H&E slides skip spatial transcriptomics: ALHENA distills cell and neighborhood similarity, reaching 0.341 patient-level signature correlation

Routine H&E-stained slides reveal cell morphology and tissue architecture but measure no molecular state, while spatial transcriptomics reads gene expression with position at higher cost and lower technical maturity. ALHENA trains on paired H&E and spatial transcriptomics and, at inference, takes only H&E to produce slide and cell representations carrying transcriptomic information. Earlier approaches mostly aligned slide representations to bulk expression or predicted local gene expression directly; SEAL also adapts UNI2-h with spatial transcriptomics but uses spot-level H&E–transcriptome alignment plus gene reconstruction. Across 110 combinations of 10 cancer types and 11 transcriptomic signatures from 750 TCGA patients, ALHENA reached a mean Spearman correlation of 0.341, against 0.161 for unadapted UNI2-h, 0.211 for SEAL and 0.284 for Virchow2; ALHENA exceeded UNI2-h on 95 of the 110 combinations.

The training target is not for images to reproduce a gene's readout but for the H&E representation to reproduce how close samples are in the transcriptomic models' view. Each cell pools the image tokens overlapping its nucleus, each tile contributes a global representation, and these are paired with cell-level and neighborhood-level transcriptomic representations. Both transcriptomic teachers stay frozen throughout: scConcept supplies cell-level representations and Novae supplies local-neighborhood representations, and neither sees H&E images. Adaptation used 25 Xenium sections spanning 13 organs, with 327,955 H&E tiles and 8,042,782 cells with measured transcriptomes; the UNI2-h backbone stayed frozen while rank-64 LoRA modules and four projection heads were trained, costing 12.9 GPU-hours on four H100 GPUs.

After adaptation the same encoder still handles conventional pathology tasks, so transcriptomic readouts and standard histopathology readouts can be analyzed on one representation. The Thunder benchmark tests pathology tasks unrelated to the molecular training objective across 20 datasets and seven evaluation axes, against UNI2-h before adaptation. Across 99 available Thunder evaluations, ALHENA's median difference from UNI2-h was −0.004; across the six discriminative axes the median difference was −0.006, in the same range as Virchow2 (−0.008) and H-optimus-1 (−0.009) in the same comparison.

Adaptation used 25 Xenium sections from 13 organs, with two further sections held out entirely for evaluation; the encoder stays frozen in all downstream evaluations and only readouts such as ridge regression or classifiers are fitted. Cell typing was evaluated on Xenium test stripes and the two fully held-out sections, and transferred without retraining or retuning to the 11 PanNuke tissue types with a matching atlas.

The gene-level gain appeared only on Xenium: on 170 HEST-1k slides across five tissues, ALHENA reached a mean correlation of 0.2117, UNI2-h 0.2123 and the permuted-target control 0.2134, so there is no evidence the gain transfers across sequencing technologies. Cell-typing errors concentrate in immune populations, for example 21.0% of B/plasma cells assigned to T/NK versus 4.0% when retrieval used the measured transcriptome. The study does not establish how performance scales with additional paired data.