Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SEPAL: Spatial Gene Expression Prediction from Local Graphs

Gabriel Mejia, Paula Cárdenas, Daniela Ruiz, Angela Castillo, Pablo Arbeláez

TL;DR

SEPAL addresses the challenge of predicting spatial gene expression from histology by introducing a delta-based supervision scheme and a two-stage local-graph framework. A local image encoder first predicts gene expression deltas relative to the training mean, which is then refined by a graph neural network operating on a neighborhood graph to produce spatial corrections; the final expression is the delta plus the training mean. The method leverages $\ar{y}_{train}$, $\Delta y = y - \\bar{y}_{train}$, and a graph-based aggregation to capture local spatial dependencies, with genes selected via Moran's I and a robust denoising/benchmarking pipeline. Across two breast cancer datasets (Visium and STNet), SEPAL achieves state-of-the-art performance on multiple metrics, demonstrating the value of integrating local spatial context for spatial transcriptomics and enabling more accurate interpretation of tissue morphology and molecular profiles.

Abstract

Spatial transcriptomics is an emerging technology that aligns histopathology images with spatially resolved gene expression profiling. It holds the potential for understanding many diseases but faces significant bottlenecks such as specialized equipment and domain expertise. In this work, we present SEPAL, a new model for predicting genetic profiles from visual tissue appearance. Our method exploits the biological biases of the problem by directly supervising relative differences with respect to mean expression, and leverages local visual context at every coordinate to make predictions using a graph neural network. This approach closes the gap between complete locality and complete globality in current methods. In addition, we propose a novel benchmark that aims to better define the task by following current best practices in transcriptomics and restricting the prediction variables to only those with clear spatial patterns. Our extensive evaluation in two different human breast cancer datasets indicates that SEPAL outperforms previous state-of-the-art methods and other mechanisms of including spatial context.

SEPAL: Spatial Gene Expression Prediction from Local Graphs

TL;DR

SEPAL addresses the challenge of predicting spatial gene expression from histology by introducing a delta-based supervision scheme and a two-stage local-graph framework. A local image encoder first predicts gene expression deltas relative to the training mean, which is then refined by a graph neural network operating on a neighborhood graph to produce spatial corrections; the final expression is the delta plus the training mean. The method leverages , , and a graph-based aggregation to capture local spatial dependencies, with genes selected via Moran's I and a robust denoising/benchmarking pipeline. Across two breast cancer datasets (Visium and STNet), SEPAL achieves state-of-the-art performance on multiple metrics, demonstrating the value of integrating local spatial context for spatial transcriptomics and enabling more accurate interpretation of tissue morphology and molecular profiles.

Abstract

Spatial transcriptomics is an emerging technology that aligns histopathology images with spatially resolved gene expression profiling. It holds the potential for understanding many diseases but faces significant bottlenecks such as specialized equipment and domain expertise. In this work, we present SEPAL, a new model for predicting genetic profiles from visual tissue appearance. Our method exploits the biological biases of the problem by directly supervising relative differences with respect to mean expression, and leverages local visual context at every coordinate to make predictions using a graph neural network. This approach closes the gap between complete locality and complete globality in current methods. In addition, we propose a novel benchmark that aims to better define the task by following current best practices in transcriptomics and restricting the prediction variables to only those with clear spatial patterns. Our extensive evaluation in two different human breast cancer datasets indicates that SEPAL outperforms previous state-of-the-art methods and other mechanisms of including spatial context.
Paper Structure (22 sections, 5 equations, 5 figures, 3 tables)

This paper contains 22 sections, 5 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Global Methods
  4. Local Methods
  5. SEPAL
  6. Problem Formulation
  7. Architecture Overview
  8. Graph construction
  9. Spatial Learning Module
  10. Experiments
  11. Datasets
  12. Benchmark
  13. Evaluation Metrics
  14. State-of-the-art Methods
  15. Architecture Optimization
  16. ...and 7 more sections

Figures (5)

  • Figure 1: Different approaches for predicting gene expression from tissue images. The inputs of each model type are enclosed by a black frame. (A) Global methods analyze a whole slide image and make a prediction about the tissue expression in every spot at once. (B) Local methods process the image by patches and predict the expression of each individual patch, one at a time. (C) SEPAL uses graphs that contain information from multiple patches to represent spatial information and predict gene expression for the central node of each graph.
  • Figure 2: (A)First stage of our proposal. Pretraining of the Image Encoder $I(\cdot)$ and a linear layer $L(\cdot)$ to output the Image Embedding $(I_{\text{emb}})$ of a patch $X$, along with a preliminar prediction $\Delta \hat{y}_{i}$ of the difference between the expression in the patch and the mean expression in the train dataset. (B) The Graph Construction process begins with an image patch of interest and its spatial neighbors to build the graph representation based on the patch embeddings returned by the frozen $I(\cdot)$ and the positional encoding of each neighbor. (C) Architecture of the spatial learning module, which receives as input a Spatial Graph of the patch neighborhood and applies a GNN to predict the spatial correction $\hat{s}$ that further improves the $\Delta \hat{y}_{i}$ to get the $\Delta \hat{y}$ associated to the center patch of the graph and obtain the final gene expression prediction $\hat{y}$.
  • Figure 3: Example of the pepper denoising for a specific gene map in the Visium dataset.
  • Figure 4: Visualization of the two genes with the highest (left) and lowest (right) Pearson Correlation Coefficient. At the top is the Ground-Truth of the expression and at the bottom is the qualitative prediction of our method with its respective PCC.
  • Figure 5: Histogram of the Pearson correlation between the ground-truth and the prediction of each gene. The X-axis displays the values of the Pearson correlation coefficient, while the Y-axis shows the number of genes that have that particular correlation.