A versatile deep-learning based strategy for multi-omics integration¶

A Python package for integrating multiple genomic data modalities (e.g., RNA-seq, CNV, and DNA methylation) using a hierarchical deep-learning architecture. It supports classification, survival outcome prediction, and SHAP-based explainability, all in a single scikit-learn-style API.

Overview¶

CustOmics is designed to provide a modern and research-friendly framework for computational biology and precision medicine.

It relies on mudata, a core scverse data structure for multimodal data.

Architecture¶

flowchart LR
    %% Input omics
    RNA["RNA-Seq"] --> AE1
    CNV["Copy Number Variations (CNV)"] --> AE2
    METH["DNA Methylation"] --> AE3

    %% Source-specific autoencoders
    subgraph Phase1["Phase 1 — Representation Learning"]
        direction TB
        AE1["RNA-Seq Autoencoder"]
        AE2["CNV Autoencoder"]
        AE3["Methylation Autoencoder"]
    end

    %% Central integration
    AE1 --> CVAE
    AE2 --> CVAE
    AE3 --> CVAE

    subgraph Phase2["Phase 2 — Mixed Integration"]
        CVAE["Central Variational Autoencoder\nShared Latent Representation"]
    end

    %% Downstream tasks
    CVAE --> CLS["Tumor / Subtype Classification"]
    CVAE --> SURV["DeepSurv Survival Prediction"]

    %% Styling
    classDef omics fill:#ebe1ff,stroke:#8c52ff,stroke-width:2,color:#000;
    classDef ae fill:#EAF7EA,stroke:#34A853,stroke-width:2,color:#000;
    classDef latent fill:#FFF4D6,stroke:#FBBC05,stroke-width:3,color:#000;
    classDef task fill:#ffe6d8,stroke:#ff914d,stroke-width:2,color:#000;

    class RNA,CNV,METH omics;
    class AE1,AE2,AE3 ae;
    class CVAE latent;
    class CLS,SURV task;

• Phase 1 trains per-source autoencoders jointly with the task heads.
• Phase 2 additionally trains the central VAE to consolidate the integrated representation.

Why use CustOmics¶

CustOmics is built for high-dimensional and heterogeneous omics datasets.

1. Unified API¶

• Classification tasks
• Survival outcome prediction
• Latent representation learning
• SHAP-based explainability
• Modular source-specific autoencoders for flexible experimentation
• Scalable training using PyTorch and GPU acceleration
• Compatible with scikit-learn-style workflows

2. Visualization Utilities¶

• Latent space exploration
• Kaplan-Meier survival stratification
• Feature attribution analysis
• Easily extensible to custom architectures, tasks, and omics sources

Start using CustOmics by reading the getting started guide.