HEST

Spatial gene-expression prediction from H&E with a frozen encoder, reproducing HEST-Benchmark (Jaume et al., NeurIPS 2024). A closed-form probe on the frozen features predicts the expression of 50 genes at every spot, so the benchmark measures the encoder alone.

Data and task

HEST-Benchmark pairs H&E images with spatial transcriptomics. Each spot is a 112 µm tile with a measured expression vector. Nine tasks are registered, one per organ or cancer type, as hest/<task>: CCRCC, COAD, IDC, LUNG, LYMPH_IDC, PAAD, PRAD, READ and SKCM.

Setting	Value
Task	Predict the log1p counts of each task’s 50 most variable genes from the spot’s tile.
Splits	HEST’s patient-level folds, read from the downloaded benchmark; soma never partitions the data itself.
Metric	mean_pearson: the Pearson correlation of each gene over the test spots, averaged over the 50 genes and then over the folds. Each gene’s own correlation is reported as pearson_geneNN_<symbol>.

Protocol

The benchmark trains no network. It extracts every spot’s tile features once into the feature cache, then fits a Ridge probe per fold, exactly as HEST does.

Setting	Value
Encoder	Any frozen tile encoder; the encoder axis of the benchmark.
Features	StandardScaler then PCA to 256 components, both fitted on the fold’s train spots only.
Probe	Ridge(solver='lsqr', fit_intercept=False) with alpha = 100 / (256 · 50), the HEST default for 256 components and 50 genes.
Seeds	0. The probe is closed-form, so one seed suffices.

Each fold writes metrics.json and predictions_<split>.parquet, and the run writes summary.json, like any soma run; the probe is soma.benchmark.hest.probe.RidgePCAProbe.

Results

soma’s mean_pearson on each task next to the value HEST publishes for the same encoder. HEST extracts its features with TRIDENT and soma with slide2vec, so the two are not expected to agree to the digit: the difference is shown as it is, without a tolerance band, and reproduction is judged by whether the encoders keep their published ranking.

Task	Encoder	Mean Pearson	HEST reference
PAAD	uni2	0.501	0.500
PAAD	virchow2	0.472	0.478
PAAD	h-optimus-1	0.492	0.496
COAD	uni2	0.310	0.301
COAD	virchow2	0.262	0.258
COAD	h-optimus-1	0.319	0.320
LUNG	uni2	0.559	0.559
LUNG	virchow2	0.569	0.569
LUNG	h-optimus-1	0.577	0.578

The published values are the HEST-Benchmark leaderboard.

Reproduce

Prepare the data. Install the optional HEST readers, then download one task from the Hugging Face Hub, leaving out HEST’s precomputed features (soma extracts its own):

pip install 'soma-pathology[hest]'
hf download MahmoodLab/hest-bench --include 'IDC/*' --exclude 'fm_v1/*' \
    --repo-type dataset --local-dir /path/to/hest-bench

Run one task or all nine; curated manifests are written under <raw-root>/curated:

soma reproduce hest/IDC --encoder virchow2 --raw-root /path/to/hest-bench/IDC
soma reproduce hest --encoder virchow2 --raw-root /path/to/hest-bench

Compare encoders with --encoders, and append a score to the results ledger with --record. Benchmarking describes the options shared by every benchmark.