PPubMed21 Jun 2023Cited 8×

Deep multi-omics integration by learning correlation-maximizing representation identifies prognostically stratified cancer subtypes

Yanrong Ji, Pratik Dutta, Ramana Davuluri, Magnus Rattray

Abstract

Advances in high-throughput technologies have facilitated generation of multiple -omics datasets, such as genomic, epigenomic, transcriptomic and proteomic measures, on same biosamples. Indeed, analyses of -omics data on hundreds of cancer tissues, profiled by The Cancer Genome Atlas (TCGA), International Cancer Genome Consortium (ICGC) and many other laboratories around the world, have led to the identification of clinically relevant subgroups of various cancers. While majority of the studies utilized microarray or RNA-Seq-based expression profiling data (; ; ; ; ; ), several others have empl

A figure from Deep multi-omics integration by learning correlation-maximizing representation identifies prognostically stratified cancer subtypes — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

DeepMOIS-MC uses deep canonical correlation to integrate multi‑omics data, yielding more prognostic cancer subtypes than traditional methods.

What’s new

i.Developed DeepMOIS-MC, a deep learning framework that maximizes correlation across multi‑omics views for cancer subtyping.
ii.Implemented a PyTorch version of Deep Generalized Canonical Correlation Analysis (DGCCA) and released the code on GitHub.
iii.Showed superior patient stratification on TCGA lung and liver cancer datasets compared to traditional methods.
iv.Created robust XGBoost classifiers validated on independent ICGC and GEO cohorts.
v.Provided a two‑stage pipeline (clustering and classification) that handles missing omics views.

How they did it

High‑dimensional omics views are fed into separate two‑layer fully connected networks and optimized with a GCCA loss to learn a shared 100‑dimensional representation. Survival‑related features are selected via univariate Cox‑PH and used for k‑means clustering. For classification, the original omics matrix is scaled, discretized by equal‑frequency binning, feature‑selected with Random Forest, and an XGBoost model is trained with 10‑fold cross‑validation.

What they found

DeepMOIS-MC identified two HCC subtypes with a Cox‑PH log‑rank P = 8.89e-7, outperforming autoencoder baselines (best P = 9.58e-5).
Runtime was under 30 seconds for 20‑epoch full‑batch training, second fastest among five compared methods.
ICGC LIRI‑JP classifier (150 variables, 30 bins) achieved AUC = 0.8838 and accuracy = 0.7995; GSE14520 classifier (190 variables, 15 bins) achieved AUC = 0.8571 and accuracy = 0.7615.
Kaplan‑Meier survival differences in validation cohorts were significant (P = 7.17e-4 and P = 3.66e-3).

Caveats

✕Limited public validation data for omics types other than expression that match training platforms.
✕Method relies on high‑quality, comprehensive survival data for outcome‑guided training.
✕Balancing preservation of single‑omics features versus shared features remains a challenge.

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Similar papers in all four archives

The same idea searched in arXiv, PubMed, bioRxiv and chemRxiv at once, so a method can lead you to where it is being applied.

AarXiv4 found

PPubMed4 found

BbioRxiv4 found

CchemRxiv2 found

Searching the other archives…