Ziwei Yang, Ph.D.
02 · Understand Biology · Multimodal Biomedical AI

GeSubNet

Subtype-specific gene networks from patient data and prior knowledge

GeSubNet learns a separate gene interaction network for each disease subtype by combining patient gene expression with general gene networks from knowledge bases, so that differences between subtypes show up as differences in gene interactions.

Cancer GenomicsGraph Learning Ziwei Yang, Zheng Chen, Xin Liu, Rikuto Kotoge, Peng Chen, Yasuko Matsubara, Yasushi Sakurai, Jimeng Sun · International Conference on Learning Representations (ICLR), 2025 · Oral
Left: a breast cancer gene network from the STRING database, with genes of interest and a shaded gene set with consistent behaviour. Right: the same network as a disease network with a dashed potential sub-network and question marks linking it to subtypes Luminal A, Luminal B and Basal-like.
The problem. A breast cancer gene network from STRING describes interactions shared across patients. A gene set with consistent behaviour points to a sub-network, but the general network cannot say which interactions belong to which subtype, such as Luminal A, Luminal B or Basal-like. Figure adapted from Yang et al., ICLR 2025.

Overview

Cancers such as breast cancer split into subtypes that differ in prognosis and treatment response. Knowledge bases such as STRING and KEGG describe how genes interact in general, but those networks are the same for every patient and do not show which interactions matter in a given subtype.

GeSubNet brings the two sources together. It learns patient representations that separate subtypes from gene expression data, learns gene representations from a prior gene network, and then combines them to infer one gene interaction network per subtype.

The inferred networks also rank genes for each subtype. In simulated knockouts, removing GeSubNet's top-ranked genes moved patients away from their subtype far more often than removing the genes chosen by the baselines.

Method

Three modules, trained in three phases.

GeSubNet overview. Step 1 Patient-M encodes gene expression data with subtype information into a patient representation in a discrete latent space. Step 2 Graph-M trains a GNN encoder and decoder by link prediction on an input gene network from a knowledge database. Step 3 Infer-M integrates both representations and retrains the GNN encoder with a reconstruction loss to infer subtype-specific networks.
How GeSubNet works: Patient-M (step 1), Graph-M (step 2) and Infer-M (step 3). Figure adapted from Yang et al., ICLR 2025.
  1. 01

    Patient-M: separate the patients

    A VQ-VAE learns patient representations from gene expression profiles in a discrete latent space where subtypes can be told apart.

  2. 02

    Graph-M: learn the prior network

    A graph neural network is trained by link prediction on a known gene network for the cancer, giving one representation per gene.

  3. 03

    Infer-M: generate subtype networks

    Patient and gene representations are combined to reconstruct expression data. The reconstruction loss retrains the GNN encoder, and the decoder then outputs a gene network for each subtype.

Key Results

Simulated gene knockout, breast cancer (BRCA, 11,327 genes)

83%
shift rate for top-ranked genes
best of seven baselines: 34%
12%
shift rate for low-ranked genes
lower is better; baselines 22–29%
12 / 12
best scores in network inference
3 graph metrics × 4 cancers

The shift rate is the share of patient samples whose profile shifts significantly when a set of genes is knocked out. GeSubNet also scored best on all three graph metrics (CDV, GED and DCS) in all four cancers.

Benchmark

Shift rate after simulated knockout: top- vs low-ranked genes High-ranking genes knocked out (higher is better)Low-ranking genes knocked out (lower is better)
WGCNA 0.24 / 0.22
wTO 0.24 / 0.24
ARACNe 0.23 / 0.25
LEAP 0.28 / 0.24
GAERF 0.32 / 0.28
LR-GNN 0.23 / 0.23
CSGNN 0.34 / 0.29
GeSubNet 0.83 / 0.12
Shift rate (ΔSR) when the high- or low-ranking genes found by each method are knocked out in silico, BRCA. A useful ranking separates the two dots: high for top-ranked genes, low for low-ranked ones. Data from the table in Fig. 4b of the paper (arXiv v3, ICLR 2025 camera-ready); redrawn.

Resources

Paper

GeSubNet: Gene Interaction Inference for Disease Subtype Network Generation

ICLR 2025 (oral) · preprint on arXiv

Code

chenzRG/GeSubNet

Source code