Catalyst initiative testing whether gut microbiome composition mediates stromal tissue proportion in CD, explaining why biopsy microbiome (AUC=0.90) outperforms transcriptomics (0.71) for treatment prediction. Uses HMP2 paired microbiome + host transcriptomics with NNLS deconvolution. Bridges 5 projects: epithelial-reserve (stromal AUC=0.667), ferroptosis (GPX4-fibrosis, A.muciniphila→GPX4), microbiome-subtyping (subtypes), anti-TNF (classifier ceiling), and literature (Zafeiropoulou biopsy microbiome).

Pan-cancer SL dependency atlas for histone modifiers KMT2C (MLL3, H3K4me1) and KDM6A (UTX, H3K27me3). Maps actionable vulnerabilities in tumors with epigenetic modifier loss. Extends portfolio beyond SWI/SNF into histone-modifying enzymes. Focus on bladder, endometrial, breast cancers.

Gene expression-based classifier to predict response to anti-TNF therapy in Crohn's disease using pre-treatment mucosal or blood transcriptomic profiles.

Cross-project investigation testing whether GPX4-regulated ferroptosis bridges epithelial cell death, ILC2 dysfunction, and fibrotic progression in CD. Convergent signal from IL-23 atlas (ILC2 ferroptosis), anti-TNF (GSDMC/NFS1/ACAD9 in non-responder signature), fibrosis literature (GPX4 predicts post-op recurrence), and flare prediction (metabolic stress). Catalyst initiative.

PIVOT: Epithelial reserve hypothesis REFUTED (0/4 criteria). Project reframed around emergent finding: stromal proportion (fibroblast + myofibroblast + endothelial) is the strongest cell-type predictor of anti-TNF response (AUC=0.667, consistent across 3 cohorts). Tissue remodeling state determines biologic efficacy regardless of immune pathway targeted. 9-gene fibrosis/EMT score (AUC=0.642) outperforms 24-feature CPA3/EMT model. Unifying framework linking treatment stratification, fibrosis, and stricture progression.

Identify KRAS allele-specific therapeutic vulnerabilities in colorectal cancer using DepMap 25Q3, comparing G12D, G13D, G12V, G12C, Q61H, and A146T dependencies to inform patient selection for allele-specific (sotorasib, zoldonrasib/MRTX1133) and pan-RAS (ERAS-0015, elironrasib) inhibitors entering CRC clinical trials.

Map synthetic lethal dependencies specific to KEAP1-loss and NRF2-GOF cancers across all DepMap tumor types to identify novel therapeutic targets for oxidative-stress-reprogrammed tumors. KEAP1 mutations affect ~23% of lung adeno, ~30% lung squamous, ~12% head/neck. These patients have poor immunotherapy outcomes (ORR ~28%) and no targeted therapies. No comprehensive pan-cancer KEAP1/NRF2 SL atlas exists.

Multi-feature predictive model for stricturing (B2) and penetrating (B3) disease behavior in Crohn's disease integrating genomic, transcriptomic, and clinical features from the RISK cohort for early risk stratification at diagnosis.

Catalyst initiative: multi-task acoustic model testing whether a single 3-minute smartphone recording (vowels, cough, counting, reading) can simultaneously screen for Parkinson's disease, respiratory conditions (TB/COPD), and mild cognitive impairment. Builds on voice-biomarkers-parkinsons and acoustic-tb-screening shared pipelines. Targets AUROC >= 0.80 for 2+ conditions without degrading single-disease performance.

Computational meta-analysis of cytokine/chemokine profiles across published PANDAS/PANS studies to define the immune signature of disease flares and build a predictive model for flare onset.

Cross-cancer MSI prediction from routine H&E histopathology slides using pathology foundation models (UNI 2, CONCH). Low-cost pre-screening tool to expand MSI testing to settings without molecular pathology. Uses TCGA pan-cancer WSIs and CPTAC for validation. Targets AUROC >=0.85 in 3+ cancer types with cross-cancer generalization.

Map synthetic lethal dependencies specific to NF1-loss cancers across all DepMap tumor types to identify novel therapeutic targets beyond MEK inhibition, with focus on MPNST (no effective targeted therapy, 5-yr survival ~40%), melanoma, glioma, and NSCLC NF1-mutant subtypes. NF1 loss activates RAS via a distinct mechanism from KRAS mutations — dependencies may differ.

Identify actionable therapeutic targets and synthetic lethal pairs in NSCLC by mapping DepMap cancer dependencies to TCGA patient molecular profiles, stratified by oncogenic driver (KP/KL/KOnly) and immune context. Reproduces and extends TCGADEPMAP (Nature Cancer 2024) with DepMap 25Q3, SL integration, PRISM drug mapping, and LUCA scRNA-seq validation.

Catalyst initiative testing the two-hit compound genetic vulnerability model: lectin complement deficiency (MBL2/MASP1/MASP2 → impaired GAS clearance) PLUS cGAS-STING dysregulation (TREX1/SAMHD1 → excessive type I IFN to persistent pathogen DNA) creates episodic infection-triggered neuropsychiatric disease. Based on convergence of two high-confidence PPI modules identified in pans-genetic-variant-pathway-analysis. Borrows from Aicardi-Goutières syndrome interferonopathy framework.

Cost-optimized tri-modal classifier combining cfDNA methylation, protein biomarkers, and cfDNA fragmentomics for early-stage colorectal cancer detection from blood. Fragmentomics (fragment lengths, end motifs, nucleosome footprints) added as third modality after Zeng et al. Nature Cancer showed methylome+fragmentome integration outperforms either alone. Targets stage I-II sensitivity >=70% at 95% specificity with estimated per-test cost under . Uses TCGA-COAD/READ, GEO cfDNA methylation, and circulating protein datasets.

Integrate metagenomic and metabolomic profiles from Crohn's patients to identify reproducible disease subtypes with distinct biological mechanisms and therapeutic implications.

Deep re-analysis of the first single-cell transcriptomic and epigenomic atlas of TS basal ganglia (Wang et al. 2025) to characterize interneuron subtype-specific vulnerability, disrupted regulatory networks, and druggable targets at single-cell resolution.

Develop a computational transcriptomic classifier that distinguishes PANDAS/PANS from primary pediatric OCD, tic disorders, and healthy controls using peripheral blood gene expression signatures.

Test whether pairwise co-loss of tumor suppressor genes from our atlas series creates non-additive dependency interactions. Every published atlas treats its target in isolation, but real tumors carry 4-7 drivers. PTEN atlas flags 21% RB1 co-loss unresolved; breast stratification found 39.4% ARID1A×PIK3CA overlap; TP53 co-mutation exceeds 60% in multiple atlas targets. Tests top 5 co-alteration pairs in DepMap double-mutant lines for synergistic, antagonistic, or redirective interactions. If non-additive: identifies novel drug combinations. If additive: validates atlas framework.

Single-cell transcriptomic atlas of the IL-23/Th17 axis in Crohn's disease mucosa to identify cell-type-specific gene programs predictive of response to IL-23-targeting biologics (ustekinumab, risankizumab).

Profile spatial and cell-type-specific expression of TS risk genes across the cortico-striato-thalamo-cortical circuit using Allen Brain Atlas, BrainSpan, and single-cell RNA-seq to identify circuit nodes and cell types most disrupted in TS.

Long-tail classification system for rare thoracic conditions in chest X-rays. RESCOPED: Using NIH ChestX-ray14 (14 findings, 112K CXRs) after credentialed data access (MIMIC-CXR, MIDRC, CXR-LT 2026) was cancelled. Evaluates LDAM-DRW, decoupled training, class-balanced losses, BiomedCLIP zero-shot, and ConvNeXt-Large baseline on 14-class long-tail split. Cross-institutional experiments dropped. Targets >=10pp tail-class AUROC improvement over baselines. PadChest under investigation as supplementary dataset.

Systematic functional annotation of Tourette syndrome GWAS risk loci using eQTL mapping, chromatin interactions, MAGMA gene-set analysis, and cell-type enrichment to identify convergent biological pathways and prioritize causal genes.

Predict Crohn's disease flares before clinical onset using longitudinal multi-omic trajectory features from HMP2/IBDMDB, enabling preemptive treatment intensification through a validated pre-flare multi-omic signature.