Sparsity is all you need: rethinking biologically informed neural networks

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Biologically informed neural networks (BINNs) integrate biological knowledge into AI models for improved accuracy and interpretability.
• Pathway-informed models are a subset of BINNs aiming to decode complex biological systems.
• The true contribution of biological knowledge versus architectural biases in BINNs remains underexplored.

Purpose of the Study:

• To systematically evaluate whether biological pathway information enhances neural network performance or if other factors are responsible.
• To decouple the effects of biological annotations from network architecture in BINNs.
• To provide a framework for assessing the impact of biological priors in AI models.

Main Methods:

• A structure-matched randomization protocol was developed to compare biologically informed models with randomized counterparts.
• Pathway annotations were replaced with random associations while maintaining architectural constraints and sparsity.
• The methodology was applied across 29 state-of-the-art pathway-informed neural networks and various prediction tasks.

Main Results:

• Randomized models consistently matched or surpassed the performance of their biologically informed counterparts across diverse tasks and datasets.
• No significant advantage in interpretability was observed; randomized models identified biomarkers comparably.
• Feature rankings from randomized and informed models were highly correlated, suggesting similar underlying mechanisms.
• Performance gains were predominantly attributed to sparsity-induced regularization, not biological pathway integration.

Conclusions:

• The predictive power of pathway-informed neural networks largely derives from sparsity regularization inherent in their architecture, rather than explicit biological knowledge.
• Biological pathway integration does not inherently improve model performance or interpretability compared to well-regularized, randomized architectures.
• A generalizable evaluation workflow is proposed to discern the true contribution of biological priors beyond sparsity in AI model development.