Enabling interpretable machine learning for biological data with reliability scores.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Life Sciences

Background:

• Machine learning (ML) offers powerful tools for analyzing large biological datasets.
• However, ML models can be opaque, optimizing performance over biological insight and sometimes relying on biased data.
• Developing interpretable ML models is crucial for scientific rigor and discovery.

Purpose of the Study:

• To introduce the SWIF(r) reliability score (SRS) as a method to quantify the trustworthiness of ML classifications.
• To demonstrate the SRS's utility in addressing common ML challenges in biological data analysis.
• To promote the development of interpretable scientific ML.

Main Methods:

• The study describes the SWIF(r) reliability score (SRS), derived from the SWIF(r) generative framework.
• The SRS is evaluated on its ability to assess classification reliability for individual data instances.
• Applications are demonstrated using diverse biological datasets, including agricultural, UK Biobank, and population genetics data.

Main Results:

• The SRS effectively handles challenges like unknown classes in test data, training-testing data mismatch, and missing attribute values.
• It provides a measure of classification trustworthiness, aiding researchers in interrogating their data and models.
• The SRS shows comparable performance to outlier detection tools but can handle missing data.

Conclusions:

• The SRS enhances the interpretability and trustworthiness of machine learning models in biological research.
• It empowers researchers to integrate domain knowledge with ML, ensuring rigor and biological insight.
• The SRS represents a step towards more interpretable and reliable scientific machine learning.