Assessment and Optimization of Explainable Machine Learning Models Applied to Transcriptomic Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Explainable artificial intelligence (AI) methods are crucial for understanding machine learning (ML) model decisions.
• Existing AI explainability tools are primarily developed for computer vision, with limited application to biological data.
• Interpreting ML models in biology, particularly from transcriptomic data, remains a challenge.

Purpose of the Study:

• To comprehensively evaluate the applicability of various AI model explainers to biological data, specifically transcriptomic data.
• To identify top contributing genes impacting tissue type prediction from transcriptomic data using different AI models.
• To propose optimization strategies for enhancing the reproducibility and interpretability of AI explainer results in biological contexts.

Main Methods:

• Interpreted pre-trained ML models (MLP and CNN) for tissue type prediction using transcriptomic data.
• Applied and optimized multiple model explainers to assess their performance and reliability.
• Identified top contributing genes for model predictions across different tissue samples.

Main Results:

• Identified three groups of explainer and model architecture combinations demonstrating high reproducibility.
• Group II, utilizing aggregated MLP models with specific explainers, pinpointed top contributing genes with tissue-specific expression patterns.
• These identified genes showed potential as biomarkers for different tissues and in cancer research.

Conclusions:

• The study provides a systematic evaluation of AI explainability methods for transcriptomic data analysis.
• Optimization strategies improve the reproducibility and interpretability of AI-driven biological insights.
• The findings offer valuable guidance for leveraging explainable AI to uncover biological mechanisms and identify potential biomarkers.