Investigating the Impact of Histopathological Foundation Models on Regressive Prediction of Homologous Recombination Deficiency

Foundation models pretrained on large-scale histopathology data have found great success in various fields of computational pathology, but their impact on regressive biomarker prediction remains underexplored. In this work, we systematically evaluate histopathological foundation models for regression-based tasks, demonstrated through the prediction of homologous recombination deficiency (HRD) score - a critical biomarker for personalized cancer treatment. Within multiple instance learning frameworks, we extract patch-level features from whole slide images (WSI) using five state-of-the-art foundation models, and evaluate their impact compared to contrastive learning-based features. Models are trained to predict continuous HRD scores based on these extracted features across breast, endometrial, and lung cancer cohorts from two public medical data collections. Extensive experiments demonstrate that models trained on foundation model features consistently outperform the baseline in terms of predictive accuracy and generalization capabilities while exhibiting systematic differences among the foundation models. Additionally, we propose a distribution-based upsampling strategy to mitigate target imbalance in these datasets, significantly improving the recall and balanced accuracy for underrepresented but clinically important patient populations. Furthermore, we investigate the impact of different sampling strategies and instance bagsizes by ablation studies. Our results highlight the benefits of large-scale histopathological pretraining for more precise and transferable regressive biomarker prediction, showcasing its potential to advance AI-driven precision oncology.

💡 Research Summary

This paper investigates how large‑scale histopathology foundation models influence the regression‑based prediction of homologous recombination deficiency (HRD) scores, a continuous biomarker critical for guiding platinum‑based chemotherapy and PARP inhibitor therapy across multiple cancer types. The authors assembled whole‑slide image (WSI) datasets from TCGA (breast invasive carcinoma, uterine corpus endometrial carcinoma, lung adenocarcinoma) and external validation cohorts from CPTAC, totaling over 5,000 patients and thousands of WSIs. Each slide was tessellated, background‑filtered, and color‑normalized before being fed to six feature extractors: a contrastive‑learning baseline (RetCCL) and five state‑of‑the‑art foundation models—UNI, UNI‑2, Virchow‑2, GPFM, and CONCH— all transformer‑based and pretrained on millions of WSIs using self‑supervised strategies such as masked image modeling and teacher‑student distillation.

The extracted patch embeddings (≈1024‑dimensional) were organized into patient‑level bags via K‑means clustering (k = 50). Three sampling strategies were explored to create fixed‑size bags (600, 800, 1 000, 12 000 instances): (1) cluster‑size‑weighted sampling, (2) clustered random sampling, and (3) pure random sampling. Two multiple‑instance learning (MIL) aggregators were employed: an attention‑based MIL (attMIL) and the SuRe Transformer, which uses clustering‑aware self‑attention to handle long sequences. For attMIL, an “all‑features” configuration (using every extracted patch) was also evaluated, albeit at higher computational cost.

To address the pronounced target imbalance—most cohorts contain many HRD‑negative (low score) patients and few HRD‑positive (high score) cases—the authors introduced a distribution‑based upsampling algorithm. The continuous HRD distribution was discretized into seven bins; for each bin, a sampling budget was computed based on the difference from the largest bin, capped by a factor α (0.65) and scaled by β (0.25). This procedure repeatedly samples whole‑patient bags from under‑represented bins, generating additional training instances without mixing patches across patients, thereby improving recall and balanced accuracy for the rare high‑HRD subgroup.

Five‑fold cross‑validation on the internal TCGA cohorts showed that every foundation model outperformed the RetCCL baseline across all cancer types. UNI‑2 consistently achieved the highest median AUROC (BRCA 0.8304, LUAD 0.7191, UCEC 0.8479), followed by UNI, Virchow‑2, GPFM, and CONCH. Even the lowest‑performing CONCH surpassed RetCCL, confirming the value of large‑scale pretraining. The superiority held for both attMIL and SuRe Transformer aggregators, though SuRe Transformer benefitted more from larger bag sizes, while attMIL’s “all‑features” mode offered marginal gains at substantially higher memory usage. Cluster‑size‑weighted sampling proved the most stable across configurations, whereas random strategies yielded higher variance.

External validation on CPTAC LUAD and UCEC cohorts confirmed the generalizability of the findings: models built on foundation‑model features retained their performance edge despite domain shift, and the upsampling scheme continued to boost metrics for the minority HRD‑positive group.

In summary, the study demonstrates that histopathology foundation models provide robust, transferable representations that enhance regression‑type biomarker prediction beyond traditional contrastive‑learning features. By coupling these representations with carefully designed MIL pipelines and a novel distribution‑aware upsampling technique, the authors achieve more accurate, balanced, and clinically relevant HRD score predictions. The work paves the way for image‑only, cost‑effective alternatives to genomic assays and suggests future extensions toward multimodal foundation models and other continuous oncologic biomarkers.