Ziwei Yang, Ph.D.
01 · Model Biology · Mechanistic Biological Modeling

GraphGDel

Graph representation learning for growth-coupled gene deletion prediction

GraphGDel turns a genome-scale metabolic model into a metabolite graph and learns from it, together with gene and metabolite sequences, to predict which genes to delete so that a cell's growth is coupled to the production of a target metabolite.

Metabolic ModelingMetabolic Engineering Ziwei Yang, Takeyuki Tamura · IEEE Transactions on Computational Biology and Bioinformatics, 2026
GraphGDel framework: gene deletion strategy data from MetNetComp and gene and metabolite sequences from KEGG and MetaNetX feed four modules (Meta-M, Gene-M, Graph-M, Pred-M) that output a predicted gene deletion strategy.
GraphGDel at a glance. Metabolite and gene sequences are encoded by Meta-M and Gene-M, Graph-M refines metabolite features on the metabolic graph, and Pred-M predicts the deletion status of every gene for a target metabolite. Figure adapted from Yang & Tamura, IEEE TCBB 2026.

Overview

Metabolic engineers use genome-scale metabolic models to design microbial strains. A central goal is growth-coupled production: deleting genes so that a cell can only grow while also making a target compound. Because faster-growing strains take over during repeated passaging, coupling keeps production stable. Computing such gene deletion strategies from scratch is expensive for genome-scale models.

Our earlier DeepGDel turned this into a prediction task learned from databases of known strategies, but it saw genes and metabolites only as sequences and ignored how metabolites are connected through reactions. Metabolic models do not come with a ready-made graph: they are written as stoichiometric matrices, gene–protein–reaction rules and flux bounds.

GraphGDel addresses both. It builds a metabolite graph from the model, filtering out hub and currency metabolites such as ATP and NADH, and combines graph learning with sequence learning to predict which genes to delete or keep for each target metabolite.

Method

Two contributions: a pipeline that constructs graphs from constraint-based metabolic models, and a learning framework that uses them.

  1. Graph construction pipeline: a BiGG metabolic model is turned into graphs G0, G1, G2 using a currency metabolite set and node filtering and restoring, giving the final graph representation.
    01

    Build the metabolite graph

    Read the model as a network of metabolites, reactions and genes, link each substrate to each product, then remove the ten highest-degree hubs and known currency metabolites, restoring targets and non-currency hubs.

  2. LSTM encoder–decoder used in Meta-M and Gene-M: tokens are embedded, encoded into a latent vector Z, and decoded back into the input sequence.
    02

    Encode genes and metabolites

    Meta-M and Gene-M are LSTM autoencoders that compress metabolite SMILES strings and gene amino-acid sequences into latent representations.

  3. Graph convolution network used as the encoder in Graph-M: node embeddings are updated from their neighbours.
    03

    Learn on the graph and predict

    Graph-M, a graph convolutional network trained by link prediction, refines metabolite features in their network context. Pred-M combines all three representations to predict whether each gene is deleted or kept.

Figure adapted from Yang & Tamura, IEEE TCBB 2026.

Key Results

Overall accuracy

77.35%
e_coli_core
E. coli core model
85.54%
iMM904
S. cerevisiae (yeast)
73.22%
iML1515
genome-scale E. coli

Overall accuracy of gene deletion status prediction on three metabolic models (mean of 5 × 10-fold cross-validation): 14.04, 16.26 and 13.18 points above a deep neural network baseline, and ahead of DeepGDel and a graph-aggregation baseline on all five metrics.

Benchmark

Overall accuracy against baselines GraphGDelBaselines
e_coli_core
DNN 63.31 ±2.70
DeepGDel 71.18 ±0.44
NMA 72.25 ±0.52
GraphGDel 77.35 ±1.45
iMM904
DNN 69.28 ±1.58
DeepGDel 80.58 ±0.44
NMA 81.18 ±0.50
GraphGDel 85.54 ±1.04
iML1515
DNN 60.04 ±1.04
DeepGDel 67.91 ±0.34
NMA 68.52 ±0.38
GraphGDel 73.22 ±1.02
Overall accuracy (%), mean ± s.d. over 5 × 10-fold cross-validation. DNN: deep feedforward network; DeepGDel: sequence-learning baseline; NMA: neighbourhood mean aggregation on the same metabolite graph. Data from Table V of the paper (arXiv v7); redrawn.

Resources

Paper

GraphGDel: Constructing and Learning Graph Representations of Genome-Scale Metabolic Models for Growth-Coupled Gene Deletion Prediction

IEEE TCBB 2026 · preprint on arXiv

Code

MetNetComp/GraphGDel

Source code and example datasets