Frozen pathology foundation models stumble on single-cell gene expression from H&E, while lightweight adaptation lifts 5K-panel correlation from 0.065 to 0.174

Frozen pathology foundation-model embeddings work markedly less well at single-cell resolution than at pseudo-spot resolution, indicating that these tile-level representations suit regional rather than cell-resolved expression prediction. On the same Xenium data, mean per-gene Pearson correlation on the 5K panel is 0.375 at pseudo-spot level and drops to 0.065 at single-cell level; the HVG50 pseudo-spot value of 0.428 is comparable to the spot-level performance reported by HEST-1K. The comparison comes from two spatial resolutions of the same Xenium data under the same linear-probe protocol; other gene panels show similar declines, and correlation falls as gene expression sparsity rises, more steeply at single-cell level.

Restricting the representation to the target cell and its immediate surroundings works better than using the global class token of the whole tile. With the class token as baseline, 5K-panel correlation is 0.128; pooling tokens inside the Xenium cell mask raises it to 0.161. Distance-weighted pooling within a 5-micron radius of the centroid, which needs no cell boundaries, gives 0.154 on the 5K panel, 4.6% below mask pooling but 20.2% above the class-token baseline.

Stacked adaptations improve every gene panel cumulatively, and the final model attains the highest per-gene correlation, per-cell correlation and gene-detection AUROC on the 5K panel together. Relative to the pre-adaptation baseline, the gain is 166.7% on 5K, 124.2% on HVG50, 95.6% on IBD-relevant genes and 37.1% on SVG50. On the 5K panel the final model with LoRA reaches per-gene correlation 0.174±0.014, per-cell correlation 0.280±0.033 and AUROC 0.800±0.009, against 0.094±0.009, 0.242±0.031 and 0.642±0.005 for the strongest competing method, SpatialEx.

Predictive performance varies widely by biological context, with spatially variable and disease-relevant genes correlating clearly above the full gene panel. The final model reaches per-gene correlation 0.643 on SVG50 and 0.356 on IBD-relevant genes, versus 0.174 on the full 5K panel. Stratified by cell type, stromal cells gain most from baseline to final model and epithelial cells least; pathway enrichment places the most improved genes in extracellular matrix organization, cell junction and adhesion pathways.

The demonstrated scope is colon tissue: 176 ulcerative colitis TMA cores from 117 donors, about 8.6 million cells, were used for evaluation, while two ulcerative colitis, two Crohn's disease and two healthy control resection samples, about 2.9 million cells, were used for fine-tuning only; gene panels are the 5,101 genes plus HVG50, SVG50 and 76 IBD-relevant genes. Adaptation was done on H-Optimus-0, and the improvement trend reproduced on Prov-GigaPath and DINOv3.

Single-cell measurements are sparse and subject to transcript spillover, imperfect cell segmentation and cell-assignment errors, all of which limit observable correlation; whether predictions capture subtle cell states and within-cell-type variation beyond broad cell identity remains to be tested, and broader validation across tissues and cohorts is needed. The authors frame pathway enrichment as hypothesis-generating rather than definitive, and characterizing model uncertainty per gene and per cell type is left to future work.