Pathway-guided architectures for interpretable AI in biological research.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Deep learning (DL) models excel at analyzing complex biological multi-omics data.
• The "black box" nature of DL hinders biological and clinical translation.
• Knowledge-guided DL, specifically Pathway-Guided Interpretable Deep Learning Architectures (PGI-DLA), addresses this by integrating prior pathway knowledge.

Purpose of the Study:

• To review the advancements in PGI-DLA for multi-omics data analysis.
• To compare major pathway databases (KEGG, GO, Reactome, MSigDB) for their suitability in PGI-DLA.
• To provide guidance on selecting and optimizing pathway databases for PGI-DLA implementation.

Main Methods:

• Review of current literature on PGI-DLA methods.
• Comparative analysis of pathway databases based on knowledge scope, structure, detail, and curation.
• Discussion on the impact of database choice on model design, performance, and interpretability.

Main Results:

• PGI-DLA enhances DL model performance and interpretability by incorporating biological pathway information.
• Significant differences exist among pathway databases, influencing their application in PGI-DLA.
• The selection of an appropriate pathway database is crucial for successful PGI-DLA implementation.

Conclusions:

• PGI-DLA offers a promising approach for translating multi-omics data into biological and clinical insights.
• Careful consideration and optimization of pathway databases are essential for maximizing the benefits of PGI-DLA.
• This review serves as a valuable resource for researchers applying PGI-DLA to complex diseases.