SEAL

SEAL (Spatial Expression-Aligned Learning) is a method for upgrading existing patch-level pathology foundation models by aligning their histology representations with spatial transcriptomics (ST). Modern pathology encoders such as UNI, Virchow2, and CONCH learn powerful features from tissue morphology alone, but the routine H&E image carries no explicit molecular signal. SEAL closes that gap by infusing localized gene expression into the vision encoder, so that morphological features become predictive of the underlying molecular state of the tissue.

Rather than pretraining a new encoder from scratch, SEAL is a parameter-efficient vision-omics finetuning method that acts as a drop-in replacement for the vision backbone in existing computational pathology pipelines. It exploits the morphomolecular coupling between local expression and local morphology: paired spot-region examples teach the model which visual patterns correspond to which expression profiles. The published framework adapts five pathology foundation models — CONCH, UNIv2, Virchow2, H-optimus-0-mini, and Phikon-v2 — leaving downstream code unchanged while enriching the features it consumes.

SEAL was introduced in February 2026 by Konstantin Hemker, Andrew H. Song, Cristina Almagro-Pérez, Guillaume Jaume, Sophia J. Wagner, Anurag Vaidya, Nikola Simidjievski, Mateja Jamnik, and Faisal Mahmood, a collaboration led by the Mahmood Lab (Mass General Brigham, Harvard Medical School) with the University of Cambridge and Télécom Paris. It is described in an arXiv preprint awaiting peer review.

#Key Features

• Vision-omics alignment: SEAL injects localized spatial-transcriptomic signal into a pathology vision encoder, coupling morphology with gene expression instead of relying on image features alone.
• Drop-in backbone replacement: The adapted encoder substitutes directly for the original foundation model in existing pipelines, requiring no change to downstream multiple-instance-learning or classification code.
• Parameter-efficient finetuning: SEAL adds lightweight adapters on top of established backbones rather than retraining from scratch, making molecular alignment inexpensive to apply across multiple encoders.
• Slide- and patch-level capabilities: The aligned features improve slide-level tasks (molecular status, subtyping, pathway activity, treatment response) and enable patch-level gene expression prediction.
• Morphomolecular retrieval: SEAL supports image-to-gene and gene-to-image retrieval, linking histological regions to expression profiles for interpretation and discovery.

#Technical Details

SEAL is trained on over 700,000 paired gene-expression-spot and tissue-region examples spanning tumor and normal samples from 14 organs, drawn from the HEST spatial transcriptomics resource. The method adds omics-aligned components to the frozen pathology backbone, so a single procedure transfers across encoders of different scale, from a ViT-Base/16 (CONCH) to ViT-Huge (UNIv2, Virchow2). It was evaluated across 38 slide-level and 15 patch-level downstream tasks, where the aligned encoders improved over their morphology-only counterparts on molecular status prediction, pathway activity, and treatment response, and showed robust generalization on out-of-distribution evaluations. The public release provides gated CONCH and UNIv2 checkpoints, each comprising separate vision and omics weight files.

#Applications

SEAL targets computational pathology researchers who want molecular-aware representations without collecting paired transcriptomics for every cohort. Once a backbone is aligned, its features can be used for biomarker and molecular status prediction from routine H&E, pathway-activity estimation, subtyping, treatment-response modeling, and direct patch-level gene expression prediction. The retrieval capabilities support morphomolecular discovery, letting researchers query tissue regions by expression signature or find the expression profile associated with a morphology of interest.

#Impact

SEAL reframes how pathology foundation models can be improved: instead of scaling morphology-only pretraining, it uses spatial transcriptomics as a supervisory signal to make existing encoders molecularly grounded, and does so cheaply enough to apply across five published backbones. By building on the HEST corpus, it connects the growing body of paired histology-transcriptomics data to the widely used UNI, CONCH, and Virchow families. As a recent preprint, its results await peer review, and the released weights are gated under a non-commercial, no-derivatives research license, so it has not been validated for clinical diagnostic use.