Abstract 2476: Prediction of clinically relevant biomarker protein levels from histopathology images for breast cancer subtyping

Abstract Background: AI is transforming precision oncology, especially with pretrained deep learning (DL) models (foundation models; FM) extracting tumor-related genomic and transcriptomic patterns from digitized whole-slide histopathology images (WSIs). However, while proteomics is closer to cell function and phenotypes, few studies focus on deriving protein-level insights from WSIs, partly due to limited matched proteomics cohorts. To address this, we offer Path2Prots, a weakly-supervised DL method to estimate relative abundance of 393 clinically relevant protein biomarkers in breast cancer (BC) from tumor histopathology. We show its utility via BC clinical subtype prediction across multiple cohorts. Methods: Path2Prots is a two-stage pipeline: First, each WSI is divided into 512 x 512 tiles at 20x magnification, followed by image processing to filter relevant tiles that are fed to a transformer-based FM to extract morphological features. Next, using the features and matched patient-level proteomics, a multilayer perceptron is trained to predict the patient’s protein levels. To train, we use 2, 074 WSIs from 841 patients in TCGA-BRCA and their proteomics from The Cancer Proteome Atlas, that uses reverse-phase protein array (RPPA) to quantify 393 proteins (313 total and 80 post-translationally modified). To leverage two WSI types with distinct benefits, we built two models: FFPE model, trained on 893 formalin-fixed paraffin embedded (FFPE) slides (used for diagnosis), and FF model, trained on 1, 181 fresh-frozen (FF) slides (better RNA quality). Results: We first assess models with Pearson correlation (R) between measured and predicted proteomics in cross-validation (CV). The FFPE model gains mean R = 0.258 with 55 well-predicted proteins (WPPs; R ≥ 0.4, P ≤ 0.05), FF model gains mean R = 0.281 and 73 WPPs and combining model predictions yields mean R = 0.305 and 96 WPPs. As currently no other cohort contains matched WSIs and RPPA, we test our models via their ability to identify BC clinical marker status: HER2, ER and PR, and clinical subtypes: HER2+, HR+ and TNBC. For status prediction, CV models yield area under the curve values (AUCs) of 0.67, 0.87, 0.76 (FFPE); 0.67, 0.89, 0.78 (FF); 0.69, 0.90, 0.79 (combined) with the upper bound using measured RPPA as 0.77, 0.96, 0.81. In external validation, CPTAC-BRCA (FF, n = 89) gains 0.68, 0.87, 0.83; TransNEO (FF, n = 160) gains 0.69, 0.78 (no PR); and IMPRESS (FFPE, n = 126) yields 0.72, 0.74, 0.77. For subtype prediction, CV AUCs are 0.67, 0.75, 0.84 (FFPE); 0.67, 0.77, 0.84 (FF); 0.69, 0.77, 0.88 (combined) and validation AUCs are 0.68, 0.72, 0.85 (CPTAC-BRCA); 0.69, 0.72, 0.83 (TransNEO); 0.72, 0.87 (IMPRESS; no HR+). Conclusion: We present a novel computational pathology AI method to infer proteomics from histopathology. Our analysis reveals a clinically important subset of proteins can be robustly predicted from routine WSIs for clinical application. Citation Format: Saugato Rahman Dhruba, Danh-Tai Hoang, Eldad D. Shulman, Ranjan K. Barman, Sanna Madan, Sanju Sinha, Kenneth Aldape, Eytan Ruppin. Prediction of clinically relevant biomarker protein levels from histopathology images for breast cancer subtyping [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 2476.