Abstract
Inferring Gene Regulatory Networks (GRNs) is essential for understanding developmental biology. While traditional methods rely on bulk RNA-sequencing, they often overlook the spatial context of gene expression. We present MIGGRI, a two-stage deep learning framework that leverages spatial expression images to reconstruct GRNs. By integrating contrastive learning with a multi-instance graph neural network (GNN), MIGGRI captures complex spatio-temporal patterns and outperforms existing models, even in the presence of missing data.
Introduction
Gene Regulatory Networks (GRNs) act as the biological circuitry governing cellular functions and development. Traditional inference methods primarily utilize transcriptomic data, which provides an averaged view of gene expression but lacks spatial resolution. Spatial transcriptomics, visualized through in situ hybridization (ISH) images, offers a richer perspective. However, utilizing these images for GRN inference presents challenges: the multi-instance nature of gene views, the lack of extensive supervised labels, and the need to account for the interconnected topology of genetic interactions.
Methodology: The MIGGRI Framework
MIGGRI (Multi-Instance Graph neural network for GRN Reconstruction and Inference) addresses these challenges through a two-stage architecture.
Stage I: Contrastive Feature Extraction
The first stage employs a Siamese CNN (based on VGG-16) to process pairs of spatial expression images. Using a contrastive loss function, the model learns to map genes with known interactions closer together in a latent feature space while separating non-interacting pairs. This self-supervised approach allows the model to identify regulatory features without requiring pixel-level annotations.
Stage II: Multi-Instance GNN
To handle multiple views per gene (e.g., dorsal, lateral, ventral), MIGGRI utilizes a shuffled LSTM-based aggregator to fuse various image instances into a single representative vector.
This vector is then processed by a GraphSAGE backbone. Unlike standard pairwise models, the GNN incorporates the broader network topology, learning from both a gene's own expression pattern and those of its neighbors in the regulatory graph.
Key Insights and Results
Robustness to Missing Data
MIGGRI demonstrates remarkable robustness. In experiments where 10% of gene image data was replaced with noise, the model maintained high accuracy by leveraging the surrounding network topology—effectively "borrowing" information from neighboring genes.
Generative Latent Space
The learned embeddings are biologically informative enough to allow for image reconstruction. A decoder trained on these embeddings can recreate the original expression patterns, preserving essential anatomical features.
Performance Metrics
MIGGRI achieved an accuracy of 82.9% on the eye development dataset, significantly outperforming the previous state-of-the-art model, ConGRI (76.7%). Furthermore, it reached an Area Under the Curve (AUC) of 0.88, demonstrating the power of combining spatial image data with graph-based topological information.
Conclusion
MIGGRI represents a significant advancement in spatial GRN inference. By treating gene expression as a multi-instance graph problem, it captures the nuances of developmental regulation more effectively than previous methods. Future work will focus on adapting this framework to high-resolution platforms like Visium and MERFISH to further refine our understanding of the genetic blueprint.
Summary updated