SPADE

Computational pathology foundation models learn from H&E-stained whole-slide images alone, and H&E is a two-color stain: it renders nuclei and cytoplasm beautifully but says nothing directly about molecular state. Spatial transcriptomics fills that gap, giving gene expression at known coordinates on the same tissue section. SPADE, from the Biomedical AI Research Lab at UCLA, uses that pairing as a training signal — not to predict expression, but to reshape the image encoder's latent space so that morphology and molecular identity sit together.

The method is contrastive, in the CLIP mould, matching an H&E patch embedding to the gene expression profile of its co-registered Visium spot. The complication is that a multi-organ training set makes contrastive learning too easy: a batch drawn from brain, heart, and lung tissue offers only trivial negatives, and the model can drive its loss down without learning anything discriminative. SPADE addresses this with a mixture-of-data-experts design adapted from MoDE, partitioning the corpus by two-step K-means clustering in image feature space so that each expert trains against hard negatives from morphologically similar tissue.

#Key Features

• Two-step hard-negative clustering: Patches are first clustered within each of 32 organ types, then the resulting centroids are clustered again across organs, producing a final set of expert conditions that span tissue types without collapsing them together.
• Distance-weighted expert routing: At inference, patch embeddings are combined across all experts with softmax weights based on distance to each expert centroid, which consistently outperforms hard nearest-centroid routing in the paper's ablations.
• Encoder-agnostic framework: SPADE wraps a frozen pathology backbone rather than replacing it; the released model builds on UNI features, and the authors show the same recipe applied to a Virchow2 backbone.
• Attention that tracks tumor: In pathologist review of prostate and colorectal cases, SPADE-based attention maps concentrated on malignant glands and high-grade dysplasia where the baseline encoder's attention drifted onto stroma, smooth muscle, and inflammation.

#Technical Details

Pretraining uses the 515 human Visium profiles within HEST-1k, patched at 224 x 224 pixels at 20x with fixed stain normalization. Expression is normalized per spot and log-transformed, and the union of the top 50 highly variable genes per sample gives a 7,986-gene vector. Patch features come from UNI; the contrastive objective uses a similarity-adjusted target so that spots with genuinely similar expression are not pushed apart. The released configuration uses 16 clusters at each of the two K-means stages.

Evaluation covers 20 downstream tasks — eight cancer subtyping tasks, five disease-specific survival tasks plus prostate biochemical recurrence, and six gene-alteration and receptor-status prediction tasks — with attention-based multiple-instance learning aggregating patches to the slide level. SPADE improves on the best baseline in 17 of the 20. It reaches the top AUC on seven of eight subtyping tasks (98.5 on PLCO lung, 94.9 on PANDA ISUP grading, 98.6 on Camelyon16), the top c-index on five of six survival tasks, and the highest mean biomarker AUROC (79.4). Baselines are UNI, CONCH, Prov-GigaPath, and TANGLE, the last of which pairs slides with bulk rather than spatial RNA-seq.

#Applications

SPADE is a drop-in patch encoder for slide-level prediction pipelines: subtyping, grading, survival and recurrence modelling, and inference of molecular status from morphology alone. Its practical appeal is that spatial transcriptomics is needed only once, during pretraining — downstream users supply ordinary H&E slides and get embeddings that carry molecular context for free, which matters in settings where sequencing every case is not viable.

#Impact

The work shows that a modest amount of paired spatial data can measurably improve a pathology encoder that was itself trained on orders of magnitude more images, and that the gain survives across organs, cohorts, and task types. The honest constraint is scale: 515 Visium profiles is small next to the slide counts behind current pathology foundation models, and spatial data remain expensive to acquire, so the ceiling here is set by the corpus rather than the method. Code and checkpoints are released through the project's GitHub repository, with weights distributed via a linked Google Drive folder.