Graph Structured Neural Networks for Perturbation Biology.

Area of Science:

• Computational biology
• Systems biology
• Pharmacology

Background:

• Traditional deep learning models struggle to capture the sequential nature of molecular interactions in perturbation biology.
• Accurate modeling of molecular mechanisms is crucial for advancing precision medicine and drug discovery.

Approach:

• Introduced Graph Structured Neural Networks (GSNN), a novel deep learning architecture incorporating cell signaling pathways as inductive biases.
• Applied GSNN to the LINCS L1000 dataset and curated molecular interaction data for perturbation biology tasks.

Key Points:

• GSNNs demonstrated superior performance over baseline algorithms in predicting perturbed gene expression and cell viability for drug combinations.
• The method also excelled in disease-specific drug prioritization, highlighting its potential in drug repurposing.
• Developed GSNNExplainer for biologically interpretable explanations of model predictions.

Conclusions:

• GSNNs offer a more trustworthy and mechanistically informed approach to modeling biological systems compared to generic deep learning methods.
• This work provides a foundation for developing reliable computational models for drug response prediction, potentially aiding clinical decision-making.