from Boehringer Ingelheim, personal fees and non-financial support from Celltrion, personal fees from Equilium, personal fees from Ensho Therapeutics, personal fees from Enveda Biosciences, personal fees from Evommune, personal fees from Ferring, personal fees from Fzata, personal fees from Galapagos, personal fees from Genentech (Roche), personal fees from Gilead Sciences, personal fees from GlaxoSmithKline, personal fees from Gossamer Bio, personal fees from Imhotex, personal fees from Immunyx Therapeutics, personal fees from Index Pharmaceuticals, personal fees from Innovation Pharmaceuticals, grants, personal fees and non-financial support from Janssen/J&J Innovative Medicine, personal fees from Kaleido, personal fees from Kallyope, personal fees and non-financial support from Merck & Co., personal fees and non-financial support from Merck Sharp & Dohme, personal fees from Microba, personal fees from Microbiotica, personal fees from Mitsubishi Tanabe Pharma, personal fees from Mobius Care, personal fees from Morphic Therapeutics, personal fees and non-financial support from Eli Lilly & Sons, personal fees from MRM Health, personal fees from Nexus Therapeutics, personal fees from Nimbus Discovery, personal fees from Odyssey Therapeutics, personal fees from Palisade Bio, personal fees and non-financial support from Prometheus Biosciences, personal fees from Prometheus Laboratories, personal fees and non-financial support from Pfizer, personal fees from Protagonist Therapeutics, personal fees from Q32 Bio, personal fees from Rasayana Therapeutics, personal fees from Recludix Therapeutics, personal fees from Reistone Biotherapeutics, personal fees from Sanofi, personal fees from Sorriso Pharmaceuticals, personal fees from Surrozen, personal fees from Target RWE, personal fees and non-financial support from Takeda, personal fees from Teva, personal fees from TLL Pharmaceutical, personal fees from TR1X, personal fees from Union Therapeutics, personal fees and non-financial support from Abivax, grants, personal fees and non-financial support from Bristol Myers Squibb, personal fees from Theravance Biopharma, personal fees, non-financial support and other from Ventyx Biopharma, outside the submitted work. Data Availability The datasets used in this study, including real-world data (RWD) derived from electronic medical records and insurance claims, as well as components of the biomedical knowledge graph, were obtained under license from proprietary vendors and are not publicly available. Due to contractual and privacy restrictions, we are unable to share the underlying data. Researchers interested in accessing the datasets may contact the corresponding author to discuss the possibility of data access, subject to licensing agreements and confidentiality obligations. Ethics Approval and Consent to Participate All methods were carried out in accordance with relevant guidelines and regulations. This study utilized retrospective, de-identified patient-level data from the PurpleLab® insurance open claims database and electronic health records (EHR) from the EVERSANA EHR database (EVERSANA Life Sciences Inc.), which were linked via a third-party tokenization service (Datavant Inc., San Francisco, CA, USA). The datasets contain anonymized information on healthcare encounters, procedures, medications, diagnoses, and demographic characteristics, with no direct patient identifiers available to the researchers. As a retrospective analysis of de-identified data, this study was exempt from institutional review board (IRB) review under 45 CFR 46.104(d)(4), and the requirement for informed consent was waived. Consent for Publication Not applicable. References Hwang TJ, Carpenter D, Lauffenburger JC, Wang B, Franklin JM, Kesselheim AS. Failure of Investigational Drugs in Late-Stage Clinical Development and Publication of Trial Results. JAMA internal medicine. 2016;176(12):1826-33.doi:10.1001/jamainternmed.2016.6008. Farid SS, Baron M, Stamatis C, Nie W, Coffman J. Benchmarking biopharmaceutical process development and manufacturing cost contributions to R&D. mAbs. 2020;12(1):1754999.doi:10.1080/19420862.2020.1754999. Mullard A. Parsing clinical success rates. Nature Reviews Drug Discovery. 2016;15(7):447-.doi:10.1038/nrd.2016.136. Schlander M, Hernandez-Villafuerte K, Cheng CY, Mestre-Ferrandiz J, Baumann M. How Much Does It Cost to Research and Develop a New Drug? A Systematic Review and Assessment. PharmacoEconomics. 2021;39(11):1243-69.doi:10.1007/s40273-021-01065-y. Augustine EF, Adams HR, Mink JW. Clinical trials in rare disease: challenges and opportunities. Journal of child neurology. 2013;28(9):1142-50.doi:10.1177/0883073813495959. Rehman AU, Li M, Wu B, Ali Y, Rasheed S, Shaheen S, et al. Role of Artificial Intelligence in Revolutionizing Drug Discovery. Fundamental Research. 2024.doi:https://doi.org/10.1016/j.fmre.2024.04.021. Fu T, Huang K, Xiao C, Glass LM, Sun J. HINT: Hierarchical interaction network for clinical-trial-outcome predictions. Patterns. 2022;3(4):100445.doi:https://doi.org/10.1016/j.patter.2022.100445. Wang Z, Xiao C, Sun J. SPOT: Sequential Predictive Modeling of Clinical Trial Outcome with Meta-Learning. Proceedings of the 14th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics; Houston, TX, USA: Association for Computing Machinery; 2023. p. Article 53. Al-Mohaya M, Demiralp B, Kurt A, Çelik Y. In silico approaches which are used in pharmacy. Journal of Applied Pharmaceutical Science. 2024.doi:10.7324/JAPS.2024.154854. Duruisseaux M, Masson P, Nativel A, Girard N, Cadranel J, Swalduz A, et al. Prospective validation of in silico clinical trials prediction using an EGFR-mutated NSCLC mechanistic computational disease model. Journal of Clinical Oncology. 2024;42(16_suppl):8614-.doi:10.1200/JCO.2024.42.16_suppl.8614. Angoulvant D, Granjeon-Noriot S, Amarenco P, Bastien A, Bechet E, Boccara F, et al. In-silico trial emulation to predict the cardiovascular protection of new lipid-lowering drugs: an illustration through the design of the SIRIUS programme. European journal of preventive cardiology. 2024;31(15):1820-30.doi:10.1093/eurjpc/zwae254. Brbić M, Yasunaga M, Agarwal P, Leskovec J. Predicting drug outcome of population via clinical knowledge graph. medRxiv : the preprint server for health sciences. 2024.doi:10.1101/2024.03.06.24303800. Joshi P, V M, Mukherjee A. A knowledge graph embedding based approach to predict the adverse drug reactions using a deep neural network. Journal of biomedical informatics. 2022;132:104122.doi:10.1016/j.jbi.2022.104122. Carracedo-Reboredo P, Liñares-Blanco J, Rodríguez-Fernández N, Cedrón F, Novoa FJ, Carballal A, et al. A review on machine learning approaches and trends in drug discovery. Computational and Structural Biotechnology Journal. 2021;19:4538-58.doi:https://doi.org/10.1016/j.csbj.2021.08.011. Franklin J, Schneeweiss S. When and How Can Real World Data Analyses Substitute for Randomized Controlled Trials? Clinical Pharmacology & Therapeutics. 2017;102.doi:10.1002/cpt.857. Raman K, Kumar R, Musante CJ, Madhavan S. Integrating Model-Informed Drug Development With AI: A Synergistic Approach to Accelerating Pharmaceutical Innovation. Clinical and translational science. 2025;18(1):e70124.doi:10.1111/cts.70124. Sharma A, Lysenko A, Jia S, Boroevich KA, Tsunoda T. Advances in AI and machine learning for predictive medicine. Journal of Human Genetics. 2024;69(10):487-97.doi:10.1038/s10038-024-01231-y. Nicholson DN, Greene CS. Constructing knowledge graphs and their biomedical applications. Comput Struct Biotechnol J. 2020;18:1414-28.doi:10.1016/j.csbj.2020.05.017. Chandak P, Huang K, Zitnik M. Building a knowledge graph to enable precision medicine. Scientific Data. 2023;10(1):67.doi:10.1038/s41597-023-01960-3. Gao Z, Ding P, Xu R. KG-Predict: A knowledge graph computational framework for drug repurposing. Journal of biomedical informatics. 2022;132:104133.doi:10.1016/j.jbi.2022.104133. FDA. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products 2025 [Available from: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological; Li P, Wu Y, Xiong W, Cao J, Chen M, Yuan Z, et al. Association between the immune-inflammation index and the severity and clinical outcomes of patients with inflammatory bowel disease: a systematic review and meta-analysis. BMC Gastroenterology. 2025;25(1):414.doi:10.1186/s12876-025-04033-4. Sands BE, Peyrin-Biroulet L, Loftus EV, Danese S, Colombel J-F, Törüner M, et al. Vedolizumab versus Adalimumab for Moderate-to-Severe Ulcerative Colitis. New England Journal of Medicine. 2019;381(13):1215-26.doi:doi:10.1056/NEJMoa1905725. Naegeli AN, Hunter T, Dong Y, Hoskin B, Middleton-Dalby C, Hetherington J, et al. Full, Partial, and Modified Permutations of the Mayo Score: Characterizing Clinical and Patient-Reported Outcomes in Ulcerative Colitis Patients. Crohn's & colitis 360. 2021;3(1):otab007.doi:10.1093/crocol/otab007. Lewis JD, Chuai S, Nessel L, Lichtenstein GR, Aberra FN, Ellenberg JH. Use of the noninvasive components of the Mayo score to assess clinical response in ulcerative colitis. Inflamm Bowel Dis. 2008;14(12):1660-6.doi:10.1002/ibd.20520. Louis E, Schreiber S, Panaccione R, Bossuyt P, Biedermann L, Colombel J-F, et al. Risankizumab for Ulcerative Colitis: Two Randomized Clinical Trials. JAMA. 2024;332(11):881-97.doi:10.1001/jama.2024.12414. Sandborn WJ, Feagan BG, D’Haens G, Wolf DC, Jovanovic I, Hanauer SB, et al. Ozanimod as Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2021;385(14):1280-91.doi:doi:10.1056/NEJMoa2033617. Feagan BG, Rutgeerts P, Sands BE, Hanauer S, Colombel J-F, Sandborn WJ, et al. Vedolizumab as Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2013;369(8):699-710.doi:doi:10.1056/NEJMoa1215734. Rutgeerts P, Sandborn WJ, Feagan BG, Reinisch W, Olson A, Johanns J, et al. Infliximab for Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2005;353(23):2462-76.doi:doi:10.1056/NEJMoa050516. Cohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. New York: Routledge; 1988. Franklin JM, Patorno E, Desai RJ, Glynn RJ, Martin D, Quinto K, et al. Emulating Randomized Clinical Trials With Nonrandomized Real-World Evidence Studies: First Results From the RCT DUPLICATE Initiative. Circulation. 2021;143(10):1002-13.doi:10.1161/circulationaha.120.051718. Franklin JM, Schneeweiss S. When and How Can Real World Data Analyses Substitute for Randomized Controlled Trials? Clinical pharmacology and therapeutics. 2017;102(6):924-33.doi:10.1002/cpt.857. Franklin JM, Pawar A, Martin D, Glynn RJ, Levenson M, Temple R, et al. Nonrandomized Real-World Evidence to Support Regulatory Decision Making: Process for a Randomized Trial Replication Project. Clinical pharmacology and therapeutics. 2020;107(4):817-26.doi:10.1002/cpt.1633. Ungaro R, Colombel JF, Lissoos T, Peyrin-Biroulet L. A Treat-to-Target Update in Ulcerative Colitis: A Systematic Review. The American journal of gastroenterology. 2019;114(6):874-83.doi:10.14309/ajg.0000000000000183. Ferretti F, Cannatelli R, Monico MC, Maconi G, Ardizzone S. An Update on Current Pharmacotherapeutic Options for the Treatment of Ulcerative Colitis. Journal of clinical medicine. 2022;11(9).doi:10.3390/jcm11092302. Sands BE, Dubinsky MC, Kotze PG, Vermeire S, Panaccione R, Long MD, et al. Efficacy and Safety of Etrasimod in Patients With Moderately to Severely Active Ulcerative Colitis Stratified by Baseline Modified Mayo Score: A Post Hoc Analysis From the Phase 3 ELEVATE UC Clinical Program. Inflammatory Bowel Diseases. 2025.doi:10.1093/ibd/izaf036. Sands BE, Leung Y, Rubin DT, Gecse KB, Panés J, Goetsch M, et al. Etrasimod Corticosteroid-Free Efficacy, Impact of Concomitant Corticosteroids on Efficacy and Safety, and Corticosteroid-Sparing Effect in Ulcerative Colitis: Analyses of the ELEVATE UC Clinical Program. Journal of Crohn's and Colitis. 2024;19(3).doi:10.1093/ecco-jcc/jjae150. Additional Declarations Competing interest reported. Flavio Dormont, Annie Kruger, and Ramon Hernandez are employees and shareholders of Sanofi (Morristown, NJ, USA). Amina Alaskarov, Amichai Perlman, Omri Matalon, and Michael Shapiro are employees of QuantHealth (Tel Aviv, Israel), which was contracted by Sanofi to conduct this study. Dr. Sands reports personal fees and non-financial support from Abbvie, personal fees from Alimentiv, personal fees from Adiso Therapeutics, personal fees from Agomab Therapeutics, personal fees from Amgen, personal fees from AnaptysBio, personal fees and non-financial support from AstraZeneca, personal fees from Biolojic Design, personal fees from Biora Therapeutics, personal fees from Boehringer Ingelheim, personal fees and non-financial support from Celltrion, personal fees from Equilium, personal fees from Ensho Therapeutics, personal fees from Enveda Biosciences, personal fees from Evommune, personal fees from Ferring, personal fees from Fzata, personal fees from Galapagos, personal fees from Genentech (Roche), personal fees from Gilead Sciences, personal fees from GlaxoSmithKline, personal fees from Gossamer Bio, personal fees from Imhotex, personal fees from Immunyx Therapeutics, personal fees from Index Pharmaceuticals, personal fees from Innovation Pharmaceuticals, grants, personal fees and non-financial support from Janssen/J&J Innovative Medicine, personal fees from Kaleido, personal fees from Kallyope, personal fees and non-financial support from Merck & Co., personal fees and non-financial support from Merck Sharp & Dohme, personal fees from Microba, personal fees from Microbiotica, personal fees from Mitsubishi Tanabe Pharma, personal fees from Mobius Care, personal fees from Morphic Therapeutics, personal fees and non-financial support from Eli Lilly & Sons, personal fees from MRM Health, personal fees from Nexus Therapeutics, personal fees from Nimbus Discovery, personal fees from Odyssey Therapeutics, personal fees from Palisade Bio, personal fees and non-financial support from Prometheus Biosciences, personal fees from Prometheus Laboratories, personal fees and non-financial support from Pfizer, personal fees from Protagonist Therapeutics, personal fees from Q32 Bio, personal fees from Rasayana Therapeutics, personal fees from Recludix Therapeutics, personal fees from Reistone Biotherapeutics, personal fees from Sanofi, personal fees from Sorriso Pharmaceuticals, personal fees from Surrozen, personal fees from Target RWE, personal fees and non-financial support from Takeda, personal fees from Teva, personal fees from TLL Pharmaceutical, personal fees from TR1X, personal fees from Union Therapeutics, personal fees and non-financial support from Abivax, grants, personal fees and non-financial support from Bristol Myers Squibb, personal fees from Theravance Biopharma, personal fees, non-financial support and other from Ventyx Biopharma, outside the submitted work. Supplementary Files SupplementaryFile1.pdf Appendix1.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9272163","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":634666557,"identity":"774a137d-209c-4d14-9b04-898041b48845","order_by":0,"name":"Flavio Dormont","email":"","orcid":"","institution":"Sanofi (United States)","correspondingAuthor":false,"prefix":"","firstName":"Flavio","middleName":"","lastName":"Dormont","suffix":""},{"id":634666558,"identity":"53de10aa-a0ff-4578-a53c-81588f418882","order_by":1,"name":"Michael Shapiro","email":"","orcid":"","institution":"QuantHealth","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"","lastName":"Shapiro","suffix":""},{"id":634666560,"identity":"a98c9eaf-678c-424e-9fda-1507c98ebfa9","order_by":2,"name":"Amichai Perlman","email":"","orcid":"","institution":"QuantHealth","correspondingAuthor":false,"prefix":"","firstName":"Amichai","middleName":"","lastName":"Perlman","suffix":""},{"id":634666561,"identity":"ff75e530-5b72-470c-afa2-de077493f192","order_by":3,"name":"Amina Alaskarov","email":"","orcid":"","institution":"QuantHealth","correspondingAuthor":false,"prefix":"","firstName":"Amina","middleName":"","lastName":"Alaskarov","suffix":""},{"id":634666562,"identity":"21959098-46aa-4ea3-ab86-a9b297231d7b","order_by":4,"name":"Annie J. Kruger","email":"","orcid":"","institution":"Sanofi (United States)","correspondingAuthor":false,"prefix":"","firstName":"Annie","middleName":"J.","lastName":"Kruger","suffix":""},{"id":634666563,"identity":"83aeb8a5-a5f3-45d3-953a-fd20e2905c6e","order_by":5,"name":"Bruce E. Sands","email":"","orcid":"","institution":"Icahn School of Medicine at Mount Sinai","correspondingAuthor":false,"prefix":"","firstName":"Bruce","middleName":"E.","lastName":"Sands","suffix":""},{"id":634666564,"identity":"97034ccd-849b-4cce-b45f-265a23bb8b9b","order_by":6,"name":"Ramon Hernandez","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDUlEQVRIiWNgGAWjYBAC+wYwdRjC42FgkAMzKgxwazE4gKbFGMw4Q4qWRLC1Z/A4zOD46cTHBQyH5cxnNz978KaiLn1++9mDDw4UMMjzix3A7pee3M3GMxj+G8vcOWZuOOcMW+6GM3nJBgcMGAxnzk7AqsWOIXebNA/D4cQZEglm0rxtPLkbGHLMpD8YMCQY3MauxZj/LVhL/QyJ9G/SvP8k0uX735hJHMCjxXAGxJYECQmg4bwNBgkMN3LwazG48XazMY/BYcMZEjllknOOJRhuuPHGGOgXCZx+MTifu/ExT8VheQmJ9G0Sb2rq5OX7cwwfHPhjI88vjV0LVCOmkAQe5aNgFIyCUTAKCAEAWx1eHzjMmcAAAAAASUVORK5CYII=","orcid":"","institution":"Sanofi (France)","correspondingAuthor":true,"prefix":"","firstName":"Ramon","middleName":"","lastName":"Hernandez","suffix":""},{"id":634666565,"identity":"eba7d9cd-c22b-4b07-a5d0-0f476ce5bebf","order_by":7,"name":"Omri Matalon","email":"","orcid":"","institution":"QuantHealth","correspondingAuthor":false,"prefix":"","firstName":"Omri","middleName":"","lastName":"Matalon","suffix":""}],"badges":[],"createdAt":"2026-03-30 22:23:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9272163/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9272163/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108841219,"identity":"192d0196-a383-4ed1-8eeb-3284d02ce139","added_by":"auto","created_at":"2026-05-09 01:07:25","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2724150,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the predictive model architecture and workflow\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 1 provides an overview of the predictive model architecture and workflow.\u003c/p\u003e\n\u003cp\u003ea) Prediction Workflow.\u003c/p\u003e\n\u003cp\u003eb) Model Architecture: covering patient drug-journey embeddings, ClinBoost model and time-to-event outcome prediction.\u003c/p\u003e","description":"","filename":"Figure1BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/9239bae0a4cd550ac646c28d.png"},{"id":108976915,"identity":"5215077c-59cc-4316-a00e-9f43ee73c446","added_by":"auto","created_at":"2026-05-11 11:29:36","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":4340412,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eQuantHealth Knowledge Graph (QHKG) \u0026amp; Embedding Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ea) Example subgraph from the QuantHealth Knowledge Graph (QHKG) is visualized, showing nodes of type drug, gene, disease, side effect, and pathway, along with their corresponding edges. Major ulcerative colitis drugs and their main molecular targets are highlighted to emphasize key components in the network.\u003c/p\u003e\n\u003cp\u003eb) A 2D t-SNE projection of drug embeddings learned from the QHKG, where drugs are colored by pharmacological subclass to assess whether embeddings capture functional groupings.\u003c/p\u003e\n\u003cp\u003ec) A scatter plot demonstrates a strong negative correlation (Pearson = –0.83) between structural distance in the KG and cosine similarity of drug embeddings, indicating that the embeddings preserve the underlying graph topology.\u003c/p\u003e\n\u003cp\u003ed) A heatmap displays average cosine similarity between drug subgroups and within each drug subgroup (the diagonal cells), with warmer colors indicating higher similarity and suggesting pharmacological or mechanistic relatedness between drug classes.\u003c/p\u003e","description":"","filename":"Figure2BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/6e8249553e3038c03916f275.png"},{"id":108841220,"identity":"8a43824f-92c5-4512-8ce6-32d8aecb0c6d","added_by":"auto","created_at":"2026-05-09 01:07:25","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1310482,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRemission Rates in Clinical Trials: Observed vs. Predicted Outcomes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ea) Comparison of treatment vs. control group outcomes across 21 clinical trials. Points represent mean remission rates. Horizontal lines with end caps represent 95% confidence intervals. Trials grayed out represent predictions that failed to reproduce the observed binary outcome. Left panel represent observed results, and right panel represents predicted results.\u003c/p\u003e\n\u003cp\u003eb) Confusion matrix of model binary (success/failure) performance.\u003c/p\u003e\n\u003cp\u003ec) Quantitative model performance metrics on the binary level (success/failure).\u003c/p\u003e\n\u003cp\u003ed) ROC-AUC curve depicting the model's performance on the patient-level RWD test set. Predictions and evaluations were performed on a per-time-bin basis.\u003c/p\u003e","description":"","filename":"Figure3BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/1d92eb1334ab5f16d3fa7b9a.png"},{"id":109067678,"identity":"ede6848b-c047-4a1a-9f32-e3c2d39562ce","added_by":"auto","created_at":"2026-05-12 09:59:37","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1482653,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTrial Prediction and Model Validation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ea) Forest Plot of Absolute Differences:\u003cstrong\u003e \u003c/strong\u003eThis forest plot presents the absolute difference between observed and predicted treatment effect, expressed in the original units of measurement (percentage points). Each trial is represented by a square; horizontal lines represent 95% confidence intervals. This provides a direct, non-standardized measure of the discrepancy between observed and predicted results. The vertical dashed line represents no difference. An aggregate “overall” measure is provided, summarizing the inverse-variance weighted difference between the observed and predicted results across all trials.\u003c/p\u003e\n\u003cp\u003eb) Forest Plot of Standardized Mean Differences:\u003cstrong\u003e \u003c/strong\u003eThis forest plot displays the Cohen’s d statistic for the difference between observed and predicted effect sizes across validation clinical trials. Because population characteristics and measurement methodologies can vary across trials - even when identical clinical scores are applied - the prediction error is best expressed using Cohen’s \u003cem\u003ed\u003c/em\u003e, which expresses the discrepancy in units of pooled standard deviation. Values of 0.2, 0.5, and 0.8 are interpreted as small, moderate, and large effects, respectively, with higher values indicating greater deviation between predicted and observed effect sizes.\u003c/p\u003e\n\u003cp\u003ec) Bland-Altman Plot of Agreement between Predicted and Observed Study Arm Remission results: Numbers correspond to Trial number in Table 2; C: control arm; T: treatment arm. The Bland-Altman plot indicates no systematic bias in predicting study arms. However, prediction accuracy decreases for arm size average remission rates greater than 0.2.\u003c/p\u003e\n\u003cp\u003ed) Bland-Altman Plot of Agreement of Study effect size: The Bland-Altman plot reveals a slight negative bias in predicting the study effect, with consistent prediction accuracy across the range of effect sizes.\u003c/p\u003e","description":"","filename":"Figure4BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/05b17331ad3d3d20760971f9.png"},{"id":108841223,"identity":"8d5b5721-0aee-409f-adad-a53182851204","added_by":"auto","created_at":"2026-05-09 01:07:25","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":350520,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStratified Performance of the Predictive Model by Clinical Trial Design\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis bar graph illustrates that our model performs robustly across both Phase 2 and Phase 3 clinical trials, as well as with drugs targeting diverse mechanisms of action. Notably, our unique knowledge graph enrichment technique enables accurate prediction of compounds that are absent from both real-world data and the original biomedical KG.\u003c/p\u003e\n\u003cp\u003eNumber of trials per group: Biologics- 10 trials, Non-Biologics - 11 trials; Phase II – 6 trials, Phase III – 15 trials; Anti-integrin trials – 5 trials, JAK inhibitor – 6 trials, Other mechanisms – 10 trials; KG enriched compounds – 6 trials, Non-Enriched Compounds – 15 trials.\u003c/p\u003e","description":"","filename":"Figure5BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/60c7c94b1917befe21d24549.png"},{"id":108841225,"identity":"85734647-7d55-461d-bfe4-83f75e50a988","added_by":"auto","created_at":"2026-05-09 01:07:25","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1068370,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel-guided Trial Design Optimization – ELEVATE UC-12 Trial\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea)\u003c/strong\u003e The base cohort from the ELEVATE UC 12 trial was analyzed using our multivariate prediction engine to identify sub-cohorts of interest. Each prediction involved modifying the original cohort by applying additional inclusion/exclusion criteria. The panel displays predicted effect sizes (remission rates in treatment minus control) for each sub-cohort, compared to an estimated cohort size representing U.S. ulcerative colitis patients meeting the respective criteria. Dots are shown in grey if the cohort failed to achieve statistical significance based on the original trial sample size. In addition to the original cohort, two cohorts of interest are highlighted: an \u003cem\u003eOptimized\u003c/em\u003ecohort with an increased predicted effect size and a \u003cem\u003eLow Effect\u003c/em\u003e cohort with reduced efficacy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb)\u003c/strong\u003e Comparison of predicted outcomes for the Optimized and Low Effect cohorts relative to the original cohort over a period of 52 weeks. Blue curve - treatment arm; Red curve - placebo.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec)\u003c/strong\u003e Power analysis using the prediction engine, showing predicted statistical significance levels across varying sample sizes, while maintaining the original treatment-to-control ratio.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ed)\u003c/strong\u003e Monte Carlo simulation results based on random sampling of patients matching the inclusion/exclusion criteria for each cohort. Distributions of predicted treatment effects are shown for the three cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLow effect cohort:\u003c/strong\u003e Previously treated with anti-TNF and without high inflammatory markers.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOptimized cohort:\u003c/strong\u003e No previous biological therapy, no systemic steroid treatment of any kind at baseline (oral/IV).\u003c/p\u003e","description":"","filename":"Figure6BMCMIDM.png","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/102c2b87741337e840f1a954.png"},{"id":109069295,"identity":"6cfddeac-012b-435e-9c50-563e1d35b566","added_by":"auto","created_at":"2026-05-12 10:22:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":12197818,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/2985439d-bbc2-4c54-8d82-3da941576f28.pdf"},{"id":108976848,"identity":"46ee6630-ae47-4163-9b6f-0c12b1e37416","added_by":"auto","created_at":"2026-05-11 11:29:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":421263,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFile1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/9dc97ef288284d15fdcb8f98.pdf"},{"id":108977042,"identity":"8d2fafe2-b91e-4270-860b-1e36bd5a2aee","added_by":"auto","created_at":"2026-05-11 11:30:06","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":23582,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix1.docx","url":"https://assets-eu.researchsquare.com/files/rs-9272163/v1/09bfffca84e2042bfaa58a68.docx"}],"financialInterests":"Competing interest reported. Flavio Dormont, Annie Kruger, and Ramon Hernandez are employees and shareholders of Sanofi (Morristown, NJ, USA). Amina Alaskarov, Amichai Perlman, Omri Matalon, and Michael Shapiro are employees of QuantHealth (Tel Aviv, Israel), which was contracted by Sanofi to conduct this study. Dr. Sands reports personal fees and non-financial support from Abbvie, personal fees from Alimentiv, personal fees from Adiso Therapeutics, personal fees from Agomab Therapeutics, personal fees from Amgen, personal fees from AnaptysBio, personal fees and non-financial support from AstraZeneca, personal fees from Biolojic Design, personal fees from Biora Therapeutics, personal fees from Boehringer Ingelheim, personal fees and non-financial support from Celltrion, personal fees from Equilium, personal fees from Ensho Therapeutics, personal fees from Enveda Biosciences, personal fees from Evommune, personal fees from Ferring, personal fees from Fzata, personal fees from Galapagos, personal fees from Genentech (Roche), personal fees from Gilead Sciences, personal fees from GlaxoSmithKline, personal fees from Gossamer Bio, personal fees from Imhotex, personal fees from Immunyx Therapeutics, personal fees from Index Pharmaceuticals, personal fees from Innovation Pharmaceuticals, grants, personal fees and non-financial support from Janssen/J\u0026J Innovative Medicine, personal fees from Kaleido, personal fees from Kallyope, personal fees and non-financial support from Merck \u0026 Co., personal fees and non-financial support from Merck Sharp \u0026 Dohme, personal fees from Microba, personal fees from Microbiotica, personal fees from Mitsubishi Tanabe Pharma, personal fees from Mobius Care, personal fees from Morphic Therapeutics, personal fees and non-financial support from Eli Lilly \u0026 Sons, personal fees from MRM Health, personal fees from Nexus Therapeutics, personal fees from Nimbus Discovery, personal fees from Odyssey Therapeutics, personal fees from Palisade Bio, personal fees and non-financial support from Prometheus Biosciences, personal fees from Prometheus Laboratories, personal fees and non-financial support from Pfizer, personal fees from Protagonist Therapeutics, personal fees from Q32 Bio, personal fees from Rasayana Therapeutics, personal fees from Recludix Therapeutics, personal fees from Reistone Biotherapeutics, personal fees from Sanofi, personal fees from Sorriso Pharmaceuticals, personal fees from Surrozen, personal fees from Target RWE, personal fees and non-financial support from Takeda, personal fees from Teva, personal fees from TLL Pharmaceutical, personal fees from TR1X, personal fees from Union Therapeutics, personal fees and non-financial support from Abivax, grants, personal fees and non-financial support from Bristol Myers Squibb, personal fees from Theravance Biopharma, personal fees, non-financial support and other from Ventyx Biopharma, outside the submitted work.","formattedTitle":"Simulating Ulcerative Colitis Clinical Trials Using Knowledge Graph–Enhanced Real-World Data Modeling: Validation Across 21 Studies","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe development of new therapeutic agents is resource-intensive and often hindered by clinical trial challenges. Traditional randomized controlled trials, despite being considered the gold standard for evaluating drug efficacy and safety, are limited by high costs, lengthy durations, and recruitment difficulties [1-5]. Predictive modeling has sought to address these challenges by shifting from traditional \u003cem\u003ein vitro\u003c/em\u003e and\u003cem\u003e\u0026nbsp;in vivo\u003c/em\u003e studies to advanced \u003cem\u003ein silico\u003c/em\u003e simulations, improving translational efficiency and accelerating drug discovery [6-8]. Computational \u003cem\u003ein silico\u003c/em\u003e models[9, 10]\u0026nbsp;can simulate biological processes and help to predict drug responses, providing a cost-effective and efficient exploration of new therapeutic mechanisms of action (MoA). These models help identify promising\u0026nbsp;compounds based on their structure and anticipated interactions with other target biomarkers, simulate pharmacokinetic and pharmacodynamic behaviors, optimize dosing, and detect potential adverse effects early\u0026nbsp;[11-13]. Such approaches can markedly reduce reliance on extensive \u003cem\u003ein vivo\u0026nbsp;\u003c/em\u003estudies and expedite drug candidate identification\u0026nbsp;[7, 8]\u0026nbsp; yet their application to clinical-trial modeling and protocol optimization remains limited.\u003c/p\u003e\n\u003cp\u003eRecent legislative and technological advances, such as the European Health Data Space (EHDS) and the US HITECH and Cures Acts, have facilitated the integration of clinical real-world data (RWD) in drug development. Coupled with advanced machine learning (ML) methods [14], RWD-based models offer dynamic time-dependent predictions of patients\u0026rsquo; trajectories, crucial for management and decision-making in progressive or fluctuating diseases [15]. These models leverage sophisticated ML techniques, enhancing predictions by parsing complex patient data and the causal frameworks influencing treatment response [16, 17]. Integration of knowledge graphs, encoding biological relationships further enhances ML-driven predictive modeling. Knowledge graphs enable the representation of complex and interconnected biomedical processes in a way amenable to machine learning frameworks [18, 19]. This facilitates discovering novel drug-target interactions, repositioning of existing drug, and enhancing the causal inference of clinical outcomes [12, 20].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe integration of real-world data and knowledge graph\u0026ndash;driven modeling holds particular promise for predicting clinical outcomes and optimizing trial protocols. Building on these advances, we developed ClinBoost, a modeling framework that combines RWD and knowledge graph drug embeddings to simulate clinical trial outcomes and guide trial design.\u0026nbsp;The unique capabilities of ClinBoost enable scalable trial designs by modeling clinical responses using endpoints aligned with those used in prospective trials. It supports comparative simulations between investigational therapies, standard-of-care, and competitors, including scenarios absent from historical data such as first-in-class or repurposed drugs. The framework can simulate diverse treatment settings, including placebo, monotherapies, and combination therapies (regimens), while capturing drug\u0026ndash;patient interactions to enable subgroup-level predictions. It also supports protocol optimization through flexible evaluation of cohorts, treatments, and endpoints.\u003c/p\u003e\n\u003cp\u003eAlthough predictive models of drug efficacy based on RWD have advanced considerably, few studies subject them to rigorous validation, and those that do typically include only one or two comparisons, limiting generalizability. To address this gap, we applied a robust validation framework to compare model predictions with outcomes from 21 clinical trials\u0026nbsp;assessing\u0026nbsp;13 different drugs, through rigorous statistical and Bland\u0026ndash;Altman analyses to ensure technical reliability and clinical relevance in line with contemporary regulatory standards\u0026nbsp;[21]. \u0026nbsp;Our approach assessed not only binary trial outcomes (e.g., success vs. failure), but also continuous treatment effect estimates by precisely simulating trial conditions.\u0026nbsp;\u0026nbsp;This included matching clinical endpoints, eligibility criteria, time points (e.g., 12 weeks, 24 weeks) and sample sizes to replicate trial designs using real-world patient cohorts.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn this study, we demonstrated the practical applicability of our ML process by developing and validating a RWD predictive drug response model using ulcerative colitis (UC) as a use case. UC is a chronic inflammatory disease with significant unmet medical needs and where patient heterogeneity hampers drug development efforts [22, 23]. By integrating ML, RWD, and knowledge graph methodologies, we demonstrate accurate predictions of drug efficacy against past clinical trials and demonstrate how such modeling frameworks can be applied to optimize future trial design, thereby optimizing recruited patient populations, increasing the probability of technical success, and accelerating therapeutic development.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eData Sources\u003c/h2\u003e\n\u003ch3\u003eReal-world Clinical Data\u003c/h3\u003e\n\u003cp\u003eFor modeling patient outcomes, we used deidentified patient-event level data from the PurpleLab\u0026reg; insurance open claims database and electronic health records (EHR) from Eversana\u0026rsquo;s Electronic Health Record database (EVERSANA Life Sciences Inc.), which were linked via a third-party tokenization service (Datavant Inc., San Francisco, California). The PurpleLab\u0026reg; open claims database covers approximately 350 million patients representing 98% of healthcare claims filed in the United States between 2014 and 2023. The database contains anonymized information on surgeries, procedures, medications, diagnoses, and various healthcare encounters, and includes comprehensive demographic and death information. The Eversana EHR dataset covers 120 million patients who were treated primarily in the community setting in the United States and includes data on medical procedures, diagnoses, healthcare encounters,\u003cem\u003e\u0026nbsp;\u003c/em\u003emedications, laboratory results, vitals, and socio-demographic factors.\u0026nbsp;The contents of the linked dataset used in this study were standardized to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) prior to model development.\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003eBio-pharmacological Data for Knowledge Graph Construction\u003c/h3\u003e\n\u003cp\u003eA comprehensive biomedical dataset compiled from a combination of open-source and proprietary data sources was used to build a biomedical knowledge graph. This dataset contained information on a wide range of small molecules and biologics and their pharmacological profiles, namely pharmacokinetics/-dynamics (PK/PD), absorption, metabolism, excretion, toxicity, mechanism of action, cellular targets and known drug-drug interactions. We also included data for relevant biological processes, diseases, anatomical structures, genes, biological pathways, and molecular functions affected by these agents. This knowledge graph consists of approximately 100,000 nodes representing biomedical entities and over 5 million edges representing the relationship between these entities. To represent novel drugs lacking mechanistic information in the bio-pharmacological dataset, we manually enriched the knowledge graph by adding connections to newly created compound nodes, based on published preclinical studies. This included, for example, their mechanism of action, additional indications for the drug under investigation, and pharmacodynamics components such as metabolism and excretion. For this project, we ensured that all medications indicated for UC patients were thoroughly represented in the graph, and we enriched the graph further whenever necessary, as described below.\u003c/p\u003e\n\u003ch2\u003eFramework for Model Training and Prediction\u003c/h2\u003e\n\u003ch3\u003eOverall Cohort Definition\u003c/h3\u003e\n\u003cp\u003eThe study cohort included patients with a confirmed diagnosis of UC. To emulate this criterion using RWD, patients eligible for the cohort were required to have at least three documented diagnosis records of ulcerative colitis and at least one prescription for a UC-related medication. The disease start date was defined as the earliest recorded date of a diagnosis of UC. For the training set, patients were permitted to enter the training process multiple times. Entry into the cohort was based on the first occurrence of the disease being classified as mild, moderate, or severe after the disease start date defined above, thus allowing for up to 3 separate cohort entries per patient. The classification of severity was done using the electronic Mayo score depicted in the next section.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis overall cohort was used to train the machine learning model (described in the following section) to be able to predict the joint effect of patient features and treatments on patient outcomes. However, when predicting the outcomes of the clinical trials in the validation set, we used a distinct cohort for each trial, consisting only of patients who met the eligibility criteria for that specific trial. Importantly, during the prediction process, we used the patient characteristics and assign to them the intervention or control arm drugs, rather than using the actual drugs they received in the RWD (described further in the section \u003cem\u003ePrediction of RCT outcomes,\u0026nbsp;\u003c/em\u003ebelow).\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003eEngineering of Patient Outcomes \u0026ndash; EHR-based Mayo Score (eMS)\u003c/h3\u003e\n\u003cp\u003eThe Mayo Score (MS) is a clinical measure of UC activity that is extensively used to evaluate treatment response. The MS has four components: stool frequency, rectal bleeding, endoscopic findings of mucosal inflammation, and physician\u0026rsquo;s global assessment. The partial MS (pMS) composed of stool frequency, bleeding components and physician assessment on a 9-point score has been shown to perform on par with MS in measuring response to UC therapy\u0026nbsp;[24, 25].\u0026nbsp;It has also been used in several clinical trials to monitor responses\u0026nbsp;[23, 26-29].\u0026nbsp;To model UC disease severity outcomes based on the MS, we constructed a retrospective EHR-based Mayo score (eMS) composed of elements that can be found in \u003cstrong\u003eTable 1\u003c/strong\u003e.\u0026nbsp;The weights contributing to the eMS were determined through a combination of expert consultation and empirical\u0026nbsp;calibration. A gastroenterologist specializing in inflammatory bowel disease initialized the weights to reflect clinical assessment of severity. To further calibrate the weights, monthly eMS scores were generated by summing the weighted records and compared to Mayo score distributions from both the literature and curated datasets containing actual scores.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1:\u003c/strong\u003e Retrospective EHR-based Mayo Score\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eElement\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecord Weights\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAbdominal pain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOral steroid usage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFrequent endoscopy procedures (\u0026gt;1 procedure within 6 months)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAnemia due to blood loss\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eHospitalization due to UC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eDiarrhea\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRectal bleeding\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe total eMS score ranges from 0 to 12. A 0-12 scale was first employed and subsequently harmonized to a 0-9 range for consistency with the pMS scoring framework (\u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e). Our validation results using curated data showed a high concordance between the electronic Mayo Score (eMS) and pMS (\u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e). For this study, the harmonized eMS scoring system was categorized into three disease severity levels: mild (eMS 0-2), moderate (3-6), and severe (7-9). The eMS for each patient were calculated on a monthly basis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDisease progression was defined as an increase in eMS by one point in a given month relative to the previous 3-month window. This definition of progression was validated against published retrospective cohort analyses of disease relapses among patients with moderate or severe UC. The eMS score successfully captured UC disease course over time and was internally consistent and concordant with previously published pMS estimates. The eMS\u0026rsquo;s capacity to monitor disease progression was validated through the analysis of the following elements: absolute progression rates relative to literature, relapse rates 1 year after diagnosis, and progression by neutrophil/lymphocyte ratio (NLR) and C-reactive protein (CRP) strata.[22] Results from these validations of the eMS can be seen on \u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e. The predictive model, described further below, was trained to predict progression in patients with UC.\u003c/p\u003e\n\u003ch3\u003eEngineering of Patient Outcomes \u0026ndash; Progression and Remission\u003c/h3\u003e\n\u003cp\u003eAs clinical visits in RWD often reflect episodes of heightened disease activity, they provide a granular view of disease progression and exacerbation, while episodes of low activity in the data can result from missing data rather than \u0026ldquo;real\u0026rdquo; remission. Thus, disease progression events are more directly measurable than remission in real-world settings.\u0026nbsp;We developed a two-stage model to predict remission in line with clinical trial benchmarks, despite the lack of directly observed remission in RWD. First, as described above, we trained our machine learning model to predict UC progression over time using real-world clinical data. Second, we developed an additional model to\u0026nbsp;infer\u0026nbsp;expected patient remission rates in clinical trials based on the predicted treatment effect on UC progression. To do this, we constructed a reference sigmoidal remission curve over time using reported placebo remission rates from clinical trials involving patients with moderate to severe UC, extracted at time points from 5 to 52 weeks post-randomization. We then created a transformation function that maps this reference curve to predicted placebo progression rates, capturing the inverse relationship between progression and remission in UC. Finally, we applied this function to the model\u0026rsquo;s predicted treatment effect on progression (as described below) to estimate corresponding remission rates.\u003c/p\u003e\n\u003ch3\u003eDrug-centered Knowledge Graph Embeddings\u003c/h3\u003e\n\u003cp\u003eThe knowledge graph, constructed from a combination of bio-pharmacological data and manual enrichment, was converted to graph embeddings for model training. We tested several embedding methods, including TransE, TransR, MURE, metapath2vec and graph neural networks, and evaluated the quality of embeddings using a scoring metric that assessed the relative cosine similarity of drugs with similar versus dissimilar mechanisms of action. The embeddings that achieved the highest evaluation score were then used for model training.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eKnowledge Graph Enrichment for Novel Drug Representation\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo enable modeling of treatment effects for novel drugs not represented in the knowledge graph (KG), a process termed KG enrichment is applied. In this process, relevant information is extracted from the literature, including mechanisms of action, drug-drug interactions, molecular targets, side effects, and approved or investigated indications. This information is structured to match the KG format and used to construct a corresponding subgraph. The subgraph is then integrated into the main KG, and updated drug embeddings are generated to include the novel drug, relevant standards of care (SoC), and comparators. These embeddings are subsequently used to train models that predict the effect of the novel drug in comparison to SoC and alternative treatments, leveraging both the newly integrated data and the existing KG context. Throughout the study, such novel drugs whose information was added using this mechanism are referred to as enriched drugs.\u003c/p\u003e\n\u003ch3\u003ePatient-Drug Journey Representation\u003c/h3\u003e\n\u003cp\u003eThe index date for representing patient-drug journey within the predictive model was defined as the first instance when the patient\u0026rsquo;s disease was classified as mild, moderate, or severe during the course of the patient\u0026rsquo;s condition, occurring after the UC start date specified in the cohort definition paragraph above. \u0026ldquo;Patient journey\u0026rdquo; was defined as the time series consisting of clinical events recorded in their real-world clinical data, starting six months prior to their index date, the end date was determined according to the corresponding clinical trial matching date \u0026ndash; therefore allowing to predict remission over different period durations. Data for patient demographics, current and prior treatments, evidence of hospitalization within the last year, high levels of inflammatory markers (namely, C-reactive protein, erythrocyte sedimentation rate and fecal calprotectin levels) were also included as predictors. This data for patient-drug journey was converted into a vector embedding that was then used in the prediction model.\u003c/p\u003e\n\u003ch3\u003eTemporal Prediction Framework \u0026ndash; Model Training and Hyperparameter Optimization\u003c/h3\u003e\n\u003cp\u003eWe developed the ClinBoost algorithm that combines patient-drug journey embeddings and the knowledge graph embedding into the XGBoost algorithm, followed by per-time-window logistic regression. This algorithm includes trainable weights to combine patient clinical data, demographics, and drug embeddings into a single model to predict outcome probabilities for each patient for each time window. This model was trained using real-world clinical data and graph embeddings, with a train/validation/test split of 80:10:10 used for tuning model hyperparameters.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFor hyperparameter optimization, we considered a three-dimensional grid of hyperparameters covering tunable components of (1) drug-centered knowledge graph embedding, (2) clinical data definitions and filters for selection of the overall cohort used for model training, and (3) the prediction model hyperparameters. For hyperparameter optimization, we evaluated the area under the receiver operating characteristics curve (ROC-AUC) or \u003cem\u003ec\u003c/em\u003e-statistics by comparing the predicted patient outcomes to those observed in the validation set of the RWD. After selecting and validating the best hyperparameters using the RWD we tested the best model on a set of validation clinical trials.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eModel Validation using Randomized Controlled Trial (RCT) Data\u003c/h2\u003e\n\u003ch3\u003eSelection of RCTs for Validation\u003c/h3\u003e\n\u003cp\u003eInformation relating to study design and results of the clinical trials used in this study for validation was sourced from Clinicaltrials.gov and PubMed (pubmed.ncbi.nlm.nih.gov). Clinical trials were included if they were phase II or phase III RCTs evaluating a pharmacological treatment for moderate and/or severe ulcerative colitis compared to placebo or standard of care. Trials had to be registered at clinicaltrials.gov, with well-documented inclusion and exclusion criteria that could be modeled using RWD; and to report results on disease remission. Trials without published results (accessible at clinicaltrials.gov or pubmed.ncbi.nlm.nih.gov) were excluded. The 21 trials ultimately included for validation spanned 2007 to 2022.\u003c/p\u003e\n\u003cp\u003eThe following elements were extracted for each trial for the purposes of modeling and validation: inclusion and exclusion criteria, investigational and comparator treatments, endpoint definitions, sample sizes per treatment arm, baseline patient characteristics, and trial results including point estimates and confidence intervals for the relevant endpoints and treatment effect estimates, as well as reported p-values.\u003c/p\u003e\n\u003ch3\u003ePrediction of RCT Outcomes\u003c/h3\u003e\n\u003cp\u003eFor prediction, eligibility criteria for each validation RCT were first applied to select an eligible population of patients from the study\u0026rsquo;s real-world dataset. For each trial, all clinical parameters that could be replicated in RWD were applied; a full list of these parameters is supplied in the supplementary materials (\u003cstrong\u003eAppendix 1\u003c/strong\u003e). Patient cohorts with age and sex distributions, and sample sizes matching those from the corresponding RCT were then sampled to create the prediction cohort. The trained prediction model was used to predict patient-level outcomes for the prediction cohort separately under the treatment and control regimen from the respective trial, regardless of what treatment the patients may have received previously in the real-world clinical data. The prediction of remission rates was performed at the same time point as the corresponding RCT to enable comparability. For subgroup-level simulations, the set of patients sampled were based on additional eligibility criteria for a variety of sample sizes, as required.\u003c/p\u003e\n\u003ch3\u003ePerformance Metrics for Model Validation\u003c/h3\u003e\n\u003cp\u003eTo measure the trial-level validation performance of our model, we evaluated the following criteria:\u0026nbsp;\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eTrial success agreement: A binary analysis examining the accuracy of predicting trial success or failure based on a direction and statistical significance of effect, which is estimated using confidence intervals and \u003cem\u003ep\u003c/em\u003e-values. In other words, the predicted trial should reach the same conclusion as the original RCT. We report four levels of agreement:\u003col style=\"list-style-type: lower-alpha;\"\u003e\n \u003cli\u003e\u003cstrong\u003eConcordant Alignment:\u0026nbsp;\u003c/strong\u003eThe prediction aligns with the original trial in both direction (treatment better than control) and statistical significance (p \u0026lt; 0.05).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eDirectionally Aligned with Threshold Misalignment:\u0026nbsp;\u003c/strong\u003eBoth the original trial and the predicted trial outcome indicate the same direction of effect (treatment better than control for instance), but the prediction fails to reach statistical significance while the original trial was statistically significant.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eParadoxical Agreement:\u0026nbsp;\u003c/strong\u003eBoth the original and emulated trials are statistically non-significant (p \u0026gt; 0.05), but the direction of the effect differs (treatment better in one, control better in the other).\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eComplete Divergence:\u0026nbsp;\u003c/strong\u003eThe original trial shows a statistically significant positive outcome for the treatment (treatment better than control), while the prediction indicates a statistically significant negative outcome (control better than treatment).\u003c/li\u003e\n \u003c/ol\u003e\n \u003c/li\u003e\n \u003cli\u003eAbsolute outcome agreement: This was achieved when the estimate from the\u0026nbsp;predicted\u0026nbsp;RWE study fell within the 95% confidence interval of the corresponding observed RCT estimate. The results were presented visually as forest plots and a Bland-Altman plot.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eCohen\u0026rsquo;s\u0026nbsp;d statistic [30]: The standardized difference between observed and predicted effect sizes was calculated using Cohen\u0026rsquo;s d, which expresses the magnitude of difference in standard deviation units. Values of 0.2, 0.5, and 0.8 were interpreted as small, moderate, and large differences, respectively. This metric enabled comparison of predictive accuracy across trials with varying populations and measurement methodologies. The average Cohen\u0026rsquo;s d across all validation trials was calculated to summarize overall model performance.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eTrial success and failure across 21 trials was evaluated using a confusion matrix and quantified by standard metrics including F1 score, sensitivity, and specificity. Additionally, model performance was also evaluated on trial strata by drug mechanism of action (e.g., biological versus non-biological).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel-guided Trial Design Optimization.\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the utility of the multivariate prediction engine in informing clinical trial design, a series of simulation analyses were conducted. These analyses aimed to identify patient sub-cohorts with differential treatment responses, optimize eligibility criteria, and assess power and sample size requirements for achieving statistically significant results.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCohort Variant Generation and Effect Prediction:\u0026nbsp;\u003c/strong\u003eThe analysis began with a base cohort derived from an existing clinical trial population. Using a multivariate predictive modeling framework, multiple cohort variants were generated by systematically modifying inclusion and exclusion criteria. These criteria included prior exposure to biological therapies (e.g., anti-TNFs), baseline inflammatory marker levels, and systemic steroid use. For each variant, the expected treatment effect, defined as the difference in remission rates between treatment and control arms at a pre-specified time point, was predicted, and the total cohort size was estimated by extrapolating to a representative real-world population.\u003c/p\u003e\n\u003cp\u003ePredicted treatment effects and corresponding cohort sizes were visualized in a scatter plot. Gray markers indicated cohort variants unlikely to achieve statistical significance given the original trial\u0026rsquo;s sample size. From the set of predictions, two cohorts of interest were selected for in-depth comparison: an \u0026ldquo;Optimized\u0026rdquo; cohort with enhanced predicted response, and a \u0026ldquo;Low Effect\u0026rdquo; cohort with diminished efficacy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLongitudinal Outcome Projection:\u0026nbsp;\u003c/strong\u003eWeekly remission rates over a 52-week period were predicted for each selected cohort using the model\u0026rsquo;s time-to-event outputs. These projections enabled longitudinal comparison of remission dynamics across treatment arms and cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePower and Sample Size Estimation:\u0026nbsp;\u003c/strong\u003eAn analysis was performed to evaluate the total sample size needed to reach statistical significance, while maintaining the original treatment-to-control allocation ratio. For each cohort, p-values corresponding to the predicted treatment effect were calculated, and the minimum sample size required to achieve statistical significance (\u0026alpha; = 0.05) was determined.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMonte Carlo Simulation for Probability of Success:\u0026nbsp;\u003c/strong\u003eTo assess variability and robustness in trial outcomes, Monte Carlo simulations were conducted. Patient profiles meeting the inclusion/exclusion criteria for each cohort were repeatedly sampled. For each simulation, the predicted treatment effect was calculated and aggregated into a distribution. The probability of technical success (PTS) was defined as the proportion of simulations in which the treatment effect exceeded a predefined threshold (e.g., 10% remission difference between treatment and control). Resulting effect-size distributions and PTS estimates were visualized.\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003eModeling Framework Overview\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 1\u003c/strong\u003e provides an overview of the predictive model architecture and workflow used in the study. The model was trained using data from a bio-pharmacological database in the form of a knowledge graph and longitudinal real-world clinical data from a linked claims/electronic health records database of patients with UC. Embeddings of the knowledge graph and patient-drug profiles from these two data sources were combined into a gradient-boosted tree model to predict patient-level outcomes on therapy over time.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe study cohort included 482,870 unique patients, which could potentially be recruited up to 3 times in relation to a change in their clinical status. The median age of patients in this cohort was 56 years (ranging 18 to 85 years), 59.5% of whom were women. For the model training and predictions, patients were permitted to enter the simulation multiple times. The demographic, clinical history, and treatment characteristics of this overall study cohort, including the multiple entries (n = 723,263) used for model development, including outcome rates, are provided in \u003cstrong\u003eSupplementary Table S1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003eThe drug-centered knowledge graph and derived drug embeddings are shown in \u003cstrong\u003eFigure 2\u003c/strong\u003e. The graph included both approved therapies represented in the real-world clinical data, as well as novel therapies with known mechanisms of action which were manually added to the graph (\u003cstrong\u003eFigure 2, a\u003c/strong\u003e). The quality of the knowledge graph embeddings was assessed by comparing the cosine similarity of drugs sharing similar mechanisms of action and therapeutic indications. Figure 2b illustrates a two-dimensional t-SNE visualization of the compound embeddings, color-coded by drug groups. We found that the learned embeddings preserve the underlying graph topology (Figure 2c), and that agents with similar mechanisms of action within the graph embedding clustered together preferentially and demonstrated high cosine similarity, as expected (\u003cstrong\u003eFigure 2, b and d\u003c/strong\u003e).\u003c/p\u003e\n\u003ch2\u003eModel Validation Results\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e provides an overview of the 21 randomized controlled trials that were included for the validation of our predictive model. Of these 21, 7 (33.3%) were unsuccessful trials, 5 (23.8%) were phase 2 trials and 6 (29%) included a novel agent that was not commonly administered in the real-world at the time of this study. The trials also encompassed a range of sample sizes (from 28 to 476 patients per arm) and varied timings of outcome measurements (6 to 52 weeks post-randomization). The treatments included diverse mechanisms of action such as biologics (e.g., tumor necrosis factor (TNF) inhibitors) and small molecule drugs (\u003cstrong\u003eTable 2\u003c/strong\u003e). Therefore, we expected this set of 21 trials to provide a sufficiently diverse cohort for evaluating our machine learning model\u0026rsquo;s ability to predict remission rate per study arm, and the treatment effects (i.e., difference in outcomes between treatment arms)\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrial Number\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrial Short Name\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNCT ID\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSuccess/ Failure\u0026nbsp;\u003c/strong\u003e\u003cbr\u003e\u003cstrong\u003e(1/0)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTime of outcome\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eMeasurement (weeks)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment Arm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment Arm Remission (% Patients)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 8px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eControl Arm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eControl Arm Remission (% Patients)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment Arm Target / MoA\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment /Control Arm Size (N)\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge (mean\u0026plusmn;std)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 4px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eStudy phase\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 4px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNovel dug?\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eELEVATE UC 12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03996369\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrasimod\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e24.8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e15.2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eS1PR1 agonist\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e222/112\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03934216\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03934216\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eDeucravacitinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e14.8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e16.3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eselective TYK2 inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e88/43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e42 (\u0026plusmn; 15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eU-Accomplish\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03653026\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eUpadacitinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e33.5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e4.1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e341/174\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e42 (\u0026plusmn; 15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03482635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT03482635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eSpesolimab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e7.1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eAnti IL-36\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e28/23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e44 (\u0026plusmn; 15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eSELECTION\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02914522\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eFilgotinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e26.1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e15.3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e245/137\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eNA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02819635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02819635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eUpadacitinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e26.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e319/154\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e44 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e7\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eVARSITY\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02497469\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e52\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eVedolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e31.3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eAdalimumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e22.5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker/ Anti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e386/383\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e41 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02435992\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02435992\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eOzanimod\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e18.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eS1PR1 agonist\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e429/216\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eNA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02289417\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02289417\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eApremilast\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e21.8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e12.1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003ePDE 4 inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e55/58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e43 (\u0026plusmn; 15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eHIBISCUS 2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02171429\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e18.2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e11.1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e143/72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e41 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e11\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eHIBISCUS 1(A)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02163759\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e19.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eAdalimumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e22.5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e144/142\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eHIBISCUS 1(B)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02163759\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e19.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e6.9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e144/72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e13\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eGARDENIA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02136069\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e20.6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eInfliximab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e32.8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e199/198\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e14\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eHICKORY\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT02100696\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e14\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eEtrolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e18.5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e6.3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e384/95\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e41 (\u0026plusmn; 13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e15\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT01694485\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT01694485\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eAbrilumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e13.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e4.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e79/116\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 12)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e16\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eOCTAVE(A)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT01465763\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eTofacitinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e18.5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e8.2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e476/112\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eNA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e17\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eOCTAVE(B)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT01458951\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eTofacitinib\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e16.6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e3.6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e429/112\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003eNA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e18\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eGEMINI 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT00783718\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eVedolizumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e16.9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e5.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e225/149\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e19\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT00488774\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT00488774\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eGolimumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e13\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e11\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e77/73\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e40 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e20\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePURSUIT-SC\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT00487539\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eGolimumab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e17.9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e6.4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e257/251\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e41 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e21\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eACT 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003eNCT00096655\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 9px;\"\u003e\n \u003cp\u003eInfliximab\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e38.8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 8px;\"\u003e\n \u003cp\u003ePlacebo\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e14.9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 6px;\"\u003e\n \u003cp\u003e121/121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e42 (\u0026plusmn; 14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 4px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2:\u0026nbsp;\u003c/strong\u003eUC Validation Trials and Experimental Drug MoA\u003c/p\u003e\n\u003cp\u003eThe 21 trials were chosen to assess model performance in predicting trial outcome. A diverse validation set was composed, including trials with different drug modalities, phase 2 or 3, novel vs. approved drugs, and different study sizes. Prediction cohorts include age and sex distributions similar to original trial, and the prediction of remission rates were done at the actual time point of the trials\u0026rsquo; outcome measure and with similar arm size.\u003c/p\u003e\n\u003cp\u003eModel validation results are presented in \u003cstrong\u003eTable 3\u003c/strong\u003e and \u003cstrong\u003eFigure 3\u003c/strong\u003e. Our model correctly predicted the effect size \u0026ndash; defined as the difference in remission rates between treatment and control groups \u0026ndash; for 16 of the 21 trials (76.2%), correctly matching the direction and statistical significance of the corresponding RCT. These concordant predictions would have led to the same conclusions regarding trial outcome (\u003cstrong\u003eTable 3\u003c/strong\u003e). Among the five trials where predictions did not fully align with the observed trial outcomes, four (80%) showed effect sizes in the same direction as the RCTs, but differed in statistical significance, which affected the conclusion. Only one trial, VARSITY (evaluating adalimumab vs vedolizumab), showed a misclassification in the direction of the effect size. This trial had the longest follow-up duration (52 weeks), and the discrepancy stemmed primarily from inaccurate predictions in the control arm. Notably, in another adalimumab trial HIBISCUS-1 (NCT02163759), the model accurately predicted the effect size at week 10, fully consistent with the corresponding RCT (\u003cstrong\u003eFigure 3a\u003c/strong\u003e). Importantly, for all six trials involving novel agents, the model\u0026rsquo;s predicted effect sizes aligned with those from the RCTs, supporting the same regulatory decisions. Overall, the model exhibits an F1 score of 0.8 when compared to RCT data on the binary level of success/failure agreement (\u003cstrong\u003eFigure 3, b and c\u003c/strong\u003e). Furthermore, the trained model performance was evaluated on the patient level real world data, and a high performing model was selected using the ROC-AUC (\u003cstrong\u003eFigure 3d\u003c/strong\u003e). The model performed slightly better when predicting outcomes at shorter follow-up lengths of time. Nonetheless, the AUC declined by only 7.2% when choosing a simulated follow-up length of 10 days and up to 400 days (AUC 0.83 and 0.77, respectively (\u003cstrong\u003eFigure 3d\u003c/strong\u003e)\u003cs\u003e.\u0026nbsp;\u003c/s\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3:\u003c/strong\u003e Effect Size Estimates and Agreement Metrics\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrial N\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrial name\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTreatment Arm Target / MoA\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eObserved Treatment Effect*\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 17px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePredicted Treatment Effect**\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDifference\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAgreement\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eELEVATE UC 12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eS1PR1 agonist\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.1 (0.04, 0.15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.15 (0.09, 0.2)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.05 (-0.03, 0.13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT03934216\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eselective TYK2 inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.01 (-0.1, 0.07)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.06 (-0.03, 0.14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.07 (-0.05, 0.19)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e3\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eU-Accomplish\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.29 (0.25, 0.34)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.11 (0.07, 0.15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.18 (-0.24, -0.12)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e4\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT03482635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eAnti IL-36\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.07 (-0.02, 0.15)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.04 (-0.12, 0.04)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.11 (-0.21, -0.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e5\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eSELECTION\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.11 (0.05, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.19 (0.13, 0.24)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.08 (0.00, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e6\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT02819635\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.21 (0.17, 0.26)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.12 (0.08, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.09 (-0.15, -0.03)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e7\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eVARSITY\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker/ Anti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.09 (0.04, 0.13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.15 (-0.19, -0.1)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.23 (-0.30, -0.17)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e8\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT02435992\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eS1PR1 agonist\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.12 (0.09, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.1 (0.06, 0.14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.02 (-0.07, 0.03)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e9\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT02289417\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003ePDE 4 inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.1 (-0.0, 0.19)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.25 (0.13, 0.35)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.15 (0.00, 0.30)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e10\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHIBISCUS 2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.07 (0.0, 0.14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.05 (-0.02, 0.12)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.02 (-0.11, 0.07)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e11\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHIBISCUS 1(A)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.03 (-0.1, 0.04)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.13 (-0.2, -0.07)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.10 (-0.20, -0.01)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e12\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHIBISCUS 1(B)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.12 (0.06, 0.19)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.05 (-0.02, 0.11)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.08 (-0.17, 0.01)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e13\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eGARDENIA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.12 (-0.18, -0.06)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e-0.12 (-0.18, -0.06)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.00 (-0.08, 0.09)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e14\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eHICKORY\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.12 (0.08, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.01 (-0.03, 0.04)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.11 (-0.17, -0.06)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e15\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT01694485\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.09 (0.03, 0.15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.04 (-0.02, 0.1)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.05 (-0.13, 0.03)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e0\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e16\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eOCTAVE(A)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.1 (0.06, 0.14)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.08 (0.04, 0.11)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.03 (-0.08, 0.02)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e17\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eOCTAVE(B)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eJAK inhibitor\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.13 (0.09, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.08 (0.04, 0.11)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.05 (-0.10, -0.00)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e18\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eGEMINI 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eIntegrin blocker\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.11 (0.07, 0.16)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.04 (0.0, 0.08)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.07 (-0.13, -0.01)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e19\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eNCT00488774\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.02 (-0.05, 0.08)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.06 (-0.0, 0.12)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.04 (-0.05, 0.13)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e20\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003ePURSUIT-SC\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.12 (0.08, 0.15)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.06 (0.03, 0.09)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.06 (-0.11, -0.00)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 7px;\"\u003e\n \u003cp\u003e21\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eACT 1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003eAnti TNF-alpha\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.24 (0.16, 0.31)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 17px;\"\u003e\n \u003cp\u003e0.11 (0.05, 0.17)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e-0.13 (-0.22, -0.03)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11px;\"\u003e\n \u003cp\u003e1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*Observe Treatment Effect is the difference in remission rates between the investigational and control groups of the actual clinical trials; 95% CI is provided around this estimate.\u003c/p\u003e\n\u003cp\u003e**Predicted Treatment Effect is the difference in remission rates between the investigational and control groups in the predicted results; 95% CI is provided around this estimate.\u003c/p\u003e\n\u003cp\u003eColumn \u0026ldquo;Difference\u0026rdquo; shows the difference between the clinical trial Observed and the Predicted Treatment Effect sizes; 95% CI is provided around this estimate.\u003c/p\u003e\n\u003cp\u003eColumn \u0026ldquo;Agreement\u0026rdquo; corresponds to the evaluation metrics \u0026ldquo;Concordant alignment\u0026rdquo; described in section Performance Metrics for Model Validation of the method section.\u003c/p\u003e\n\u003cp\u003eWe also showed visually the difference in point estimates between predicted versus observed treatment effects (\u003cstrong\u003eFigure 4 a and b\u003c/strong\u003e). The forest plots present the absolute difference between observed and predicted treatment effect, expressed in the original units of measurement (percentage points), with the related 95% CI (Figure 4 a). These differences were relatively small on average, ranging from 0 for the GARDENIA trial to a maximum of \u0026ndash;0.23 for the VARSITY trial, with a median difference of -0.05\u003cstrong\u003e.\u0026nbsp;\u003c/strong\u003eThe standardized mean difference (\u003cstrong\u003eFigure 4b)\u0026nbsp;\u003c/strong\u003eshows a narrower dispersion\u003c/p\u003e\n\u003cp\u003eTo further evaluate model performance, we applied Bland\u0026ndash;Altman analysis to quantify agreement between predicted and observed remission outcomes. Figure 4c presents the Bland\u0026ndash;Altman plot comparing predicted and observed remission rates at the arm level (treatment and control). The mean difference across arms was minimal (\u0026ndash;0.009), indicating no substantial systematic bias, with most trial arms falling within the 95% limits of agreement (LoA: \u0026ndash;0.148 to 0.129), suggesting good concordance. However, a trend of increasing prediction error was observed for arms with higher average remission rates (\u0026gt;0.2), indicating that model accuracy may decline slightly in high-response scenarios.\u003c/p\u003e\n\u003cp\u003eFigure 4d shows a Bland\u0026ndash;Altman plot comparing predicted versus observed effect sizes. The average difference was \u0026ndash;0.045, indicating a slight negative bias, i.e., modest underestimation of treatment effect. Despite this bias, prediction accuracy was consistent across the range of effect sizes, with the majority of values well within the LoA bounds (\u0026ndash;0.219 to 0.129), demonstrating stable model behavior across trials with both small and large treatment effects.\u003c/p\u003e\n\u003cp\u003eThese analyses confirm that ClinBoost provides accurate and unbiased estimates of both individual arm outcomes and relative treatment effects, with persistent performance across a diverse range of trial conditions\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSub-analyses on trial strata were conducted to compare the prediction accuracy according to the type of the administered drugs, (biological products vs. small molecules) and the drug development phase (phase II or III clinical trials). We found that the predictive model performed slightly better on average for\u0026nbsp;trials involving non-biologic agents (F1 score 0.94 versus 0.80 for trials with biologics) and those predicting the results of phase III trials compared to phase II trials (F1 score of 0.90 versus 0.75; \u003cstrong\u003eFigure 5\u003c/strong\u003e). Furthermore, analysis comparing the prediction accuracy according to the drug\u0026rsquo;s mechanism of actions showed the best results for JAK inhibitors (5 trials) compared to anti-integrins (5 trials) showing F1 values of 1 and 0.75, respectively; a third group that included other mechanism of actions revealed an F1 score of 0.84 (\u003cstrong\u003eFigure 5\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFinally, the model\u0026rsquo;s flexibility was evaluated by processing trials involving pharmacological agents that required manual graph integration, and that were withheld from the training RWD dataset. To enable this, the KG was updated to map the mechanistic attributes of these agents, followed by the generation of corresponding embeddings. The subset of trials utilizing this manual enrichment process (6 trials) showed robust alignment with observed results (F1 = 1.0) compared to the standard automated pipeline (15 trials; F1 = 0.80). These findings validate the consistency of the graph embedding technique, demonstrating that mechanistic signal can be effectively preserved even when integrating pharmacological profiles that were not part of the primary training data.\u0026nbsp;\u003cstrong\u003eFigure 5:\u0026nbsp;\u003c/strong\u003eStratified Performance of the Predictive Model by Clinical Trial Design\u003c/p\u003e\n\u003ch2\u003eModel-guided Trial Design\u0026nbsp;Optimization\u003c/h2\u003e\n\u003cp\u003eA key application of drug response modelling is to inform clinical trial design by predicting responder populations, optimizing eligibility criteria and improving sample size estimation. To demonstrate this, we first evaluated predicted treatment outcomes for the ELEVATE UC-12 trial base cohort, as well as various sub-cohorts after applying additional inclusion and exclusion criteria. ELEVATE UC-12 was selected as a recent clinical trial with modestly positive treatment efficacy findings of a novel MoA (not observed in RWD)\u0026nbsp;and for which enrichment for specific responder subpopulations could have afforded improved clinical outcomes. Treatment effect predictions were plotted alongside cohort size to assess treatment effect heterogeneity within these sub-cohorts, which revealed distinct response subgroups of patients with higher and lower effect sizes than those in the base cohort (\u003cstrong\u003eFigure 6\u003c/strong\u003e). Specifically, \u003cstrong\u003eFigure 6a\u003c/strong\u003e shows how adding or omitting different combinations of inclusion/exclusion criteria change the results of the trial in terms of treatment effect and sample size (i.e., the available number of patients in the US). The predicted remission rates over time for selected cohorts, including the base cohort, a responding optimized cohort, and a low treatment effect cohort, are presented with their associated p-values over a range of sample sizes (\u003cstrong\u003eFigures 6b\u0026nbsp;\u003c/strong\u003eand\u003cstrong\u003e\u0026nbsp;6c)\u003c/strong\u003e. \u0026nbsp;Comparing the original validation cohort to the optimized sub-cohort in a phase 3 trial setting, we observed a significant improvement in remission rates at 24 weeks post-randomization in our simulations (\u003cstrong\u003eFigure 6b\u003c/strong\u003e). Additionally, it was possible to estimate the minimum sample size that would be required to achieve statistically significant results using the model (\u003cstrong\u003eFigure 6c\u003c/strong\u003e). Patient-level simulations across varying arm sizes demonstrated that this optimized sub-cohort required fewer participants to achieve statistical significance, as determined by a standard p-value threshold of 0.05 (\u003cstrong\u003eFigure 6c\u003c/strong\u003e). Finally, we simulated the distribution of possible trial results for each cohort and computed the probability of technical success (PTS) for a specific threshold of remission difference (in this case 10% more remission in the treatment vs. the control arm). \u003cstrong\u003eFigure 6d\u0026nbsp;\u003c/strong\u003eshows in how many simulations the treatment effect exceeded the threshold with PTS variations according to the type of cohort. These findings emphasize the potential of model-guided patient selection in enhancing clinical trial efficiency and success rates.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur study demonstrates that a predictive model leveraging a combined knowledge graph and RWD-based machine learning framework can effectively generate out-of-sample drug response predictions that align well with clinical trial findings in UC. This is evidenced by the strong convergence between observed and model-predicted trial results, with 16 of 21 comparisons showing an agreement regarding the direction, the effect sizes, and the statistical significance of the results. This robust performance extended across a diverse set of trials encompassing various drug mechanisms (11 small molecules and 10 biologics), phases (Phase II and III), and endpoint duration (6 to 52 weeks), showcasing the model\u0026apos;s adaptability. Notably, the model showed high concordance with outcomes for all seven trials involving manually enriched agents, highlighting the utility of the knowledge graph framework in estimating effects based on pharmacological attributes, even when historical utilization data is unavailable. While the alignment was broadly strong, some discrepancies were observed.\u0026nbsp;\u003cbr\u003e\u0026nbsp;In four cases out of the five where full concordance was not achieved, the effect sizes remained directionally consistent with the RCTs, but conclusions diverged due to differences in statistical significance. Only one comparison -the VARSITY trial\u0026mdash;showed a contradictory result between the original RCT that showed a positive effect for vedolizumab [0.09 (0.03-0.15)] in contrast to the model-generated effect size that indicated a negative effect compared to adalimumab [-0.15 (-0.21, -0.08)], both statistically significant but in opposite directions. A plausible explanation for this outlier is the extended follow-up duration (52 weeks) in VARSITY, which is substantially longer than most other trials (8 to 14 weeks, with most spanning 8\u0026ndash;10 weeks) and can introduce increased uncertainty from informative censoring in RWD-derived estimates and potential for uncontrolled concomitant interventions. The differences in effect size between observed and predicted results were generally small, ranging from 0 to -0.23 with a median of -0.05 and a mean of -0.045 (SD: 0.09; \u003cstrong\u003eTable 3\u003c/strong\u003e). \u003cstrong\u003eFigure 4 a \u0026amp; b\u003c/strong\u003e visually confirm the alignment of predictive model results with actual RCTs through forest plots of treatment versus control. The Bland-Altman plots (\u003cstrong\u003eFigure 4 c \u0026amp; d\u003c/strong\u003e) further demonstrate minimally significant differences and little systematic bias, particularly at lower remission rates (\u0026lt;20%) or for novel agents (in red). Our efforts build upon and expand previous work that uses real-world evidence to emulate clinical trial findings, including target trial emulation frameworks and initiatives like RCT DUPLICATE [31-33]. Unlike replication-focused studies, our approach incorporates knowledge graphs to enable outcome estimation for mechanisms not yet represented in the RWD \u0026ndash; a significant advantage in scenarios where historical utilization data is unavailable.. The successful validation across 21 trials represents a more extensive comparison than typically reported in studies validating results in cohort-level predicted trial results [11]. Importantly, our validation went beyond binary success/failure predictions to assess the accuracy of precise treatment effect estimates, matching clinical endpoints, patient eligibility criteria, evaluation time points, and sample sizes with the original trials. The use of multiple performance metrics, including trial success agreement, absolute outcome agreement, Cohen\u0026rsquo;s d, and F1 scores, provided a comprehensive evaluation of the model\u0026rsquo;s performance against contemporary regulatory standards. This rigorous, multi-faceted validation provides a strong basis for confidence in the model\u0026apos;s outputs.\u003c/p\u003e\n\u003cp\u003eConsidering the substantial heterogeneity among patients with UC, characterized by varying clinical expressions and disease trajectories including frequent relapses versus more stable courses, achieving a strong concordance between predicted and RCT observed results is noteworthy. This heterogeneity is driven by multiple factors, including environmental factors, diet, life stress, psychological influences, and genetics. Additionally, the diversity of studied drugs further contributes to the variability [34, 35]. Our analysis encompassed seven drug types, each with distinct mechanisms of action, spanning two major categories: small molecules (JAK inhibitors, S1P receptor modulators, TYK2 inhibitors, PDE4 inhibitors) and biologics (monoclonal antibodies targeting TNF-\u0026alpha;, \u0026beta;7 integrin, and IL-36). The high consistency of results across this heterogeneity underscores the model\u0026rsquo;s robustness, predictive reliability, and validity in the context of UC treatment. Several factors likely contribute to the model\u0026rsquo;s robustness: the comprehensive knowledge graph capturing diverse mechanisms of action, rigorous validation of individual outcome model components and, critically, the large RWD sample size (N=723,263 patient events) facilitating well-matched cohort creation.\u003c/p\u003e\n\u003cp\u003eA critical application of a validated drug response model lies in its capacity to inform clinical trial design. Focusing predictions on specific patient subgroups allow to preemptively identify responder subgroups and estimate treatment effects across diverse patient profiles, paving the way for personalized therapeutic strategies.\u0026nbsp;\u003cstrong\u003eSupplementary Figure S2 and Table S2\u003c/strong\u003e show factors associated with modified treatment effects in response to S1P modulator treatment. For instance, men under 40 without prior biologic exposure exhibited enhanced response.\u0026nbsp;Subgroup identification also directly informs cohort optimization (\u003cstrong\u003eFigure 6 a \u0026amp; b\u003c/strong\u003e) and increases the probability of technical success in clinical trials (\u003cstrong\u003eFigure 6 c \u0026amp; d\u003c/strong\u003e).\u0026nbsp;Notably, our drug response model predictions for optimized patient cohorts in ELEVATE-UC-12 aligned closely with post-hoc analyses of the trial - analyses that were not available at the time of the study. These retrospective analyses confirmed that patients with low inflammatory markers who were na\u0026iuml;ve to biologic therapy consistently showed higher response rates to etrasimod (an S1P modulator)\u0026nbsp;[36]. Additionally, patients not receiving concomitant steroid therapy demonstrated increased response to S1P modulator treatment in the post-hoc analyses\u0026nbsp;[37].\u0026nbsp;Future research should elucidate whether the identified predictive markers of response are specific to S1P modulator therapy or represent generally predictive factors for better response to induction therapy in ulcerative colitis. Overall, these convergent findings validate the use of real-world data-informed predictive drug response modeling to guide future clinical trial design and advance personalized therapy development.\u003c/p\u003e\n\u003cp\u003eIt is important to emphasize that predictive models should be applied within the context in which they were validated. In our case, induction therapy trials may be more amenable to implementing these insights compared to maintenance therapy trials, given the specific patient populations and treatment paradigms used in model development. Beyond trial optimization, the validation and modelling frameworks presented here aim to reinforce the evolving role of RWE in drug development and evaluation. While RCTs remain the gold standard, our findings suggest that well-designed RWE studies, enhanced by appropriate ML approaches, can confirm and extend RCT findings to broader, more heterogeneous real-world populations where adherence and concomitant care differ, and provide out-of-sample estimations for novel drug classes. This offers a robust framework for comparative research in a context where head-to-head trials are scarce, and where assessing efficacy in smaller subpopulations is often impractical. Here, validated computational approaches can provide key signals to warrant investment in larger studies involving primary collection of healthcare data. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe strong concordance between our real-world predictions and RCT findings, established through a comprehensive validation process, reinforces the reliability of such model-based RWE approaches and has significant implications for regulatory and clinical decision making. As healthcare practitioners, drug developers and regulatory agencies increasingly consider the role of AI and RWD in their processes, robust, patient level, validation frameworks offer tangible pathways to evaluate and integrate these data sources effectively. Demonstrating alignment between predicted outcomes and trial data builds confidence in using predictive models for regulatory and clinical decision-making. Gradual implementations of such models could support prioritizing resource allocation, optimizing designs, and assessing the feasibility of new indications or stratifications before undertaking expensive RCTs. As regulatory agencies increasingly integrate RWE into decision-making processes, RWD models have the potential to support adaptive approvals, label expansions, and health technology assessments where traditional trials are impractical. Given its demonstrated validity across multiple drug classes, this approach could be extended to other disease areas, broadening its applicability in real-world drug evaluation and accelerating the translation of biomedical knowledge into tangible patient benefit.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study demonstrates the value of integrating RWD, biomedical knowledge graphs, and machine learning to improve the prediction of clinical trial outcomes in UC. The ClinBoost framework showed strong concordance with observed results across 21 RCTs, accurately reproducing treatment effects in diverse settings, including those involving novel agents not represented in existing datasets. By leveraging enriched knowledge graph embeddings and longitudinal patient journeys, the model provided unbiased estimates of remission outcomes and treatment effects while identifying responder subgroups that could enhance trial efficiency. Importantly, ClinBoost not only emulated past trials but also offered actionable insights to optimize trial design, such as refining eligibility criteria, guiding sample size estimation, and improving probability of technical success. These findings highlight the potential of data-driven modeling to complement traditional randomized trials by informing trial strategies, reducing uncertainty, and accelerating drug development. Broader application of this approach may extend to other therapeutic areas, supporting more efficient, personalized, and cost-effective pathways for evaluating novel interventions.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbbreviation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDefinition\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eCDM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eCommon Data Model\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eCI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eConfidence Interval\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eCRP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eC-Reactive Protein\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eEHDS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eEuropean Health Data Space\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eEHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eElectronic Health Records\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eeMS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eElectronic Mayo Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eJAK\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eJanus Kinase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eKG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eKnowledge Graph\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eLoA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eLimits of Agreement\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eMachine Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eMoA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eMechanism of Action\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eMS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eMayo Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eNLR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eNeutrophil/Lymphocyte Ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eOMOP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eObservational Medical Outcomes Partnership\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003ePDE4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003ePhosphodiesterase 4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003epMS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003ePartial Mayo Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003ePTS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eProbability of Technical Success\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eRCT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eRandomized Controlled Trial\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eRWD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eReal-World Data\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eRWE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eReal-World Evidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eStandard Deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eSoC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eStandard of Care\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eSphingosine-1-Phosphate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eSphingosine-1-Phosphate\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eTNF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eTumor Necrosis Factors\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003et-SNE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003et-distributed Stochastic Neighbor Embedding\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eTYK2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eTyrosine Kinase 2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 62px;\"\u003e\n \u003cp\u003eUlcerative Colitis\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis study was funded by Sanofi.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank Jean-Paul Collet and Maninder Anand of Evidinno Outcomes Research Inc. (Vancouver, BC, Canada) for medical writing assistance.\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eF.D., O.M., and R.H. conceived and supervised the study,\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eF.D. and O.M. led the study design and the interpretation of results.\u003c/p\u003e\n\u003cp\u003eM.S., O.M., A.A., and A.P. led the data curation, preprocessing, cohort definition, and development of the modeling framework and predictive models, including the integration of real-world data and knowledge graph components.\u003c/p\u003e\n\u003cp\u003eF.D., O.M., M.S., and A.P. led the design of the statistical validation and simulation analyses.\u003c/p\u003e\n\u003cp\u003eA.K. and B.S. contributed domain expertise in ulcerative colitis, assessed clinical relevance, and supported interpretation of the study findings.\u003c/p\u003e\n\u003cp\u003eF.D., O.M., M.S., and A.P. jointly drafted, wrote, and revised the manuscript. All authors critically reviewed the manuscript, contributed to its final version, and approved it for submission.\u003c/p\u003e\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eFlavio Dormont, Annie Kruger, and Ramon Hernandez are employees and shareholders of Sanofi (Morristown, NJ, USA). Amina Alaskarov, Amichai Perlman, Omri Matalon, and Michael Shapiro are employees of QuantHealth (Tel Aviv, Israel), which was contracted by Sanofi to conduct this study.\u0026nbsp;Dr. Sands reports personal fees and non-financial support from Abbvie, personal fees from Alimentiv, personal fees from Adiso Therapeutics, personal fees from Agomab Therapeutics, personal fees from Amgen, personal fees from AnaptysBio, personal fees and non-financial support from AstraZeneca, personal fees from Biolojic Design, personal fees from Biora Therapeutics, personal fees from Boehringer Ingelheim, personal fees and non-financial support from Celltrion, personal fees from Equilium, personal fees from Ensho Therapeutics, personal fees from Enveda Biosciences, personal fees from Evommune, personal fees from Ferring, personal fees from Fzata, personal fees from Galapagos, personal fees from Genentech (Roche), personal fees from Gilead Sciences, personal fees from GlaxoSmithKline, personal fees from Gossamer Bio, personal fees from Imhotex, personal fees from Immunyx Therapeutics, personal fees from Index Pharmaceuticals, personal fees from Innovation Pharmaceuticals, grants, personal fees and non-financial support from Janssen/J\u0026amp;J Innovative Medicine, personal fees from Kaleido, personal fees from Kallyope, personal fees and non-financial support from Merck \u0026amp; Co., personal fees and non-financial support from Merck Sharp \u0026amp; Dohme, personal fees from Microba, personal fees from Microbiotica, personal fees from Mitsubishi Tanabe Pharma, personal fees from Mobius Care, personal fees from Morphic Therapeutics, personal fees and non-financial support from Eli Lilly \u0026amp; Sons, personal fees from MRM Health, personal fees from Nexus Therapeutics, personal fees from Nimbus Discovery, personal fees from Odyssey Therapeutics, personal fees from Palisade Bio, personal fees and non-financial support from Prometheus Biosciences, personal fees from Prometheus Laboratories, personal fees and non-financial support from Pfizer, personal fees from Protagonist Therapeutics, personal fees from Q32 Bio, personal fees from Rasayana Therapeutics, personal fees from Recludix Therapeutics, personal fees from Reistone Biotherapeutics, personal fees from Sanofi, personal fees from Sorriso Pharmaceuticals, personal fees from Surrozen, personal fees from Target RWE, personal fees and non-financial support from Takeda, personal fees from Teva, personal fees from TLL Pharmaceutical, personal fees from TR1X, personal fees from Union Therapeutics, personal fees and non-financial support from Abivax, grants, personal fees and non-financial support from Bristol Myers Squibb, personal fees from Theravance Biopharma, personal fees, non-financial support and other from Ventyx Biopharma, outside the submitted work.\u003c/p\u003e\n\u003cp\u003eData Availability\u003c/p\u003e\n\u003cp\u003eThe datasets used in this study, including real-world data (RWD) derived from electronic medical records and insurance claims, as well as components of the biomedical knowledge graph, were obtained under license from proprietary vendors and are not publicly available. Due to contractual and privacy restrictions, we are unable to share the underlying data. Researchers interested in accessing the datasets may contact the corresponding author to discuss the possibility of data access, subject to licensing agreements and confidentiality obligations.\u003c/p\u003e\n\u003cp\u003eEthics Approval and Consent to Participate\u003c/p\u003e\n\u003cp\u003eAll methods were carried out in accordance with relevant guidelines and regulations. This study utilized retrospective, de-identified patient-level data from the PurpleLab\u0026reg; insurance open claims database and electronic health records (EHR) from the EVERSANA EHR database (EVERSANA Life Sciences Inc.), which were linked via a third-party tokenization service (Datavant Inc., San Francisco, CA, USA). The datasets contain anonymized information on healthcare encounters, procedures, medications, diagnoses, and demographic characteristics, with no direct patient identifiers available to the researchers. As a retrospective analysis of de-identified data, this study was exempt from institutional review board (IRB) review under 45 CFR 46.104(d)(4), and the requirement for informed consent was waived.\u003c/p\u003e\n\u003cp\u003eConsent for Publication\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eHwang TJ, Carpenter D, Lauffenburger JC, Wang B, Franklin JM, Kesselheim AS. Failure of Investigational Drugs in Late-Stage Clinical Development and Publication of Trial Results. JAMA internal medicine. 2016;176(12):1826-33.doi:10.1001/jamainternmed.2016.6008.\u003c/li\u003e\n\u003cli\u003eFarid SS, Baron M, Stamatis C, Nie W, Coffman J. Benchmarking biopharmaceutical process development and manufacturing cost contributions to R\u0026amp;D. mAbs. 2020;12(1):1754999.doi:10.1080/19420862.2020.1754999.\u003c/li\u003e\n\u003cli\u003eMullard A. Parsing clinical success rates. Nature Reviews Drug Discovery. 2016;15(7):447-.doi:10.1038/nrd.2016.136.\u003c/li\u003e\n\u003cli\u003eSchlander M, Hernandez-Villafuerte K, Cheng CY, Mestre-Ferrandiz J, Baumann M. How Much Does It Cost to Research and Develop a New Drug? A Systematic Review and Assessment. PharmacoEconomics. 2021;39(11):1243-69.doi:10.1007/s40273-021-01065-y.\u003c/li\u003e\n\u003cli\u003eAugustine EF, Adams HR, Mink JW. Clinical trials in rare disease: challenges and opportunities. Journal of child neurology. 2013;28(9):1142-50.doi:10.1177/0883073813495959.\u003c/li\u003e\n\u003cli\u003eRehman AU, Li M, Wu B, Ali Y, Rasheed S, Shaheen S, et al. Role of Artificial Intelligence in Revolutionizing Drug Discovery. Fundamental Research. 2024.doi:https://doi.org/10.1016/j.fmre.2024.04.021.\u003c/li\u003e\n\u003cli\u003eFu T, Huang K, Xiao C, Glass LM, Sun J. HINT: Hierarchical interaction network for clinical-trial-outcome predictions. Patterns. 2022;3(4):100445.doi:https://doi.org/10.1016/j.patter.2022.100445.\u003c/li\u003e\n\u003cli\u003eWang Z, Xiao C, Sun J. SPOT: Sequential Predictive Modeling of Clinical Trial Outcome with Meta-Learning. Proceedings of the 14th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics; Houston, TX, USA: Association for Computing Machinery; 2023. p. Article 53.\u003c/li\u003e\n\u003cli\u003eAl-Mohaya M, Demiralp B, Kurt A, \u0026Ccedil;elik Y. In silico approaches which are used in pharmacy. Journal of Applied Pharmaceutical Science. 2024.doi:10.7324/JAPS.2024.154854.\u003c/li\u003e\n\u003cli\u003eDuruisseaux M, Masson P, Nativel A, Girard N, Cadranel J, Swalduz A, et al. Prospective validation of in silico clinical trials prediction using an EGFR-mutated NSCLC mechanistic computational disease model. Journal of Clinical Oncology. 2024;42(16_suppl):8614-.doi:10.1200/JCO.2024.42.16_suppl.8614.\u003c/li\u003e\n\u003cli\u003eAngoulvant D, Granjeon-Noriot S, Amarenco P, Bastien A, Bechet E, Boccara F, et al. In-silico trial emulation to predict the cardiovascular protection of new lipid-lowering drugs: an illustration through the design of the SIRIUS programme. European journal of preventive cardiology. 2024;31(15):1820-30.doi:10.1093/eurjpc/zwae254.\u003c/li\u003e\n\u003cli\u003eBrbić M, Yasunaga M, Agarwal P, Leskovec J. Predicting drug outcome of population via clinical knowledge graph. medRxiv : the preprint server for health sciences. 2024.doi:10.1101/2024.03.06.24303800.\u003c/li\u003e\n\u003cli\u003eJoshi P, V M, Mukherjee A. A knowledge graph embedding based approach to predict the adverse drug reactions using a deep neural network. Journal of biomedical informatics. 2022;132:104122.doi:10.1016/j.jbi.2022.104122.\u003c/li\u003e\n\u003cli\u003eCarracedo-Reboredo P, Li\u0026ntilde;ares-Blanco J, Rodr\u0026iacute;guez-Fern\u0026aacute;ndez N, Cedr\u0026oacute;n F, Novoa FJ, Carballal A, et al. A review on machine learning approaches and trends in drug discovery. Computational and Structural Biotechnology Journal. 2021;19:4538-58.doi:https://doi.org/10.1016/j.csbj.2021.08.011.\u003c/li\u003e\n\u003cli\u003eFranklin J, Schneeweiss S. When and How Can Real World Data Analyses Substitute for Randomized Controlled Trials? Clinical Pharmacology \u0026amp; Therapeutics. 2017;102.doi:10.1002/cpt.857.\u003c/li\u003e\n\u003cli\u003eRaman K, Kumar R, Musante CJ, Madhavan S. Integrating Model-Informed Drug Development With AI: A Synergistic Approach to Accelerating Pharmaceutical Innovation. Clinical and translational science. 2025;18(1):e70124.doi:10.1111/cts.70124.\u003c/li\u003e\n\u003cli\u003eSharma A, Lysenko A, Jia S, Boroevich KA, Tsunoda T. Advances in AI and machine learning for predictive medicine. Journal of Human Genetics. 2024;69(10):487-97.doi:10.1038/s10038-024-01231-y.\u003c/li\u003e\n\u003cli\u003eNicholson DN, Greene CS. Constructing knowledge graphs and their biomedical applications. Comput Struct Biotechnol J. 2020;18:1414-28.doi:10.1016/j.csbj.2020.05.017.\u003c/li\u003e\n\u003cli\u003eChandak P, Huang K, Zitnik M. Building a knowledge graph to enable precision medicine. Scientific Data. 2023;10(1):67.doi:10.1038/s41597-023-01960-3.\u003c/li\u003e\n\u003cli\u003eGao Z, Ding P, Xu R. KG-Predict: A knowledge graph computational framework for drug repurposing. Journal of biomedical informatics. 2022;132:104133.doi:10.1016/j.jbi.2022.104133.\u003c/li\u003e\n\u003cli\u003eFDA. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products 2025 [Available from: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological; \u003c/li\u003e\n\u003cli\u003eLi P, Wu Y, Xiong W, Cao J, Chen M, Yuan Z, et al. Association between the immune-inflammation index and the severity and clinical outcomes of patients with inflammatory bowel disease: a systematic review and meta-analysis. BMC Gastroenterology. 2025;25(1):414.doi:10.1186/s12876-025-04033-4.\u003c/li\u003e\n\u003cli\u003eSands BE, Peyrin-Biroulet L, Loftus EV, Danese S, Colombel J-F, T\u0026ouml;r\u0026uuml;ner M, et al. Vedolizumab versus Adalimumab for Moderate-to-Severe Ulcerative Colitis. New England Journal of Medicine. 2019;381(13):1215-26.doi:doi:10.1056/NEJMoa1905725.\u003c/li\u003e\n\u003cli\u003eNaegeli AN, Hunter T, Dong Y, Hoskin B, Middleton-Dalby C, Hetherington J, et al. Full, Partial, and Modified Permutations of the Mayo Score: Characterizing Clinical and Patient-Reported Outcomes in Ulcerative Colitis Patients. Crohn\u0026apos;s \u0026amp; colitis 360. 2021;3(1):otab007.doi:10.1093/crocol/otab007.\u003c/li\u003e\n\u003cli\u003eLewis JD, Chuai S, Nessel L, Lichtenstein GR, Aberra FN, Ellenberg JH. Use of the noninvasive components of the Mayo score to assess clinical response in ulcerative colitis. Inflamm Bowel Dis. 2008;14(12):1660-6.doi:10.1002/ibd.20520.\u003c/li\u003e\n\u003cli\u003eLouis E, Schreiber S, Panaccione R, Bossuyt P, Biedermann L, Colombel J-F, et al. Risankizumab for Ulcerative Colitis: Two Randomized Clinical Trials. JAMA. 2024;332(11):881-97.doi:10.1001/jama.2024.12414.\u003c/li\u003e\n\u003cli\u003eSandborn WJ, Feagan BG, D\u0026rsquo;Haens G, Wolf DC, Jovanovic I, Hanauer SB, et al. Ozanimod as Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2021;385(14):1280-91.doi:doi:10.1056/NEJMoa2033617.\u003c/li\u003e\n\u003cli\u003eFeagan BG, Rutgeerts P, Sands BE, Hanauer S, Colombel J-F, Sandborn WJ, et al. Vedolizumab as Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2013;369(8):699-710.doi:doi:10.1056/NEJMoa1215734.\u003c/li\u003e\n\u003cli\u003eRutgeerts P, Sandborn WJ, Feagan BG, Reinisch W, Olson A, Johanns J, et al. Infliximab for Induction and Maintenance Therapy for Ulcerative Colitis. New England Journal of Medicine. 2005;353(23):2462-76.doi:doi:10.1056/NEJMoa050516.\u003c/li\u003e\n\u003cli\u003eCohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. New York: Routledge; 1988.\u003c/li\u003e\n\u003cli\u003eFranklin JM, Patorno E, Desai RJ, Glynn RJ, Martin D, Quinto K, et al. Emulating Randomized Clinical Trials With Nonrandomized Real-World Evidence Studies: First Results From the RCT DUPLICATE Initiative. Circulation. 2021;143(10):1002-13.doi:10.1161/circulationaha.120.051718.\u003c/li\u003e\n\u003cli\u003eFranklin JM, Schneeweiss S. When and How Can Real World Data Analyses Substitute for Randomized Controlled Trials? Clinical pharmacology and therapeutics. 2017;102(6):924-33.doi:10.1002/cpt.857.\u003c/li\u003e\n\u003cli\u003eFranklin JM, Pawar A, Martin D, Glynn RJ, Levenson M, Temple R, et al. Nonrandomized Real-World Evidence to Support Regulatory Decision Making: Process for a Randomized Trial Replication Project. Clinical pharmacology and therapeutics. 2020;107(4):817-26.doi:10.1002/cpt.1633.\u003c/li\u003e\n\u003cli\u003eUngaro R, Colombel JF, Lissoos T, Peyrin-Biroulet L. A Treat-to-Target Update in Ulcerative Colitis: A Systematic Review. The American journal of gastroenterology. 2019;114(6):874-83.doi:10.14309/ajg.0000000000000183.\u003c/li\u003e\n\u003cli\u003eFerretti F, Cannatelli R, Monico MC, Maconi G, Ardizzone S. An Update on Current Pharmacotherapeutic Options for the Treatment of Ulcerative Colitis. Journal of clinical medicine. 2022;11(9).doi:10.3390/jcm11092302.\u003c/li\u003e\n\u003cli\u003eSands BE, Dubinsky MC, Kotze PG, Vermeire S, Panaccione R, Long MD, et al. Efficacy and Safety of Etrasimod in Patients With Moderately to Severely Active Ulcerative Colitis Stratified by Baseline Modified Mayo Score: A Post Hoc Analysis From the Phase 3 ELEVATE UC Clinical Program. Inflammatory Bowel Diseases. 2025.doi:10.1093/ibd/izaf036.\u003c/li\u003e\n\u003cli\u003eSands BE, Leung Y, Rubin DT, Gecse KB, Pan\u0026eacute;s J, Goetsch M, et al. Etrasimod Corticosteroid-Free Efficacy, Impact of Concomitant Corticosteroids on Efficacy and Safety, and Corticosteroid-Sparing Effect in Ulcerative Colitis: Analyses of the ELEVATE UC Clinical Program. Journal of Crohn\u0026apos;s and Colitis. 2024;19(3).doi:10.1093/ecco-jcc/jjae150.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":true,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Machine learning, real-world data, knowledge graphs, graph neural networks, predictive modeling","lastPublishedDoi":"10.21203/rs.3.rs-9272163/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9272163/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003cbr\u003e\nClinical trials are costly, time-consuming, and often yield uncertain outcomes. Predictive modeling using machine learning (ML), real-world data (RWD), and biomedical knowledge graphs offers new opportunities to improve translational efficiency and trial outcomes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003cbr\u003e\nThis study validated an ML-based modeling framework, \u003cem\u003eClinBoost\u003c/em\u003e, which integrates de-identified RWD from insurance claims and electronic health records with a drug-centered biomedical knowledge graph. Patient severity was modeled using an electronic Mayo Score, and longitudinal data inputs were transformed into patient-drug journey embeddings. Model performance was evaluated against 21 randomized controlled trials using a robust validation framework with multiple performance metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003cbr\u003e\nClinBoost a achieved high concordance with outcomes for 16 of the 21 trials, achieving an F1 score of 80%. The model demonstrated minimal bias in estimating treatment effect sizes. Additionally, the framework was able to refine trial design by identifying sub-cohorts with better response and increased statistical power.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003cbr\u003e\nThese findings demonstrate the potential of the ClinBoost modeling approach, which combines RWD and knowledge graph-informed ML, to improve clinical trial outcomes, optimize study design, and accelerate drug development.\u003c/p\u003e","manuscriptTitle":"Simulating Ulcerative Colitis Clinical Trials Using Knowledge Graph–Enhanced Real-World Data Modeling: Validation Across 21 Studies","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-09 01:07:20","doi":"10.21203/rs.3.rs-9272163/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"53eec47d-48bc-400e-8cb0-67426e8dbefd","owner":[],"postedDate":"May 9th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-07T07:11:01+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-05T09:09:47+00:00","index":76,"fulltext":""},{"type":"reviewerAgreed","content":"90799013883165387312871076765667720421","date":"2026-05-04T07:34:49+00:00","index":74,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-04T01:46:20+00:00","index":73,"fulltext":""},{"type":"reviewerAgreed","content":"255818046417682905282194281923498064464","date":"2026-05-03T23:59:05+00:00","index":72,"fulltext":""},{"type":"reviewerAgreed","content":"196517697510522839066136457588850776721","date":"2026-05-02T10:43:33+00:00","index":69,"fulltext":""},{"type":"reviewersInvited","content":"7","date":"2026-05-02T10:40:13+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-02T10:31:21+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67540400,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":67540401,"name":"Health sciences/Diseases"},{"id":67540402,"name":"Health sciences/Health care"},{"id":67540403,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-05-11T11:29:35+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-09 01:07:20","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9272163","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9272163","identity":"rs-9272163","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}