1 October 2025 in Nature Biomedical Engineering
Numerous pathology foundation models have been developed to extract clinically relevant information. There is currently limited literature independently evaluating these foundation models on external cohorts and clinically relevant tasks to uncover adjustments for future improvements. Here we benchmark 19 histopathology foundation models on 13 patient cohorts with 6,818 patients and 9,528 slides from lung, colorectal, gastric and breast cancers. The models were evaluated on weakly supervised tasks related to biomarkers, morphological properties and prognostic outcomes. We show that a vision-language foundation model, CONCH, yielded the highest overall performance when compared with vision-only foundation models, with Virchow2 as close second, although its superior performance was less pronounced in low-data scenarios and low-prevalence tasks. The experiments reveal that foundation models trained on distinct cohorts learn complementary features to predict the same label, and can be fused to outperform the current state of the art. An ensemble combining CONCH and Virchow2 predictions outperformed individual models in 55% of tasks, leveraging their complementary strengths in classification scenarios. Moreover, our findings suggest that data diversity outweighs data volume for foundation models.
1 July 2026 in Nature Communications
Artificial intelligence has transformed digital pathology by enabling biomarker prediction from high-resolution whole-slide images. However, current methods are computationally inefficient, processing thousands of redundant tiles per slide and requiring complex aggregation models. We introduce EAGLE (Efficient Approach for Guided Local Examination), a deep learning framework that emulates pathologists by selectively analyzing informative regions. EAGLE combines task-agnostic tile selection with detailed feature extraction and is benchmarked against leading slide- and tile-level foundation models across 43 tasks from nine cancer types spanning morphology, biomarker prediction, treatment response and prognosis. EAGLE outperforms patch aggregation methods by up to 23% and achieves the highest overall classification performance. It processes one slide in 2.27 s, reducing computational time by more than 99% compared with existing models. This efficiency supports rapid and auditable workflows by enabling review of the exact tiles used for each prediction and reducing dependence on high-performance computing. By reliably identifying informative regions and minimizing artifacts, EAGLE provides robust and auditable outputs, supported by systematic negative controls and attention concentration analyses. Its unified embedding enables rapid slide search, integration into multi-omics pipelines and emerging clinical foundation models.
23 October 2026 in Science Advances
Digital twin models can accelerate therapeutic development by enabling low-risk testing of candidate interventions. In preterm labor (PTL), a major pregnancy complication where clinical trials face ethical and financial barriers, digital twins are especially valuable for evaluating new therapies targeting immune dysfunctions driving PTL. Yet, current models lack single-cell resolution, limiting detection of cell type–specific mechanisms, off-target effects, and the design of personalized interventions. We present Simulated Immunome Modeling of Clinical Outcomes (SIMCO), a single-cell–level digital twin framework that models immunomodulatory treatment effects on the timing of labor using immunome-wide, single-cell simulations. SIMCO’s digital twins are trained and validated on a newly generated mass cytometry atlas of the pregnant immunome exposed to nine candidate drugs preselected for PTL prevention. Applying SIMCO to an independent cohort of pregnant individuals, we simulate treatment effects on gestational length, screening for candidate drugs that delay labor timing and providing system-level mechanistic insight for each drug candidate. Tetrahydrofolate and the combination of aspirin and lansoprazole emerged as top candidates for PTL prevention, delaying labor onset primarily through enhanced mTOR signaling in innate immune cells and attenuated JAK/STAT signaling in naïve CD4 + T cells.