Abstract Foundation models trained on large histopathology datasets offer broad inference capabilities, yet their performance in Merkel cell carcinoma has not been characterized. We investigated whether a whole slide imaging foundation model could extract clinically informative signals from immunohistochemistry biopsies stained for Merkel cell polyomavirus T antigen. Thirty one digitized slides (12 Stage I, 3 Stage II, 11 Stage III, 3 Stage IV, 2 indeterminate) were processed using a custom tissue masking, tiling, and stain normalization pipeline. High-dimensional features were extracted from the standardized output using a UNI2, a pretrained frozen ViT G/14 encoder developed by the Mahmood Lab. Each tile yielded a 1536 dimensional embedding that was summarized at the slide level by feature means and at the zone level by Moran’s I to capture spatial structure. Low dimensional representations were computed with PCA and UMAP. Clinical annotations included recurrence, viral status, immunosuppressed status, and AJCC stage.Across 40 tested correlations, slide level summaries showed the largest associations. Viral status served as a positive control and produced the strongest effects, with slide level PC1 correlating at r equals −0.684 and UMAP1 at r equals 0.594. Binarized American Joint Committee on Cancer (AJCC) stage showed consistent separation in PCA and UMAP spaces, and four individual embedding dimensions correlated with advanced stage at magnitudes up to r equals 0.58. Recurrence also aligned with several embedding axes, including representative correlations of r equals 0.68, 0.63, 0.57, and 0.56. Zone level summaries captured additional biology, with immunosuppressed status reflected in zone level UMAP and PCA components that reached r equals 0.337. The mean absolute correlation across all outcomes was 0.184, and nine correlations exceeded an absolute value of 0.3.These results show that foundation model representations derived from polyomavirus targeted immunohistochemistry contain measurable structure related to viral status, stage, recurrence, and host immunologic state. The ability of slide level and zone level embeddings to recover these signals in a modest cohort suggests that pre trained encoders capture histoarchitectural and spatial cues relevant to Merkel cell carcinoma biology. Larger multi site cohorts with longitudinal follow up will enable validation of these features and support integration with circulating or genomic markers for risk stratification in prospective studies. Citation Format: Roshan Lodha, Kelsey Ouyang, Claire Reynolds, Allison Vidimos, Bryan Carroll. Foundation model-derived features from immunohistochemistry correlate with recurrence and stage in Merkel cell carcinoma abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1439.
Lodha et al. (Fri,) studied this question.