Deep learning model flags bacterial proteins likely to drive infection

The work

Technical · from the record

DeepGVS predicts whether a bacterial protein is a virulence factor by combining hand-computed features of its coding DNA sequence with learned protein representations: a graph attention network over an ESMFold-predicted residue graph, a bidirectional Mamba encoder over ESM-2 residue embeddings, and a global ProtT5 embedding. Four base classifiers were fitted to the combined features and their probabilities were merged by a neural additive model meta-learner. On the unchanged accession-level independent test set of the DeepVF benchmark (Dataset_B) the model reached an accuracy of 87.50%, which the authors report as 6.30, 2.60 and 1.40 percentage points above published values for DeepVF, GTAE-VF and PLMVF, while noting that PLMVF reported a higher recall and that the baseline numbers are cited rather than reimplemented; accuracy on the authors' own Dataset_A test set was 0.9262. A taxonomy-only logistic regression reached AUROC 0.933 on Dataset_A and 0.868 on Dataset_B against 0.9795 and 0.9275 for DeepGVS, which the authors read as taxonomic composition contributing to but not fully explaining the discrimination.

How AI was used

Two paired protein–CDS datasets were assembled, one newly curated with CD-HIT at a 0.5 identity cutoff and one retaining the DeepVF benchmark's 0.3-cutoff partitioning with its independent test set left unchanged. For each coding sequence, four groups of non-learned nucleotide descriptors were computed: global statistics, k-mer frequencies for k of 2, 3 and 4, 128-bp sliding-window heterogeneity, and repeat or low-complexity features. In parallel, ESMFold predicted Cα coordinates for every protein, from which a residue graph was built by linking each residue to its 30 nearest spatial neighbours, with nodes initialised by 256-dimensional ESM-2 embeddings; a graph attention network and a bidirectional Mamba encoder were run in parallel over this input, combined by gated fusion, pooled, and concatenated with a global ProtT5 embedding, with sequences truncated at 1000 residues. The concatenated CDS and protein vectors were supplied to Random Forest, SVM, XGBoost and MLP base classifiers, whose predicted probabilities were integrated by a Neural Additive Model meta-learner trained on five-fold stratified out-of-fold predictions, after which the base classifiers were refitted on the full development set for independent-test prediction at a fixed 0.5 threshold. The same trained models were then run over an E. coli O157:H7 proteome under a dual-model consensus rule, and SHAP attribution over the base-learner probabilities and features, plus sliding-window sequence masking on the top candidate, were used to inspect model behaviour.