Unsupervised construction of computational graphs for gene expression data with explicit structural inductive biases.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• Gene expression data is crucial for understanding cancer but is high-dimensional and noisy.
• Traditional machine learning models often struggle with overfitting and capturing biological relevance in gene expression data.
• Integrating external biological knowledge, like protein-protein interaction networks, can improve model construction.

Purpose of the Study:

• To develop an unsupervised method for constructing computational graph models for gene expression data.
• To guide model construction using prior knowledge from gene interaction networks, specifically protein-protein interaction (PPI) networks.
• To enhance cancer phenotype prediction by incorporating biological network structures.

Main Methods:

• Gene Interaction Network Constrained Construction (GINCCo): an unsupervised method for automated computational graph model construction.
• Utilizing topological clustering algorithms on PPI networks to structurally build computational graphs.
• Representing graph entities as biological components (genes, protein complexes, phenotypes) for biologically relevant regularization.

Main Results:

• GINCCo successfully incorporates PPI networks for cancer phenotype prediction.
• The method significantly reduces model complexity and parameters compared to standard machine learning models.
• GINCCo demonstrates superior predictive performance over support vector machines and multi-layer perceptrons on various cancer phenotypes.

Conclusions:

• GINCCo offers a biologically relevant and effective approach for predictive modeling with gene expression data.
• The method enhances model interpretability and reduces overfitting by leveraging network biology.
• GINCCo provides a robust framework for cancer research, improving predictive accuracy and enabling targeted biological insights.