Fairness-aware supervised hierarchical contrastive semantic learning for sexual dimorphism analysis.

Area of Science:

• Genomics
• Artificial Intelligence
• Precision Medicine
• Computational Biology

Background:

• Sexual dimorphism significantly impacts disease susceptibility and outcomes.
• Current AI genomic models often exhibit bias, failing to account for sex-specific biological mechanisms.
• This limits the development of unbiased precision medicine.

Purpose of the Study:

• To develop a fairness-aware AI model for discovering unbiased sex-common and sex-specific genomic features.
• To address algorithmic bias in AI-based genomic models concerning sexual dimorphism.
• To advance inclusive precision medicine by elucidating sex-specific molecular heterogeneity.

Main Methods:

• Proposed a fairness-aware supervised hierarchical contrastive learning approach (FairHICON).
• Utilized transcriptomic datasets for cancer and asthma.
• Evaluated model performance against state-of-the-art benchmarks.

Main Results:

• FairHICON significantly outperformed existing benchmarks, improving predictive performance by up to 9%.
• The model effectively reduced the performance gap between male and female sexes.
• Identified sex-specific pathways significantly improved patient survival stratification within sex groups.

Conclusions:

• FairHICON successfully elucidates the molecular heterogeneity of sexual dimorphism.
• The approach advances the development of unbiased and inclusive precision medicine.
• The identified sex-specific features hold prognostic value for patient stratification.