Skip to main content
Back to timeline
bioRxivSource publication:

Metabolites as nodes, reactions as hyperedges: HyperGenie predicts essential genes with 13% higher PR-AUC than objective-dependent methods

Synopsis

Predicting bacterial essential genes has relied on flux balance analysis, which assumes knockout strains still optimize for maximum growth, an assumption that does not always hold. HyperGenie represents the E. coli metabolic network as a directed hypergraph, with metabolites as nodes and reactions as hyperedges, and takes wild-type flux features directly as input, removing the post-knockout optimality assumption. On the iML1515 model under aerobic glucose, its PR-AUC is 13% above the objective-dependent FlowGAT and random forest, and it matches the flux-sampling FCL on comparable gene subsets.

Source-provided article image: HyperGenie: A method for predicting enzymatic gene essentiality using hypergraph neural networks and genome-scale metabolic models
Fig. 1 · bioRxiv · Page 3

Interpretation

Essential-gene prediction no longer needs the assumption that a knockout strain still optimizes for maximum growth. HyperGenie takes wild-type flux features as input and treats reaction essentiality as the classification target, removing the post-knockout optimality assumption from the pipeline. Earlier approaches either simulated knockouts with flux balance analysis while assuming mutants optimize the same objective, or bypassed that assumption with random forests or graph neural networks on mass flow graphs; random forests discard network topology, and mass flow graphs split multi-reactant, multi-product reactions into pairwise links. For E. coli grown aerobically in M9 glucose medium, over 50 runs from 10 repeats of 5-fold cross-validation, HyperGenie reached 96.1±2.6% PR-AUC against 82.5% for FlowGAT and 83.0% for random forest, a 13% gain; on two comparable gene subsets of 136 and 137 genes its accuracy matched FCL.

The model treats the metabolic network as a hypergraph, with metabolites as nodes and reactions as hyperedges that join several reactants to several products, so one message-passing step covers a reaction's full stoichiometry. To isolate the contribution of hypergraph message passing, the same flux features were fed to MLPGenie, a multilayer perceptron without topological propagation that shares input features, network scale and the gene-mapping pipeline. On glucose minimal medium HyperGenie's accuracy exceeds MLPGenie's by more than 2%, and by more than 4% across six alternative carbon sources, including over 10% on xylose; with FBA fluxes alone the gap is significant, while with the full feature set including flux variability bounds the PR-AUC difference is not significant.

Perspective

The demonstrated scope is reaction-level essentiality prediction for the E. coli K-12 iML1515 reconstruction across 12 sole carbon sources at uptake rates of 10 to 40 mmol gDW−1 h−1, plus growth curves for the panB, bioF and bioH knockout strains in glucose minimal medium and colony PCR verification for two of them.

In zero-shot cross-carbon-source prediction the model judges genes that gain essentiality in a new environment well but genes that lose essentiality poorly, with an overall tendency to over-predict essentiality; below 50% training data HyperGenie degrades faster than MLPGenie; at the lowest uptake rate MLPGenie performs no worse than HyperGenie on three-carbon substrates such as pyruvate, alanine and D-lactate; SHAP attributions depend on the chosen background baseline.