Biologically informed neural network models are robust to spurious interactions via self-pruning

Area of Science:

• Computational biology
• Systems biology
• Bioinformatics

Background:

• Computational models of cellular networks are crucial for understanding disease mechanisms and developing therapies.
• Biology-informed neural networks (BINNs) integrate deep learning with biological prior knowledge for enhanced model interpretability.
• Evaluating the reliability of BINNs in complex cellular systems remains a significant challenge.

Purpose of the Study:

• To develop and evaluate a holistic approach for assessing the reliability of BINNs.
• To quantify the self-pruning capability of BINNs in response to spurious interactions.
• To enhance the computational efficiency of existing BINN frameworks for large-scale analysis.

Main Methods:

• Introduced the Relative Residual Area (RRA) metric to measure the self-pruning extent of BINNs.
• Updated the LEMBAS (Large-scale knowledge-EMBedded Artificial Signaling-networks) framework with GPU acceleration for a >7-fold speedup.
• Evaluated BINN self-pruning performance across three datasets with varying levels of introduced spurious interactions.

Main Results:

• BINNs demonstrated significant self-pruning of randomly introduced spurious interactions compared to those from the prior knowledge network (PKN).
• The self-pruning effectiveness was enhanced when the BINN was regularized with a sufficiently large L2 norm.
• The updated LEMBAS framework achieved substantial speedup while maintaining predictive accuracy.
• The RRA metric effectively distinguished between perfect self-pruning and failures in pruning.

Conclusions:

• BINNs exhibit robustness to uncertainty within prior knowledge networks through effective self-pruning mechanisms.
• The LEMBAS-GPU framework provides an efficient tool for large-scale BINN analysis and reliability assessment.
• This work offers a novel method for evaluating the trustworthiness of computational models in systems biology.