MIND: multimodal integration with neighbourhood-aware distributions.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Multimodal data integration enhances predictive performance in biological applications like cancer patient stratification.
• Integrating multi-omics data is challenging due to missingness and heterogeneity, often leading to information loss with traditional methods.

Purpose of the Study:

• To propose a novel method, MIND (Multimodal Integration with Neighbourhood-aware Distributions), for robust integration of incomplete multi-omics data.
• To improve patient stratification and other downstream tasks by learning accurate patient-specific embeddings.

Main Methods:

• Developed a multimodal Variational Autoencoder with a data-driven prior to learn patient-specific embeddings from incomplete multi-omics data.
• Incorporated neighbourhood structure from the observed data into the prior to penalize divergence between data and latent spaces.
• MIND is designed to handle high missing rates, unbalanced missingness, and low signal-to-noise ratios.

Main Results:

• MIND demonstrated robust performance across various data challenges, including high missingness and heterogeneity.
• Achieved superior performance on downstream tasks compared to existing integration methods on both synthetic and real datasets.
• Successfully generated patient-specific embeddings from incomplete multi-omics profiles.

Conclusions:

• MIND offers a powerful and robust solution for multi-omics data integration, addressing key limitations of existing methods.
• The neighbourhood-aware approach enhances the reliability of learned embeddings for biological applications.
• This method has significant potential for improving cancer patient stratification and precision medicine.