Interpretable multi-omics integration across mixed-order tensors with MANTRA.

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Understanding complex diseases requires integrating multi-modal molecular data.
• Existing methods struggle with modern experimental designs generating mixed-order tensor datasets.
• Integrating drug-response tensors and transcriptomics matrices presents analytical challenges.

Purpose of the Study:

• To present MANTRA (Multi-view ANalysis with Tensor and matRix Alignment), a probabilistic framework for integrating multi-modal molecular data.
• To develop a method capable of handling datasets with mixed-order tensors.
• To uncover hidden biological variation and identify novel patient subgroups in complex diseases.

Main Methods:

• MANTRA is a probabilistic framework combining group factor analysis and tensor decomposition.
• It integrates collections of tensors of different orders, including drug-response tensors and transcriptomics matrices.
• A Bayesian approach with structured sparsity priors handles missing data and learns interpretable latent factors.

Main Results:

• MANTRA identified clinically relevant patient subgroups in a Chronic Lymphocytic Leukemia (CLL) dataset that were missed by single-view analyses.
• In an Acute Lymphoblastic Leukemia (ALL) study, MANTRA identified a novel patient subgroup linked to a specific cell type (pDCs).
• The framework explicitly models higher-order data structures to reveal hidden biological variation.

Conclusions:

• MANTRA provides an interpretable tool for integrating multi-modal molecular data with mixed-order tensors.
• The framework successfully identified novel patient subgroups and biological insights in CLL and ALL.
• MANTRA advances the analysis of complex experimental data for disease understanding.