A Genomics-Guided Multimodal Contrastive Learning Framework for Clinically Significant Prostate Cancer Risk...

Area of Science:

• Biomedical Informatics
• Machine Learning
• Genomics

Background:

• Heterogeneous data integration is challenging in intelligent systems, especially with missing data and cross-domain issues.
• Current multimodal fusion methods lack robustness due to reliance on complete datasets and weak alignment.

Purpose of the Study:

• Develop a genomics-guided multimodal representation learning framework for robust data fusion.
• Enable reliable cross-modal correspondence and accurate prediction with incomplete data.
• Address limitations of existing approaches in handling missing modalities and cross-domain scenarios.

Main Methods:

• Proposed a multimodal learning architecture using genomics as a biological anchor.
• Modeled conditional projections to imaging modalities like multiparametric MRI and whole-slide histopathology (WSI).
• Formulated fusion as genomics-guided contrastive learning with domain-specific constraints for a shared latent representation.

Main Results:

• Achieved an AUROC of 0.985 for prostate cancer risk discrimination in the Genomics+WSI cohort.
• Demonstrated high accuracy (92.1%) and reduced critical errors by 58% with strong calibration.
• Showcased robustness to domain shifts and maintained high performance with genomics alone.

Conclusions:

• The proposed framework offers a scalable and generalizable solution for multimodal data fusion.
• It supports robust prediction and handles missing modalities effectively.
• Applicable to complex information systems beyond the studied domain.