Mapping the Genome’s Geography: A New Foundation for Spatial Transcriptomics

by

Author: Denis Avetisyan

Researchers have developed a lightweight model to better understand gene expression within the physical context of tissues, paving the way for more accurate biological insights.

SAGE-FM is a graph convolutional network-based foundation model for spatial transcriptomics that learns interpretable gene embeddings and excels at downstream tasks like spot annotation and in silico perturbation.

Capturing spatially-conditioned gene regulatory relationships remains a challenge in translating spatial transcriptomics data into biological insight. Here, we introduce SAGE-FM: A lightweight and interpretable spatial transcriptomics foundation model, a graph convolutional network trained with a masked central spot prediction objective to learn robust, spatially coherent gene embeddings. These embeddings outperform existing methods in unsupervised clustering and demonstrate strong generalization to downstream tasks, including accurate spot annotation and biologically consistent in silico perturbation experiments. Can this parameter-efficient approach unlock a new era of interpretable foundation models for large-scale spatial omics data?

The Erosion of Averaged Data: Why Location Matters

Historically, the study of gene expression has largely treated tissues as homogenous mixtures, averaging the activity of cells across an entire sample-a practice that obscures critical biological details. This approach overlooks the fundamental reality that cells within a tissue are not uniformly distributed; their functions and gene expression patterns are deeply influenced by their precise location and interactions with neighboring cells. Consequently, traditional methods often fail to capture the nuanced spatial organization that governs processes like embryonic development, immune responses, and tumor progression. This simplification can lead to inaccurate models of biological systems and hinder the development of effective therapies, as treatments designed without considering spatial context may target the wrong cells or fail to account for localized variations in disease mechanisms. A growing body of research demonstrates that incorporating spatial information is not merely a refinement of existing techniques, but a necessary paradigm shift for a more complete understanding of life at the molecular level.

The precise location of a cell within a tissue profoundly influences its gene expression, revealing that a gene’s activity isn’t solely dictated by its sequence but also by its neighborhood. This spatial context is increasingly recognized as fundamental to understanding complex biological processes – from embryonic development and immune responses to tumor progression and neurological function. Variations in gene expression across even short distances can delineate distinct cellular states and functional roles, creating a nuanced landscape previously obscured by bulk tissue analysis. Consequently, pinpointing how location-specific gene expression contributes to cellular behavior is proving essential for deciphering disease mechanisms, as malfunctions in spatial regulation can disrupt normal tissue architecture and contribute to pathological conditions. This realization is driving the development of innovative technologies aimed at mapping gene expression with unprecedented spatial resolution, promising a more complete and accurate picture of biological systems.

Current computational techniques for analyzing gene expression, such as Multi-Omics Factor Analysis (MOFA), often treat tissues as homogenous mixtures, thereby losing vital information encoded by cellular architecture. While MOFA excels at deconvoluting complex datasets and identifying underlying factors driving gene expression, it traditionally lacks the capacity to explicitly incorporate where genes are expressed within the tissue itself. This limitation hinders a complete understanding of biological processes; genes expressed in different anatomical locations can have drastically different functions, even within the same cell type. Consequently, analyses relying solely on bulk or averaged data may obscure crucial spatial patterns and misinterpret the true regulatory landscape, demanding new computational strategies that prioritize and effectively integrate spatial transcriptomics data with multi-omic information.

SAGE-FM: Mapping the Transcriptional Landscape

SAGE-FM employs Graph Convolutional Networks (GCN) to represent spatial transcriptomics data as a graph, where genes are nodes and spatial proximity defines edges. This allows the model to directly incorporate positional information during analysis; each gene’s expression is contextualized by the expression levels of its neighboring genes. The GCN architecture enables the aggregation of feature information from spatially connected genes, effectively capturing location-specific expression patterns that traditional methods may overlook. By modeling these spatial relationships, SAGE-FM can infer gene expression based on its surrounding transcriptional environment, providing a more nuanced understanding of tissue architecture and gene regulation.

SAGE-FM is trained utilizing a Masked Spot Prediction (MSP) objective, a self-supervised learning approach where the model predicts gene expression values at a given spatial location, termed the “central spot.” This prediction is based solely on the expression data from neighboring spatial locations, effectively learning spatial dependencies between genes. During training, a proportion of spots are masked, and the model is tasked with reconstructing the expression profile of these masked locations. The loss function quantifies the difference between the predicted and actual gene expression levels in the masked spots, driving the model to learn robust representations of spatial gene expression patterns. This methodology avoids reliance on labeled data, enabling pre-training on large-scale datasets like HEST1k.

SAGE-FM is designed to process data acquired through Visium Spatial Transcriptomics technology, a platform that enables spatially-resolved transcriptomic profiling. Visium utilizes a spotted array to capture gene expression data while preserving the spatial context of the tissue sample. This methodology facilitates high-resolution analysis of both tissue architecture and corresponding gene expression patterns, allowing researchers to identify location-specific gene expression and understand how gene expression relates to tissue organization. The resulting data provides a detailed map of gene activity within the tissue, with each spot on the array representing the average gene expression from a defined area of the tissue section.

Pre-training of SAGE-FM utilized the HEST1k dataset, a large collection of spatial transcriptomics data, to enhance the model’s ability to generalize to diverse biological contexts. This pre-training process resulted in a 91% accuracy rate in predicting masked gene expression levels based on spatial correlations with neighboring genes. Statistical significance was established with a p-value of less than 0.05, indicating that the observed predictive performance is unlikely due to random chance. The extensive dataset and demonstrated accuracy suggest SAGE-FM can effectively infer gene expression patterns in novel tissues and experimental conditions without requiring task-specific training.

Validation & Application: From Prediction to Biological Insight

SAGE-FM demonstrates robust performance in Glioblastoma (GBM) subtype identification through the analysis of spatial transcriptomic data. The model accurately classifies distinct cancer subtypes by leveraging spatially resolved gene expression patterns, enabling a more granular understanding of tumor heterogeneity. This capability is achieved by integrating spatial information directly into the factorization process, allowing SAGE-FM to identify subtype-specific expression signatures that may be obscured in bulk RNA sequencing analyses. Validation studies demonstrate the model’s ability to consistently differentiate between established GBM subtypes, providing a reliable tool for both research and potential clinical applications in personalized medicine.

SAGE-FM demonstrates improved performance in predicting pathologist annotations of Oral Squamous Cell Carcinoma (OSCC) tissue samples compared to the MOFA model. This enhancement is quantified through Adjusted Rand Index (ARI) scores obtained from spot clustering analysis, indicating SAGE-FM’s superior ability to capture clinically relevant tissue features. Higher ARI values signify greater agreement between the model’s clustering and the expert pathologist’s annotations, suggesting that SAGE-FM effectively identifies and categorizes distinct cellular patterns within OSCC samples as defined by human expert review.

In Silico Perturbation is employed to model biological effects by simulating alterations in gene expression and assessing their impact on downstream regulatory networks and ligand-receptor interactions. This computational approach allows for the prediction of phenotypic changes resulting from specific gene expression modifications. Validation of these predictions is performed by comparing the simulated changes with observed changes in biological systems, with statistical tests used to determine the degree of concordance between predicted and empirical results. This methodology facilitates the investigation of gene regulatory mechanisms and potential therapeutic targets by enabling the assessment of causal relationships between gene expression and biological outcomes.

Expanding the Scope of Spatial Omics: A Future Perspective

SAGE-FM establishes a robust computational framework designed to overcome the limitations of analyzing single omics datasets in isolation. By effectively integrating spatial transcriptomics – which maps gene expression within tissue architecture – with other omics layers such as genomics, proteomics, and metabolomics, the model generates a significantly more comprehensive depiction of cellular processes. This integrative approach allows researchers to move beyond simply identifying genes or proteins associated with a disease, and instead, begin to unravel the complex interplay of molecular events occurring within specific spatial contexts. The ability to correlate gene expression patterns with protein abundance, metabolic activity, and genomic variations, all while accounting for cellular location, provides a holistic understanding crucial for deciphering disease mechanisms and ultimately, informing targeted therapeutic strategies.

The predictive power of SAGE-FM extends significantly into the realms of pharmaceutical development and tailored healthcare. By accurately forecasting gene expression and its regulatory influences, the model facilitates the in silico screening of potential drug candidates, reducing the need for extensive and costly laboratory experiments. This capability allows researchers to pinpoint compounds likely to elicit a desired therapeutic response with greater efficiency. Furthermore, SAGE-FM’s capacity to model individual cellular responses within a spatial context holds promise for personalized medicine; it could potentially predict how a patient will respond to a specific treatment based on their unique molecular profile and the spatial organization of their tissues. Consequently, treatment strategies can be optimized, maximizing efficacy and minimizing adverse effects, ultimately ushering in a new era of precision healthcare.

Current iterations of SAGE-FM offer a static snapshot of spatial gene expression, yet biological processes are inherently dynamic. Future development prioritizes extending the model to incorporate temporal data, allowing researchers to track how gene expression patterns evolve over time and in response to stimuli. This will involve integrating time-series spatial transcriptomics data and refining the model’s algorithms to accurately predict these dynamic changes. Crucially, this expanded capability will enable investigations into disease progression, capturing how spatial gene expression shifts during initiation, development, and potential therapeutic intervention – offering a more nuanced understanding than currently possible and facilitating the identification of novel biomarkers and therapeutic targets.

By applying SAGE-FM to the study of downstream gene regulation, researchers anticipate gaining crucial insights into the molecular events that govern disease progression. This analytical framework moves beyond simply identifying genes that are differentially expressed; it aims to map the cascading effects of regulatory changes, revealing how initial molecular disturbances propagate through cellular networks. Understanding these downstream consequences is particularly vital, as they often dictate the phenotypic characteristics of disease and represent potential therapeutic targets. The model’s predictive capacity, combined with spatial omics data, promises to delineate the precise regulatory pathways altered in disease states, allowing for the identification of key nodes susceptible to intervention and potentially leading to the development of more effective, targeted therapies.

The pursuit of foundational models, as exemplified by SAGE-FM, acknowledges an inherent truth about complex systems: their eventual drift from initial ideals. This model, leveraging graph convolutional networks to map spatial transcriptomics data, attempts to create a robust representation, yet even the most carefully constructed embedding will ultimately require refinement. As Ken Thompson observed, “Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.” This sentiment echoes the need for interpretable models like SAGE-FM; understanding the underlying mechanisms-the ‘code’-allows for graceful adaptation and correction as the system encounters the inevitable decay of real-world data and evolving biological understanding. Versioning, in this context, becomes a form of memory, preserving the lineage of the model’s evolution.

What Lies Ahead?

The introduction of SAGE-FM, while a step toward more coherent spatial transcriptomics analyses, merely clarifies the contours of what remains unknown. Every commit is a record in the annals, and every version a chapter – this model is not a destination, but a waypoint. The very act of creating ‘foundation’ models presupposes a stable substrate upon which to build, yet biological systems are defined by their inherent instability. The question isn’t whether these models will decay, but how gracefully they will do so as the data landscape shifts and the nuances of spatial biology are revealed.

Current limitations-the reliance on specific graph construction methods, the inherent challenges in extrapolating beyond the training data-are not bugs, but features of the system. They highlight the need to move beyond purely data-driven approaches and integrate prior biological knowledge more effectively. The true test will lie in the model’s ability to generate predictive insights, not merely reconstruct existing knowledge. Spot annotation and in silico perturbation are useful exercises, but they are ultimately retrospective.

Delaying fixes is a tax on ambition. The field must now confront the fundamental problem of scale – both in terms of data volume and biological complexity. Can these models truly capture the dynamic interplay of cells and their microenvironment, or will they remain static snapshots of a fleeting moment in time? Future work should focus on developing methods for incorporating temporal data and accounting for individual cellular heterogeneity. The goal isn’t to build a perfect model, but one that acknowledges its own imperfections and provides a framework for continual refinement.

Original article: https://arxiv.org/pdf/2601.15504.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-25 16:52