TL;DR
This paper introduces SAGE-FM, a novel foundation model for spatial transcriptomics that uses graph convolutional networks (GCNs) to learn biologically interpretable and spatially aware gene expression embeddings. SAGE-FM excels at predicting missing gene expression, preserving biological heterogeneity, and generalizing to downstream tasks like cancer subtype classification and pathologist-guided annotation, outperforming existing methods.
What Problem Does This Paper Address?
Spatial transcriptomics (ST) offers an unprecedented view of cellular organization by profiling gene expression while preserving spatial information. However, analyzing this high-dimensional, sparse, and noisy data requires sophisticated computational models. Current approaches often focus on reconstruction accuracy or lack biological interpretability, failing to fully leverage the crucial spatial context. There's a growing need for "foundation models" – like those in NLP and computer vision – that can learn generalizable, interpretable representations from ST data, enabling a wide range of downstream biological analyses. This paper aims to fill this gap by developing a lightweight, interpretable, and spatially aware foundation model for ST.
Methods
SAGE-FM is built upon a Graph Convolutional Network (GCN) architecture. Each spatial spot in an ST dataset is treated as a node, with its gene expression profile as node features. Edges between nodes represent the physical distance between spots. The model is trained using a masked gene prediction objective on local subgraphs (15 spots). During training, a percentage of genes in a central spot are randomly masked, and the GCN learns to predict these masked values using the unmasked genes in the central spot and the gene expression of its neighboring spots. This self-supervised approach allows the model to learn spatially conditioned gene-gene dependencies. The authors emphasize the use of a simple GCN architecture, arguing it offers parameter efficiency and interpretability without sacrificing biological fidelity, unlike more complex transformer-based models.
Datasets
The primary dataset used for pre-training SAGE-FM is the HEST1k dataset, a large-scale collection of human spatial transcriptomics samples from 15 organs, specifically focusing on 10x Visium data. This dataset comprises 416 samples, which were partitioned into training (80%), validation (10%), and test (10%) sets.
For downstream evaluation, the authors utilized:
- Head and Neck Oral Squamous Cell Carcinoma (OSCC) Visium dataset: This dataset features pathologist-defined annotations, serving as independent biological ground truth for evaluating spot pathology annotation.
- Glioblastoma (GBM) spatial transcriptomics datasets: Curated from three public resources, these datasets contain deconvolution-based estimates of dominant cell types, used for subtype identification.
Key Insights and Ideas
- Simplicity is Powerful: The paper demonstrates that a straightforward GCN architecture, when combined with a spatially aware masked gene prediction objective, can achieve state-of-the-art performance and biological interpretability, challenging the notion that only complex models can capture intricate biological patterns.
- Spatial Context is Key: By explicitly incorporating spatial neighborhood information into the GCN, SAGE-FM learns embeddings that are not just gene-centric but also spatially contextualized, leading to better preservation of biological heterogeneity and more accurate downstream predictions.
- Interpretability Matters: The authors go beyond standard performance metrics to rigorously evaluate the biological interpretability of their model's embeddings, showing how they align with tissue structures and clinical annotations.
- Foundation Models for ST: SAGE-FM establishes a blueprint for developing foundation models in spatial transcriptomics, showcasing how self-supervised learning on large-scale ST data can yield versatile representations for diverse biological tasks.
Key Results
- Gene Prediction Accuracy: SAGE-FM robustly recovers masked genes, with 91% showing significant correlations (p < 0.05) in the validation set.
- Clustering Performance: GNN embeddings from SAGE-FM significantly outperform MOFA and raw ST data in unsupervised spot clustering, achieving higher Adjusted Rand Index (ARI) and lower Davies-Bouldin Index (DBI).
- Biological Heterogeneity Preservation: SAGE-FM embeddings better preserve global inter-sample and tissue geometry compared to MOFA.
- Pathology Annotation: In OSCC, models trained on SAGE-FM embeddings achieved significantly superior pathologist annotation classification performance across multiple metrics, outperforming MOFA and raw ST features.
- Glioblastoma Subtyping: SAGE-FM embeddings enabled accurate GBM subtype prediction, achieving accuracy close to raw ST features and substantially outperforming MOFA.
- In Silico Perturbation: The model accurately captures directional ligand-receptor and upstream-downstream regulatory effects, demonstrating its ability to learn gene regulatory networks.
Limitations and Future Work
- Platform Specificity: The current framework is based on 10x Visium data. Future work could explore its scalability to higher-resolution and multimodal spatial platforms (e.g., Stereo-seq, CosMx, Xenium).
- Data Modalities: Integrating additional data modalities like proteomics, histopathology, or chromatin accessibility could enhance the model's ability to capture comprehensive spatial gene-protein interactions.
- Architecture Exploration: While a simple GCN proved effective, future research could systematically compare more advanced architectures like graph transformers or heterophilic GNNs to assess their added value.
- Cross-Tissue Generalization: Expanding benchmarking across diverse tissues, species, and disease contexts is crucial for evaluating model generalizability and identifying universal spatial principles.
- Clinical Integration: Moving spatial foundation models towards translational and diagnostic applications will require integration into spatially resolved clinical pipelines.
Summary updated