Simulation and empirical evaluation of biologically-informed neural network performance.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Biologically-informed neural networks (BiNNs) provide interpretable deep learning for biological data.
• Understanding dataset requirements for optimal BiNN performance is crucial.
• Previous models like P-NET predict prostate cancer metastasis using somatic data.

Purpose of the Study:

• To develop simulation frameworks for evaluating factors influencing BiNN performance.
• To assess the impact of integrating germline and somatic data on predicting prostate cancer metastatic status.
• To provide a principled approach for benchmarking BiNNs and understanding their data dependencies.

Main Methods:

• Developed two simulation frameworks to test BiNN performance under varying conditions (signal type, strength, sparsity, sample size).
• Empirically integrated germline and somatic data into the P-NET model.
• Evaluated model prediction accuracy, gene prioritization, and interpretability.

Main Results:

• BiNN performance is limited by small sample size, weak signal strength, and high feature sparsity.
• BiNNs preferentially utilize linear over nonlinear signals.
• P-NET showed poor performance on sparse germline data; integrating germline data did not improve prediction but enhanced interpretation.

Conclusions:

• Simulation frameworks enable systematic evaluation of dataset characteristics affecting BiNNs.
• Dataset properties significantly influence the success of biologically-informed neural networks.
• Further research is needed to optimize BiNNs for complex biological data, particularly sparse genomic information.