OmiXAI: An ensemble XAI pipeline for interpretable deep learning in omics data.

Area of Science:

• Genomics
• Artificial Intelligence
• Bioinformatics

Background:

• Deep learning models excel in genomic data analysis but require interpretable feature importance.
• Leveraging multi-omics data enhances deep learning performance, necessitating methods to understand predictive factors.
• Existing explainable AI (XAI) methods are often computationally intensive for complex deep learning models.

Purpose of the Study:

• To introduce OmiXAI, an integrated pipeline of ensemble model-aware XAI methods for deep learning models trained on omics data.
• To provide a framework for identifying critical features contributing to the predictive power of deep learning models in genomics.
• To benchmark and assess the efficacy of various XAI techniques within a multi-omics context.

Main Methods:

• OmiXAI integrates gradient-based XAI techniques (Integrated Gradients, InputXGradients, Guided Backpropagation, Deconvolution) and GNN-specific methods (Saliency Maps, GNNExplainer).
• The framework is designed for deep learning models trained on explicit omics feature matrices.
• Evaluation involved a case study on functional genomic element prediction using epigenomic features.

Main Results:

• OmiXAI successfully identified important features, enabling significant feature engineering by reducing the critical feature set from nearly 2,000 to 50.
• Benchmarking demonstrated the efficacy of the ensemble approach and highlighted limitations of individual XAI methods.
• The pipeline's modular design facilitates integration of new attribution methods and applicability beyond omics.

Conclusions:

• OmiXAI offers an efficient solution for interpreting deep learning models in multi-omics data analysis.
• The framework aids in understanding feature importance and facilitates feature engineering for improved model performance.
• OmiXAI is a versatile and adaptable tool for explainable AI in diverse scientific domains.