Abstract Background Metastatic renal cell carcinoma (mRCC) is a molecularly heterogeneous disease commonly driven by somatic mutations in genes such as VHL, PBRM1, and BAP1. Although next-generation sequencing (NGS) is the gold standard for identifying such mutations, it remains costly and logistically challenging. Given that genetic alterations often lead to morphological changes, we hypothesized that an artificial intelligence (AI) model applied to hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) could predict underlying somatic mutations. Methods We identified consecutive patients with mRCC who had undergone clinical NGS and had available H&E-stained histopathology slides. Patients with somatic alterations in at least one target gene were annotated. WSIs were tiled into non-overlapping 256 × 256-pixel patches. Background tiles were excluded based on brightness. Across 160 total WSIs, 3 slides had no valid tissue tiles and were excluded. The final dataset included 157 slides with usable content. From 69 672 tile attempts, 28 924 tiles (41.5%) passed quality filters and were used for model development. Two modeling pipelines were implemented: (1) slide-level, aggregating global features from WSIs, and (2) tile-level, embedding and classifying each tile with prediction aggregation at the slide level. Embeddings were generated using DINOv2 and the GigaPath foundation model. To evaluate model performance, we used receiver operating characteristic (ROC) curves and computed the area under the curve (AUC) for both slide-level and tile-level pipelines. ROC analysis was applied at both the tile level and the patient-aggregated level. Principal component analysis (PCA) was used for embedding visualization. Violin plots summarized patient-level AUCs across genes. Results A total of 157 WSIs (one per patient) were analyzed. The cohort included 83 VHL with pathogenic variants (PV) (52.9%) and 74 VHL-wild type (47.1%), along with PVs in: PBRM1 (48/157, 30.6%), BAP1 (9/157, 5.7%), PTEN (8/157, 5.1%), TSC1 (7/157, 4.5%), KDM5C (14/157, 8.9%), SETD2 (27/157, 17.2%), MTOR (8/157, 5.1%), TERT (19/157, 12.1%), and NF2 (2/157, 1.3%). Slide-level modeling demonstrated limited performance in mutation prediction. However, tile-level modeling substantially improved performance. The ROC curve for tile-level VHL prediction achieved AUC = 0.7206. Other tile-based AUCs included PBRM1 (0.76), PTEN (0.80), BAP1 (0.69), TSC1 (0.70), and NF2 (0.62), while lower performance persisted for MTOR (0.25), TERT (0.46), and KDM5C (0.47). PCA of tile embeddings revealed partial clustering by mutation status, and heatmaps localized high-probability regions consistent with expected histologic features. Conclusions Tile-level AI modeling enables accurate, interpretable prediction of somatic mutations in mRCC using H&E-stained WSIs. This approach may offer a scalable and accessible tool for expanding molecular profiling in settings without widespread access to NGS.
Barragán-Carrillo et al. (Wed,) studied this question.