Interpretable genotype-to-phenotype classifiers with performance guarantees.

Area of Science:

• Genomics
• Machine Learning
• Precision Medicine

Background:

• Genotype-phenotype prediction is crucial for precision medicine but faces challenges with high-dimensional data, hindering generalization and interpretability of machine learning models.
• Existing machine learning algorithms struggle with scalability and producing easily understandable models for complex genomic data.

Purpose of the Study:

• To develop highly interpretable rule-based learning algorithms for genotype-to-phenotype prediction with strong performance guarantees.
• To address the limitations of current machine learning approaches in terms of generalization, scalability, and model interpretability.

Main Methods:

• Utilized sample compression theory to establish performance guarantees for rule-based learning algorithms.
• Developed and validated an open-source, disk-based implementation for efficient memory and computation.

Main Results:

• Achieved highly accurate genomic prediction of antimicrobial resistance across 12 species and 56 antibiotics.
• The interpretable models revealed known and potentially novel antimicrobial resistance mechanisms.

Conclusions:

• The proposed approach enhances genotype-to-phenotype prediction by providing accurate, interpretable, and scalable machine learning models.
• This work facilitates the application of machine learning in precision medicine, particularly for public health concerns like antimicrobial resistance.