Abstract 5003: Machine-learning informed multifactorial biomarkers as an approach to stratify survival outcomes among ADC treated patients

Abstract As oncology treatments increase in molecular complexity, the need for biomarkers continues to grow. This becomes critical in treatments such as antibody-drug conjugates (ADC’s), which are a promising targeted anti-cancer therapy, that not all patients respond equally well to. Though prior work suggests that patients who over-express ADC targets relative to those that do not are more likely to benefit from ADC therapy, a robust biomarker signature that predicts response has yet to be defined for most ADCs. To address this unmet need, we developed an approach to impute composite warhead-target ADC biomarkers from digitized histopathology (H&E) slides. The first component of this biomarker signature is a machine learning model that imputes ADC target RNA expression from histopathology slides. Flash frozen and FFPE histopathology slides from 11, 115 patients in The Cancer Genome Atlas (TCGA) and 2, 732 patients from Aster across 32 tumor types were embedded to a lower dimensional representation via the DINO ViT model. A multilayer perceptron (MLP) was trained within each tumor type to predict transcriptome-wide gene expression (19, 421 protein-coding genes) from these embedding vectors, achieving a mean spearman r = 0.20+/-0.17 across all genes (mean pearson r = 0.21+/-0.18), and a mean spearman r = 0.32 +/- 0.17 across the top 1000 genes across cohorts. Improved performance was observed when imputing a weighted sum of expression for transcriptomically correlated genes (as ascertained via differential expression analysis) for a suite of 346 known candidate ADC targets. The second component of the biomarker signature focuses on stratifying longitudinal outcomes in patients treated with topoisomerase 1 (TOP1) inhibitors. We developed a gene signature (weighted mean expression of 52 genes) based on treatment response in cell lines (GDSC2) and therapy-agnostic patient gene expression profiles (TCGA, Aster). Additional weighting based on cross-modality evidence was employed. The resulting signatures were associated with longitudinal endpoints in primary and metastatic colorectal cancer patients. An MLP was subsequently trained within the TCGA and Aster cohorts to impute the TOP1 signature from H&E datasets, achieving cross-validation spearman correlation of 0.42 within the COAD-READ cohort and a spearman R of 0.29 in a held-out MRC FOCUS test set (N=73). In summary, the combined warhead-target imputation signature shows promise in stratifying likely ADC treatment responders from non-responders. Given that assays directly measuring gene expression at the bulk and single cell level are still seldom used in the clinic, this imputation model may provide a useful tool to leverage H&E imaging routinely performed as part of patient diagnosis as a tool to inform treatment selection. Citation Format: Anna Shcherbina, Yajas Shah, Zachary McCaw, Yang Su, Serra Elliott, David Light, Angela Oliveira Pisco, Sanjana Narayanan, Chaya Wurman, Vincent Luckzow, Christopher Probert, Sacha Holland, Virginia Donovan, Peter Szabo. Machine-learning informed multifactorial biomarkers as an approach to stratify survival outcomes among ADC treated patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5003.