Allosteric pocket predictor that fuses sequence- and structure-based protein language model embeddings with Fpocket features in a stacked ensemble.
No providers recorded yet. Browse all providers
An allosteric pocket can sit tens of ångströms from a protein's active site and still switch its activity on or off, which is what makes such pockets attractive drug targets: a ligand there need not compete with the native substrate and can be more selective than an orthosteric inhibitor. The difficulty is finding them. Pocket-detection software such as Fpocket readily enumerates every cavity on a structure, but most of those cavities are not allosteric, and the ones that are share no single shape or conservation signature. AlloPoc, from Huiling Zhang's group at South China Agricultural University with Shenzhen Institutes of Advanced Technology, is a classifier that decides which of those candidate pockets are allosteric.
Earlier language-model approaches to the problem, including DeepAllo and PLM Allosteric Classification, draw their learned signal from sequence-only protein language models. AlloPoc's premise is that sequence embeddings miss part of the picture, so it pairs them with embeddings from a structure-aware protein language model, the authors' description of the first use of dual-modality PLMs for allosteric pocket prediction. Both embedding types are concatenated with Fpocket's physicochemical pocket descriptors, and a hierarchical ensemble of conventional classifiers makes the final call.
AlloPoc succeeds the same group's AlloPED, whose pocket module used physicochemical features alone, and reports higher test-set performance than AlloPED, DeepAllo and AlloFusion. It was published in IEEE Transactions on Computational Biology and Bioinformatics in September 2026.
The framework has three modules: feature extraction, feature processing (selection and enhancement) and ensemble prediction. In the released training code, the base learners are CatBoost, elastic-net logistic regression, a balanced random forest, RUSBoost and a multilayer perceptron, stacked with five-fold internal cross-validation. Feature selection uses recursive feature elimination with a random forest scored by ROC AUC and protein-grouped cross-validation, retaining at least 100 features. The augmentation step trains a variational autoencoder with separate encoder branches for the language-model and Fpocket features on the positive pockets only, samples perturbed latent codes, and keeps the synthetic pockets that lie closest to real allosteric pockets and furthest from non-allosteric ones, raising positives to 40% of the negative count. Model selection runs five-fold cross-validation grouped by PDB entry.
Training uses the ASBench Core Diversity Set, 142 distinct protein structures, and evaluation uses a separate independent test set of 22 structures. On that test set AlloPoc reaches an AUC of 0.927 and a Matthews correlation coefficient of 0.524. The GitHub repository publishes the training and inference scripts, the conda environment and both PDB ID lists without a license file; the fitted ensemble checkpoint and the code that computes the language-model embeddings and Fpocket features are not included.
AlloPoc is aimed at the target-assessment stage of allosteric drug discovery, where a structural biologist or medicinal chemist has a structure and needs to decide which of its many cavities merits docking, molecular dynamics or mutagenesis. Because the classifier scores every pocket Fpocket reports, it can be run across a panel of targets to triage candidate sites, and the same scores can nominate distal pockets when investigating how a mutation far from the active site alters function.
AlloPoc extends the finding that pretrained protein language model representations carry allosteric signal by showing that adding a structure-aware language model on top of a sequence model helps further. Its evidence rests on a small benchmark: the training set spans under 150 proteins and the independent test set 22, so the margins over DeepAllo and AlloPED are measured on few examples and in silico only. Reproducing the reported numbers requires retraining, since the fitted model and the feature-extraction pipeline are not public.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.