Explainable rule-based prediction of cultivation media for microbes.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• Microbial growth preference knowledge is fragmented, hindering experimental design.
• Existing computational methods lack transparency and can use biased features.

Purpose of the Study:

• To develop an interpretable model for predicting microbial growth media preferences.
• To compare explainable methods with black-box models for microbial trait prediction.

Main Methods:

• Utilized the KG-Microbe knowledge graph for microbial traits.
• Developed a rule-based classifier and a black-box model.
• Applied SHAP and rule-based methods for feature importance analysis.

Main Results:

• The rule-based system provides transparent, biologically plausible rules for growth media prediction.
• Black-box models offered slightly higher predictive performance but lacked interpretability.
• Feature importance analysis revealed insights from both explainable and black-box approaches.

Conclusions:

• Explainable AI, particularly rule-based systems, offers a sustainable and insightful framework for microbial research.
• Integrating knowledge graphs, LLMs, and domain expertise can advance microbial discovery.