Evaluating recommender systems for AI-driven biomedical informatics.

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence

Background:

• Many domain experts lack machine learning (ML) and coding skills, hindering bioinformatics data analysis.
• Existing automated ML methods often require programming expertise and deep knowledge for algorithm tuning.

Purpose of the Study:

• To develop a web-based AI platform for automating biomedical data science.
• To enable easy construction of sophisticated models for biological processes.
• To provide an AI agent that recommends and conducts experiments based on user data and prior knowledge.

Main Methods:

• A novel AI framework was developed for automated ML in bioinformatics.
• The framework was validated on 165 classification problems, comparing against state-of-the-art automated approaches.
• Matrix factorization-based recommendation systems were employed for model selection.

Main Results:

• The AI platform demonstrated competitiveness with state-of-the-art automated ML methods.
• Matrix factorization outperformed meta-learning approaches in automating ML.
• An AI-driven model for septic shock prediction achieved an AUROC of 0.85±0.02, comparable to deep learning models with reduced computational cost.

Conclusions:

• The developed AI platform effectively automates ML for bioinformatics, lowering the barrier for domain experts.
• The AI system can identify optimal algorithm configurations and generate high-performing predictive models.
• The approach offers a computationally efficient alternative for complex biomedical data analysis and prediction tasks.