Multi-omics integration and batch correction using a modality-agnostic deep learning framework

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Modern biotechnologies generate complex, high-dimensional multi-modal datasets from single biological samples.
• Integrating diverse omics data is crucial for understanding complex biological processes like oncogenesis and aging.
• Technical artifacts, such as batch effects, complicate multi-modal data analysis and interpretation.

Purpose of the Study:

• To present MIMA, a modular, unsupervised AI framework designed for multi-omics data integration and batch correction.
• To demonstrate MIMA's capability in handling complex spatial and single-cell datasets.
• To establish a foundation for AI-driven digital pathology frameworks using integrated multi-omics data.

Main Methods:

• Development of MIMA, a modular and unsupervised AI framework.
• Application of MIMA to spatial and single-cell multi-omics datasets.
• Evaluation of MIMA's performance in batch effect removal, biological information preservation, and prediction of pathologist annotations.

Main Results:

• MIMA effectively removes batch effects while preserving biologically relevant information in multi-modal datasets.
• Learned representations from MIMA are predictive of expert pathologist annotations.
• MIMA enables cross-modal translation and uncovers novel molecular patterns.
• MIMA performs comparably to specialized tools despite being modality-agnostic.

Conclusions:

• MIMA is a flexible and scalable tool for multi-modal data analysis, crucial for advancing digital pathology.
• The framework facilitates AI-based integration of high-dimensional molecular data and histopathological imaging.
• MIMA offers new avenues for enhanced patient stratification and precision medicine.