NCI Division of Cancer Biology Junior Investigators Meeting (Abstract): Submission #12

Submission information
Submission Number: 12
Submission ID: 194189
Submission UUID: 3012179f-6abd-4927-b198-0a63cdc1c96b
Submission URI: /dcb/ji-meeting/abstract

Created: Tue, 09/08/2026 - 13:24
Completed: Tue, 09/08/2026 - 13:24
Changed: Tue, 09/08/2026 - 13:24

Remote IP address: 10.208.24.147
Submitted by: Anonymous
Language: English

Is draft: No
First Name Ruohan
Middle Initial
Last Name Wang
Degree(s) Ph.D.
Position/Title/Career Status
Organization Stanford University
Organization Address Stanford
Email ruohwang@stanford.edu
Abstract Category Use my abstract for team formation only (do not consider it for a presentation)
Abstract Keywords Cancer Immunotherapy; Tumor microenvironment; Deep learning
Abstract Title Deep-learning–driven tumor microenvironment profiling improves immunotherapy response prediction
Abstract Immunotherapy benefits only a fraction of cancer patients, and we still lack reliable ways to predict who will respond. A key reason is that response depends not on any single biomarker but on the collective behavior of diverse cell populations within the tumor microenvironment (TME), communicating through structured signaling networks that existing methods do not capture.
We developed EcoNet, a framework that first profiles the TME as co-occurring cell-state ecosystems (ecotypes), then reconstructs the intercellular and intracellular signaling networks connecting them, and feeds these networks into a graph attention model to predict immunotherapy outcome from bulk RNA-seq. Pretrained on 2,532 pan-cancer immunotherapy samples, EcoNet outperformed existing prediction tools by 7.5 to 19.9% and marker-based signatures by 4.4 to 21.2% in AUC. Its high-attention genes were independently validated by a CRISPR screen for T-cell-mediated tumor killing. Applied to clear cell renal cell carcinoma with disease-specific fine-tuning, EcoNet stratified an independent cohort by progression-free survival (P=0.024) and identified an interferon-responsive program linking antigen presentation, immune infiltration, and checkpoint regulation. We further validated these findings using CosMx single-cell spatial transcriptomics on a 46-sample clinical cohort (~450,000 cells), confirming that the therapy-beneficial TME configuration localizes to immune-enriched tissue niches.
EcoNet provides a generalizable, biologically grounded approach to predicting immunotherapy response and is readily extensible to other cancers and therapeutic settings.