Drylab generate model predicting patient response to etanercept drug for rheumatoid arthritis.

Abstract

Original article: Deep molecular profiling of synovial biopsies in the STRAP trial identifies signatures predictive of treatment response to biologic therapies in rheumatoid arthritis (Nature Communication, 2025)

• RNA-seq data from 208 pre-treatment joint tissue samples, collected in the STRAP trial.

• Studied patient response to drug etanercept (TNF inhibitor)

Comparison result:

Performance from Drylab:

• Surpassed baseline model within 12 hours.

• Best model is generated within 24 hours

• Model is biological driven and fully explainable

Solution from Drylab

Drylab AI agent is both creative and robust in finding the right method. The agent succeeds by replacing high-dimensional gene features with compact, biologically informed pathway and cell-type features, and by leveraging XGBoost to model non-linear interactions under rigorous cross-validation.

1. Pathway activity (GSVA)

   • Engineer pathway-PCA specific for RA
     (TNF signaling, NF-κB, JAK–STAT, interferon, T/B cell signaling)

   • Primary scoring via GSVA/ssGSEA; Secondary to z-scored pathway

   • Reduces thousands of genes → dozens of interpretable pathway scores

   • Example of pathway activity range

   • tnf_signaling : -3.76 → 1.31

   • Top contributed pathways: complement, TNF, inflammatory cytokines, matrix remodeling, interferon

     
     

2. Cell-type deconvolution

   • Correlation-based method using compact immune signatures (CD4/CD8 T cells, monocytes, etc.)

   • Captures baseline immune composition differences between responders and non-responders

   
   

3. XGBoost cross-validation

   • Nested CV: outer fold for evaluation, inner xgb.cv for tuning 

   • Early stopping to pick optimal rounds

   • Shallow trees (depth 3–6), moderate regularization to prevent overfitting

   
   

Drylab AI agent performed post-training analysis with literature search. This ensure each solution is scientifically rigorous and hypothesis is verified before continually evolve the solution

Reference

Based on feature importance analysis and literature search:

• TNF → NF-κB → inflammatory cytokines: Etanercept blocks TNF-α, which sits upstream of NF-κB and the pro-inflammatory cascade (IL-1β, IL-6, etc.). Seeing these pathways tied to response is mechanistically on-point. (NCBI)

• B-cell & T-cell programs: In STRAP synovial RNA-seq, etanercept (and rituximab) responders showed higher baseline B-cell genes; adaptive immunity modules commonly stratify response. (Nature)

• JAK–STAT: Many RA cytokines signal via JAK/STAT; baseline activity often tracks inflammatory load and drug response across biologics. (PubMed Central)

• Interferon: Type I/II IFN signatures have repeatedly associated with anti-TNF responsiveness/non-responsiveness. It’s reasonable that your AutoML pulled this in. (Frontiers)

• Matrix remodelling (MMPs/ECM): STRAP reported collagen genes and MMP9 associated with non-response to etanercept/rituximab—your “matrix remodelling” signal fits that observation. (Nature)

• Complement: Complement activation is a well-documented RA driver and plausibly separates synovial phenotypes related to TNFi response. (PubMed Central)

• Apoptosis: TNF/NF-κB regulate survival of synovial cells; TNF blockade can shift cell-death programs in synovium—so apoptosis pathway features are expected. (PubMed Central)

• Oxidative stress: RA synovium and SF are redox-stressed; neutrophils and FLS amplify ROS—seeing oxidative stress as a predictor is common. (PubMed Central)