Language models trained on protein sequences used to predict sugar-attachment sites

The work

Technical · from the record

StackGlyEmbed predicts which N-X-[S/T] sequons in human proteins carry N-linked glycosylation. Sites were represented with embeddings from three pre-trained protein language models (ProteinBERT, ESM-2 window-averaged, ProtT5-XL-U50 per-residue), selected by incremental feature selection, and classified by a stacking ensemble of SVM, XGBoost and KNN base learners feeding an SVM meta learner. On the N-GlyDE independent test set the authors report 98.2% sensitivity, 92.5% balanced accuracy, 89.1% F1-score, 82.6% MCC and 93.1% AUPR, and on the N-GlycositeAtlas independent test set 83.1% sensitivity, 69.3% F1-score, 51.5% MCC and 65.9% AUPR; the paper reports higher sensitivity, F1, MCC, AUROC and AUPR than the compared published methods on both test sets, while cross-dataset training and testing performed poorly. A SHAP analysis on a PCA-reduced feature space found the base learner probability group had the highest mean absolute SHAP value.

How AI was used

Two published datasets of N-X-[S/T] sequons (N-GlyDE and N-GlycositeAtlas) were re-checked against UniProt, sequons whose protein sequences had changed were dropped, and each training set was split into a 60% part for base learners and a 40% part for the meta learner. For every sequon, features were computed as per-residue, window-averaged (31 residues) and global embeddings from three frozen pre-trained protein language models (ProtT5-XL-U50, ESM-2 esm2_t33_650M_UR50D, ProteinBERT) together with averaged physicochemical parameters. Incremental feature selection, scored by the F1 of SVM models, chose ProteinBERT global, ESM-2 window and ProtT5-XL-U50 per-residue embeddings. Ten candidate classifiers were fitted and compared; incremental mutual information between learner outputs and labels selected SVM, XGBoost and KNN as base learners, and cross-validation on the BLP-augmented validation set selected SVM as meta learner, with grid search used for hyperparameters and window size. Features were Yeo–Johnson transformed; base learners were trained on multiple random undersamples of the majority class, and the meta learner was trained on randomly undersampled validation data whose feature vectors were augmented with base learner probabilities. The trained ensemble was then run over the held-out independent test sequons and a case-study protein, and SHAP values were computed on a PCA-reduced version of the feature space to attribute importance to each feature group.