Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study utilized a self-supervised AI foundation model to analyze routine pretreatment H&E images, aiming to predict the efficacy of the BOT+BAL immunotherapy combination in treatment-refractory solid tumors. By inferring spatial transcriptomics from these images, the model identified features linked to therapeutic response in microsatellite-stable colorectal cancer, sarcoma, and ovarian cancer. The results showed that the model achieved cross-validated AUROC scores of 0.61, 0.67, and 0.77 across the tumor types, indicating its potential as a predictive biomarker for clinical benefit from the treatment.
AI-driven analysis of H&E images can enhance the prediction of immunotherapy responses in previously treatment-resistant tumors, with up to 77% accuracy in ovarian cancer.
2535 Background: Predicting immunotherapy response is challenging in treatment-refractory/resistant (R/R) tumors where heterogeneity limits conventional biomarker utility. BOT (Fc-enhanced anti–CTLA-4) augments T-cell priming, depletes Tregs, and activates antigen-presenting cells. BOT+BAL (anti–PD-1) has shown activity across “cold” and R/R tumors including PD-L1–low and tumor mutational burden–low disease; thus, non-conventional predictive biomarkers are needed. A self-supervised AI foundation model applied to routine pretreatment H&E images was used to infer spatial transcriptomics and identify features associated with therapeutic benefit from BOT+BAL in microsatellite-stable colorectal cancer (MSS CRC), sarcoma, and ovarian cancer. Methods: A self-supervised AI foundation model was trained to infer spatial transcriptome expression from H&E images using purpose-built multimodal training data from thousands of human tumors profiled with multimodal assays. Pretreatment H&E images (not used for training) from 121 BOT+BAL–treated patients (pts) with R/R MSS metastatic CRC (included 20 pts with liver metastases), sarcoma, and ovarian cancer from the phase 1b C-800-01 trial (NCT03860272) were analyzed. Pt-specific embeddings were used to fit cross-validated classifiers predicting BOT+BAL clinical benefit (defined as complete or partial response [CR/PR] or stable disease [SD]). Results: Fitted logistic regression classifiers separated responders and non-responders in all tumor types, as measured by area under the receiver operating characteristic curve (AUROC; ranges from 0.5 [chance] to 1.0 [perfect separation]). Cross-validated AUROC was 0.61 in MSS CRC, 0.67 in sarcoma, and 0.77 in ovarian cancer (table; shows all findings). The model-recommended population is predicted to have a higher rate of clinical benefit, as measured by cross-validated precision. Beyond this binary analysis, the concordance index (C-index; measures accuracy and ranges from 0.5 [chance] to 1.0 [perfect]) for predicting overall survival (OS; via Cox proportional hazards model) was >0.5 in all tumor types, and highest in ovarian cancer. Conclusions: A self-supervised AI foundation model applied to routine pretreatment H&E images predicted BOT+BAL responses in MSS CRC, sarcoma, and ovarian cancer. These findings support AI-derived, H&E–based biomarker strategies for BOT+BAL and warrant prospective validation. Sample characteristics and model predictions. MSS CRC Sarcoma Ovarian Sampled population a No. of pt samples 67 30 24 Pts with CR/PR or SD 65% 57% 54% Modeled population AUROC 0.61 0.67 0.77 OS prediction, C-index 0.58 0.69 0.78 Model recommended population Population size 43% 50% 46% Estimated pts with CR/PR or SD 76% 67% 73% a Data cutoff: Mar 13, 2025; analyses ongoing.