Deep-learning-driven tumor microenvironment profiling improves immunotherapy response prediction
- 1Stanford University
Abstract
Immunotherapy has transformed the management of many cancers, such as clear cell renal cell carcinoma (ccRCC), yet durable clinical benefit is achieved in only a subset of patients. Immunotherapy outcomes are governed by the multicellular tumor microenvironment (TME), through the structured and directional interplay of regulatory programs and ligand-receptor communications. Yet most existing methods for immunotherapy response prediction rely on independent signatures or cell types and fail to model the structured and directional signaling network within the TME. To address this gap, we introduce EcoNet, a computational framework that first models ecosystems consisting of co-occurring cell states, then constructs intercellular and intracellular signaling networks based on the ecosystems, and applies a graph attention network to derive embeddings from the signaling network for immunotherapy response prediction. Pretrained on 2,532 pan-cancer immunotherapy samples across multiple cohorts, EcoNet outperformed existing prediction methods by 7.5-19.9% and marker-based signatures by 4.4-21.2% in AUC, and its high-attention genes were enriched among those required for T-cell-mediated tumor killing in an independent CRISPR screen; applied to ccRCC with disease-specific fine-tuning, it stratified an independent ccRCC cohort by clinical outcome (P=0.024). Biologically, high-attention genes capture an interferon-responsive program that enhances antigen presentation and immune infiltration. To validate these findings, we curated a ccRCC clinical cohort profiled with CosMx single-cell spatial transcriptomics (46 samples, ~450,000 cells) and found the response-associated TME configuration identified by EcoNet localized to immune-enriched niches, supporting the response stratification. These results establish EcoNet as a generalizable framework for TME-informed therapeutic response prediction, readily extensible to other cancer types and therapeutic contexts.
Related
- An automated ranking platform for machine learning regression models for meat spoilage prediction using multi-spectral imaging and metabolic profiling
- Comprehensive interaction profiling and machine learning prediction of bacteriophage infectivity across clinically diverse Pseudomonas aeruginosa
- Targeted tumor desialylation by engineered probiotic reprograms the tumor microenvironment to potentiate immunotherapy
- Targeting Dendritic Cells in Cancer Immunotherapy: From Immune Priming to Tumor Microenvironment Reprogramming
- Ultrasound-responsive nanoplatforms for cancer immunotherapy: engineering spatiotemporal immune activation in the tumor microenvironment
BibTeX
@misc{wang2026deeplearningdriven,
title = {{Deep-learning-driven tumor microenvironment profiling improves immunotherapy response prediction}},
author = {R. H. Wang and B. A. Benard and I. Ilerten and S. Zdravkovic and R. Al-Humadi and S. Nasr and B. Howitt and A. J. Gentles},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.06.757143},
url = {https://www.biorxiv.org/content/10.64898/2026.10.06.757143v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Deep-learning-driven tumor microenvironment profiling improves immunotherapy response prediction
AU - R. H. Wang
AU - B. A. Benard
AU - I. Ilerten
AU - S. Zdravkovic
AU - R. Al-Humadi
AU - S. Nasr
AU - B. Howitt
AU - A. J. Gentles
PY - 2026
DA - 2026/10/06
PB - bioRxiv
DO - 10.64898/2026.10.06.757143
UR - https://www.biorxiv.org/content/10.64898/2026.10.06.757143v1
AB - Immunotherapy has transformed the management of many cancers, such as clear cell renal cell carcinoma (ccRCC), yet durable clinical benefit is achieved in only a subset of patients. Immunotherapy outcomes are governed by the multicellular tumor microenvironment (TME), through the structured and directional interplay of regulatory programs and ligand-receptor communications. Yet most existing methods for immunotherapy response prediction rely on independent signatures or cell types and fail to model the structured and directional signaling network within the TME. To address this gap, we introduce EcoNet, a computational framework that first models ecosystems consisting of co-occurring cell states, then constructs intercellular and intracellular signaling networks based on the ecosystems, and applies a graph attention network to derive embeddings from the signaling network for immunotherapy response prediction. Pretrained on 2,532 pan-cancer immunotherapy samples across multiple cohorts, EcoNet outperformed existing prediction methods by 7.5-19.9% and marker-based signatures by 4.4-21.2% in AUC, and its high-attention genes were enriched among those required for T-cell-mediated tumor killing in an independent CRISPR screen; applied to ccRCC with disease-specific fine-tuning, it stratified an independent ccRCC cohort by clinical outcome (P=0.024). Biologically, high-attention genes capture an interferon-responsive program that enhances antigen presentation and immune infiltration. To validate these findings, we curated a ccRCC clinical cohort profiled with CosMx single-cell spatial transcriptomics (46 samples, ~450,000 cells) and found the response-associated TME configuration identified by EcoNet localized to immune-enriched niches, supporting the response stratification. These results establish EcoNet as a generalizable framework for TME-informed therapeutic response prediction, readily extensible to other cancer types and therapeutic contexts.
ER -