A deep manifold-regularized learning model for improving phenotype prediction from multi-modal data.

Area of Science:

• Computational biology
• Systems biology
• Machine learning

Background:

• Biological phenotypes arise from complex multi-scale mechanisms.
• Multi-modal data, like single-cell multi-omics, offers deeper insights into these mechanisms.
• Existing methods may not fully leverage cross-modal interactions for phenotype prediction.

Purpose of the Study:

• To develop an interpretable regularized learning model, deepManReg, for phenotype prediction from multi-modal data.
• To improve phenotype prediction accuracy by integrating information across different data modalities.
• To identify and prioritize key multi-modal features and their interactions driving specific phenotypes.

Main Methods:

• Deep neural networks are used to learn cross-modal manifolds and align multi-modal features into a common latent space.
• Cross-modal manifolds serve as a feature graph to regularize classifiers, enhancing prediction.
• The model was validated on a handwritten digit dataset and mouse brain single-cell Patch-seq data (transcriptomics and electrophysiology).

Main Results:

• deepManReg demonstrated improved phenotype prediction accuracy on both benchmark and biological datasets.
• The model successfully prioritized relevant multi-modal features and cross-modal interactions.
• Key genes and electrophysiological features driving neuronal cell phenotypes were identified.

Conclusions:

• deepManReg provides an effective and interpretable approach for phenotype prediction using multi-modal data.
• The method enhances understanding of complex biological systems by integrating diverse data sources.
• This framework facilitates feature discovery and interpretation in multi-omics studies.