Introduction

GSNN Overview

What is GSNN?

Graph Structured Neural Networks (GSNN) is a novel approach that incorporates prior knowledge of latent variable interactions directly into neural network architecture. Unlike traditional neural networks that learn relationships from data alone, GSNN leverages domain-specific knowledge to guide the learning process, making it particularly powerful for applications in perturbation biology and other domains where prior knowledge about variable relationships is available.

The GSNN method was introduced in the paper “Graph Structured Neural Networks for Perturbation Biology” (Evans et al., 2024) and provides a framework for building interpretable, structured neural networks that respect known biological or domain relationships.

Key Concepts

Prior Knowledge Integration

GSNN allows you to encode domain-specific knowledge about how variables interact through a graph structure. This graph defines which variables can directly influence each other, constraining the neural network to learn only biologically or physically plausible relationships.

Three-Node Architecture

GSNN uses a specialized three-node type architecture:

  • Input nodes: Represent observed variables

  • Function nodes: Represent latent variables

  • Output nodes: Represent target variables

Sparse Connectivity

The connections between nodes are sparse and determined by your prior knowledge graph, leading to more interpretable models and better generalization, especially in data-limited scenarios.

Perturbation Biology Applications

GSNN is particularly well-suited for perturbation biology studies where you want to understand how interventions (perturbations) affect biological systems. The graph structure can encode known biological pathways, protein-protein interactions, or regulatory networks.

Why Use GSNN?

Interpretability

Unlike black-box neural networks, GSNN models are interpretable because the learned weights correspond to specific relationships in your prior knowledge graph. You can directly examine which connections are important for predictions.

Data Efficiency

By incorporating prior knowledge, GSNN can learn meaningful relationships from smaller datasets than would be required for traditional neural networks.

Domain Knowledge Constraints

GSNN ensures that your model respects known biological or physical constraints, preventing it from learning spurious correlations that violate domain knowledge.

Flexible Architecture

GSNN supports various normalization strategies and activation functions, and can be combined with advanced techniques like uncertainty quantification, reinforcement learning, and Bayesian optimization.

Core Features

Graph-Based Architecture

  • Define custom graph structures representing your domain knowledge

  • Automatic handling of sparse connectivity patterns

Training Options

  • Gradient checkpointing for memory efficiency

  • Multiple normalization strategies (Layer, Batch)

  • Residual connections to improve training stability

  • Support for various activation functions and weight initialization strategies

Optimization and Inference

  • Reinforcement learning for graph structure optimization

  • Uncertainty quantification through hypernetworks

Interpretation Tools

  • Model explanation and visualization (GSNNExplainer)

  • Entity function extraction

How Are GSNNs Different from Graph Neural Networks?

While both GSNNs and Graph Neural Networks (GNNs) use graphs, they serve fundamentally different purposes and operate in distinct ways:

GNNs: Learning From Graph Structure

Traditional GNNs treat the graph as data to learn from. They use permutation-invariant aggregation functions to learn local patterns and node representations, and can often generalize to new, unseen graphs.

GSNNs: Constraining With Graph Structure

GSNNs use the graph structure as a constraint mechanism rather than a learning target. The graph defines which variables can directly influence each other, applying inductive biases through feature constraints. GSNNs are transductive—they are trained on a single graph and cannot be applied to new graphs.

Example

This distinction is critical when choosing between approaches. In biological signaling, for example, similar local network structures may produce drastically different signaling patterns. While the network structure is useful for understanding causal interactions, the graph patterns themselves are not necessarily predictive of signaling behaviors. GSNNs leverage this domain knowledge to constrain the model architecture, while GNNs will attempt to map similar local graph structures to signaling patterns.

Getting Started

The GSNN library provides comprehensive tutorials to help you get started:

  • Basic Usage: Learn to build and train your first GSNN model

  • Simulation: Use Bayesian networks to generate synthetic data

  • Comparison: Compare GSNN performance against baseline methods

  • Advanced Features: Explore reinforcement learning, Bayesian optimization, and uncertainty quantification

Installation

Create the conda/mamba environment and install GSNN:

mamba env create -f environment.yml
conda activate gsnn
pip install -e .

Citation

If you use GSNN in your research, please cite:

@article{Evans2024.02.28.582164,
    author = {Nathaniel J. Evans and Gordon B. Mills and Guanming Wu and Xubo Song and Shannon McWeeney},
    title = {Graph Structured Neural Networks for Perturbation Biology},
    elocation-id = {2024.02.28.582164},
    year = {2024},
    doi = {10.1101/2024.02.28.582164},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2024/02/29/2024.02.28.582164},
    journal = {bioRxiv}
}

Next Steps