Source-linked AI summary
A technical review of multi-omics data integration methods: from classical statistical to deep generative approaches
Ana R. Baião, Zhaoxiang Cai, Rebecca C Poulos, Phillip J. Robinson, Roger R Reddel, Qing Zhong, Susana Vinga, Emanuel Gonçalves
TL;DR
Multi-omics integration must handle high dimensionality, heterogeneity, and incompleteness in datasets relevant to precision medicine. This review synthesizes integration methods with emphasis on VAEs, their training and regularisation strategies, and emerging foundation-model and modality directions. It concludes that VAEs can integrate complex incomplete data and produce biologically meaningful representations, while often requiring large sample sizes for effective training.
Problem
Multi-omics integration faces high-dimensional, heterogeneous, and incomplete data, limiting the analysis of datasets relevant to precision medicine.
Method
The review categorizes integration methods and technically examines deep generative models, especially VAEs, including their training and regularisation strategies.
Results
VAEs can integrate high-dimensional, incomplete, complex, and heterogeneous data while producing biologically meaningful representations.
Takeaways & Limitations
The review provides a technical framework for multi-omics integration and identifies foundation models and additional modalities as future directions.
Takeaways & Limitations
Deep learning approaches often require large sample sizes for effective training.
Abstract
from arXiv · showhide
The rapid advancement of high-throughput sequencing and other assay technologies has resulted in the generation of large and complex multi-omics datasets, offering unprecedented opportunities for advancing precision medicine strategies. However, multi-omics data integration presents significant challenges due to the high dimensionality, heterogeneity, experimental gaps, and frequency of missing values across data types. Computational methods have been developed to address these issues, employing statistical and machine learning approaches to uncover complex biological patterns and provide deeper insights into our understanding of disease mechanisms. Here, we comprehensively review state-of-the-art multi-omics data integration methods with a focus on deep generative models, particularly variational autoencoders (VAEs) that have been widely used for data imputation and augmentation, joint embedding creation, and batch effect correction. We explore the technical aspects of loss functions and regularisation techniques including adversarial training, disentanglement and contrastive learning. Moreover, we discuss recent advancements in foundation models and the integration of emerging data modalities, while describing the current limitations and outlining future directions for enhancing multi-modal methodologies in biomedical research.
Introduction
Multi-omics integration is motivated by the biological value of combining heterogeneous, high-dimensional, and incomplete measurements, but these same properties create substantial computational challenges. This review organizes integration methods by their underlying architectures and computational innovations, with particular emphasis on deep generative approaches and VAEs.
- Challenges: Multi-omics datasets contain thousands of features, heterogeneous distributions, and missing data at both sample and modality levels.These properties create dimensionality, heterogeneity, sparsity, and incompleteness challenges for integration.
- Computational responses: Statistical and machine learning methods address these challenges through dimensionality reduction, batch effect correction, and imputation.These tasks support joint embeddings, technical-bias mitigation, denoising, and dataset augmentation.
- Review scope: The review categorizes methods as correlation-based, matrix factorization, probabilistic, network- or kernel-based, and deep learning approaches.This architecture-oriented taxonomy provides a technical perspective distinct from classifications based mainly on data types or applications.
- Deep generative approaches: Deep learning has shifted attention toward generative methods, especially VAEs for imputation, denoising, and joint embedding creation.The review also examines training and regularisation strategies including adversarial, cycle-consistent, contrastive, and disentangled representation learning.
- Future directions: The review standardizes terminology and discusses foundation models and additional data modalities as future directions for multi-modal biomedical research.These directions are presented in connection with potential applications in biomedical research and precision medicine.
Classical statistical and machine-learning approaches
Classical and machine-learning approaches integrate multi-omics data through correlation, factorization, probabilistic, network, kernel, and deep-learning frameworks. These methods reduce dimensionality, model shared and modality-specific structure, support clustering or phenotype prediction, and address challenges such as high dimensionality, missing values, and heterogeneity.
- Correlation/Covariance-based methods: Canonical Correlation Analysis maximises correlation between linear combinations of two omics datasets collected from the same samples.CCA produces canonical variables from projections X1w1 and X2w2 and serves as a joint dimensionality-reduction and information-extraction method.
- Correlation/Covariance-based methods: When features greatly outnumber observations (p_i ≫ n), CCA becomes ill-defined because its optimisation requires covariance-matrix inversion.Sparse CCA extensions, including sCCA, CCA-EN, and CCA-sparse group, were proposed to address this setting.
- Correlation/Covariance-based methods: Supervised and nonlinear CCA extensions support multi-omics prediction and integration beyond paired linear relationships.DIABLO maximises common or correlated information while minimising outcome-prediction error, whereas SDGCCA uses neural networks to capture nonlinear cross-data correlations.
- Matrix factorization methods: Matrix-factorization methods decompose omics datasets into shared, modality-specific, and residual components for dimensionality reduction and pattern discovery.Examples include jNMF, intNMF, LIGER, UINMF, and JIVE; LIGER can identify shared cell types, while JIVE quantifies joint variation.
- Probabilistic-based methods: Probabilistic and kernel-based methods address heterogeneous data structures through uncertainty modelling, flexible distributions, and feature-space transformations.MOFA accommodates different distributions and missing values while providing uncertainty estimates; kernel integration moves optimisation into sample space, making it independent of feature count.
Deep learning approaches
Deep learning methods integrate diverse omics data and learn nonlinear patterns, with VAEs receiving particular attention for probabilistic and generative multi-omics analysis. Reviewed models support tasks including joint representation learning, modality translation, imputation, batch correction, and supervised prediction.
- Deep learning approaches offer flexibility for integrating high-dimensional, diverse omics data and learning nonlinear, complex patterns.
- Autoencoders: Autoencoder-based models support cross-modality translation, joint dimensionality reduction, disease prognosis, clustering, drug-combination prediction, and batch correction.
- Generative approaches: VAEs are prominent in multi-omics because their probabilistic framework and generative capability support latent representations and new-data generation.
- Supervised models: VAE-based supervised modules enhance biologically relevant feature extraction and optimise classification or regression performance.
- Reported results: Reviewed applications report correlations of 0.9 or higher for scATAC-seq generated from scRNA-seq across two datasets and better cancer outcome prediction than original data.
- Multimodal integration: Multimodal generative models can create joint embeddings, reconstruct missing modalities, impute unobserved values, denoise observed features, and translate between omics modalities.
Promising perspectives
Multi-omics integration is expanding beyond molecular assays to include phenotypic, imaging, electronic-health, wearable, and other biological data. The review highlights classical methods, deep learning—especially VAEs—and foundation models as complementary directions for integrating heterogeneous datasets and advancing biological and clinical insight.
- Emerging integrations combine molecular omics with phenotypic screens, imaging, electronic health records, and wearable signals to connect molecular profiles with functional effects and patient outcomes.
- Foundation models use pretraining and transfer learning to generalise across tasks, with biological applications spanning multi-omics integration, perturbation prediction, drug-response prediction, annotation, and regulatory inference.
- Applying foundation models to omics remains constrained by the lack of inherent sequential structure, motivating strategies such as gene-order ranking and further model refinement.
- Classical approaches remain effective with limited sample sizes, providing robust and interpretable tools for dimensionality reduction, clustering, and pathway analysis.
- Deep learning enables integration of high-dimensional, incomplete, complex, and heterogeneous data, although effective training often requires large sample sizes.
- VAEs flexibly condition latent spaces on discrete genomic data and use adversarial, disentanglement, or contrastive regularisation to represent heterogeneous and biologically meaningful structure.