🧬 biology

HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data

This paper introduces HEIST, a hierarchical graph transformer foundation model pretrained on 22.3 million spatially resolved cells that integrates spatial context with intra-cellular gene and protein networks to achieve state-of-the-art performance in cell annotation, gene imputation, and clinical outcome prediction while generalizing to unseen data types like spatial proteomics.

Original authors: Hiren Madhu, João Felipe Rocha, Tinglin Huang, Siddharth Viswanath, Smita Krishnaswamy, Rex Ying

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hiren Madhu, João Felipe Rocha, Tinglin Huang, Siddharth Viswanath, Smita Krishnaswamy, Rex Ying

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand a bustling, crowded city. You have two types of data about this city:

The People List: A massive spreadsheet telling you what every single person is thinking and doing (their "gene expression").
The Map: A GPS showing exactly where every person is standing and who they are standing next to (their "spatial location").

For a long time, scientists had to choose between looking at the People List or the Map. If they looked at the list, they missed the context of the neighborhood. If they looked at the map, they missed the internal thoughts of the individuals. Furthermore, most computer models treated every person's "vocabulary" as fixed, meaning they couldn't understand a new city if it used different words (genes or proteins) than the one they were trained on.

Enter HEIST.

HEIST is a new "foundation model" (a super-smart AI base) designed to understand both the internal thoughts of cells and their neighborhood context all at once. Here is how it works, using simple analogies:

1. The Two-Level City Map (Hierarchical Graphs)

Most models look at a city in just one way. HEIST looks at it like a Russian nesting doll or a two-story building:

The Top Floor (The Neighborhood): This is a map of the whole tissue. It connects cells based on who is standing next to whom. It knows that a cell in a "tumor neighborhood" is different from one in a "healthy neighborhood."
The Bottom Floor (The Internal Office): Inside every single cell, HEIST builds a mini-network of that cell's genes. It doesn't just list the genes; it connects them based on who talks to whom (co-expression).

The Magic Trick: HEIST doesn't just look at the top floor or the bottom floor. It lets them talk to each other.

If a cell is standing in a noisy, crowded neighborhood (a specific tissue environment), HEIST updates that cell's internal "thoughts" to reflect that stress.
Conversely, if the genes inside a cell are screaming a specific message, HEIST uses that to update how it sees the cell's place in the neighborhood.

2. No Fixed Dictionary (Generalization)

Imagine you trained a translator to speak English, but then you handed it a book written in French. A standard model would crash because it only knows the "English vocabulary."

HEIST is different. Instead of memorizing a fixed list of words (genes), it learns how words relate to each other.

If it sees a new protein it has never seen before (like in proteomics data), it doesn't panic. It looks at how that new protein is "talking" to its neighbors and infers its meaning based on those relationships.
This allows HEIST to jump from studying RNA (transcriptomics) to studying Proteins (proteomics) without needing to be retrained from scratch. It's like a polyglot who can learn a new language just by understanding the grammar and sentence structure, even if they've never heard the words before.

3. How It Learned (The Training)

HEIST was trained on a massive "city" dataset containing 22.3 million cells from 15 different organs and 124 different tissues.

The Teacher's Game: To learn, the AI played two games:
1. The "Spot the Difference" Game (Contrastive Learning): It was shown two cells and asked, "Are these neighbors similar or different?" It learned to group similar cells together and push different ones apart.
2. The "Fill in the Blanks" Game (Masked Autoencoding): The AI was shown a cell with some of its genes or location data hidden (masked). It had to guess the missing pieces based on what it knew about the cell's neighbors and internal gene network.

4. What HEIST Can Do (The Results)

The paper claims HEIST is a "superpower" for biology in four specific areas:

Finding Hidden Clusters: It can find tiny, hidden sub-groups of cells that other models miss. For example, it can spot a specific type of immune cell that only exists in a very specific "micro-environment" within a tumor, which previous models treated as just "generic immune cells."
Predicting Health Outcomes: It can look at a tissue sample and predict if a cancer patient will respond to immunotherapy or if a placenta is healthy, often beating other top models.
Filling in the Gaps: If a gene measurement is missing or noisy (a common problem in labs), HEIST can accurately guess what it should be.
Speed: It is incredibly fast—8 times faster than some competitors and 48 times faster than others—because it uses a smart, sparse way of looking at the data rather than trying to connect every single cell to every other single cell.

Summary

Think of HEIST as the ultimate biological detective. It doesn't just read the ID cards of the cells (genes); it looks at the crime scene (the tissue map), understands the relationships between the suspects (gene networks), and uses that context to solve mysteries that other detectives (models) were too blind to see. It is the first model to successfully combine the "who" (genes), the "where" (space), and the "how they talk" (networks) into one unified, adaptable system.

Technical Summary: HEIST - A Graph Foundation Model for Spatial Transcriptomics and Proteomics

Problem Statement

Single-cell transcriptomics and proteomics have enabled deep insights into cellular heterogeneity, yet traditional single-cell RNA sequencing (scRNA-seq) lacks the spatial context necessary to understand tissue organization, microenvironment interactions, and how gene co-expression influences tissue-level behaviors. While spatial transcriptomics and proteomics address this by preserving physical locations, existing computational models face significant limitations:

Ignoring Spatial or Molecular Complexity: Many models either ignore spatial information or fail to model the complex genetic and proteomic programs within cells, preventing the inference of how internal regulation adapts to microenvironmental cues.
Fixed Vocabularies: Most foundation models utilize fixed gene vocabularies, hindering their generalizability to datasets containing genes or proteins not present during pretraining.
Lack of Hierarchical Modeling: Existing approaches often treat cells and genes as flat structures or fully connected graphs, failing to capture the inductive bias of sparse, biologically plausible co-expression networks interacting with spatial neighborhoods.

Methodology: HEIST Architecture

The authors propose HEIST (Hierarchical EmbeddIngs for Spatial Transcriptomics), a hierarchical graph transformer foundation model designed to jointly model spatial proximity and internal gene co-expression networks.

1. Hierarchical Graph Construction

HEIST models tissues as a two-level hierarchical graph:

High-Level (Spatial Graph): A cell-cell graph ( $G_c$ ) constructed by computing Voronoi polygons from cell coordinates and connecting cells in adjacent polygons to capture spatial proximity and tissue architecture.
Low-Level (Gene Co-expression Networks): Each cell is represented by a lower-level gene co-expression network ( $G^k_g$ ). These networks are constructed per cell type by computing pairwise mutual information between denoised genes (using MAGIC) and connecting pairs above a data-driven, adaptive threshold $\tau$ .

2. Model Architecture

HEIST employs a Hierarchical Graph Transformer that performs two distinct types of message passing:

Intra-level Message Passing: Within each layer, the model updates cell embeddings using a CellGraphTransformer (capturing long-range spatial interactions) and gene embeddings using a GeneGraphTransformer (capturing fine-grained co-expression patterns).
Cross-level Message Passing: A novel directional attention mechanism integrates the two modalities.
- Gene-to-Cell: Cell embeddings are refined using pooled summaries of their constituent genes, ensuring transcriptional states shape spatial identity.
- Cell-to-Gene: Gene embeddings are updated based on the spatial context of their parent cell. This allows the same gene to be encoded differently depending on its cellular context, enabling the model to generalize to unseen genes or proteomic features without retraining.

3. Pre-training Objectives

HEIST is pre-trained on 22.3 million cells from 124 tissues across 15 organs using a combination of self-supervised objectives:

Contrastive Learning: Aligns similar cells and genes (e.g., neighboring cells of the same type, co-expressed genes) while pushing dissimilar pairs apart. It includes cross-level contrastive alignment to ensure consistency between gene and cell representations.
Masked Autoencoding (MAE): The model reconstructs masked gene expression values and spatial coordinates, forcing it to infer missing data and learn robust representations of gene signals and spatial cues.
Orthogonality Regularization: Encourages embedding dimensions to be decorrelated, promoting diverse and non-redundant representations.

Key Contributions

First Hierarchical Foundation Model for Spatial Omics: HEIST is the first model to explicitly incorporate co-expression networks alongside spatial graphs in a unified hierarchical framework, allowing local gene programs to influence tissue-level organization and vice versa.
Context-Aware, Dynamic Gene Embeddings: Unlike models relying on fixed vocabularies, HEIST computes gene embeddings dynamically from co-expression networks and cellular context. This enables generalization to novel datatypes, including spatial proteomics, without retraining.
Computational Efficiency: By utilizing sparse modeling and avoiding full self-attention over all genes, HEIST achieves significant speedups, reported as 8× faster than SCGPT-SPATIAL and 48× faster than SCFOUNDATION for embedding extraction.
Biological Inductive Bias: The architecture captures fine-grained gene co-expression within cells and long-range cellular interactions through cross-level message passing, producing biologically contextualized embeddings.

Experimental Results

HEIST was evaluated on four downstream tasks across multiple technologies (MERFISH, Xenium, CODEX, MIBI) and organs:

Cell Clustering: HEIST outperformed baselines (STAGATE, GraphST, CellPLM, scGPT-spatial) in Normalized Mutual Information (NMI) scores. Crucially, it successfully identified spatially-informed subpopulations that previous models collapsed, revealing microenvironment-driven heterogeneity (e.g., distinct L4-IT neuron subclusters).
Cell Type Annotation: Achieved state-of-the-art F1 scores in four out of five datasets, with substantial improvements (e.g., +48.5% on UPMC neck data) compared to baselines.
Clinical Outcome Prediction: HEIST achieved the highest AUC-ROC in six out of seven scenarios, including predicting immunotherapy response, remission, and placental conditions. Notably, it outperformed proteomics-specific baselines by up to 44.3% by leveraging all available markers without retraining.
Gene Imputation: After fine-tuning, HEIST surpassed all baselines (including MAGIC and fine-tuned foundation models) in Pearson correlation for gene imputation on placenta and skin datasets.
Ligand-Receptor Prediction: HEIST achieved an AUC-ROC of 0.995 in predicting ligand-receptor interactions, significantly outperforming other foundation models, indicating its embeddings capture spatial signatures of cell communication effectively.

Significance and Claims

The paper claims that HEIST represents a significant advancement in spatial omics modeling by unifying spatial and molecular hierarchies. Its primary significance lies in its ability to:

Discover Novel Biology: By capturing microenvironment-specific subpopulations missed by prior models, HEIST provides deeper insights into tissue organization and disease progression (e.g., Braak stage progression in Alzheimer's).
Enable Cross-Modal Transfer: The dynamic embedding mechanism allows the model to generalize to proteomics data and datasets with different gene sets, addressing a major limitation of current fixed-vocabulary foundation models.
Provide a Scalable Foundation: With state-of-the-art performance and high computational efficiency, HEIST offers a practical, general-purpose foundation for diverse downstream clinical and biological tasks, from cell annotation to predicting therapeutic responses.

The authors conclude that modeling both molecular (co-expression) and spatial information is essential for developing transferable models for spatial omics, a gap HEIST aims to fill.