Phenotypic subtyping via contrastive learning.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Disease Subtyping

Background:

• Subphenotypic structure is crucial for understanding complex disease heterogeneity and increasing statistical power.
• Current subtyping methods often capture dominant variations (e.g., population structure, technical noise) unrelated to mechanistic disease heterogeneity.
• There is a need for methods that can detect weak, specific subphenotypic variations masked by stronger, unrelated effects.

Approach:

• Introduced Phenotype Aware Components Analysis (PACA), a contrastive learning approach utilizing canonical correlation analysis.
• PACA is designed to robustly capture weak sources of subphenotypic variation unique to specific sample groups (e.g., disease cases).
• Leverages control samples to account for biological and technical confounders, ensuring focus on disease-specific variation.

Key Points:

• PACA successfully identifies and quantifies weak, previously undetected subphenotypic variations.
• Demonstrated robust performance across diverse data types including genotypes, transcriptomics, and DNA methylation.
• Outperformed alternative methods in simulation studies and real-world subtyping tasks.

Conclusions:

• PACA is a state-of-the-art tool for defining de novo disease subtypes.
• The method is particularly effective in complex cases where phenotypic heterogeneity is obscured by confounding factors.
• PACA offers a powerful approach to uncovering the molecular basis of disease heterogeneity.