HEIST

Spatial transcriptomics and spatial proteomics record molecular counts while preserving each cell's position within the tissue, but most single-cell foundation models throw one of those two signals away. Models that treat a cell as a flat expression vector cannot express how a cell's internal regulatory program responds to its microenvironment, and models built on a fixed gene vocabulary cannot be applied to a panel — let alone a modality — they were never trained on. HEIST, from Yale University, addresses both constraints by representing a tissue as a two-level hierarchical graph.

The upper level is a spatial cell graph, connecting each cell to its physical neighbors. The lower level is a gene co-expression network inside every individual cell. Intra-level message passing propagates information among cells and among genes, while cross-level attention lets the two levels condition each other, so that a cell embedding carries microenvironmental context and a gene embedding carries the cell it sits in. Because genes enter the model through their position in a co-expression network rather than through a fixed vocabulary index, the pretrained encoder can be applied to gene panels it has never seen — and to protein markers measured by spatial proteomics platforms such as CODEX and MIBI — without retraining.

HEIST was first posted as a preprint in June 2025 by Hiren Madhu, João Felipe Rocha, Tinglin Huang, Siddharth Viswanath, Smita Krishnaswamy, and Rex Ying, and was subsequently accepted at ICLR 2026. Code is on GitHub and the pretrained checkpoint is on Hugging Face under an MIT license.

#Key Features

• Hierarchical tissue graph: Cells are nodes in a spatial neighbor graph, and each cell is itself a graph of co-expressed genes, so intercellular organization and intracellular regulation are modeled jointly rather than in separate stages.
• Vocabulary-free gene handling: Genes are encoded by rank-based positional encodings over the co-expression network instead of a fixed lookup table, which is what allows transfer to new panels and new assays.
• Cross-modality generalization: The same pretrained weights produce useful embeddings for spatial proteomics data despite pretraining only on spatial transcriptomics.
• Frozen-encoder downstream use: Clinical outcome prediction, cell type annotation, and cell clustering are performed on frozen embeddings; only the decoder is fine-tuned for gene imputation.
• Sparse and fast: Avoiding dense full self-attention makes embedding extraction roughly 8× faster than scGPT-spatial and 48× faster than scFoundation.

#Technical Details

HEIST stacks ten hierarchical graph transformer layers with 128-dimensional embeddings and eight attention heads per layer. Cell-level message passing combines a transformer with a Graph Isomorphism Network; gene-level message passing uses sparse attention restricted to co-expression edges; a directional cross-attention mechanism links the levels. Spatial coordinates enter through sinusoidal positional encodings, and a three-layer GIN decoder handles reconstruction. The released checkpoint holds roughly 10.4M parameters. Pretraining covers 22.3M cells from 124 tissue slices spanning 15 organs, all measured by MERFISH or Xenium: 13.3M cells from 10x Genomics datasets, 8.7M from Vizgen, and 360K from the Seattle Alzheimer's Brain Atlas. The objective combines spatially aware contrastive learning (cell–cell, gene–gene, and cross-level) with masked autoencoding of coordinates and expression. Training ran on four NVIDIA L40S GPUs at roughly three hours per epoch. On downstream benchmarks HEIST records 0.769–0.937 AUC-ROC across seven tissue-level classification scenarios — winning six of seven — takes the top spot on four of five cell type annotation datasets, and lifts fine-tuned gene imputation Pearson correlation to 0.821 on placenta and 0.807 on skin, against baselines including scGPT and CellPLM.

#Applications

HEIST is aimed at groups running CODEX, MIBI, MERFISH, or Xenium experiments who want a general-purpose embedding rather than a bespoke pipeline per study. Frozen cell embeddings support patient-level tasks such as predicting immunotherapy response in melanoma or remission in head-and-neck cancer, as well as routine annotation and clustering work. Because the encoder is panel-agnostic, a lab can embed a new proteomics panel or a new tissue without collecting the pretraining-scale data such models usually demand, and the gene imputation head can extend a limited targeted panel toward transcriptome-wide coverage.

#Impact

HEIST is the first spatial-omics foundation model to model intracellular co-expression networks and intercellular spatial structure as one hierarchical object, and it shows that doing so recovers spatially informed cell subpopulations that flat single-cell models miss. Its vocabulary-free design is the more portable idea: it converts a transcriptomics-pretrained encoder into something usable on proteomics panels, which is otherwise a retraining problem. Adoption remains early — the checkpoint appeared on Hugging Face in 2026 — and the model's limits are inherited from its corpus, which covers only two single-cell-resolution technologies and excludes spot-based platforms such as Visium.