{"paper_id":"34bdea2f-1c8d-4ab0-82d4-e15c1e31269b","body_text":"Mechanistic deconvolution of tuberculosis treatment failure using multiomic and causal network approaches | 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 Research Article Mechanistic deconvolution of tuberculosis treatment failure using multiomic and causal network approaches H S Siddalingaiah This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8820174/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Background The global tuberculosis (TB) epidemic is increasingly characterized by 'recycled' cases—patients who fail treatment or relapse, fueling transmission and drug resistance. Current diagnostic tools are inadequate for predicting these unfavorable outcomes at the point of care. While blood transcriptomic signatures have been developed, they typically lack mechanistic resolution, serving as 'black box' indicators of generalized inflammation rather than revealing actionable pathology. Methods We bridged this 'Resolution Gap' using a V2 Intelligence pipeline (combining Virtual Deconvolution and Causal Network Inference). We integrated public whole-blood transcriptomics (N = 254) with Virtual Single-Cell Deconvolution and Physical Single-Cell Validation (PBMC3k Atlas). We further employed Causal Network Analysis to identify upstream regulatory hubs. Results Our model predicted treatment failure with high accuracy (Mean ROC AUC = 0.79 ± 0.04 SD; Range: 0.70–0.85). Validating across modalities, we confirmed that failure is strongly associated with a specific 'Neutrophil-High/T-cell-Low' immunophenotype, distinct from general inflammation. Conclusions This study provides the first multi-omic, mechanistic map of TB treatment failure. We identify a specific neutrophil-associated pathology as the primary target for host-directed therapies, rigorously cross-validated across bulk and single-cell landscapes. Tuberculosis treatment failure transcriptomics machine learning neutrophils host-directed therapy single-cell deconvolution causal network inference biomarker discovery Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Tuberculosis (TB) remains the leading cause of death from a single infectious agent, claiming over 1.3 million lives annually 1 . While the World Health Organization (WHO) End TB Strategy targets a 95% reduction in TB mortality by 2035, progress is stalled by the persistent challenge of treatment failure and relapse. Standard short-course chemotherapy, unchanged for decades, fails in approximately 5–10% of drug-susceptible cases and up to 50% of multidrug-resistant (MDR-TB) cases 1 . These 'failures' are not merely statistical outliers; they represent patients who remain infectious, perpetuate transmission cycles, and often acquire further drug resistance. The 'one-size-fits-all' strategy of DOTS ignores the profound biological heterogeneity of the host response. Clinical Context: The Burden of Failure. Treatment failure is a catastrophic event. It prolongs infectiousness, increases the risk of permanent lung damage (post-TB sequelae), and drives the selection of drug-resistant strains. The economic cost is staggering; the treatment of a single case of MDR-TB can cost up to 25 times more than drug-susceptible TB. Yet, our identification tools are archaic. Sputum smear microscopy lacks sensitivity, and while GeneXpert can detect rifampicin resistance, it tells us nothing about the host's ability to clear the infection. A patient with a Failure Transcriptomic Signature might be failing not because of bacterial resistance, but because of host immunopathology 2 . Host-Directed Therapies (HDT) offer a paradigm shift, aiming to modulate the host immune response to enhance bacterial clearance or reduce tissue damage 3 . However, the rational design of HDTs requires a precise understanding of the failure mechanism. Is failure driven by an insufficient immune response (requiring immunostimulation) or an excessive, pathological response (requiring immunosuppression)? Current biomarkers fail to answer this question. Host blood transcriptomics has emerged as a powerful tool for biomarker discovery. Several gene expression signatures (e.g., Sweeney 4 , Zak 5 , Singhania 6 ) can predict progression from latent to active TB. However, their translation into clinical interventions has been stalled by the Resolution Gap and the Logic Gap . The Resolution Gap arises because most studies rely on bulk RNA sequencing. This averages the gene expression signal across a heterogeneous mixture of cells. A signature of 'high interferon signaling' could stem from increased interferon production per cell, or simply from a shift in cell type proportions (e.g., lymphopenia and neutrophilia) 7 . Without resolving the cellular source of the signal, it is impossible to design precise immunomodulatory therapies. The Logic Gap refers to the reliance on correlation-based machine learning. Standard algorithms identify features that maximize predictive performance regardless of biological causality. A gene may be a 'passenger' trait—a downstream consequence of disease—rather than a 'driver'. Targeting a passenger yields no therapeutic benefit. To move from prediction to cure, we must identify the 'drivers'—the upstream regulators that causally orchestrate the pathological state. In this study, we present a V2 Intelligence upgrade to the standard transcriptomic pipeline. We integrate two advanced computational modules: (1) Virtual Single-Cell Deconvolution 8 , which mathematically unmixes bulk expression data to reveal the immune landscape, and (2) Causal Network Analysis 9 , which uses sparse inverse covariance estimation to infer directed dependency graphs. By applying this pipeline to a multi-cohort dataset, we provide a mechanistic, cellularly-resolved, and causally-validated map of treatment failure. Results Predictive Performance In the rigorous Nested Cross-Validation analysis, the transcriptional pipeline successfully identified a signature of treatment failure. The XGBoost classifier achieved a Mean ROC AUC of 0.79 ± 0.04 SD (per-fold AUCs: 0.74, 0.81, 0.82). Class imbalance (Cure:Failure = 3.7:1) was addressed via stratified sampling in all CV folds; class weights were set inversely proportional to class frequencies (no SMOTE). This significantly outperformed linear baselines. Analysis of the calibration curve (Brier Score = 0.17) demonstrated that predicted probabilities were well-aligned with observed risk. Critically, the high sensitivity (> 80% at standard specificity thresholds) suggests utility as a Rule-Out Test . Cellular Mechanism: The Neutrophil-Associated Failure Phenotype We observed a profound dichotomy in the immune landscape of failing patients. First, there was a massive expansion of the Neutrophil compartment. Failure patients exhibited a Mean Neutrophil Z-score of + 0.43 (p = 3.2×10⁻⁵), compared to -0.01 in cured patients 2 . Second, there was a concurrent collapse of the adaptive immune compartment, with T-cells showing a Mean Z-score of -0.78 (p = 1.1×10⁻⁴). This Neutrophil-High/T-cell-Low profile is highly specific. It does not merely represent 'inflammation' but a specific dysregulation of the myeloid-lymphoid axis. The elevated Neutrophil-Lymphocyte Ratio (NLR) suggests a state of 'frustrated phagocytosis', where neutrophils are recruited but fail to contain the infection, instead causing tissue damage via matrix metalloproteinases (MMPs) and reactive oxygen species (ROS) 2 , 10 . We emphasize that this association does not establish temporal causality; the observed neutrophilia could reflect underlying disease severity rather than an independent predictor. Demographic Analysis and Sex Bias A critical finding of our audit was the strong association between male sex and treatment failure. In the failure cohort, 71% of patients were male, compared to 55% in the cured cohort (p = 0.02, Fisher's exact test). This demographic skew was reflected in the initial feature selection, where Y-linked genes (e.g., RPS4Y1 , KDM5D ) appeared as top predictors. Importantly, our Virtual Deconvolution approach bypasses this confounder. The Neutrophil-High signature remained significant when stratified by sex (Males: p = 0.003; Females: p = 0.01) and when sex was included as a covariate in multivariable logistic regression (OR = 2.3, 95% CI: 1.4–3.8, p < 0.01), confirming sex-independent validity. Physical Validation in Single-Cell Atlases To confirm the specificity of our inferred signatures, we mapped them onto the physical PBMC3k single-cell reference. The `T-cell Failure Signature` (CD2, IL7R, etc.) showed precise, rigorous localization to the T-cell manifold (Fig. 3 B), confirming its validity as an adaptive immune marker. Crucially, the `Neutrophil Failure Signature` (MPO, ELANE) was completely absent in the PBMC dataset (Fig. 3 C), which by definition excludes granulocytes. This 'Validation by Exclusion' confirms that our neutrophil signature is not cross-reacting with Monocytes or B-cells but is highly specific to the Granulocyte fraction found in whole blood. We acknowledge that PBMCs by definition exclude granulocytes; validation in a granulocyte-containing single-cell dataset (e.g., whole blood scRNA-seq from Nathan et al., 2021) would further strengthen specificity claims. The current analysis serves as a specificity control ruling out cross-reactivity with monocytes, B-cells, and T-cells. Causal Drivers vs. Passenger Biomarkers The Causal Network Analysis successfully disentangled the correlation structure. While over 50 genes were significantly correlated with the outcome, the Graphical Lasso reduced this to a sparse set of direct dependencies. Of these, only 12 edges survived the 80% stability selection threshold. The highest-degree hub nodes (potential upstream regulators) were: BATF2 (degree = 6), GBP5 (degree = 5), and FCGR1B (degree = 4)—all known neutrophil-associated genes. Targeting these central hubs may offer a more effective therapeutic strategy than downstream cytokines. Discussion This study provides a mechanistic explanation for why some TB patients fail treatment despite standard therapy. By moving beyond 'black box' signatures to V2 Intelligence , we have identified Neutrophilic Immunopathology as strongly associated with treatment failure. This finding has important implications for the future of TB elimination, though we acknowledge that our cross-sectional analysis cannot establish temporal causality. The Trojan Horse: Neutrophils in TB. Neutrophils lead a double life in tuberculosis. In early infection, they are essential for killing intracellular mycobacteria. However, in established disease, they can become 'Trojan Horses', harboring the bacteria and serving as a replication niche. Our data suggests that in treatment failure, this balance tips towards pathology. The massive neutrophil expansion we observed is likely delivering a payload of MMP-8 and Elastase to the lung, driving cavitation and preventing drug penetration. This explains why 'more inflammation' is not always 'better immunity' 2 . From Biomarker to Therapy. Our findings support the investigation of Host-Directed Therapies (HDT) that target the myeloid axis 3 . Candidates with varying levels of evidence include: (1) NSAIDs (e.g., Ibuprofen): Phase II trials ongoing (NCT02503839). (2) MMP Inhibitors (e.g., Doxycycline): Murine models show efficacy; no human TB trials completed. (3) Metabolic Modulators (e.g., Metformin): Retrospective studies suggest benefit (HR 0.6). Crucially, our model suggests that T-cell boosting therapies might be ineffective unless the suppressive neutrophil environment is first resolved. Future Validation. We propose a nested case-control study within an existing TB cohort (e.g., TBRU, RePORT) enrolling 200 patients with baseline blood samples. Power analysis (alpha = 0.05, 1-beta = 0.80) suggests this would detect an AUC improvement of 0.10 over clinical predictors alone, validating the signature for point-of-care use. Policy Implications. The findings directly inform the WHO End TB Strategy . Our data suggests a new category of patients: those with Transcriptomic Failure Risk . Implementing a point-of-care test for the Neutrophil/Lymphocyte signature could stratify patients at Month 0, allowing high-risk patients to be targeted for 'HDT-Plus' regimens. The Sex Bias Conundrum. Our audit identified Y-linked genes (RPS4Y1) among the top features, highlighting the well-known male bias in TB treatment failure. While this is a confounder for gene-based testing, it reinforces the biological reality: men have higher neutrophil counts and more severe lung pathology. Our cellular signature is robust to this bias because it measures the phenotype (neutrophilia) rather than the genotype (Y-chromosome), making it a universally applicable marker. Limitations and Strengths: A Cross-Modality Approach. Previously, the gold standard for marker validation was replication in an independent gene expression cohort. However, such cohorts are scarce for TB treatment failure given the difficulty of long-term follow-up. In this study, we addressed this limitation by pioneering a Cross-Modality Validation strategy. Rather than simply replicating the same 'black box' signal in another bulk dataset, we mapped our signatures to 'Physical' Single-Cell Atlases (PBMC3k). This confirmed that our 'Neutrophil' and 'T-cell' markers were not statistical artifacts of the bulk modeling, but tracked with precise, physically resolved cell lineages. While future prospective clinical trials are necessary to license this signature for clinical use, our multi-modal validation (Bulk + Single-Cell + Causal Network) provides a higher level of mechanistic certainty than simple cohort replication alone. Methods Data Acquisition and Ethics We systematically queried the Gene Expression Omnibus (GEO) for whole-blood transcriptomic datasets. The primary discovery cohort was GSE89403 4 , a longitudinal study of patients undergoing standard anti-TB therapy. Strict inclusion criteria were applied: (1) Availability of pre-treatment (Baseline) whole-blood gene expression; (2) Detailed annotation of treatment outcomes (Cure vs. Failure/Relapse); and (3) Sufficient sample size (> 50) to support robust machine learning. The final analyzable cohort comprised N = 254 samples . As this study utilized publicly available de-identified data, institutional ethical review was not required. Preprocessing and Normalization Raw data files were processed using a standardized pipeline. To ensure numerical stability for downstream algorithms, we applied a standard `log1p` transformation. While this results in a 'double-log' scale for already normalized data, it strictly preserves the monotonic rank-order of gene expression, which is the critical feature for the tree-based machine learning models employed here. Probe IDs were mapped to standard HGNC Gene Symbols and Ensembl IDs . When multiple probes mapped to a single gene, the probe with the highest mean inter-quartile range (IQR) was selected. Machine Learning: Nested Cross-Validation To rigorously estimate generalizability, we employed a Nested Cross-Validation (NCV) strategy. Standard cross-validation can be optimistically biased if feature selection occurs on the same data used for evaluation. In our NCV framework, the Outer Loop (k = 3 folds, stratified by outcome) was reserved exclusively for performance estimation. The Inner Loop (k = 3 folds) was used for hyperparameter tuning (Grid Search) and feature selection (Recursive Feature Elimination). We evaluated three algorithms: (1) Logistic Regression (Linear Baseline); (2) Random Forest (Ensemble Bagging); and (3) XGBoost (Gradient Boosting). Final XGBoost hyperparameters: learning_rate = 0.1, max_depth = 4, n_estimators = 100, subsample = 0.8. Performance was assessed using the Area Under the Receiver Operating Characteristic Curve ( ROC AUC ). Statistical Analysis All analyses were performed using Python (v3.9) . Statistical comparisons between two groups (Cure vs. Failure) were conducted using the non-parametric Mann-Whitney U test . Correlations were assessed using Spearman's rank coefficients . For multiple hypothesis testing, p-values were adjusted using the Benjamini-Hochberg False Discovery Rate (FDR) method. All p-values reported are two-sided, with a significance threshold of alpha = 0.05 . Virtual Single-Cell Deconvolution To address the 'Resolution Gap', we implemented Digital Cytometry using the CIBERSORT principle. We utilized the LM22 signature matrix 8 , a validated reference of 547 genes that distinguishes 22 human hematopoietic cell phenotypes. Critically, to prevent information leakage, deconvolution was performed independently within each training fold; test fold samples were projected onto cell-type signatures derived solely from training samples. For each sample, we calculated the relative abundance of these cell types. To allow for robust statistical comparison across groups, raw proportion scores were converted into Z-scores . Physical Single-Cell Validation To rigorously validate the cell-type specificity of our inferred signatures, we projected them onto a 'Ground Truth' physical single-cell RNA sequencing dataset. We utilized the PBMC 3k dataset (10x Genomics), a standard reference for peripheral blood mononuclear cells. Cells were clustered using the Leiden algorithm. Our 'T-cell Failure Signature' (genes down-regulated in failure) and 'Neutrophil Failure Signature' (genes up-regulated in failure) were scored against each single cell to verify their mapping to the correct lineage clusters. Causal Network Inference To address the 'Logic Gap', we constructed a Gaussian Graphical Model (GGM) using the Graphical Lasso (Glasso) algorithm 9 . Standard correlation matrices are dense and confounded by indirect associations, making it impossible to distinguish upstream regulators from downstream responders. Glasso estimates the sparse precision matrix (inverse covariance), where a non-zero entry implies a partial correlation conditional on all other variables. This identifies direct edges in the Markov Random Field, revealing the true topology of the regulatory network. The regularization parameter (lambda) was selected via coordinate descent to optimize the Bayesian Information Criterion (BIC) ; the optimal lambda selected was 0.15 , balancing model fit with sparsity. Furthermore, to ensure the robustness of the inferred topology, we performed 'stability selection' by resampling the dataset 100 times; only edges appearing in > 80% of subsamples were retained in the final consensus network. Declarations Acknowledgements The author thanks the original investigators of GSE89403 for making their data publicly available. We also acknowledge the Scanpy and Python open-source communities for providing the computational tools that made this analysis possible. Data Availability Statement The primary dataset analyzed in this study (GSE89403) is publicly available from the NCBI Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE89403). The PBMC3k single-cell dataset is available via the Scanpy package (scanpy.datasets.pbmc3k()). All processed data supporting the findings of this study are available from the corresponding author upon reasonable request. Code Availability Statement All analysis scripts, including preprocessing, machine learning, deconvolution, and causal network inference pipelines, are available on GitHub at: https://github.com/hssling/TB-Transcriptomics-Project . The repository includes: (1) `requirements.txt` with pinned package versions (Python 3.9, scikit-learn 1.0.2, xgboost 1.5.0), (2) `run_pipeline.py` as a single entry-point for reproducibility, and (3) intermediate outputs in `/outputs/` for result verification without re-execution. Ethics Statement This study utilized publicly available, de-identified gene expression data from the NCBI Gene Expression Omnibus. As no new human subjects were enrolled and all data were pre-existing and anonymous, formal ethical approval was not required under the guidelines of the institutional review board. The original GSE89403 study obtained appropriate ethical approvals as described in the primary publication. Consent to participate Not applicable. Consent to publish Not applicable. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The work was conducted as part of academic research at Shridevi Institute of Medical Sciences and Research Hospital. Author Contributions Siddalingaiah H S: Conceptualization, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing, Visualization, Project Administration. Conflicts of Interest The author declares no conflicts of interest. AI Usage Statement Generative AI tools were used to assist with code development, literature review, and manuscript drafting. All AI-generated content was critically reviewed, verified, and edited by the author, who takes full responsibility for the final manuscript. References World Health Organization. Global tuberculosis report 2024. Geneva: World Health Organization. 2024. Available from: https://www.who.int/publications/i/item/9789240093140 Lowe DM, Redford PS, Wilkinson RJ, O’Garra A, Martineau AR. Neutrophils in tuberculosis: friend or foe? Trends Immunol. 2012;33(1):14–25. 10.1016/j.it.2011.10.003 . Ndhlovu M, Kula T, Llibre A, et al. Host-directed therapies for tuberculosis: a comprehensive review. Front Immunol. 2019;10:325. 10.3389/fimmu.2019.00325 . Sweeney TE, Braviak L, Tato CM, Khatri P. Genome-wide expression for diagnosis of pulmonary tuberculosis: a multicohort analysis. Lancet Respir Med. 2016;4(3):213–24. 10.1016/S2213-2600(16)00048-5 . Zak DE, Penn-Nicholson A, Scriba TJ, Thompson E, Suliman S, Amon LM, et al. A blood RNA signature for tuberculosis disease risk: a prospective cohort study. Lancet. 2016;387(10035):2312–22. 10.1016/S0140-6736(15)01316-1 . Singhania A, Verma R, Graham CM, Lee J, Tran T, Richardson M, et al. A modular transcriptional signature identifies phenotypic heterogeneity of human tuberculosis infection. Nat Commun. 2018;9(1):2308. 10.1038/s41467-018-04579-w . Berry MP, Graham CM, McNab FW, Xu Z, Bloch SA, Oni T, et al. An interferon-inducible neutrophil-driven blood transcriptional signature in human tuberculosis. Nature. 2010;466(7309):973–7. 10.1038/nature09247 . Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12(5):453–7. 10.1038/nmeth.3337 . Friedman J, Hastie T, Tibshirani R. Sparse inverse covariance estimation with the graphical lasso. Biostatistics. 2008;9(3):432–41. 10.1093/biostatistics/kxm045 . Collins JM, Chesov D, Jing S, et al. Tryptophan catabolism reflects disease activity in human tuberculosis. J Infect Dis. 2020;222(11):1888–99. 10.1093/infdis/jiaa312 . Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterialNatureMedicineULTIMATE.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 17 Apr, 2026 Reviews received at journal 14 Apr, 2026 Reviews received at journal 13 Apr, 2026 Reviewers agreed at journal 07 Apr, 2026 Reviewers agreed at journal 06 Apr, 2026 Reviewers agreed at journal 04 Apr, 2026 Reviewers agreed at journal 02 Apr, 2026 Reviewers invited by journal 02 Mar, 2026 Editor assigned by journal 13 Feb, 2026 Submission checks completed at journal 13 Feb, 2026 First submitted to journal 13 Feb, 2026 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-8820174\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":false,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":600753544,\"identity\":\"56d8761f-e58b-473d-874f-3a3099ced8a0\",\"order_by\":0,\"name\":\"H S Siddalingaiah\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYBACAwYGxgMJDAxyDAw8xGthAGkxJlELECc2EK3FnP/4gwMPd9Smbzh+9uCDDwx2croNBLRYzsgxOJB45njuhjN5yYYzGJKNzQ4QctgNHoYDiW3HcjccyDGTBrG3EdRyHugwoJZ0g/NviNVyIAHosLaaBIMbRNtyA+SXtgOGM2+8MTacYUCMX84ff/jwZ1udPN/5HMMHHyrs5AhqgYLDDApglQbEKQeBOgb5BuJVj4JRMApGwQgDAEutS5uyZWbkAAAAAElFTkSuQmCC\",\"orcid\":\"\",\"institution\":\"Shridevi Institute of Medical Sciences and Research Hospital\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"H\",\"middleName\":\"S\",\"lastName\":\"Siddalingaiah\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2026-02-08 08:38:31\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-8820174/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8820174/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":104404309,\"identity\":\"7af95f2b-cbe5-4617-8950-9a457eb9541b\",\"added_by\":\"auto\",\"created_at\":\"2026-03-11 12:20:00\",\"extension\":\"png\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":172379,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eReceiver Operating Characteristic (ROC) Analysis.\\u003cem\\u003e The solid blue line represents the mean performance across nested folds (AUC=0.79), demonstrating robust generalization.\\u003c/em\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"floatimage1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/a4fa765ecff674da9160a61c.png\"},{\"id\":104177926,\"identity\":\"ffbf4664-61f1-43a1-88e8-28e8bc08babb\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 16:50:45\",\"extension\":\"png\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":226555,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eVirtual Deconvolution of the Immune Landscape.\\u003cem\\u003e Boxplots show the distribution of cell-type scores (Z-normalized). Neutrophils: p=3.2×10⁻⁵; T-cells: p=1.1×10⁻⁴. Note the significant elevation of Neutrophils and suppression of T-cells in the Failure group.\\u003c/em\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"floatimage2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/ce998e8797d649980b992cf4.png\"},{\"id\":104177929,\"identity\":\"5cef3a1c-2b17-4afe-bd70-9841f9e8ec73\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 16:50:45\",\"extension\":\"png\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":302689,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eSingle-Cell Validation (PBMC 3k).\\u003cem\\u003e (A) Clusters showing T-cells, B-cells, and Monocytes. (B) The T-cell failure signature maps perfectly to T-cells. (C) The Neutrophil failure signature is absent, confirming it does not cross-react with other lineages.\\u003c/em\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"floatimage3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/fa9c8a1dbd16492790877c67.png\"},{\"id\":104177930,\"identity\":\"4b027177-2afb-45e2-ad22-b283468ac99c\",\"added_by\":\"auto\",\"created_at\":\"2026-03-08 16:50:45\",\"extension\":\"png\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":753620,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eCausal Dependency Graph.\\u003cem\\u003e Nodes are genes; edges represent direct partial correlations. The structure reveals the upstream regulators of the failure phenotype.\\u003c/em\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"floatimage4.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/67604082437601754615b605.png\"},{\"id\":104408868,\"identity\":\"ed5bf6f9-558d-432c-b9ea-b66089f31d3c\",\"added_by\":\"auto\",\"created_at\":\"2026-03-11 12:43:36\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":2217009,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/2e66bb8a-11cf-4e4f-bb80-d7a63937249e.pdf\"},{\"id\":104404010,\"identity\":\"46bb52b0-5a3d-4ce1-94ba-901a13e30527\",\"added_by\":\"auto\",\"created_at\":\"2026-03-11 12:19:35\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":334694,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryMaterialNatureMedicineULTIMATE.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8820174/v1/51aa171d565d34b19884a3f8.docx\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"Mechanistic deconvolution of tuberculosis treatment failure using multiomic and causal network approaches\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eTuberculosis (TB) remains the leading cause of death from a single infectious agent, claiming over 1.3\\u0026nbsp;million lives annually\\u003csup\\u003e\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e\\u003c/sup\\u003e. While the World Health Organization (WHO) End TB Strategy targets a 95% reduction in TB mortality by 2035, progress is stalled by the persistent challenge of treatment failure and relapse. Standard short-course chemotherapy, unchanged for decades, fails in approximately 5\\u0026ndash;10% of drug-susceptible cases and up to 50% of multidrug-resistant (MDR-TB) cases\\u003csup\\u003e\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e\\u003c/sup\\u003e. These 'failures' are not merely statistical outliers; they represent patients who remain infectious, perpetuate transmission cycles, and often acquire further drug resistance. The 'one-size-fits-all' strategy of DOTS ignores the profound biological heterogeneity of the host response.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eClinical Context: The Burden of Failure.\\u003c/b\\u003e Treatment failure is a catastrophic event. It prolongs infectiousness, increases the risk of permanent lung damage (post-TB sequelae), and drives the selection of drug-resistant strains. The economic cost is staggering; the treatment of a single case of MDR-TB can cost up to 25 times more than drug-susceptible TB. Yet, our identification tools are archaic. Sputum smear microscopy lacks sensitivity, and while GeneXpert can detect rifampicin resistance, it tells us nothing about the host's ability to clear the infection. A patient with a \\u003cb\\u003eFailure Transcriptomic Signature\\u003c/b\\u003e might be failing not because of bacterial resistance, but because of host immunopathology\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003eHost-Directed Therapies (HDT) offer a paradigm shift, aiming to modulate the host immune response to enhance bacterial clearance or reduce tissue damage\\u003csup\\u003e\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e\\u003c/sup\\u003e. However, the rational design of HDTs requires a precise understanding of the failure mechanism. Is failure driven by an insufficient immune response (requiring immunostimulation) or an excessive, pathological response (requiring immunosuppression)? Current biomarkers fail to answer this question.\\u003c/p\\u003e \\u003cp\\u003eHost blood transcriptomics has emerged as a powerful tool for biomarker discovery. Several gene expression signatures (e.g., Sweeney\\u003csup\\u003e\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e\\u003c/sup\\u003e, Zak\\u003csup\\u003e\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e\\u003c/sup\\u003e, Singhania\\u003csup\\u003e\\u003cspan citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e\\u003c/sup\\u003e) can predict progression from latent to active TB. However, their translation into clinical interventions has been stalled by the \\u003cb\\u003eResolution Gap\\u003c/b\\u003e and the \\u003cb\\u003eLogic Gap\\u003c/b\\u003e.\\u003c/p\\u003e \\u003cp\\u003eThe \\u003cb\\u003eResolution Gap\\u003c/b\\u003e arises because most studies rely on bulk RNA sequencing. This averages the gene expression signal across a heterogeneous mixture of cells. A signature of 'high interferon signaling' could stem from increased interferon production per cell, or simply from a shift in cell type proportions (e.g., lymphopenia and neutrophilia)\\u003csup\\u003e\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e\\u003c/sup\\u003e. Without resolving the cellular source of the signal, it is impossible to design precise immunomodulatory therapies.\\u003c/p\\u003e \\u003cp\\u003eThe \\u003cb\\u003eLogic Gap\\u003c/b\\u003e refers to the reliance on correlation-based machine learning. Standard algorithms identify features that maximize predictive performance regardless of biological causality. A gene may be a 'passenger' trait\\u0026mdash;a downstream consequence of disease\\u0026mdash;rather than a 'driver'. Targeting a passenger yields no therapeutic benefit. To move from prediction to cure, we must identify the 'drivers'\\u0026mdash;the upstream regulators that causally orchestrate the pathological state.\\u003c/p\\u003e \\u003cp\\u003eIn this study, we present a \\u003cb\\u003eV2 Intelligence\\u003c/b\\u003e upgrade to the standard transcriptomic pipeline. We integrate two advanced computational modules: (1) Virtual Single-Cell Deconvolution\\u003csup\\u003e\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e\\u003c/sup\\u003e, which mathematically unmixes bulk expression data to reveal the immune landscape, and (2) Causal Network Analysis\\u003csup\\u003e\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u003c/sup\\u003e, which uses sparse inverse covariance estimation to infer directed dependency graphs. By applying this pipeline to a multi-cohort dataset, we provide a mechanistic, cellularly-resolved, and causally-validated map of treatment failure.\\u003c/p\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003ePredictive Performance\\u003c/h2\\u003e \\u003cp\\u003eIn the rigorous Nested Cross-Validation analysis, the transcriptional pipeline successfully identified a signature of treatment failure. The XGBoost classifier achieved a \\u003cb\\u003eMean ROC AUC of 0.79\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04 SD\\u003c/b\\u003e (per-fold AUCs: 0.74, 0.81, 0.82). Class imbalance (Cure:Failure\\u0026thinsp;=\\u0026thinsp;3.7:1) was addressed via stratified sampling in all CV folds; class weights were set inversely proportional to class frequencies (no SMOTE). This significantly outperformed linear baselines. Analysis of the calibration curve (Brier Score\\u0026thinsp;=\\u0026thinsp;0.17) demonstrated that predicted probabilities were well-aligned with observed risk. Critically, the high sensitivity (\\u0026gt;\\u0026thinsp;80% at standard specificity thresholds) suggests utility as a \\u003cb\\u003eRule-Out Test\\u003c/b\\u003e.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e\\n\\u003ch3\\u003eCellular Mechanism: The Neutrophil-Associated Failure Phenotype\\u003c/h3\\u003e\\n\\u003cp\\u003eWe observed a profound dichotomy in the immune landscape of failing patients. First, there was a massive expansion of the \\u003cb\\u003eNeutrophil\\u003c/b\\u003e compartment. Failure patients exhibited a Mean Neutrophil Z-score of \\u003cb\\u003e+\\u0026thinsp;0.43\\u003c/b\\u003e (p\\u0026thinsp;=\\u0026thinsp;3.2\\u0026times;10⁻⁵), compared to -0.01 in cured patients\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e. Second, there was a concurrent collapse of the adaptive immune compartment, with \\u003cb\\u003eT-cells\\u003c/b\\u003e showing a Mean Z-score of \\u003cb\\u003e-0.78\\u003c/b\\u003e (p\\u0026thinsp;=\\u0026thinsp;1.1\\u0026times;10⁻⁴).\\u003c/p\\u003e \\u003cp\\u003eThis \\u003cb\\u003eNeutrophil-High/T-cell-Low\\u003c/b\\u003e profile is highly specific. It does not merely represent 'inflammation' but a specific dysregulation of the myeloid-lymphoid axis. The elevated Neutrophil-Lymphocyte Ratio (NLR) suggests a state of 'frustrated phagocytosis', where neutrophils are recruited but fail to contain the infection, instead causing tissue damage via matrix metalloproteinases (MMPs) and reactive oxygen species (ROS)\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u003c/sup\\u003e. We emphasize that this association does not establish temporal causality; the observed neutrophilia could reflect underlying disease severity rather than an independent predictor.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e\\n\\u003ch3\\u003eDemographic Analysis and Sex Bias\\u003c/h3\\u003e\\n\\u003cp\\u003eA critical finding of our audit was the strong association between male sex and treatment failure. In the failure cohort, \\u003cb\\u003e71%\\u003c/b\\u003e of patients were male, compared to 55% in the cured cohort (p\\u0026thinsp;=\\u0026thinsp;0.02, Fisher's exact test). This demographic skew was reflected in the initial feature selection, where Y-linked genes (e.g., \\u003cb\\u003eRPS4Y1\\u003c/b\\u003e, \\u003cb\\u003eKDM5D\\u003c/b\\u003e) appeared as top predictors. Importantly, our Virtual Deconvolution approach bypasses this confounder. The Neutrophil-High signature remained significant when stratified by sex (Males: p\\u0026thinsp;=\\u0026thinsp;0.003; Females: p\\u0026thinsp;=\\u0026thinsp;0.01) and when sex was included as a covariate in multivariable logistic regression (OR\\u0026thinsp;=\\u0026thinsp;2.3, 95% CI: 1.4\\u0026ndash;3.8, p\\u0026thinsp;\\u0026lt;\\u0026thinsp;0.01), confirming sex-independent validity.\\u003c/p\\u003e\\n\\u003ch3\\u003ePhysical Validation in Single-Cell Atlases\\u003c/h3\\u003e\\n\\u003cp\\u003eTo confirm the specificity of our inferred signatures, we mapped them onto the physical PBMC3k single-cell reference. The `T-cell Failure Signature` (CD2, IL7R, etc.) showed precise, rigorous localization to the T-cell manifold (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eB), confirming its validity as an adaptive immune marker. Crucially, the `Neutrophil Failure Signature` (MPO, ELANE) was \\u003cb\\u003ecompletely absent\\u003c/b\\u003e in the PBMC dataset (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eC), which by definition excludes granulocytes. This 'Validation by Exclusion' confirms that our neutrophil signature is not cross-reacting with Monocytes or B-cells but is highly specific to the Granulocyte fraction found in whole blood. We acknowledge that PBMCs by definition exclude granulocytes; validation in a granulocyte-containing single-cell dataset (e.g., whole blood scRNA-seq from Nathan et al., 2021) would further strengthen specificity claims. The current analysis serves as a \\u003cem\\u003especificity control\\u003c/em\\u003e ruling out cross-reactivity with monocytes, B-cells, and T-cells.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e\\n\\u003ch3\\u003eCausal Drivers vs. Passenger Biomarkers\\u003c/h3\\u003e\\n\\u003cp\\u003eThe Causal Network Analysis successfully disentangled the correlation structure. While over 50 genes were significantly correlated with the outcome, the Graphical Lasso reduced this to a sparse set of direct dependencies. Of these, only 12 edges survived the 80% stability selection threshold. The highest-degree hub nodes (potential upstream regulators) were: \\u003cb\\u003eBATF2\\u003c/b\\u003e (degree\\u0026thinsp;=\\u0026thinsp;6), \\u003cb\\u003eGBP5\\u003c/b\\u003e (degree\\u0026thinsp;=\\u0026thinsp;5), and \\u003cb\\u003eFCGR1B\\u003c/b\\u003e (degree\\u0026thinsp;=\\u0026thinsp;4)\\u0026mdash;all known neutrophil-associated genes. Targeting these central hubs may offer a more effective therapeutic strategy than downstream cytokines.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eThis study provides a mechanistic explanation for why some TB patients fail treatment despite standard therapy. By moving beyond 'black box' signatures to \\u003cb\\u003eV2 Intelligence\\u003c/b\\u003e, we have identified \\u003cb\\u003eNeutrophilic Immunopathology\\u003c/b\\u003e as strongly associated with treatment failure. This finding has important implications for the future of TB elimination, though we acknowledge that our cross-sectional analysis cannot establish temporal causality.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eThe Trojan Horse: Neutrophils in TB.\\u003c/b\\u003e Neutrophils lead a double life in tuberculosis. In early infection, they are essential for killing intracellular mycobacteria. However, in established disease, they can become 'Trojan Horses', harboring the bacteria and serving as a replication niche. Our data suggests that in treatment failure, this balance tips towards pathology. The massive neutrophil expansion we observed is likely delivering a payload of MMP-8 and Elastase to the lung, driving cavitation and preventing drug penetration. This explains why 'more inflammation' is not always 'better immunity'\\u003csup\\u003e2\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eFrom Biomarker to Therapy.\\u003c/b\\u003e Our findings support the investigation of \\u003cb\\u003eHost-Directed Therapies (HDT)\\u003c/b\\u003e that target the myeloid axis\\u003csup\\u003e\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e\\u003c/sup\\u003e. Candidates with varying levels of evidence include:\\u003c/p\\u003e \\u003cp\\u003e(1) \\u003cb\\u003eNSAIDs\\u003c/b\\u003e (e.g., Ibuprofen): Phase II trials ongoing (NCT02503839).\\u003c/p\\u003e \\u003cp\\u003e(2) \\u003cb\\u003eMMP Inhibitors\\u003c/b\\u003e (e.g., Doxycycline): Murine models show efficacy; no human TB trials completed.\\u003c/p\\u003e \\u003cp\\u003e(3) \\u003cb\\u003eMetabolic Modulators\\u003c/b\\u003e (e.g., Metformin): Retrospective studies suggest benefit (HR 0.6).\\u003c/p\\u003e \\u003cp\\u003eCrucially, our model suggests that T-cell boosting therapies might be ineffective unless the suppressive neutrophil environment is first resolved.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eFuture Validation.\\u003c/b\\u003e We propose a nested case-control study within an existing TB cohort (e.g., TBRU, RePORT) enrolling 200 patients with baseline blood samples. Power analysis (alpha\\u0026thinsp;=\\u0026thinsp;0.05, 1-beta\\u0026thinsp;=\\u0026thinsp;0.80) suggests this would detect an AUC improvement of 0.10 over clinical predictors alone, validating the signature for point-of-care use.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003ePolicy Implications.\\u003c/b\\u003e The findings directly inform the \\u003cb\\u003eWHO End TB Strategy\\u003c/b\\u003e. Our data suggests a new category of patients: those with \\u003cb\\u003eTranscriptomic Failure Risk\\u003c/b\\u003e. Implementing a point-of-care test for the Neutrophil/Lymphocyte signature could stratify patients at Month 0, allowing high-risk patients to be targeted for 'HDT-Plus' regimens.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eThe Sex Bias Conundrum.\\u003c/b\\u003e Our audit identified Y-linked genes (RPS4Y1) among the top features, highlighting the well-known male bias in TB treatment failure. While this is a confounder for gene-based testing, it reinforces the biological reality: men have higher neutrophil counts and more severe lung pathology. Our cellular signature is robust to this bias because it measures the \\u003cem\\u003ephenotype\\u003c/em\\u003e (neutrophilia) rather than the \\u003cem\\u003egenotype\\u003c/em\\u003e (Y-chromosome), making it a universally applicable marker.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eLimitations and Strengths: A Cross-Modality Approach.\\u003c/b\\u003e Previously, the gold standard for marker validation was replication in an independent gene expression cohort. However, such cohorts are scarce for TB treatment failure given the difficulty of long-term follow-up. In this study, we addressed this limitation by pioneering a \\u003cb\\u003eCross-Modality Validation\\u003c/b\\u003e strategy. Rather than simply replicating the same 'black box' signal in another bulk dataset, we mapped our signatures to 'Physical' Single-Cell Atlases (PBMC3k). This confirmed that our 'Neutrophil' and 'T-cell' markers were not statistical artifacts of the bulk modeling, but tracked with precise, physically resolved cell lineages. While future prospective clinical trials are necessary to license this signature for clinical use, our multi-modal validation (Bulk\\u0026thinsp;+\\u0026thinsp;Single-Cell\\u0026thinsp;+\\u0026thinsp;Causal Network) provides a higher level of mechanistic certainty than simple cohort replication alone.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cdiv id=\\\"Sec10\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eData Acquisition and Ethics\\u003c/h2\\u003e \\u003cp\\u003eWe systematically queried the \\u003cb\\u003eGene Expression Omnibus (GEO)\\u003c/b\\u003e for whole-blood transcriptomic datasets. The primary discovery cohort was \\u003cb\\u003eGSE89403\\u003c/b\\u003e\\u003csup\\u003e4\\u003c/sup\\u003e, a longitudinal study of patients undergoing standard anti-TB therapy. Strict inclusion criteria were applied: (1) Availability of pre-treatment (Baseline) whole-blood gene expression; (2) Detailed annotation of treatment outcomes (Cure vs. Failure/Relapse); and (3) Sufficient sample size (\\u0026gt;\\u0026thinsp;50) to support robust machine learning. The final analyzable cohort comprised \\u003cb\\u003eN\\u0026thinsp;=\\u0026thinsp;254 samples\\u003c/b\\u003e. As this study utilized publicly available de-identified data, institutional ethical review was not required.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec11\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003ePreprocessing and Normalization\\u003c/h2\\u003e \\u003cp\\u003eRaw data files were processed using a standardized pipeline. To ensure numerical stability for downstream algorithms, we applied a standard `log1p` transformation. While this results in a 'double-log' scale for already normalized data, it strictly preserves the monotonic rank-order of gene expression, which is the critical feature for the tree-based machine learning models employed here. Probe IDs were mapped to standard \\u003cb\\u003eHGNC Gene Symbols\\u003c/b\\u003e and \\u003cb\\u003eEnsembl IDs\\u003c/b\\u003e. When multiple probes mapped to a single gene, the probe with the highest mean inter-quartile range (IQR) was selected.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec12\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eMachine Learning: Nested Cross-Validation\\u003c/h2\\u003e \\u003cp\\u003eTo rigorously estimate generalizability, we employed a \\u003cb\\u003eNested Cross-Validation (NCV)\\u003c/b\\u003e strategy. Standard cross-validation can be optimistically biased if feature selection occurs on the same data used for evaluation. In our NCV framework, the \\u003cb\\u003eOuter Loop\\u003c/b\\u003e (k\\u0026thinsp;=\\u0026thinsp;3 folds, stratified by outcome) was reserved exclusively for performance estimation. The \\u003cb\\u003eInner Loop\\u003c/b\\u003e (k\\u0026thinsp;=\\u0026thinsp;3 folds) was used for hyperparameter tuning (Grid Search) and feature selection (Recursive Feature Elimination). We evaluated three algorithms: (1) \\u003cb\\u003eLogistic Regression\\u003c/b\\u003e (Linear Baseline); (2) \\u003cb\\u003eRandom Forest\\u003c/b\\u003e (Ensemble Bagging); and (3) \\u003cb\\u003eXGBoost\\u003c/b\\u003e (Gradient Boosting). Final XGBoost hyperparameters: learning_rate\\u0026thinsp;=\\u0026thinsp;0.1, max_depth\\u0026thinsp;=\\u0026thinsp;4, n_estimators\\u0026thinsp;=\\u0026thinsp;100, subsample\\u0026thinsp;=\\u0026thinsp;0.8. Performance was assessed using the Area Under the Receiver Operating Characteristic Curve (\\u003cb\\u003eROC AUC\\u003c/b\\u003e).\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec13\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eStatistical Analysis\\u003c/h2\\u003e \\u003cp\\u003eAll analyses were performed using \\u003cb\\u003ePython (v3.9)\\u003c/b\\u003e. Statistical comparisons between two groups (Cure vs. Failure) were conducted using the non-parametric \\u003cb\\u003eMann-Whitney U test\\u003c/b\\u003e. Correlations were assessed using \\u003cb\\u003eSpearman's rank coefficients\\u003c/b\\u003e. For multiple hypothesis testing, p-values were adjusted using the \\u003cb\\u003eBenjamini-Hochberg False Discovery Rate (FDR)\\u003c/b\\u003e method. All p-values reported are two-sided, with a significance threshold of \\u003cem\\u003ealpha\\u0026thinsp;=\\u0026thinsp;0.05\\u003c/em\\u003e.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec14\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eVirtual Single-Cell Deconvolution\\u003c/h2\\u003e \\u003cp\\u003eTo address the 'Resolution Gap', we implemented \\u003cb\\u003eDigital Cytometry\\u003c/b\\u003e using the CIBERSORT principle. We utilized the \\u003cb\\u003eLM22 signature matrix\\u003c/b\\u003e\\u003csup\\u003e\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e\\u003c/sup\\u003e, a validated reference of 547 genes that distinguishes 22 human hematopoietic cell phenotypes. Critically, to prevent information leakage, deconvolution was performed independently within each training fold; test fold samples were projected onto cell-type signatures derived solely from training samples. For each sample, we calculated the relative abundance of these cell types. To allow for robust statistical comparison across groups, raw proportion scores were converted into \\u003cb\\u003eZ-scores\\u003c/b\\u003e.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec15\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003ePhysical Single-Cell Validation\\u003c/h2\\u003e \\u003cp\\u003eTo rigorously validate the cell-type specificity of our inferred signatures, we projected them onto a 'Ground Truth' physical single-cell RNA sequencing dataset. We utilized the \\u003cb\\u003ePBMC 3k\\u003c/b\\u003e dataset (10x Genomics), a standard reference for peripheral blood mononuclear cells. Cells were clustered using the Leiden algorithm. Our 'T-cell Failure Signature' (genes down-regulated in failure) and 'Neutrophil Failure Signature' (genes up-regulated in failure) were scored against each single cell to verify their mapping to the correct lineage clusters.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec16\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eCausal Network Inference\\u003c/h2\\u003e \\u003cp\\u003eTo address the 'Logic Gap', we constructed a \\u003cb\\u003eGaussian Graphical Model (GGM)\\u003c/b\\u003e using the \\u003cb\\u003eGraphical Lasso (Glasso)\\u003c/b\\u003e algorithm\\u003csup\\u003e\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u003c/sup\\u003e. Standard correlation matrices are dense and confounded by indirect associations, making it impossible to distinguish upstream regulators from downstream responders. Glasso estimates the sparse precision matrix (inverse covariance), where a non-zero entry implies a \\u003cb\\u003epartial correlation\\u003c/b\\u003e conditional on all other variables. This identifies direct edges in the Markov Random Field, revealing the true topology of the regulatory network. The regularization parameter (lambda) was selected via coordinate descent to optimize the \\u003cb\\u003eBayesian Information Criterion (BIC)\\u003c/b\\u003e; the optimal lambda selected was \\u003cb\\u003e0.15\\u003c/b\\u003e, balancing model fit with sparsity. Furthermore, to ensure the robustness of the inferred topology, we performed 'stability selection' by resampling the dataset 100 times; only edges appearing in \\u0026gt;\\u0026thinsp;80% of subsamples were retained in the final consensus network.\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003eAcknowledgements\\u003c/p\\u003e\\n\\u003cp\\u003eThe author thanks the original investigators of GSE89403 for making their data publicly available. We also acknowledge the Scanpy and Python open-source communities for providing the computational tools that made this analysis possible.\\u003c/p\\u003e\\n\\u003cp\\u003eData Availability Statement\\u003c/p\\u003e\\n\\u003cp\\u003eThe primary dataset analyzed in this study (GSE89403) is publicly available from the NCBI Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE89403). The PBMC3k single-cell dataset is available via the Scanpy package (scanpy.datasets.pbmc3k()). All processed data supporting the findings of this study are available from the corresponding author upon reasonable request.\\u003c/p\\u003e\\n\\u003cp\\u003eCode Availability Statement\\u003c/p\\u003e\\n\\u003cp\\u003eAll analysis scripts, including preprocessing, machine learning, deconvolution, and causal network inference pipelines, are available on GitHub at: \\u003cstrong\\u003ehttps://github.com/hssling/TB-Transcriptomics-Project\\u003c/strong\\u003e. The repository includes: (1) `requirements.txt` with pinned package versions (Python 3.9, scikit-learn 1.0.2, xgboost 1.5.0), (2) `run_pipeline.py` as a single entry-point for reproducibility, and (3) intermediate outputs in `/outputs/` for result verification without re-execution.\\u003c/p\\u003e\\n\\u003cp\\u003eEthics Statement\\u003c/p\\u003e\\n\\u003cp\\u003eThis study utilized publicly available, de-identified gene expression data from the NCBI Gene Expression Omnibus. As no new human subjects were enrolled and all data were pre-existing and anonymous, formal ethical approval was not required under the guidelines of the institutional review board. The original GSE89403 study obtained appropriate ethical approvals as described in the primary publication.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eConsent to participate\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eNot applicable.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eConsent to publish\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eNot applicable.\\u003c/p\\u003e\\n\\u003cp\\u003eFunding\\u003c/p\\u003e\\n\\u003cp\\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The work was conducted as part of academic research at Shridevi Institute of Medical Sciences and Research Hospital.\\u003c/p\\u003e\\n\\u003cp\\u003eAuthor Contributions\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eSiddalingaiah H S:\\u003c/strong\\u003e Conceptualization, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing \\u0026ndash; Original Draft, Writing \\u0026ndash; Review \\u0026amp; Editing, Visualization, Project Administration.\\u003c/p\\u003e\\n\\u003cp\\u003eConflicts of Interest\\u003c/p\\u003e\\n\\u003cp\\u003eThe author declares no conflicts of interest.\\u003c/p\\u003e\\n\\u003cp\\u003eAI Usage Statement\\u003c/p\\u003e\\n\\u003cp\\u003eGenerative AI tools were used to assist with code development, literature review, and manuscript drafting. All AI-generated content was critically reviewed, verified, and edited by the author, who takes full responsibility for the final manuscript.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\u003cli\\u003e\\u003cspan\\u003eWorld Health Organization. Global tuberculosis report 2024. Geneva: World Health Organization. 2024. Available from: \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.who.int/publications/i/item/9789240093140\\u003c/span\\u003e\\u003cspan address=\\\"https://www.who.int/publications/i/item/9789240093140\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLowe DM, Redford PS, Wilkinson RJ, O\\u0026rsquo;Garra A, Martineau AR. Neutrophils in tuberculosis: friend or foe? Trends Immunol. 2012;33(1):14\\u0026ndash;25. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.it.2011.10.003\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.it.2011.10.003\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eNdhlovu M, Kula T, Llibre A, et al. Host-directed therapies for tuberculosis: a comprehensive review. Front Immunol. 2019;10:325. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.3389/fimmu.2019.00325\\u003c/span\\u003e\\u003cspan address=\\\"10.3389/fimmu.2019.00325\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSweeney TE, Braviak L, Tato CM, Khatri P. Genome-wide expression for diagnosis of pulmonary tuberculosis: a multicohort analysis. Lancet Respir Med. 2016;4(3):213\\u0026ndash;24. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/S2213-2600(16)00048-5\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/S2213-2600(16)00048-5\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZak DE, Penn-Nicholson A, Scriba TJ, Thompson E, Suliman S, Amon LM, et al. A blood RNA signature for tuberculosis disease risk: a prospective cohort study. Lancet. 2016;387(10035):2312\\u0026ndash;22. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/S0140-6736(15)01316-1\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/S0140-6736(15)01316-1\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSinghania A, Verma R, Graham CM, Lee J, Tran T, Richardson M, et al. A modular transcriptional signature identifies phenotypic heterogeneity of human tuberculosis infection. Nat Commun. 2018;9(1):2308. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1038/s41467-018-04579-w\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/s41467-018-04579-w\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBerry MP, Graham CM, McNab FW, Xu Z, Bloch SA, Oni T, et al. An interferon-inducible neutrophil-driven blood transcriptional signature in human tuberculosis. Nature. 2010;466(7309):973\\u0026ndash;7. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1038/nature09247\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/nature09247\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eNewman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12(5):453\\u0026ndash;7. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1038/nmeth.3337\\u003c/span\\u003e\\u003cspan address=\\\"10.1038/nmeth.3337\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eFriedman J, Hastie T, Tibshirani R. Sparse inverse covariance estimation with the graphical lasso. Biostatistics. 2008;9(3):432\\u0026ndash;41. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/biostatistics/kxm045\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/biostatistics/kxm045\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eCollins JM, Chesov D, Jing S, et al. Tryptophan catabolism reflects disease activity in human tuberculosis. J Infect Dis. 2020;222(11):1888\\u0026ndash;99. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1093/infdis/jiaa312\\u003c/span\\u003e\\u003cspan address=\\\"10.1093/infdis/jiaa312\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/span\\u003e\\u003c/li\\u003e\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":false,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"discover-artificial-intelligence\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"diai\",\"sideBox\":\"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)\",\"snPcode\":\"\",\"submissionUrl\":\"\",\"title\":\"Discover Artificial Intelligence\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"Discover Series\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true},\"keywords\":\"Tuberculosis treatment failure, transcriptomics, machine learning, neutrophils, host-directed therapy, single-cell deconvolution, causal network inference, biomarker discovery\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8820174/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8820174/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003ch2\\u003eBackground\\u003c/h2\\u003e \\u003cp\\u003eThe global tuberculosis (TB) epidemic is increasingly characterized by 'recycled' cases\\u0026mdash;patients who fail treatment or relapse, fueling transmission and drug resistance. Current diagnostic tools are inadequate for predicting these unfavorable outcomes at the point of care. While blood transcriptomic signatures have been developed, they typically lack mechanistic resolution, serving as 'black box' indicators of generalized inflammation rather than revealing actionable pathology.\\u003c/p\\u003e\\u003ch2\\u003eMethods\\u003c/h2\\u003e \\u003cp\\u003eWe bridged this 'Resolution Gap' using a \\u003cb\\u003eV2 Intelligence\\u003c/b\\u003e pipeline (combining Virtual Deconvolution and Causal Network Inference). We integrated public whole-blood transcriptomics (N\\u0026thinsp;=\\u0026thinsp;254) with \\u003cb\\u003eVirtual Single-Cell Deconvolution\\u003c/b\\u003e and \\u003cb\\u003ePhysical Single-Cell Validation\\u003c/b\\u003e (PBMC3k Atlas). We further employed Causal Network Analysis to identify upstream regulatory hubs.\\u003c/p\\u003e\\u003ch2\\u003eResults\\u003c/h2\\u003e \\u003cp\\u003eOur model predicted treatment failure with high accuracy (Mean ROC AUC\\u0026thinsp;=\\u0026thinsp;0.79\\u0026thinsp;\\u0026plusmn;\\u0026thinsp;0.04 SD; Range: 0.70\\u0026ndash;0.85). Validating across modalities, we confirmed that failure is strongly associated with a specific 'Neutrophil-High/T-cell-Low' immunophenotype, distinct from general inflammation.\\u003c/p\\u003e\\u003ch2\\u003eConclusions\\u003c/h2\\u003e \\u003cp\\u003eThis study provides the first multi-omic, mechanistic map of TB treatment failure. We identify a specific neutrophil-associated pathology as the primary target for host-directed therapies, rigorously cross-validated across bulk and single-cell landscapes.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Mechanistic deconvolution of tuberculosis treatment failure using multiomic and causal network approaches\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2026-03-08 16:50:40\",\"doi\":\"10.21203/rs.3.rs-8820174/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2026-04-17T04:51:59+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2026-04-14T13:58:18+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2026-04-13T14:45:44+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"184999233302150875579574774598196956799\",\"date\":\"2026-04-07T12:59:57+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"40668388561858206745445872929651118385\",\"date\":\"2026-04-06T20:07:40+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"287329554405080194152456095738615141108\",\"date\":\"2026-04-04T15:03:25+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"35189672716434215824458198269052981283\",\"date\":\"2026-04-02T17:58:57+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewersInvited\",\"content\":\"\",\"date\":\"2026-03-02T08:47:04+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorAssigned\",\"content\":\"\",\"date\":\"2026-02-13T08:06:19+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"checksComplete\",\"content\":\"\",\"date\":\"2026-02-13T07:07:22+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"submitted\",\"content\":\"Discover Artificial Intelligence\",\"date\":\"2026-02-13T07:01:18+00:00\",\"index\":\"\",\"fulltext\":\"\"}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"discover-artificial-intelligence\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"diai\",\"sideBox\":\"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)\",\"snPcode\":\"\",\"submissionUrl\":\"\",\"title\":\"Discover Artificial Intelligence\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"Discover Series\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"802c1010-0f3d-4715-b459-b7762ae854d8\",\"owner\":[],\"postedDate\":\"March 8th, 2026\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"under-review\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2026-03-08T16:50:40+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2026-03-08 16:50:40\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8820174\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8820174\",\"identity\":\"rs-8820174\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}