Robust identification of molecular phenotypes using semi-supervised learning.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in healthcare

Background:

• Modern molecular profiling generates extensive patient data for machine learning applications.
• Unsupervised methods identify disease subtypes but ignore clinical outcomes.
• Supervised methods struggle with unclear treatment benefits and noisy training data.

Purpose of the Study:

• To develop an iterative classification method for identifying molecular phenotypes associated with clinical outcomes.
• To address challenges in defining training classes for supervised learning when outcome data is nuanced.
• To integrate clinical data, including time-to-event endpoints, into a self-consistent refinement process.

Main Methods:

• An iterative binary classification approach refining training labels and classifiers simultaneously.
• Incorporation of clinical data, such as time-to-event endpoints, for outcome association.
• Validation using synthetic and real-world genomic datasets to assess performance and generalization.

Main Results:

• The iterative method accurately identifies outcome-related molecular phenotypes and attributes.
• Demonstrated improved accuracy and generalization on both synthetic and real-world genomic data.
• Convergence of the iterative process consistently integrates molecular data into classifiers.

Conclusions:

• The iterative approach minimizes overfitting by incorporating molecular data structure.
• Facilitates robust generalization of classification and molecular phenotypes.
• Enables reliable identification of biologically relevant features and underlying biological processes.