BioPharmaTrend
Latest News
All Topics
  • Artificial Intelligence
  • NeuroTech
  • Premium Content
  • Knowledge Center
Interviews
Companies
  • Company Directory
  • Sponsored Case Studies
  • Create Company Profile
More
  • About Us
  • Our Team
  • Advisory Board
  • Citations and Press Coverage
  • Partner Events Calendar
  • Advertise with Us
  • Write for Us
Subscribe
Login/Join

  Latest News

Harvard and Roche Develop Foundation AI Model to Predict and Explain Immunotherapy Response Across 33 Cancer Types

by Roman Kasianov   •   May 26, 2025  

Disclaimer: All opinions expressed by Contributors are their own and do not represent those of their employers, or BiopharmaTrend.com.
Contributors are fully responsible for assuring they own any required copyright for any content they submit to BiopharmaTrend.com. This website and its owners shall not be liable for neither information and content submitted for publication by Contributors, nor its accuracy.

# AI & Digital   
Share:   Share in LinkedIn  Share in Bluesky  Share in Reddit  Share in Hacker News  Share in X  Share in Facebook  Send by email

Researchers from Harvard Medical School and Roche have developed COMPASS, a foundation AI model designed to predict patient response to immune checkpoint inhibitors (ICIs) across 33 cancer types and six ICI therapies. The model addresses a long-standing challenge in oncology of identifying which patients will benefit from immunotherapy by learning biological concepts directly from tumor transcriptomes and linking them to treatment outcomes.

#advertisement
AI in Drug Discovery Report 2025

Trained on over 10,000 pre-treatment tumor RNA-seq samples and fine-tuned on 16 clinical cohorts, COMPASS outperformed 22 baseline methods and surpassed commonly used biomarkers such as PD-L1 expression and tumor mutational burden (TMB). In a phase II urothelial cancer trial, it achieved a hazard ratio of 4.7 (p < 0.0001) for survival prediction, compared to 1.7 and 1.5 for TMB and PD-L1, respectively.

Figure 1: Architecture of the COMPASS model for interpretable immunotherapy response prediction. Gene expression data is encoded into biologically grounded tumor-immune (TIME) concepts via a hierarchical projector, enabling classification of responders versus non-responders with flexible fine-tuning options for clinical adaptation.

The model is designed for clinical generalization and interpretability, integrating transfer learning with a concept bottleneck approach that maps gene expression to biologically meaningful latent variables. This enables not only improved prediction across unseen cancer types and therapies, but also insight into why patients respond or fail. For example, COMPASS uncovered distinct resistance mechanisms in immune-inflamed non-responders—including TGF-β signaling, vascular exclusion, CD4+ T cell dysfunction, and B cell deficiency.

COMPASS was evaluated using leave-one-cohort-out and cross-cohort transfer benchmarks, showing high accuracy and robustness even in small and heterogeneous datasets. Its predictions remained consistent across diverse cancer types, biopsy sites, and sequencing pipelines, supporting its potential as a pan-cancer clinical tool for immunotherapy stratification.

The model’s release marks a step toward more clinically aligned, biologically grounded AI systems in oncology, with implications for drug development, patient selection, and precision immunotherapy. An interactive tool for personalized response explanation is also publicly available for researchers.

Topics: AI & Digital   

Share:   Share in LinkedIn  Share in Bluesky  Share in Reddit  Share in Hacker News  Share in X  Share in Facebook  Send by email
#advertisement
ThermoFisher Scientific: Integrated genetic technologies for cell therapy development
#advertisement
Webinar: AI in Clinical Trials

BiopharmaTrend.com

Where Tech Meets Bio
mail  Newsletter
in  LinkedIn
x  X
gnews  Google News
rss  RSS Feed

About


  • What we do
  • Citations and Press Coverage
  • Terms of Use
  • Privacy Policy
  • Disclaimer

We Offer


  • Premium Content
  • BioTech Scout
  • Interviews
  • Partner Events
  • Case Studies

Opportunities


  • Membership
  • Advertise
  • Submit Company
  • Write for Us
  • Contact Us

© BPT Analytics LTD 2025
We use cookies to personalise content and to analyse our traffic. You consent to our cookies if you continue to use our website. Read more details in our cookies policy.