DeepAllo

Allosteric drugs act at pockets far from a protein's catalytic or orthosteric site, which makes them attractive — they can be more selective and less prone to competition with native substrate — and hard to find. Most computational allosteric-site predictors score candidate pockets using hand-engineered descriptors of shape, conservation, and dynamics. DeepAllo, from Koç University, instead asks a protein language model to supply the sequence-level signal, and combines what it learns with the geometric descriptors that pocket-detection software already produces.

The pipeline runs in two stages. A pretrained protein language model is first fine-tuned on the AlloSteric Database (ASD) in a multitask setting, learning to label residues as allosteric or not while simultaneously predicting three-state secondary structure. The fine-tuned encoder is then frozen and used as a feature extractor: its per-pocket embeddings are concatenated with FPocket's structural descriptors and handed to gradient-boosted tree models (XGBoost and an AutoML search) that make the final call on whether a candidate pocket is allosteric.

DeepAllo sits alongside other language-model-based approaches to the same problem, such as PLM Allosteric Classification, but takes a different route: pocket-level ranking with explicit geometric features rather than residue-level labelling from sequence alone. The work first appeared as a preprint in late 2024 and was published in Bioinformatics in 2025.

#Key Features

• Multitask fine-tuning: The language model is trained on allosteric-residue classification and secondary-structure prediction at once, so structural context is injected into the embeddings without requiring a solved structure at inference.
• Hybrid feature space: Learned sequence embeddings are fused with FPocket's geometric and physicochemical pocket descriptors, letting the classifier draw on both evolutionary signal and cavity shape.
• Pocket-level ranking: Rather than emitting a single binary call, DeepAllo ranks every candidate pocket on a protein; 90.5% of true allosteric pockets appear in the top three, which is the form a medicinal chemist can act on.
• Attention-based interpretation: The authors visualise the fine-tuned model's attention to contrast how it distributes weight over allosteric versus non-allosteric residues, giving a handle on what the encoder has learned.

#Technical Details

The language model backbone is ProtBert-BFD, a BERT-style encoder with 30 layers and 16 attention heads per layer. Training data comes from the AlloSteric Database, reduced after preprocessing to 207 proteins: 165 proteins in training (3,375 pockets, 64,593 residues) and 42 in test (848 pockets, 15,567 residues). The task is severely imbalanced — only 304 of 4,223 pockets, or 7.76%, are allosteric — which the pocket-level classifier must absorb. The best configuration, AutoML over multitask language-model features plus FPocket descriptors, reaches an F1 of 89.66% with 92.3% precision and 88.1% recall, against a previous best of 78.2% F1 (72.6% precision, 84.7% recall) for an ensemble baseline; top-3 pocket recovery rises from 84.9% to 90.5%.

Source code is available on GitHub under GPL-3.0 and archived on Zenodo, the trained checkpoints are on Hugging Face under AGPL-3.0, and the curated training and evaluation dataset is separately archived on Zenodo.

#Applications

DeepAllo is aimed at the early stages of allosteric drug discovery, where the practical question is which of a protein's many surface cavities is worth committing docking, molecular dynamics, or mutagenesis effort to. Because it consumes a sequence plus FPocket output rather than requiring co-crystallised allosteric ligands, it can be run across target lists to triage candidates. The same ranking is useful in mechanistic work — for instance, nominating sites for testing hypotheses about how a distal mutation propagates to an active site.

#Impact

DeepAllo's contribution is showing that protein language model representations carry usable allosteric signal, and that fusing them with the geometric descriptors classical tools already compute beats either source on its own. The main limitation is scale: the curated ASD subset covers only a couple of hundred proteins, so generalisation to protein families absent from the database is not established by the reported benchmarks, and the heavy class imbalance means precision on genuinely novel targets may differ from held-out test performance. Releasing code, weights, and the curated dataset under open licences makes the comparison reproducible for the groups building the next generation of allosteric predictors.