CustOmics: A versatile deep-learning based strategy for multi-omics integration

Posted by mhb on 2025-11-12 10:14:24 |

Share: Facebook | Twitter | Whatsapp | Linkedin Visits: 219

Introduction

Modern biology generates multiple types of “omics” data—genomics, transcriptomics, epigenomics, etc.—that describe different biological layers. Integrating these heterogeneous datasets is crucial for understanding diseases like cancer but is challenging due to differences in scale, distribution, and dimensionality.
CustOmics is a new deep-learning framework designed to integrate multi-omics data efficiently and interpretably using a two-phase mixed-integration approach based on variational autoencoders (VAEs).

Methods

CustOmics works in two main phases:

1. Phase 1 – Independent Training:
   Each omic dataset (e.g., RNA-Seq, DNA methylation, CNV) is encoded separately using its own autoencoder to learn compact “sub-representations.”
   This allows each data type to be normalized and learned according to its specific structure and signal.

2. Phase 2 – Joint Integration:
   The learned sub-representations from all omics are combined in a central variational autoencoder that learns cross-omic interactions.
   The training uses Maximum Mean Discrepancy (MMD) instead of Kullback–Leibler divergence for better generalization and includes task-specific losses for classification and survival prediction.

The framework also includes SHAP-based interpretability, showing which genes or features contribute most to predictions.

Results

• Tested on TCGA multi-omics datasets (CNV, RNA-Seq, methylation) across seven cancer types.

• Outperformed other integration methods (e.g., PCA, NMF, joint/late VAEs).

• Achieved accuracy ≈ 97.8% in pan-cancer classification and C-index ≈ 0.68 in survival prediction.

• Demonstrated strong biological interpretability, identifying known biomarkers such as TFF1 for breast cancer.

• Robust even with smaller datasets and fewer samples.

Conclusion

CustOmics introduces a hybrid, interpretable, and flexible approach to multi-omics data integration.
It effectively handles heterogeneity, improves predictive accuracy for cancer type and survival outcomes, and provides biological insight through explainable AI tools.
The method and code are freely available on GitHub (https://github.com/HakimBenkirane/CustOmics).

Read Full Paper Here