XModNN: Explainable Modular Neural Network to Identify Clinical Parameters and Disease Biomarkers in Transcriptomic...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Transcriptomic data analysis is crucial for disease classification and biomarker discovery.
• Standard methods often require large datasets and can yield numerous candidate biomarkers.
• Integrating biological knowledge into machine learning models can enhance efficiency and interpretability.

Purpose of the Study:

• To introduce the Explainable Modular Neural Network (XModNN) for identifying biomarkers in transcriptomic datasets.
• To demonstrate XModNN's ability to classify diseases and clinical parameters efficiently.
• To provide robust post hoc explanations for model predictions.

Main Methods:

• Developed XModNN with modules representing functional hierarchies (pathways/genes).
• Incorporated biological insights into the neural network architecture to reduce parameters.
• Employed weighted multi-loss progressive training and layer-wise relevance propagation for explanation.
• Applied XModNN to predict sex and neuroblastoma cell states.

Main Results:

• XModNN identified fewer candidate biomarkers compared to standard statistical approaches.
• Achieved comparable performance and robustness to Support Vector Machines and Random Forests with limited data.
• Integrated pathway relevance analysis improved upon standard gene set overrepresentation analysis.
• Identified key genes and pathways for sex classification and neuroblastoma cell state discrimination.

Conclusions:

• XModNN offers an effective and explainable method for biomarker identification and classification in transcriptomic data.
• The architecture's biological integration and training strategy enhance efficiency and reduce data requirements.
• XModNN provides valuable insights into biological pathways and gene contributions for disease states.