Semi-supervised meta-learning elucidates understudied molecular interactions.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Biological research is often limited by experimental constraints and human biases.
• Deep learning struggles with scarce labeled data and distribution shifts, hindering scientific discovery.
• Existing transfer learning methods are insufficient for out-of-distribution (OOD) data challenges.

Purpose of the Study:

• To develop a novel deep learning framework, MMAPLE, to address challenges in understudied biological problems.
• To effectively explore out-of-distribution (OOD) unlabeled data where conventional methods fail.
• To integrate meta-learning, transfer learning, and semi-supervised learning for enhanced biological data analysis.

Main Methods:

• Developed the Meta Model Agnostic Pseudo Label Learning (MMAPLE) framework.
• Integrated meta-learning, transfer learning, and semi-supervised learning into a unified approach.
• Applied MMAPLE to predict drug-target interactions, human metabolite-enzyme interactions, and microbiome metabolite-human receptor interactions in OOD settings.

Main Results:

• MMAPLE demonstrated significant improvements (11% to 242%) in prediction-recall across multiple OOD benchmarks.
• Achieved superior performance over various base models on challenging OOD datasets.
• Identified novel interspecies metabolite-protein interactions, validated by activity assays.

Conclusions:

• MMAPLE is a powerful and generalizable framework for exploring previously unrecognized biological domains.
• The framework effectively overcomes limitations of scarce data and distribution shifts in biological data analysis.
• MMAPLE facilitates the discovery of crucial biological interactions, including those in microbiome-human interactions.