A novel translational in-silico indication discovery framework identifies indications and predictive biomarkers for cenerimod

A novel translational in-silico indication discovery framework identifies indications and predictive biomarkers for cenerimod | 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 A novel translational in-silico indication discovery framework identifies indications and predictive biomarkers for cenerimod Dominik Hartl, Madeleine Suffiotti, Peter Cornelisse, Daniel S. Strasser This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6993651/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 To fully explore a drug candidate's therapeutic potential, assessing its effectiveness across all possible diseases is essential. While traditional approaches match drug mechanisms to disease pathophysiology, they are limited by the high costs and slow progress of preclinical and clinical trials. This study introduces a novel in silico framework to identify new indications for drug candidates or repurpose approved drugs by analyzing their effects on gene expression in patients or animal models compared to controls. The framework integrate data from 13,602 patient samples across 146 diseases with drug candidate tested in preclinical models and use a neural network to reduce noise and improve sensitivity. The framework was exemplified with cenerimod, a S1P1 receptor modulator, which predicted its efficacy in immune-related diseases such as SLE, Psoriasis, and Crohn’s disease and kidney transplantation complications. Additionally, it identified six genes predictive of maximal clinical response in SLE patients, validated using RNA-seq data from a phase 2b cenerimod trial. Biological sciences/Drug discovery/Biomarkers Biological sciences/Drug discovery Biological sciences/Immunology Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Drug discovery is a complex task that heavily relies on innovation of drug research and development techniques and methods. The last decades technical capabilities in drug discovery increased considerably delivering patients novel, efficacious and safe treatment options. Despite the impressive advances in technology the number of drugs approved per billion US Dollar has decreased roughly by half per decade since the 1950s 1 . A key aspect of developing successful drugs is testing the drug candidate in the appropriate indications where it demonstrates efficacy 2 . This remains difficult as evidenced by lack of efficacy in between 50–60% of phase II and III trials 3 , 4 , 5 . Additionally, drug life cycle management aims to identify novel indications for approved drugs (repurposing or repositioning) and has the advantage to start from well-characterized pharmacology and safety profiles 6 . Ranking indications in order of likelihood of success is not trivial. Traditionally drug targets are selected based on their involvement in disease pathogenesis pathways. Because several pathways can be involved in diseases, selection of an indication, where the target pathway is a key driver of pathology, is crucial to increase the likelihood of drug development success. To find a match for a drug candidate with all suitable diseases, in-silico approaches can improve efficiency and speed in the indication selection process. Here, the complementarity of the drug candidate pharmacodynamic profile with the indication pathology represents a key factor 2 . In the recent years, the use of artificial intelligence techniques such as machine learning and deep learning (DL) has gained increasing interest to accelerate and potentiate the analysis of large amounts of data 7 . Indeed, DL methods have shown their potential in identifying drug candidates based on changes in gene expression profiles linked to disease conditions 8 , 9 , 10 . With the increasing availability of patient data of a variety of sources, in-silico approaches have the potential to provide an unbiased, comprehensive evaluation of suitable indications for a drug candidate. In this study, a novel in-silico framework was established to enable the discovery of indications for a drug candidate or repurposing of an approved drug. We took advantage of gene expression data as the most abundant and widely used data type for disease characterization 11 which enables a comprehensive assessment of drug candidate effects and diseases. Additionally, gene expression data is well curated and readily available in the public domain. Here, we present the in-silico framework showcasing an application for the S1P 1 receptor modulator cenerimod. Results and discussion Integration of preclinical models and human disease data The drug candidate gene expression profile measured in mouse models was correlated with the disease gene expression profile in human patients in comparison to healthy controls. A strong anti-correlation indicated a high degree of complementarity between the drug candidate and the disease (Fig. 1 ). In contrast to previous strategies where complementarity was assessed based on gene expression signatures 14 , here the expression levels of all genes measured across samples were used to avoid biases due to gene selection cut-offs. Additionally, the use of gene expression profiles instead of gene signatures allows for a broader assessment and subtler compound effects, while gene signature-based analyses focus exclusively on the strongest effects. The impact of restricting the focus on the strongest drug effects rather than the entire expression profile changes could explain the decreased success rates of clinical trials 1 . In contrast, drug candidates that lead to changes in very few specific genes might be more selectively matched using gene signatures. Since the goal of this study is to investigate a broad variety of drug targets by considering the effects in different tissues, this framework was built to use the comprehensive gene expression profiles. The gene expression effects of cenerimod treatment were measured in three different tissues of two disease mouse models. More specifically, cenerimod blood and kidney gene expression profiles were measured in the MRL/lpr autoimmune disease model representing aspects of systemic autoimmune diseases such as systemic lupus erythematosus 15 , 16 . Whereas the cenerimod skin gene expression profile was measured in a sclerodermatous chronic graft-versus-host disease model 17 representing a T-cell transfer model that reflects autoimmune diseases. In both preclinical animal model studies, the gene expression profiles were derived from the comparison of cenerimod-treated versus vehicle-treated (i.e., control) animals using RNA-sequencing and gene expression microarrays, respectively. A comprehensive collection of patient gene expression profiles was crucial to identify potential indications for cenerimod treatment. Gene expression profiles were obtained from Genevestigator 12 , which provides curated transcriptomics studies including a large variety of diseases. Initially, samples from the same study were separated by disease status (healthy controls or patients), and subsequently by tissue and/or cell type. Then, fold-changes between disease and healthy subjects were calculated. The within-disease patient heterogeneity was maintained by deriving the patient-level gene expression profiles. The expression profile of each individual patient was compared to the mean of all healthy subjects per study and tissue and/or cell type. This resulted in 13'602 patient expression profiles from 146 diseases. The human disease gene expression profiles were then combined with the cenerimod treated mouse gene expression profiles based on orthologous genes, which resulted in ~ 8'000 genes matched across platforms and species. The mouse disease model gene expression profiles were included as a benchmark in the analysis as cenerimod showed a beneficial effect on disease pathology 16 , 17 . Noise reduction using autoencoder The gene expression data originated from different cell types and tissue origin in mouse and human, multiple gene expression platforms (RNA-sequencing and microarrays), sample types (fresh-frozen samples and formalin-fixed paraffin-embedded (FFPE) samples, laboratories, and study designs. The combination of data of such diverse origins can contain considerable noise due to technical variation 18 . A type of deep learning algorithm called autoencoder, normally used in image analysis was employed to reduce noise in these samples (Fig. 2 a) 19 . Autoencoders are a special type of neural networks that encode a dataset into a lower dimensional space and then decode it to re-expand the data to its original dimensionality. In this way, the neural network is trained to generate an output that closely resembles its input. By forcing the data through a bottleneck in dimensionality, random noise is reduced as only higher order features are considered. In image analysis such higher order features for example represent shapes, that in gene expression data might depict in-silico de novo generated pathways. The matrix of fold-changes calculated per gene from 13'602 patient samples relative to the average of healthy controls was used as input to train and test multiple autoencoder designs. The best performing model based on the mean squared error (MSE) consisted of three hidden-layers in the coding and decoding modules respectively, and a coding-layer with 100 nodes ( Supplementary Fig. 1a ). Of note, the dimensionality of the coding-layer encoded the expression levels of about 8'000 genes into only 100 dimensions while maintaining the ability to decode with marginal loss ( Supplementary Fig. 1b ). When training neural networks, the amount of data required to obtain meaningful results is a major concern. To test if the data size used was sufficiently large to train such a model, the data were subsampled into different fractions of the original data set on which the algorithm was trained. The results showed that the data set was large enough as the loss of the validation set decreased and the loss of the training set increased until approximately 6'000 samples, from where it stabilized ( Supplementary Fig. 1c ). This indicated that for this design at least 6'000 samples were required. Since the sample size for training was around 11’000 patient samples, the amount of data was considered to not be a limiting factor to this approach. To assess the autoencoder ability to reduce noise, artificial noise was randomly added to different fractions of genes (35% and 70%) and subsequently processed by the autoencoder. Intriguingly, the noise was considerably reduced even when it was added to 70% of the genes (Fig. 2 b). On a global scale, this analysis showed that down to a correlation of 0.4 between the original dataset and the data with added random noise, the autoencoder removed the noise yielding to a correlation factor above 0.9 (Fig. 2 c). To evaluate whether the noise reduction improved the sensitivity of the analysis and therefore the detected signal, different tests were performed. Because the complementarity between drug candidate and disease was the target measure, an increased (anti-)correlation was considered indicative of an improvement of the signal. Therefore, the distribution of sample correlations between patients and mouse data was compared. The hypothesis was that if the autoencoder would improve the signal, the distribution of correlation would become broader. This means that extreme values would increase after autoencoder processing as similar samples would get more similar, and complementary samples more complementary. Indeed, this was the case as shown in Fig. 2 d, providing confidence that this processing step overall improved our analysis by efficient noise reduction and signal enhancement. Taken together, these results demonstrated the autoencoder capacity of noise reduction and improved signalling efficiency. Linking cenerimod with patient signatures based on anti-correlation Patient gene expression profiles reflect aspects of the molecular disease pathology and its heterogeneity. On the gene expression level, a drug candidate can be expected to be beneficial if it normalizes the pathological gene expression profile in the patient. This implies that the drug candidate effects on gene expression is complementary to the changes in the disease. This complementarity was measured by anti-correlation, the closer the correlation to -1 the strongest the complementarity between the drug candidate and the disease. The anti-correlations were calculated between 13'600 patients and cenerimod profiles in mouse disease models (MRL/lpr, Scl-cGVHD). Initially, the analysis was performed at the disease-level and subsequently at the patient-level. To increase sensitivity, for each disease the top 20% of patients anti-correlating to the cenerimod gene expression profile in blood from SLE mouse model were selected. Then, the average anti-correlation was calculated for each treated mouse sample. The top-ranked diseases with cenerimod in blood include Asperger syndrome, COPD, coronary artery disease and asthma (Fig. 3 a). Interestingly, autism that is condition not traditionally associated with autoimmunity was found among the top 10 diseases. However, it has been shown that patients with autism presented increased S1P levels in serum 20 and that S1P 1 modulation improves the phenotype of an autism rat model 21 . The first top-ranked autoimmune disease in blood was systemic lupus erythematosus, followed by multiple sclerosis. Indeed, S1P receptor modulators are approved for relapsing remitting multiple sclerosis 22 . However, in the patient-level analysis, none of the multiple sclerosis sample was found among the top-ranked diseases (Fig. 3 b). Because cenerimod showed efficacy in a phase 2b clinical trial (CARE, NCT03742037) in patients with moderate to severe SLE, we further evaluated SLE as an indication. The correlation between patient-level profile and cenerimod profile in the MRL/lpr mouse model in blood showed that SLE patients ranking at the top (Fig. 3 b). However, SLE patients were found at different levels of the raking, which reflects the heterogeneity of the SLE population gene expression profile. Interestingly, Crohn’s disease that origins in the gut was ranked at the patient-level among the top diseases to be complementary to the cenerimod profile. Supporting evidence of the obtained ranking for Crohn’s disease patients is the testing of S1P receptor modulators in clinical studies with Crohn’s disease patients (NCT03464097) and approval for ulcerative colitis 23 . These results demonstrated the capacity of our framework in detecting complementary diseases beyond the tissue type tested preclinically. The inspection of the cenerimod profile from the MRL/lpr mouse model in kidneys revealed the kidney transplantation, diabetic nephropathy, chronic kidney disease and rapidly progressive glomerulonephritis patients ranking at the top (Fig. 3 c) and the highest complementarity (correlation = ~-0.4) was found for kidney transplantation at the disease-level analysis ( Supplementary Fig. 2a ). This result was consistent with the testing of a S1P receptor modulator in kidney transplantation patients 24 . In addition to kidney-related diseases, ulcerative colitis patients appeared in the top-ranking diseases both in the disease- and patient-level analysis for which S1P receptor modulators have been approved. Finally, the cenerimod profile in the Scl-cGVHD mouse model in skin showed that the top-ranked disease patients included rheumatoid arthritis, atopic dermatitis, and psoriasis (Fig. 3 d, Supplementary Fig. 2b ). This result was consistent with the demonstrated clinical efficacy of a S1P receptor modulator in psoriasis patients clinical trial (NCT01208090) 25 . In order to demonstrate the improved comparability of the data after the autoencoder processing, we compared the ranking results from the skin profile for rheumatoid arthritis patients before and after noise reduction. As expected, the ranking of rheumatoid arthritis samples from skin obtained different ranking before and after noise reduction (Fig. 3 d, Supplementary Fig. 2c ). In particular, without denoising, the rheumatoid arthritis samples disappear from the top-ranks, leaving only atopic dermatitis and psoriasis patients in the top position, thus making the rankings tissue-specific. These results demonstrated the importance of the autoencoder pre-processing in increasing comparability across tissues. Firstly, we presented the complementarity evaluation of cenerimod profile by considering separately the different tissue types. Then, we evaluated the complementarity results when combining the data from different tissue types and whether this improved sensitivity. In particular, SLE is a very heterogeneous disease not only from the gene expression profile but also in terms of clinical manifestations, affecting different organs depending on the degree of disease severity. Consequently, in order to obtain a more holistic view of the predicted efficacy of cenerimod across patients, we considered the heterogeneity of the phenotypic manifestations by selecting the highest complementarity (high anti-correlation) across all cenerimod tissues. As expected, the distributions changed considerably for SLE patients when considering the maximum anti-correlation across tissues, while the changes were much less pronounced for atopic dermatitis, for instance (Fig. 3 e). This observation reflected the inherent high heterogeneity of the SLE patients with a heterogeneous complementary of SLE patients with cenerimod treatment, providing important insights for SLE patients responder analysis. In addition, when combining the cenerimod profiles of both tissues and mouse models with all disease tissue types, we observed a consistently strong anti-correlation for the following top-ranked disease: Asperger syndrome (autism), kidney transplantation, SLE, multiple sclerosis and ulcerative colitis ( Supplementary Fig. 2d ). Take together these results showed the importance of integrating data from different animal models and characterizing the drug candidate effect in different tissues to get a holistic view about the potential complementary indications. In summary, our in-silico indication discovery framework confirmed SLE as an autoimmune indication for cenerimod. Furthermore, indications such as Asperger syndrome (autism), multiple sclerosis, rheumatoid arthritis, atopic dermatitis, psoriasis, kidney transplantation, Crohn’s disease and ulcerative colitis emerged to be additional potential indications for cenerimod treatment. As described previously, none of the multiple sclerosis samples was found in the top-ranked diseases in the patient-level analysis, while the disease appeared at the sixth position in the disease-level analysis. This observation could be explained by the lack of data from brain samples from cenerimod preclinical studies, which might have reduced the sensitivity for such indications. This further highlights the importance of measuring drug gene expression profile in several animal models and tissues, in order to allow for a more holistic indication discovery. Determinants of response to cenerimod Once target indications were identified, the patient-level analysis provided an opportunity to gain insights into potential factors of the magnitude of the expected response to cenerimod. As described previously, the disease-level complementarity analysis showed considerable heterogeneity even among the top-ranked SLE patients. In order to gain insight into SLE patients’ heterogeneity from their clinical characteristics, we investigated what were the characteristics associated with the predicted magnitude of response to cenerimod. Because type-1-interferon (IFN-1) has previously been used in clinical trials as a stratification biomarker for SLE patients, we tested whether an IFN-1 signature can be used to estimate the cenerimod predicted response. The comparison of the predicted response with the expression level of the IFN-1 signature in SLE patients showed a higher predicted response in patients with higher IFN-1 signature (Fig. 4 a), which is consistent with the CARE study 26 . To identify genes that are predictive of cenerimod response, a neural network classifier was implemented. The response variable was defined by applying cut-offs on patients’ complementarity (correlation coefficient r) to define responders (r 0) (Fig. 4 b). The neural network was trained on all available patient profiles to solve a binary classification problem and predict responders and non-responders. In order to identify the genes predicting cenerimod response in SLE, a Local Interpretable Model-agnostic Explanations (LIME) algorithm 13 was applied on SLE patient samples. The ranking of the most frequent genes with the highest predictive weight (i.e., high expression levels) included TUBB, EMP3, JCHAIN, TUBB6, ILR2G and EMP1, which were linked to response (Fig. 4 b). These genes were combined in term of z-scores to define a de novo discovered gene signature and assess its predictive value to cenerimod. The de novo gene expression signature showed an anti-correlation of -0.57 with the predicted response (complementarity or correlation coefficient) (Fig. 4 c). This indicated that SLE patients with high expression levels of the identified genes were likely better responding to cenerimod treatment. Subsequently, we evaluated SLE patients’ characteristics such as disease activity (SLEDAI-2K) or age that were expected to be associated with the predicted response to cenerimod. Since the predicted response to cenerimod was shown to be higher depending on the strength of the molecular endotype, we expected a stronger positive effect of the drug on patients with a more severe disease phenotype. Indeed, patients with a higher disease (SLEDAI) score were predicted to have the higher response (or lower correlation coefficient) ( P = 0.0016), providing increasing evidence of the beneficial effects of cenerimod treatment for SLE patients (Fig. 4 d). In addition, younger patients age (< 20) was significantly associated ( P < 0.001) with increased response (Fig. 4 e), which is consistent with the increased disease severity in children compared to adults 27 . Validation of cenerimod predictive biomarkers in a Phase 2b clinical trial in SLE patients To validate the ability of framework in identifying biomarkers predictive of cenerimod response, we evaluated clinical response based on mSLEDAI-2k disease activity score to the efficacious dose of cenerimod 4 mg in the CARE RNA-seq sub-study 28 . The baseline expression levels of the top-predictors were evaluated in responders and non-responders receiving placebo or cenerimod 4 mg treatment. Linear regression models were used to determine significance of JCHAIN baseline levels to cenerimod 4 mg in responders compared to non-responders (p-value = 0.019) and identified JCHAIN as a novel predictive biomarkers (Fig. 5 a). Moreover, TUBB gene was significantly decreased (p-value < 0.001) after 6 months of cenerimod 4 mg treatment in responders but not in non-responders and only in cenerimod treated patients, making TUBB a potential pharmacodynamic biomarker to monitor treatment response in the blood (Fig. 5 b). In conclusion, we demonstrated that our in-silico indication discovery framework supported the selection of SLE as an indication for cenerimod treatment and the identification of predictive biomarkers that were validated in a Phase 2b clinical trial of moderate to severe patients with SLE (Fig. 5 c). Methods Mouse models and gene expression data The sclerodermatous chronic graft-versus-house-disease (Scl-cGVHD) mouse model gene expression data from skin was obtained from Kano et al 2019 29 . For the lupus mouse model, experiment was conducted as described in Strasser et al 2020 16 . In brief, 7-week-old female MRL /lpr (Jackson Laboratories) mice were treated with vehicle or cenerimod (0.2 mg/g food) until the end of the study, which was predefined as the time point when at least 20% morbidity/mortality was reached in one group (end of treatment week 10). RNA from blood and kidney from cenerimod- and vehicle-treated mice was isolated and prepared for sequencing using NUGEN Universal Plus mRNA kit with PolyA select (and globin depletion for the blood samples) according to protocol and sequenced on Illumina NextSeq 500 mRNA-seq platform. Age matched w C57BL/6 mice were used as healthy reference. The samples were prepared and sequenced as above. All experimental procedures were conducted in accordance with the Swiss animal welfare ordinance and Idorsia Animal Welfare policy on the use of experimental animals. The study was approved by the Basel-land Cantonal Veterinary Home Office. Experiments and results were reported according to the ARRIVE guidelines 2.0. Data and processing All gene expression data was pre-processed, curated and retrieved from the Genevestigator application 12 . Data was separated by study, and within the study by cell type and disease state (healthy or disease) as defined in Genevestigator. To obtain a collapsed profile of the state "healthy", and to account for outliers, all samples from healthy subjects with the same cell type in the study were compared with each other resulting in a correlation matrix. By taking the row means of this matrix, divided by the total sum of correlation, weights for each healthy sample were obtained by calculating a weighted mean. Fold-changes were calculated by dividing each disease sample with the corresponding healthy weighted mean. Pseudocounts were added to healthy and disease samples to avoid overamplification of noise (pseudocounts were 8 for gene expression array data and 0.25 for RNA-seq data), these pseudocounts were chosen based on visual inspection of histograms of expression data across all genes and samples. Samples from mouse disease models were processed in the same way either as disease model versus healthy reference mice or cenerimod-treated versus vehicle-treated disease model. Mouse and human data were then merged based on orthologous genes obtained by the getLDS function from the biomaRt Bioconductor package. Duplicated genes were discarded. Neural networks Neural networks were engineered using the R package Keras with TensorFlow backend. The data was split 80/20 for training and testing. Multiple designs were tested with different numbers of layers and nodes as well as activation functions. The design shown in Supplementary Fig. 1 resulted in the best balance of loss and overfitting while still having a low number of nodes in the coding layer for the autoencoder. Performance metrics In-silico Identification of predictive biomarkers Neural network was trained on all patient data to predict response or no response (-0.1 and 0 as cutoff respectively). Subsequently, the LIME (Local Interpretable Model-agnostic Explanations) algorithm was used to identify features (genes) frequently used for 13 de novo signature calculation. To quantify the overall change across this set of genes the mean z-scores were calculated across the genes for each patient and compared to the correlation with the cenerimod profile. Statistical analysis Statistical analysis was performed using R software, version 4.1.2 (R Foundation for Statistical Computing). Correlations were calculated using the R cor function, in this study Pearson and Spearman correlations were calculated as indicated in the figure legends. Significance of correlations were tested when focussing on SLE patients in Fig. 4 using the R function cor.test. Differences in gene expression levels at baseline between groups (i.e., responders vs. non-responders) were tested using linear regression models. When comparing gene expression levels between time points (baseline vs. month 6), mixed effects models with a subject random effect were used (lmer function). Resulting p-values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate method. Conclusion The ability to link a drug candidate with the indication or patients it can have the greatest effect is a key determinant of the success in drug development. Taking advantage of a drug candidate gene expression profile measured in preclinical models with the wealth of patient disease gene expression profiles available in the public domain, provides great potential to support unbiased indication discovery. In this study, an in-silico discovery framework enabling the matching of drug candidates with indications based on the gene expression profiles and the identification of associated predictive biomarkers is presented. This framework is not limiting the complementarity investigation to specific gene expression signatures but allows to investigate broader effects by considering the whole gene expression profile effects and quantitatively compared changes for a holistic assessment of molecular effects. In addition, our approach has the potential of being extended to other data types such as proteomics or metabolomics data. The patient-level analysis enabled to decipher the disease patient population heterogeneity with molecularly and clinically. This approach was facilitated by focussing on „reverse curation" of patient data of interest, avoiding upfront intensive efforts for sample annotation in the entire patient dataset. Cenerimod treatment was used to exemplify our in-silico discovery framework, which predicted its potential efficacy in several autoimmune diseases such as SLE, multiple sclerosis, rheumatoid arthritis, psoriasis, Crohn’s disease, and kidney transplantation. Indeed, several of the predicted indications were previously clinically tested with S1P receptor modulators. In addition, novel indications such as Asperger syndrome were found to be complementary to the cenerimod profile, representing a potential treatment opportunity for additional central nervous system diseases with increasing evidence of immune system dysfunction 30 . Applying the framework to individual patients with SLE allowed to discover determinants linked to response, which led to the identification of predictive biomarkers. Our major findings included the identification of TUBB gene as a potential pharmacodynamic biomarker to monitor patient response in the blood and JCHAIN gene as a predictive biomarker for the clinical compound cenerimod. Declarations Authors contributions: DH: Conceptualization, Data curation, formal analysis, writing – original draft, writing – review and editing. MS: Conceptualization, Data curation, formal analysis, writing – original draft, writing – review and editing. PC: Data curation, review and editing. DS: Conceptualization, Data curation, formal analysis, writing – original draft, writing – review and editing. All authors had full access to all the data in the study. All authors were responsible for the decision to submit the article. Data availability RNA-Seq data were deposited into the Gene Expression Omnibus database under accession number GSE293404 and are available at the following URL: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE293404 . Acknowledgement: We are grateful to Estelle Gerossier-Creusat and Sylvie Froidevaux for providing animal model tissue samples and Bérengère Renault for RNA-seq library preparation and sequencing. Competing interest: DH has no competing interest MS has no competing interest PC is a holder of stocks of Idorsia Pharmaceuticals Ltd DSS is a holder of stocks and stock options of Idorsia Pharmaceuticals Ltd References Scannell JW, Blanckley A, Boldon H, Warrington B. Diagnosing the decline in pharmaceutical R&D efficiency. Nat Rev Drug Discov 11 , 191-200 (2012). Shih HP, Zhang X, Aronov AM. Drug discovery effectiveness from the standpoint of therapeutic mechanisms and indications. Nat Rev Drug Discov 17 , 19-33 (2018). Arrowsmith J. Phase II failures: 2008–2010. Nature Reviews Drug Discovery 10 , 328-329 (2011). Arrowsmith J, Miller P. Phase II and Phase III attrition rates 2011–2012. Nature Reviews Drug Discovery 12 , 569-569 (2013). Harrison RK. Phase II and phase III failures: 2013-2015. Nat Rev Drug Discov 15 , 817-818 (2016). Liu R, Wei L, Zhang P. A deep learning framework for drug repurposing via emulating clinical trials on real-world patient data. Nat Mach Intell 3 , 68-75 (2021). Blanco-Gonzalez A, Cabezon A, Seco-Gonzalez A, Conde-Torres D, Antelo-Riveiro P, Pineiro A, Garcia-Fandino R. The Role of AI in Drug Discovery: Challenges, Opportunities, and Strategies. Pharmaceuticals (Basel) 16 , (2023). Zhu J , et al. Prediction of drug efficacy from transcriptional profiles with deep learning. Nat Biotechnol 39 , 1444-1452 (2021). Lamb J , et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Science 313 , 1929-1935 (2006). Zeng X, Zhu S, Liu X, Zhou Y, Nussinov R, Cheng F. deepDR: a network-based deep learning approach to in silico drug repositioning. Bioinformatics 35 , 5191-5198 (2019). Chen B, Butte AJ. Leveraging big data to transform target selection and drug discovery. Clin Pharmacol Ther 99 , 285-297 (2016). Hruz T , et al. Genevestigator v3: a reference expression database for the meta-analysis of transcriptomes. Adv Bioinformatics 2008 , 420747 (2008). Ribeiro MT, Singh S, Guestrin C. "Why Should I Trust You?": Explaining the Predictions of Any Classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining ). Association for Computing Machinery (2016). Sirota M , et al. Discovery and preclinical validation of drug indications using compendia of public gene expression data. Sci Transl Med 3 , 96ra77 (2011). Perry D, Sang A, Yin Y, Zheng YY, Morel L. Murine models of systemic lupus erythematosus. J Biomed Biotechnol 2011 , 271694 (2011). Strasser DS , et al. Preclinical to clinical translation of cenerimod, a novel S1P(1) receptor modulator, in systemic lupus erythematosus. RMD Open 6 , (2020). Kano M , et al. Attenuation of murine sclerodermatous models by the selective S1P1 receptor modulator cenerimod. Scientific Reports 9 , 658 (2019). Finotello F, Di Camillo B. Measuring differential gene expression with RNA-seq: challenges and strategies for data analysis. Brief Funct Genomics 14 , 130-142 (2015). Jain V, Seung H. Natural Image Denoising with Convolutional Networks (2008). Wang H , et al. Potential serum biomarkers from a metabolomics study of autism. J Psychiatry Neurosci 41 , 27-37 (2016). Wu H , et al. Modulation of sphingosine 1-phosphate (S1P) attenuates spatial learning and memory impairments in the valproic acid rat model of autism. Psychopharmacology (Berl) 235 , 873-886 (2018). Kappos L , et al. A placebo-controlled trial of oral fingolimod in relapsing multiple sclerosis. N Engl J Med 362 , 387-401 (2010). Qiu, Jingyue, et al. Efficacy and safety of the S1PR modulator etrasimod in the treatment of moderately to severely active ulcerative colitis during the induction phase: a systematic review and meta-analysis of randomized controlled trials. Frontiers in Pharmacology 15 (2024): 1420455. Skerjanec A, Tedesco H, Neumayer HH, Cole E, Budde K, Hsu CH, Schmouder R. FTY720, a novel immunomodulator in de novo kidney transplant patients: pharmacokinetics and exposure-response relationship. J Clin Pharmacol 45 , 1268-1278 (2005). Vaclavkova A , et al. Oral ponesimod in patients with chronic plaque psoriasis: a randomised, double-blind, placebo-controlled phase 2 trial. Lancet 384 , 2036-2045 (2014). Askanase, Anca D., et al. Cenerimod, a sphingosine-1-phosphate receptor modulator, versus placebo in patients with moderate-to-severe systemic lupus erythematosus (CARE): an international, double-blind, randomised, placebo-controlled, phase 2 trial. The Lancet Rheumatolog y 7.1 : e21-e32(2025). das Chagas Medeiros MM , et al. Clinical and immunological aspects and outcome of a Brazilian cohort of 414 patients with systemic lupus erythematosus (SLE): comparison between childhood-onset, adult-onset, and late-onset SLE. Lupus 25 , 355-363 (2016). Suffiotti, Madeleine, et al. Pharmacodynamics of the S1P1 receptor modulator cenerimod in a phase 2b randomised clinical trial in patients with moderate to severe SLE. Annals of the Rheumatic Diseases (2025). Kano, Miyu, et al. Attenuation of murine sclerodermatous models by the selective S1P1 receptor modulator cenerimod. Scientific Reports 9.1: 658 . (2019) Enstrom AM, Van de Water JA, Ashwood P. Autoimmunity in autism. Curr Opin Investig Dr ugs 10 , 463-473 (2009). Additional Declarations Competing interest reported. DH has no competing interest MS has no competing interest PC is a holder of stocks of Idorsia Pharmaceuticals Ltd DSS is a holder of stocks and stock options of Idorsia Pharmaceuticals Ltd Supplementary Files SupplementaryMaterial.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-6993651","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":501043682,"identity":"96831c2a-3488-47cb-8e16-f6dd93d879ca","order_by":0,"name":"Dominik Hartl","email":"","orcid":"","institution":"Idorsia Pharmaceuticals Ltd","correspondingAuthor":false,"prefix":"","firstName":"Dominik","middleName":"","lastName":"Hartl","suffix":""},{"id":501043683,"identity":"f331c466-9eaf-44b8-90bd-8b6617796829","order_by":1,"name":"Madeleine Suffiotti","email":"","orcid":"","institution":"Idorsia Pharmaceuticals Ltd","correspondingAuthor":false,"prefix":"","firstName":"Madeleine","middleName":"","lastName":"Suffiotti","suffix":""},{"id":501043684,"identity":"2a16838c-9e31-455e-87d8-a88ac0d6b044","order_by":2,"name":"Peter Cornelisse","email":"","orcid":"","institution":"Idorsia Pharmaceuticals Ltd","correspondingAuthor":false,"prefix":"","firstName":"Peter","middleName":"","lastName":"Cornelisse","suffix":""},{"id":501043685,"identity":"2ebaac47-d8d1-4f90-bab7-65ee9c1e5d61","order_by":3,"name":"Daniel S. Strasser","email":"data:image/png;base64,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","orcid":"","institution":"Idorsia Pharmaceuticals Ltd","correspondingAuthor":true,"prefix":"","firstName":"Daniel","middleName":"S.","lastName":"Strasser","suffix":""}],"badges":[],"createdAt":"2025-06-27 17:38:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6993651/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6993651/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89454323,"identity":"c67ceba1-d189-4934-84dd-0952af0f4217","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":196133,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003ein-silico\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e discovery framework for indications selection.\u003c/strong\u003e Gene expression profile from drug candidate treated mice comparison with human gene expression profiles of patients to link the drug profile to indications based on complementarity. Complementarity was measured based on anti-correlation.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/0ab42a1de4ef48ffa9ad4fa2.png"},{"id":89454335,"identity":"d40a9ca5-1f61-4885-a7f1-633c4d8608b8","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":434628,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe autoencoder efficiently reduced noise and improved the signal.\u003cbr\u003e\n a)\u003c/strong\u003e Schematic representation of the autoencoder use for noise reduction. The neural network was designed to encode the input gene expression data into a low dimensional representation and forcing the selection of high-level features, thus reducing noise. \u003cstrong\u003eb)\u003c/strong\u003e Scatterplots showing the reduction of the introduced artificial noise in one randomly selected sample. Artificial noise was introduced to randomly selected (35% or 70%) genes in the dataset. The magnitude of noise was from -3 to 3. The y-axis shows the fold-change gene expression with introduced artificial noise. The x-axis shows the fold-change gene expression without artificial noise. \u003cstrong\u003ec)\u003c/strong\u003e Plot displaying the correlation between the original samples and samples with introduced artificial noise before and after the autoencoder processing of the data. Correlations were calculated on all samples; the line represents a loess curve fitted to all data points. \u003cstrong\u003ed)\u003c/strong\u003e Density plot of correlations calculated between patients’ samples and the treated mouse model samples before and after the autoencoder processing. The flanks of the distribution are extended indicating an increased signal.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/3c78d3eb36d788ef310ab3c3.png"},{"id":89454327,"identity":"a0802705-f9e0-47c0-a636-d69ef6b9c2b9","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1409684,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRanking of indications based on complementarity between the cenerimod gene expression profile and diseases.\u003c/strong\u003e \u003cstrong\u003ea)\u003c/strong\u003e Boxplots of human disease complementarity with MRL/lpr SLE mouse model with cenerimod in blood (disease-level analysis). \u003cstrong\u003eb)\u003c/strong\u003e Boxplots of human patient samples complementarity with MRL/lpr SLE mouse model with cenerimod in blood (patient-level analysis). \u003cstrong\u003ec)\u003c/strong\u003e Boxplots of human patient samples complementarity with MRL/lpr SLE mouse model with cenerimod in kidneys (patient-level analysis). \u003cstrong\u003ec)\u003c/strong\u003e Boxplots of human patient samples complementarity with Scl-cGVHD mouse model with cenerimod in skin (patient-level analysis). Spearman correlation coefficients were calculated using the average correlation of the top-20% anti-correlating samples per disease with the individual treated mouse model samples. Black squares show data points with than 1.5-fold interquartile range from the lower or upper quartile value. Diseases with less than 20 samples were discarded. The mouse model gene expression profiles were collapsed into one by calculating the mean fold-changes across all samples per animal model and tissue. Blue text highlights the first autoimmune disease in the top-ranks in blood. Orange text highlights indications with different tissue origin than the tissue profiles in the preclinical model. Green text highlights top-ranked indications from the disease-level analysis (Supplementary Figure 2). The values on the right-side of the panel in the patient-level analysis show the enrichment of the disease in the top-100 complementary patients compared to the random distribution of the samples. \u003cstrong\u003ee)\u003c/strong\u003e Density plots of correlations calculated between the mean cenerimod treated mouse model profiles with patient samples from different diseases. Blue lines show the correlations in the highest ranked mouse model tissue across diseases. Red lines show the distribution when the maximum anti-correlation was selected across tissues.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/cdf391ea45fd6c0b5a554b24.png"},{"id":89454334,"identity":"b9bdf333-2818-4d0b-82fa-d761df9fb9f0","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":145842,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSLE patient characteristics associated with cenerimod response.\u003c/strong\u003e \u003cstrong\u003ea)\u003c/strong\u003e Scatterplot displaying maximum (across tissues) correlation of SLE patient samples with mean scaled (z-score) IFN-1 signature. \u003cstrong\u003eb)\u003c/strong\u003e Strategy overview for the identification of genes linked to the cenerimod response (predictive biomarkers) using a neural network classifier. Barplots x-axis shows the fraction of positive predictors or weight for each gene, while y-axis shows the value at which a give gene is predictive. \u003cstrong\u003ec)\u003c/strong\u003e Scatterplot displaying maximum (across tissue) correlation of SLE patient samples with mean scaled (z-score) of the \u003cem\u003ede-novo\u003c/em\u003e signature. \u003cstrong\u003ed)\u003c/strong\u003e Boxplots showing the maximum (across tissues) correlation values across different ranges of the SLEDAI score (disease score). \u003cstrong\u003ee)\u003c/strong\u003e Boxplots showing the maximum (across tissues) correlation values across different age ranges.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/696cfad28eb864d3a4a85336.png"},{"id":89454329,"identity":"546848a5-547c-4c49-9230-1b8a5ae8d958","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":97441,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eValidation of the predictive biomarkers in gene expression data of a Ph2b clinical trial in SLE patients.\u003c/strong\u003e \u003cstrong\u003ea)\u003c/strong\u003e JCHAIN baseline expression levels in SLE patients treated with placebo or cenerimod 4 mg in responders (R) versus non-responders (NR). Differences between R and NR were tested using linear regression models. \u003cstrong\u003eb)\u003c/strong\u003eTUBB expression levels at baseline and month 6 in SLE patients treated with placebo or cenerimod 4 mg in responders (R) versus non-responders (NR). Differences between baseline and month 6 were tested using mixed-effects models. \u003cstrong\u003ec)\u003c/strong\u003e Summary scheme of the overall strategy to link drug candidates with indications.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/d0ec6f50fd47618ccc97914b.png"},{"id":97672899,"identity":"c005cf33-01e7-485c-8e70-9a29d975418d","added_by":"auto","created_at":"2025-12-08 09:39:03","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2940549,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/8b9e3181-e6e1-48ba-aae6-d9fbe6c038b4.pdf"},{"id":89454326,"identity":"80ea27e4-0201-4703-9f93-0fe96b09c80d","added_by":"auto","created_at":"2025-08-20 06:46:23","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":341809,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-6993651/v1/2f9bdbb67e3c040a12dafd0c.docx"}],"financialInterests":"Competing interest reported. DH has no competing interest\nMS has no competing interest\nPC is a holder of stocks of Idorsia Pharmaceuticals Ltd\nDSS is a holder of stocks and stock options of Idorsia Pharmaceuticals Ltd","formattedTitle":"A novel translational in-silico indication discovery framework identifies indications and predictive biomarkers for cenerimod","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDrug discovery is a complex task that heavily relies on innovation of drug research and development techniques and methods. The last decades technical capabilities in drug discovery increased considerably delivering patients novel, efficacious and safe treatment options. Despite the impressive advances in technology the number of drugs approved per billion US Dollar has decreased roughly by half per decade since the 1950s\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. A key aspect of developing successful drugs is testing the drug candidate in the appropriate indications where it demonstrates efficacy\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. This remains difficult as evidenced by lack of efficacy in between 50\u0026ndash;60% of phase II and III trials \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Additionally, drug life cycle management aims to identify novel indications for approved drugs (repurposing or repositioning) and has the advantage to start from well-characterized pharmacology and safety profiles\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eRanking indications in order of likelihood of success is not trivial. Traditionally drug targets are selected based on their involvement in disease pathogenesis pathways. Because several pathways can be involved in diseases, selection of an indication, where the target pathway is a key driver of pathology, is crucial to increase the likelihood of drug development success.\u003c/p\u003e\u003cp\u003eTo find a match for a drug candidate with all suitable diseases, in-silico approaches can improve efficiency and speed in the indication selection process. Here, the complementarity of the drug candidate pharmacodynamic profile with the indication pathology represents a key factor\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eIn the recent years, the use of artificial intelligence techniques such as machine learning and deep learning (DL) has gained increasing interest to accelerate and potentiate the analysis of large amounts of data\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Indeed, DL methods have shown their potential in identifying drug candidates based on changes in gene expression profiles linked to disease conditions\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eWith the increasing availability of patient data of a variety of sources, in-silico approaches have the potential to provide an unbiased, comprehensive evaluation of suitable indications for a drug candidate.\u003c/p\u003e\u003cp\u003eIn this study, a novel \u003cem\u003ein-silico\u003c/em\u003e framework was established to enable the discovery of indications for a drug candidate or repurposing of an approved drug. We took advantage of gene expression data as the most abundant and widely used data type for disease characterization \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e which enables a comprehensive assessment of drug candidate effects and diseases. Additionally, gene expression data is well curated and readily available in the public domain. Here, we present the in-silico framework showcasing an application for the S1P\u003csub\u003e1\u003c/sub\u003e receptor modulator cenerimod.\u003c/p\u003e"},{"header":"Results and discussion","content":"\u003cp\u003e\u003cem\u003eIntegration of preclinical models and human disease data\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe drug candidate gene expression profile measured in mouse models was correlated with the disease gene expression profile in human patients in comparison to healthy controls. A strong anti-correlation indicated a high degree of complementarity between the drug candidate and the disease (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn contrast to previous strategies where complementarity was assessed based on gene expression signatures \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, here the expression levels of all genes measured across samples were used to avoid biases due to gene selection cut-offs. Additionally, the use of gene expression profiles instead of gene signatures allows for a broader assessment and subtler compound effects, while gene signature-based analyses focus exclusively on the strongest effects. The impact of restricting the focus on the strongest drug effects rather than the entire expression profile changes could explain the decreased success rates of clinical trials \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. In contrast, drug candidates that lead to changes in very few specific genes might be more selectively matched using gene signatures. Since the goal of this study is to investigate a broad variety of drug targets by considering the effects in different tissues, this framework was built to use the comprehensive gene expression profiles.\u003c/p\u003e\u003cp\u003eThe gene expression effects of cenerimod treatment were measured in three different tissues of two disease mouse models. More specifically, cenerimod blood and kidney gene expression profiles were measured in the MRL/lpr autoimmune disease model representing aspects of systemic autoimmune diseases such as systemic lupus erythematosus \u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Whereas the cenerimod skin gene expression profile was measured in a sclerodermatous chronic graft-versus-host disease model \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e representing a T-cell transfer model that reflects autoimmune diseases. In both preclinical animal model studies, the gene expression profiles were derived from the comparison of cenerimod-treated versus vehicle-treated (i.e., control) animals using RNA-sequencing and gene expression microarrays, respectively.\u003c/p\u003e\u003cp\u003eA comprehensive collection of patient gene expression profiles was crucial to identify potential indications for cenerimod treatment. Gene expression profiles were obtained from Genevestigator \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e, which provides curated transcriptomics studies including a large variety of diseases. Initially, samples from the same study were separated by disease status (healthy controls or patients), and subsequently by tissue and/or cell type. Then, fold-changes between disease and healthy subjects were calculated. The within-disease patient heterogeneity was maintained by deriving the patient-level gene expression profiles. The expression profile of each individual patient was compared to the mean of all healthy subjects per study and tissue and/or cell type. This resulted in 13'602 patient expression profiles from 146 diseases. The human disease gene expression profiles were then combined with the cenerimod treated mouse gene expression profiles based on orthologous genes, which resulted in ~ 8'000 genes matched across platforms and species. The mouse disease model gene expression profiles were included as a benchmark in the analysis as cenerimod showed a beneficial effect on disease pathology \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003cem\u003eNoise reduction using autoencoder\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe gene expression data originated from different cell types and tissue origin in mouse and human, multiple gene expression platforms (RNA-sequencing and microarrays), sample types (fresh-frozen samples and formalin-fixed paraffin-embedded (FFPE) samples, laboratories, and study designs. The combination of data of such diverse origins can contain considerable noise due to technical variation \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. A type of deep learning algorithm called autoencoder, normally used in image analysis was employed to reduce noise in these samples (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea) \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Autoencoders are a special type of neural networks that encode a dataset into a lower dimensional space and then decode it to re-expand the data to its original dimensionality. In this way, the neural network is trained to generate an output that closely resembles its input. By forcing the data through a bottleneck in dimensionality, random noise is reduced as only higher order features are considered. In image analysis such higher order features for example represent shapes, that in gene expression data might depict \u003cem\u003ein-silico de novo\u003c/em\u003e generated pathways.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe matrix of fold-changes calculated per gene from 13'602 patient samples relative to the average of healthy controls was used as input to train and test multiple autoencoder designs. The best performing model based on the mean squared error (MSE) consisted of three hidden-layers in the coding and decoding modules respectively, and a coding-layer with 100 nodes (\u003cb\u003eSupplementary Fig.\u0026nbsp;1a\u003c/b\u003e). Of note, the dimensionality of the coding-layer encoded the expression levels of about 8'000 genes into only 100 dimensions while maintaining the ability to decode with marginal loss (\u003cb\u003eSupplementary Fig.\u0026nbsp;1b\u003c/b\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWhen training neural networks, the amount of data required to obtain meaningful results is a major concern. To test if the data size used was sufficiently large to train such a model, the data were subsampled into different fractions of the original data set on which the algorithm was trained. The results showed that the data set was large enough as the loss of the validation set decreased and the loss of the training set increased until approximately 6'000 samples, from where it stabilized (\u003cb\u003eSupplementary Fig.\u0026nbsp;1c\u003c/b\u003e). This indicated that for this design at least 6'000 samples were required. Since the sample size for training was around 11’000 patient samples, the amount of data was considered to not be a limiting factor to this approach.\u003c/p\u003e\u003cp\u003eTo assess the autoencoder ability to reduce noise, artificial noise was randomly added to different fractions of genes (35% and 70%) and subsequently processed by the autoencoder. Intriguingly, the noise was considerably reduced even when it was added to 70% of the genes (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). On a global scale, this analysis showed that down to a correlation of 0.4 between the original dataset and the data with added random noise, the autoencoder removed the noise yielding to a correlation factor above 0.9 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec).\u003c/p\u003e\u003cp\u003eTo evaluate whether the noise reduction improved the sensitivity of the analysis and therefore the detected signal, different tests were performed. Because the complementarity between drug candidate and disease was the target measure, an increased (anti-)correlation was considered indicative of an improvement of the signal. Therefore, the distribution of sample correlations between patients and mouse data was compared. The hypothesis was that if the autoencoder would improve the signal, the distribution of correlation would become broader. This means that extreme values would increase after autoencoder processing as similar samples would get more similar, and complementary samples more complementary. Indeed, this was the case as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed, providing confidence that this processing step overall improved our analysis by efficient noise reduction and signal enhancement. Taken together, these results demonstrated the autoencoder capacity of noise reduction and improved signalling efficiency.\u003c/p\u003e\u003cp\u003e\u003cem\u003eLinking cenerimod with patient signatures based on anti-correlation\u003c/em\u003e\u003c/p\u003e\u003cp\u003ePatient gene expression profiles reflect aspects of the molecular disease pathology and its heterogeneity. On the gene expression level, a drug candidate can be expected to be beneficial if it normalizes the pathological gene expression profile in the patient. This implies that the drug candidate effects on gene expression is complementary to the changes in the disease.\u003c/p\u003e\u003cp\u003eThis complementarity was measured by anti-correlation, the closer the correlation to -1 the strongest the complementarity between the drug candidate and the disease. The anti-correlations were calculated between 13'600 patients and cenerimod profiles in mouse disease models (MRL/lpr, Scl-cGVHD). Initially, the analysis was performed at the disease-level and subsequently at the patient-level. To increase sensitivity, for each disease the top 20% of patients anti-correlating to the cenerimod gene expression profile in blood from SLE mouse model were selected. Then, the average anti-correlation was calculated for each treated mouse sample.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe top-ranked diseases with cenerimod in blood include Asperger syndrome, COPD, coronary artery disease and asthma (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). Interestingly, autism that is condition not traditionally associated with autoimmunity was found among the top 10 diseases. However, it has been shown that patients with autism presented increased S1P levels in serum \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e and that S1P\u003csub\u003e1\u003c/sub\u003e modulation improves the phenotype of an autism rat model \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe first top-ranked autoimmune disease in blood was systemic lupus erythematosus, followed by multiple sclerosis. Indeed, S1P receptor modulators are approved for relapsing remitting multiple sclerosis \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. However, in the patient-level analysis, none of the multiple sclerosis sample was found among the top-ranked diseases (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb).\u003c/p\u003e\u003cp\u003eBecause cenerimod showed efficacy in a phase 2b clinical trial (CARE, NCT03742037) in patients with moderate to severe SLE, we further evaluated SLE as an indication. The correlation between patient-level profile and cenerimod profile in the MRL/lpr mouse model in blood showed that SLE patients ranking at the top (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). However, SLE patients were found at different levels of the raking, which reflects the heterogeneity of the SLE population gene expression profile.\u003c/p\u003e\u003cp\u003eInterestingly, Crohn’s disease that origins in the gut was ranked at the patient-level among the top diseases to be complementary to the cenerimod profile. Supporting evidence of the obtained ranking for Crohn’s disease patients is the testing of S1P receptor modulators in clinical studies with Crohn’s disease patients (NCT03464097) and approval for ulcerative colitis \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. These results demonstrated the capacity of our framework in detecting complementary diseases beyond the tissue type tested preclinically.\u003c/p\u003e\u003cp\u003eThe inspection of the cenerimod profile from the MRL/lpr mouse model in kidneys revealed the kidney transplantation, diabetic nephropathy, chronic kidney disease and rapidly progressive glomerulonephritis patients ranking at the top (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec) and the highest complementarity (correlation = ~-0.4) was found for kidney transplantation at the disease-level analysis (\u003cb\u003eSupplementary Fig.\u0026nbsp;2a\u003c/b\u003e). This result was consistent with the testing of a S1P receptor modulator in kidney transplantation patients \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. In addition to kidney-related diseases, ulcerative colitis patients appeared in the top-ranking diseases both in the disease- and patient-level analysis for which S1P receptor modulators have been approved.\u003c/p\u003e\u003cp\u003eFinally, the cenerimod profile in the Scl-cGVHD mouse model in skin showed that the top-ranked disease patients included rheumatoid arthritis, atopic dermatitis, and psoriasis (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, \u003cb\u003eSupplementary Fig.\u0026nbsp;2b\u003c/b\u003e). This result was consistent with the demonstrated clinical efficacy of a S1P receptor modulator in psoriasis patients clinical trial (NCT01208090) \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eIn order to demonstrate the improved comparability of the data after the autoencoder processing, we compared the ranking results from the skin profile for rheumatoid arthritis patients before and after noise reduction. As expected, the ranking of rheumatoid arthritis samples from skin obtained different ranking before and after noise reduction (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, \u003cb\u003eSupplementary Fig.\u0026nbsp;2c\u003c/b\u003e). In particular, without denoising, the rheumatoid arthritis samples disappear from the top-ranks, leaving only atopic dermatitis and psoriasis patients in the top position, thus making the rankings tissue-specific.\u003c/p\u003e\u003cp\u003eThese results demonstrated the importance of the autoencoder pre-processing in increasing comparability across tissues.\u003c/p\u003e\u003cp\u003eFirstly, we presented the complementarity evaluation of cenerimod profile by considering separately the different tissue types. Then, we evaluated the complementarity results when combining the data from different tissue types and whether this improved sensitivity.\u003c/p\u003e\u003cp\u003eIn particular, SLE is a very heterogeneous disease not only from the gene expression profile but also in terms of clinical manifestations, affecting different organs depending on the degree of disease severity. Consequently, in order to obtain a more holistic view of the predicted efficacy of cenerimod across patients, we considered the heterogeneity of the phenotypic manifestations by selecting the highest complementarity (high anti-correlation) across all cenerimod tissues. As expected, the distributions changed considerably for SLE patients when considering the maximum anti-correlation across tissues, while the changes were much less pronounced for atopic dermatitis, for instance (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee). This observation reflected the inherent high heterogeneity of the SLE patients with a heterogeneous complementary of SLE patients with cenerimod treatment, providing important insights for SLE patients responder analysis.\u003c/p\u003e\u003cp\u003eIn addition, when combining the cenerimod profiles of both tissues and mouse models with all disease tissue types, we observed a consistently strong anti-correlation for the following top-ranked disease: Asperger syndrome (autism), kidney transplantation, SLE, multiple sclerosis and ulcerative colitis (\u003cb\u003eSupplementary Fig.\u0026nbsp;2d\u003c/b\u003e).\u003c/p\u003e\u003cp\u003eTake together these results showed the importance of integrating data from different animal models and characterizing the drug candidate effect in different tissues to get a holistic view about the potential complementary indications.\u003c/p\u003e\u003cp\u003eIn summary, our \u003cem\u003ein-silico\u003c/em\u003e indication discovery framework confirmed SLE as an autoimmune indication for cenerimod. Furthermore, indications such as Asperger syndrome (autism), multiple sclerosis, rheumatoid arthritis, atopic dermatitis, psoriasis, kidney transplantation, Crohn’s disease and ulcerative colitis emerged to be additional potential indications for cenerimod treatment. As described previously, none of the multiple sclerosis samples was found in the top-ranked diseases in the patient-level analysis, while the disease appeared at the sixth position in the disease-level analysis. This observation could be explained by the lack of data from brain samples from cenerimod preclinical studies, which might have reduced the sensitivity for such indications. This further highlights the importance of measuring drug gene expression profile in several animal models and tissues, in order to allow for a more holistic indication discovery.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDeterminants of response to cenerimod\u003c/b\u003e\u003c/p\u003e\u003cp\u003eOnce target indications were identified, the patient-level analysis provided an opportunity to gain insights into potential factors of the magnitude of the expected response to cenerimod.\u003c/p\u003e\u003cp\u003eAs described previously, the disease-level complementarity analysis showed considerable heterogeneity even among the top-ranked SLE patients. In order to gain insight into SLE patients’ heterogeneity from their clinical characteristics, we investigated what were the characteristics associated with the predicted magnitude of response to cenerimod. Because type-1-interferon (IFN-1) has previously been used in clinical trials as a stratification biomarker for SLE patients, we tested whether an IFN-1 signature can be used to estimate the cenerimod predicted response. The comparison of the predicted response with the expression level of the IFN-1 signature in SLE patients showed a higher predicted response in patients with higher IFN-1 signature (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea), which is consistent with the CARE study \u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo identify genes that are predictive of cenerimod response, a neural network classifier was implemented. The response variable was defined by applying cut-offs on patients’ complementarity (correlation coefficient r) to define responders (r \u0026lt; -0.1) and non-responders (r \u0026gt; 0) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). The neural network was trained on all available patient profiles to solve a binary classification problem and predict responders and non-responders.\u003c/p\u003e\u003cp\u003eIn order to identify the genes predicting cenerimod response in SLE, a Local Interpretable Model-agnostic Explanations (LIME) algorithm \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e was applied on SLE patient samples. The ranking of the most frequent genes with the highest predictive weight (i.e., high expression levels) included TUBB, EMP3, JCHAIN, TUBB6, ILR2G and EMP1, which were linked to response (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). These genes were combined in term of z-scores to define a \u003cem\u003ede novo\u003c/em\u003e discovered gene signature and assess its predictive value to cenerimod. The \u003cem\u003ede novo\u003c/em\u003e gene expression signature showed an anti-correlation of -0.57 with the predicted response (complementarity or correlation coefficient) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec). This indicated that SLE patients with high expression levels of the identified genes were likely better responding to cenerimod treatment.\u003c/p\u003e\u003cp\u003eSubsequently, we evaluated SLE patients’ characteristics such as disease activity (SLEDAI-2K) or age that were expected to be associated with the predicted response to cenerimod. Since the predicted response to cenerimod was shown to be higher depending on the strength of the molecular endotype, we expected a stronger positive effect of the drug on patients with a more severe disease phenotype. Indeed, patients with a higher disease (SLEDAI) score were predicted to have the higher response (or lower correlation coefficient) (\u003cem\u003eP\u003c/em\u003e = 0.0016), providing increasing evidence of the beneficial effects of cenerimod treatment for SLE patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed). In addition, younger patients age (\u0026lt; 20) was significantly associated (\u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001) with increased response (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ee), which is consistent with the increased disease severity in children compared to adults \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eValidation of cenerimod predictive biomarkers in a Phase 2b clinical trial in SLE patients\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo validate the ability of framework in identifying biomarkers predictive of cenerimod response, we evaluated clinical response based on mSLEDAI-2k disease activity score to the efficacious dose of cenerimod 4 mg in the CARE RNA-seq sub-study \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. The baseline expression levels of the top-predictors were evaluated in responders and non-responders receiving placebo or cenerimod 4 mg treatment. Linear regression models were used to determine significance of JCHAIN baseline levels to cenerimod 4 mg in responders compared to non-responders (p-value = 0.019) and identified JCHAIN as a novel predictive biomarkers (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eMoreover, TUBB gene was significantly decreased (p-value \u0026lt; 0.001) after 6 months of cenerimod 4 mg treatment in responders but not in non-responders and only in cenerimod treated patients, making TUBB a potential pharmacodynamic biomarker to monitor treatment response in the blood (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb).\u003c/p\u003e\u003cp\u003eIn conclusion, we demonstrated that our \u003cem\u003ein-silico\u003c/em\u003e indication discovery framework supported the selection of SLE as an indication for cenerimod treatment and the identification of predictive biomarkers that were validated in a Phase 2b clinical trial of moderate to severe patients with SLE (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec).\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cem\u003eMouse models and gene expression data\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThe sclerodermatous chronic graft-versus-house-disease (Scl-cGVHD) mouse model gene expression data from skin was obtained from Kano et al 2019 \u003csup\u003e29\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eFor the lupus mouse model, experiment was conducted as described in Strasser et al 2020 \u003csup\u003e16\u003c/sup\u003e. In brief, 7-week-old female MRL\u003cem\u003e/lpr\u003c/em\u003e (Jackson Laboratories) mice were treated with vehicle or cenerimod (0.2 mg/g food) until the end of the study, which was predefined as the time point when at least 20% morbidity/mortality was reached in one group (end of treatment week 10). RNA from blood and kidney from cenerimod- and vehicle-treated mice was isolated and prepared for sequencing using NUGEN Universal Plus mRNA kit with PolyA select (and globin depletion for the blood samples) according to protocol and sequenced on Illumina NextSeq 500 mRNA-seq platform. Age matched w C57BL/6 mice were used as healthy reference. The samples were prepared and sequenced as above.\u003c/p\u003e\u003cp\u003e All experimental procedures were conducted in accordance with the Swiss animal welfare ordinance and Idorsia Animal Welfare policy on the use of experimental animals. The study was approved by the Basel-land Cantonal Veterinary Home Office. Experiments and results were reported according to the ARRIVE guidelines 2.0.\u003c/p\u003e\u003cp\u003e\u003cem\u003eData and processing\u003c/em\u003e\u003c/p\u003e\u003cp\u003eAll gene expression data was pre-processed, curated and retrieved from the Genevestigator application \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Data was separated by study, and within the study by cell type and disease state (healthy or disease) as defined in Genevestigator. To obtain a collapsed profile of the state \"healthy\", and to account for outliers, all samples from healthy subjects with the same cell type in the study were compared with each other resulting in a correlation matrix. By taking the row means of this matrix, divided by the total sum of correlation, weights for each healthy sample were obtained by calculating a weighted mean. Fold-changes were calculated by dividing each disease sample with the corresponding healthy weighted mean. Pseudocounts were added to healthy and disease samples to avoid overamplification of noise (pseudocounts were 8 for gene expression array data and 0.25 for RNA-seq data), these pseudocounts were chosen based on visual inspection of histograms of expression data across all genes and samples. Samples from mouse disease models were processed in the same way either as disease model versus healthy reference mice or cenerimod-treated versus vehicle-treated disease model.\u003c/p\u003e\u003cp\u003eMouse and human data were then merged based on orthologous genes obtained by the getLDS function from the biomaRt Bioconductor package. Duplicated genes were discarded.\u003c/p\u003e\u003cp\u003e\u003cem\u003eNeural networks\u003c/em\u003e\u003c/p\u003e\u003cp\u003eNeural networks were engineered using the R package Keras with TensorFlow backend. The data was split 80/20 for training and testing. Multiple designs were tested with different numbers of layers and nodes as well as activation functions. The design shown in Supplementary Fig.\u0026nbsp;1 resulted in the best balance of loss and overfitting while still having a low number of nodes in the coding layer for the autoencoder.\u003c/p\u003e\u003cp\u003e\u003cem\u003ePerformance metrics\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eIn-silico Identification of predictive biomarkers\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eNeural network was trained on all patient data to predict response or no response (-0.1 and 0 as cutoff respectively). Subsequently, the LIME (Local Interpretable Model-agnostic Explanations) algorithm was used to identify features (genes) frequently used for \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e \u003cem\u003ede novo\u003c/em\u003e signature calculation.\u003c/p\u003e\n\u003cp\u003eTo quantify the overall change across this set of genes the mean z-scores were calculated across the genes for each patient and compared to the correlation with the cenerimod profile.\u003c/p\u003e\n\u003ch2\u003eStatistical analysis\u003c/h2\u003e\n\u003cp\u003eStatistical analysis was performed using R software, version 4.1.2 (R Foundation for Statistical Computing). Correlations were calculated using the R cor function, in this study Pearson and Spearman correlations were calculated as indicated in the figure legends. Significance of correlations were tested when focussing on SLE patients in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e using the R function cor.test.\u003c/p\u003e\n\u003cp\u003eDifferences in gene expression levels at baseline between groups (i.e., responders vs. non-responders) were tested using linear regression models. When comparing gene expression levels between time points (baseline vs. month 6), mixed effects models with a subject random effect were used (lmer function). Resulting p-values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate method.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe ability to link a drug candidate with the indication or patients it can have the greatest effect is a key determinant of the success in drug development. Taking advantage of a drug candidate gene expression profile measured in preclinical models with the wealth of patient disease gene expression profiles available in the public domain, provides great potential to support unbiased indication discovery.\u003c/p\u003e\u003cp\u003eIn this study, an \u003cem\u003ein-silico\u003c/em\u003e discovery framework enabling the matching of drug candidates with indications based on the gene expression profiles and the identification of associated predictive biomarkers is presented.\u003c/p\u003e\u003cp\u003eThis framework is not limiting the complementarity investigation to specific gene expression signatures but allows to investigate broader effects by considering the whole gene expression profile effects and quantitatively compared changes for a holistic assessment of molecular effects. In addition, our approach has the potential of being extended to other data types such as proteomics or metabolomics data.\u003c/p\u003e\u003cp\u003eThe patient-level analysis enabled to decipher the disease patient population heterogeneity with molecularly and clinically. This approach was facilitated by focussing on \u0026bdquo;reverse curation\" of patient data of interest, avoiding upfront intensive efforts for sample annotation in the entire patient dataset.\u003c/p\u003e\u003cp\u003eCenerimod treatment was used to exemplify our \u003cem\u003ein-silico\u003c/em\u003e discovery framework, which predicted its potential efficacy in several autoimmune diseases such as SLE, multiple sclerosis, rheumatoid arthritis, psoriasis, Crohn\u0026rsquo;s disease, and kidney transplantation. Indeed, several of the predicted indications were previously clinically tested with S1P receptor modulators. In addition, novel indications such as Asperger syndrome were found to be complementary to the cenerimod profile, representing a potential treatment opportunity for additional central nervous system diseases with increasing evidence of immune system dysfunction \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. Applying the framework to individual patients with SLE allowed to discover determinants linked to response, which led to the identification of predictive biomarkers. Our major findings included the identification of TUBB gene as a potential pharmacodynamic biomarker to monitor patient response in the blood and JCHAIN gene as a predictive biomarker for the clinical compound cenerimod.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthors contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDH: Conceptualization, Data curation, formal analysis, writing \u0026ndash; original draft, writing \u0026ndash; review and editing.\u003c/p\u003e\n\u003cp\u003eMS: Conceptualization, Data curation, formal analysis, writing \u0026ndash; original draft, writing \u0026ndash; review and editing.\u003c/p\u003e\n\u003cp\u003ePC: Data curation, review and editing.\u003c/p\u003e\n\u003cp\u003eDS: Conceptualization, Data curation, formal analysis, writing \u0026ndash; original draft, writing \u0026ndash; review and editing.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;All authors had full access to all the data in the study. All authors were responsible for the decision to submit the article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRNA-Seq data were deposited into the Gene Expression Omnibus database under accession number GSE293404 and are available at the following URL: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE293404\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cu\u003eAcknowledgement:\u003c/u\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe are grateful to Estelle Gerossier-Creusat and Sylvie Froidevaux for providing animal model tissue samples and B\u0026eacute;reng\u0026egrave;re Renault for RNA-seq library preparation and sequencing.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003e\u003cu\u003eCompeting interest:\u003c/u\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDH has no competing interest\u003c/p\u003e\n\u003cp\u003eMS has no competing interest\u003c/p\u003e\n\u003cp\u003ePC is a holder of stocks of Idorsia Pharmaceuticals Ltd\u003c/p\u003e\n\u003cp\u003eDSS is a holder of stocks and stock options of Idorsia Pharmaceuticals Ltd\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eScannell JW, Blanckley A, Boldon H, Warrington B. Diagnosing the decline in pharmaceutical R\u0026amp;D efficiency. \u003cem\u003eNat Rev Drug Discov\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 191-200 (2012).\u003c/li\u003e\n\u003cli\u003eShih HP, Zhang X, Aronov AM. Drug discovery effectiveness from the standpoint of therapeutic mechanisms and indications. \u003cem\u003eNat Rev Drug Discov\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 19-33 (2018).\u003c/li\u003e\n\u003cli\u003eArrowsmith J. Phase II failures: 2008\u0026ndash;2010. \u003cem\u003eNature Reviews Drug Discovery\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 328-329 (2011).\u003c/li\u003e\n\u003cli\u003eArrowsmith J, Miller P. Phase II and Phase III attrition rates 2011\u0026ndash;2012. \u003cem\u003eNature Reviews Drug Discovery\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 569-569 (2013).\u003c/li\u003e\n\u003cli\u003eHarrison RK. Phase II and phase III failures: 2013-2015. \u003cem\u003eNat Rev Drug Discov\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 817-818 (2016).\u003c/li\u003e\n\u003cli\u003eLiu R, Wei L, Zhang P. A deep learning framework for drug repurposing via emulating clinical trials on real-world patient data. \u003cem\u003eNat Mach Intell\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 68-75 (2021).\u003c/li\u003e\n\u003cli\u003eBlanco-Gonzalez A, Cabezon A, Seco-Gonzalez A, Conde-Torres D, Antelo-Riveiro P, Pineiro A, Garcia-Fandino R. The Role of AI in Drug Discovery: Challenges, Opportunities, and Strategies. \u003cem\u003ePharmaceuticals (Basel)\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eZhu J\u003cem\u003e, et al.\u003c/em\u003e Prediction of drug efficacy from transcriptional profiles with deep learning. \u003cem\u003eNat Biotechnol\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 1444-1452 (2021).\u003c/li\u003e\n\u003cli\u003eLamb J\u003cem\u003e, et al.\u003c/em\u003e The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e313\u003c/strong\u003e, 1929-1935 (2006).\u003c/li\u003e\n\u003cli\u003eZeng X, Zhu S, Liu X, Zhou Y, Nussinov R, Cheng F. deepDR: a network-based deep learning approach to in silico drug repositioning. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 5191-5198 (2019).\u003c/li\u003e\n\u003cli\u003eChen B, Butte AJ. Leveraging big data to transform target selection and drug discovery. \u003cem\u003eClin Pharmacol Ther\u003c/em\u003e \u003cstrong\u003e99\u003c/strong\u003e, 285-297 (2016).\u003c/li\u003e\n\u003cli\u003eHruz T\u003cem\u003e, et al.\u003c/em\u003e Genevestigator v3: a reference expression database for the meta-analysis of transcriptomes. \u003cem\u003eAdv Bioinformatics\u003c/em\u003e \u003cstrong\u003e2008\u003c/strong\u003e, 420747 (2008).\u003c/li\u003e\n\u003cli\u003eRibeiro MT, Singh S, Guestrin C. \u0026quot;Why Should I Trust You?\u0026quot;: Explaining the Predictions of Any Classifier. In: \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e). Association for Computing Machinery (2016).\u003c/li\u003e\n\u003cli\u003eSirota M\u003cem\u003e, et al.\u003c/em\u003e Discovery and preclinical validation of drug indications using compendia of public gene expression data. \u003cem\u003eSci Transl Med\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 96ra77 (2011).\u003c/li\u003e\n\u003cli\u003ePerry D, Sang A, Yin Y, Zheng YY, Morel L. Murine models of systemic lupus erythematosus. \u003cem\u003eJ Biomed Biotechnol\u003c/em\u003e \u003cstrong\u003e2011\u003c/strong\u003e, 271694 (2011).\u003c/li\u003e\n\u003cli\u003eStrasser DS\u003cem\u003e, et al.\u003c/em\u003e Preclinical to clinical translation of cenerimod, a novel S1P(1) receptor modulator, in systemic lupus erythematosus. \u003cem\u003eRMD Open\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, (2020).\u003c/li\u003e\n\u003cli\u003eKano M\u003cem\u003e, et al.\u003c/em\u003e Attenuation of murine sclerodermatous models by the selective S1P1 receptor modulator cenerimod. \u003cem\u003eScientific Reports\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 658 (2019).\u003c/li\u003e\n\u003cli\u003eFinotello F, Di Camillo B. Measuring differential gene expression with RNA-seq: challenges and strategies for data analysis. \u003cem\u003eBrief Funct Genomics\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 130-142 (2015).\u003c/li\u003e\n\u003cli\u003eJain V, Seung H. \u003cem\u003eNatural Image Denoising with Convolutional Networks\u003c/em\u003e (2008).\u003c/li\u003e\n\u003cli\u003eWang H\u003cem\u003e, et al.\u003c/em\u003e Potential serum biomarkers from a metabolomics study of autism. \u003cem\u003eJ Psychiatry Neurosci\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 27-37 (2016).\u003c/li\u003e\n\u003cli\u003eWu H\u003cem\u003e, et al.\u003c/em\u003e Modulation of sphingosine 1-phosphate (S1P) attenuates spatial learning and memory impairments in the valproic acid rat model of autism. \u003cem\u003ePsychopharmacology (Berl)\u003c/em\u003e \u003cstrong\u003e235\u003c/strong\u003e, 873-886 (2018).\u003c/li\u003e\n\u003cli\u003eKappos L\u003cem\u003e, et al.\u003c/em\u003e A placebo-controlled trial of oral fingolimod in relapsing multiple sclerosis. \u003cem\u003eN Engl J Med\u003c/em\u003e \u003cstrong\u003e362\u003c/strong\u003e, 387-401 (2010).\u003c/li\u003e\n\u003cli\u003eQiu, Jingyue, et al. Efficacy and safety of the S1PR modulator etrasimod in the treatment of moderately to severely active ulcerative colitis during the induction phase: a systematic review and meta-analysis of randomized controlled trials. \u003cem\u003eFrontiers in Pharmacology\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e (2024): 1420455.\u003c/li\u003e\n\u003cli\u003eSkerjanec A, Tedesco H, Neumayer HH, Cole E, Budde K, Hsu CH, Schmouder R. FTY720, a novel immunomodulator in de novo kidney transplant patients: pharmacokinetics and exposure-response relationship. \u003cem\u003eJ Clin Pharmacol\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 1268-1278 (2005).\u003c/li\u003e\n\u003cli\u003eVaclavkova A\u003cem\u003e, et al.\u003c/em\u003e Oral ponesimod in patients with chronic plaque psoriasis: a randomised, double-blind, placebo-controlled phase 2 trial. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e384\u003c/strong\u003e, 2036-2045 (2014).\u003c/li\u003e\n\u003cli\u003eAskanase, Anca D., et al. Cenerimod, a sphingosine-1-phosphate receptor modulator, versus placebo in patients with moderate-to-severe systemic lupus erythematosus (CARE): an international, double-blind, randomised, placebo-controlled, phase 2 trial. \u003cem\u003eThe Lancet Rheumatolog\u003c/em\u003ey \u003cstrong\u003e7.1\u003c/strong\u003e: e21-e32(2025).\u003c/li\u003e\n\u003cli\u003edas Chagas Medeiros MM\u003cem\u003e, et al.\u003c/em\u003e Clinical and immunological aspects and outcome of a Brazilian cohort of 414 patients with systemic lupus erythematosus (SLE): comparison between childhood-onset, adult-onset, and late-onset SLE. \u003cem\u003eLupus\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 355-363 (2016).\u003c/li\u003e\n\u003cli\u003eSuffiotti, Madeleine, et al. Pharmacodynamics of the S1P1 receptor modulator cenerimod in a phase 2b randomised clinical trial in patients with moderate to severe SLE. \u003cem\u003eAnnals of the Rheumatic Diseases\u003c/em\u003e (2025).\u003c/li\u003e\n\u003cli\u003eKano, Miyu, et al. Attenuation of murine sclerodermatous models by the selective S1P1 receptor modulator cenerimod. \u003cem\u003eScientific Reports\u003c/em\u003e \u003cstrong\u003e9.1: 658\u003c/strong\u003e. (2019)\u003c/li\u003e\n\u003cli\u003eEnstrom AM, Van de Water JA, Ashwood P. Autoimmunity in autism. \u003cem\u003eCurr Opin Investig Dr\u003c/em\u003e\u003cem\u003eugs\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 463-473 (2009).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"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":"","lastPublishedDoi":"10.21203/rs.3.rs-6993651/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6993651/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTo fully explore a drug candidate's therapeutic potential, assessing its effectiveness across all possible diseases is essential. While traditional approaches match drug mechanisms to disease pathophysiology, they are limited by the high costs and slow progress of preclinical and clinical trials. This study introduces a novel in silico framework to identify new indications for drug candidates or repurpose approved drugs by analyzing their effects on gene expression in patients or animal models compared to controls.\u003c/p\u003e\u003cp\u003eThe framework integrate data from 13,602 patient samples across 146 diseases with drug candidate tested in preclinical models and use a neural network to reduce noise and improve sensitivity.\u003c/p\u003e\u003cp\u003eThe framework was exemplified with cenerimod, a S1P1 receptor modulator, which predicted its efficacy in immune-related diseases such as SLE, Psoriasis, and Crohn\u0026rsquo;s disease and kidney transplantation complications. Additionally, it identified six genes predictive of maximal clinical response in SLE patients, validated using RNA-seq data from a phase 2b cenerimod trial.\u003c/p\u003e","manuscriptTitle":"A novel translational in-silico indication discovery framework identifies indications and predictive biomarkers for cenerimod","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-20 06:46:19","doi":"10.21203/rs.3.rs-6993651/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":"78c0e3e4-eab3-49b8-8a09-d03ef07b8519","owner":[],"postedDate":"August 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":53226994,"name":"Biological sciences/Drug discovery/Biomarkers"},{"id":53226995,"name":"Biological sciences/Drug discovery"},{"id":53226996,"name":"Biological sciences/Immunology"}],"tags":[],"updatedAt":"2025-12-05T18:23:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-20 06:46:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6993651","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6993651","identity":"rs-6993651","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}