Generalizable AI predicts immunotherapy outcomes across cancers and treatments

Area of Science:

• Computational biology
• Cancer immunology
• Machine learning in oncology

Background:

• Immune checkpoint inhibitors (ICIs) are crucial cancer therapies, but patient response varies significantly.
• Current biomarkers for ICI response lack generalizability across different cancer types and treatments.
• Predicting immunotherapy response remains a challenge, hindering optimal patient selection and treatment strategies.

Purpose of the Study:

• To develop a pan-cancer foundation model for predicting immunotherapy response using bulk tumor transcriptomes.
• To encode gene expression into biologically interpretable immune concepts for enhanced predictive power.
• To improve patient stratification and provide mechanistic insights into treatment response and resistance.

Main Methods:

• Development of COMPASS, a concept bottleneck transformer model.
• Training on a large dataset of 10,184 tumors across 33 cancer types.
• Encoding gene expression into 44 biologically grounded immune concepts.

Main Results:

• COMPASS demonstrated superior performance compared to 22 existing methods across 16 clinical cohorts (seven cancers, six ICIs).
• Achieved an average improvement of 8.5% in accuracy and 15.7% in area under the precision-recall curve.
• Showed generalization to unseen cancer types and treatments, and identified responders with significantly longer overall survival (HR=4.7, P<0.0001).

Conclusions:

• COMPASS accurately predicts immunotherapy response and generalizes across diverse cancer types and treatments.
• The model provides mechanistic insights into response and resistance, highlighting pathways like TGFβ signaling and T cell dysfunction.
• COMPASS has the potential to guide clinical trial design, patient stratification, and translational studies in cancer immunotherapy.