Two-stage allosteric site predictor: an ensemble ranks Fpocket pockets, then a dilated CNN over ProtT5 embeddings calls allosteric residues.
No providers recorded yet. Browse all providers
Allosteric drugs bind away from a protein's active site and change its activity from a distance, which lets them avoid competing with the native substrate and often makes them more selective than orthosteric inhibitors. Most known allosteric sites were found by accident, because nothing in a structure announces them: they are less conserved than active sites, shallower, and more varied in shape. Pocket-detection software such as Fpocket lists every cavity on a protein, and on a typical structure only a handful of those dozens of cavities are allosteric. Allo-PED (written AlloPED in the paper's body) is a two-stage predictor that first decides which cavities are allosteric pockets and then decides which residues within them form the site.
It was developed by Huiling Zhang's group at South China Agricultural University with Yanjie Wei at the Shenzhen Institute of Advanced Technology, and posted to bioRxiv in April 2025. Its residue-level stage draws its learned signal from ProtTrans's ProtT5-XL-UniRef50 protein language model, joining a line of allosteric predictors built on pretrained sequence representations that includes DeepAllo and Allo-Allo. The same group's later AlloPoc is a separate model that reworks the pocket stage with language-model embeddings and uses Allo-PED as a baseline.
Training data come from ASBench, 146 allosteric proteins with X-ray resolution of 3.0 Å or better and under 30% pairwise sequence identity, on which Fpocket finds 4,756 pockets (185 allosteric, 4,571 not). Evaluation uses the 24-protein independent test set assembled for AllositePro, with 53 allosteric and 745 other pockets. The pocket stage filters features in two passes, minimum-redundancy maximum-relevance selection and then recursive feature elimination with a random forest, and trains its ensemble with five-fold cross-validation grouped by PDB entry on balanced subsets that pair each positive pocket with a negative from the same structure. On the test set it reaches an MCC of 0.544, an AUC of 0.920, precision of 0.471 and recall of 0.686, against an MCC of 0.388 for AllositePro, 0.264 for Allosite and 0.099 for PARS.
AlloPED-site is a dilated one-dimensional convolutional network. An initial convolution maps each residue's features to 512 channels, residual blocks with channel attention and a dilation-2 convolution widen the receptive field, a combined spatial-and-channel attention module reweights positions, and fully connected layers with dropout of 0.75 output a per-residue probability. Training uses focal loss (α 0.85, γ 4.0) to handle the scarcity of positive residues, with a window of 5 residues chosen from sweeps between 5 and 45. The authors compared ProtT5 against ESM-2 (650M), ESM Cambrian (600M) and ProteinBERT as the embedding source, and ProtT5 gave the best results. On the test proteins, applied within the pockets AlloPED-pocket predicts, AlloPED-site reaches a precision of 0.601, recall of 0.422 and specificity of 0.661, with an AUC of 0.563.
Allo-PED addresses the target-assessment step of allosteric drug discovery. A structural biologist or medicinal chemist with a structure of a kinase, GPCR or metabolic enzyme can use the pocket stage to rank which cavities merit docking, molecular dynamics or fragment screening, and the residue stage to choose residues for mutagenesis when testing whether a distal pocket regulates activity. The feature analysis is also of interest to researchers studying how allosteric sites differ from active sites.
Allo-PED adds residue-level prediction to a problem that earlier machine-learning tools mostly treated at the pocket level, and its comparison of four protein language models gives one data point on which pretrained representations carry allosteric signal. Its evidence is modest in scale and entirely computational. The test set covers 24 proteins, the residue-level AUC of 0.563 sits close to chance, and the pocket-stage gains over AllositePro are measured on 53 positive pockets. The work has since been peer reviewed, appearing in August 2025 as a chapter in the ISBRA proceedings volume of Springer's Lecture Notes in Computer Science. The GitHub repository named in the paper holds only a placeholder README, with no code, trained models or data released; the PNFx feature-extraction toolbox is public as a separate repository.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.