PLM Allosteric Classification

Allosteric drugs bind at sites remote from a protein's active site and shift the conformational ensemble to tune activity up or down. Because allosteric pockets are less conserved across protein families than orthosteric ones, modulators that target them can be more selective and better tolerated. The obstacle is finding the pockets: allosteric sites involve few residues, sit close to the surface where they are hard to distinguish from incidental grooves, and often only become visible once a ligand is bound. Every previously published allosteric site predictor — Allosite, AlloPred, ALLO, the PASSer family, AlloReverse — depends on geometric and physicochemical descriptors computed from a high-resolution 3D structure.

PLM Allosteric Classification, from R. C. Eccleston and N. Furnham at the London School of Hygiene & Tropical Medicine, tests whether protein language models can do the job from sequence alone. The work evaluates three escalating strategies on three frozen backbones — Ankh Large, ProtT5 XL from the ProtTrans suite, and ESM-2 3B — and lands on a structure-aware conditioning mechanism as the approach that works.

That mechanism is the paper's central contribution. Rather than adding structural coordinates, the model encodes the orthosteric binding site as a binary residue mask and injects it directly into the input embeddings, so allosteric prediction is conditioned on where the active site is. The rationale is biological: an allosteric site exists only in relation to an orthosteric one, and conditioning lets the model learn that functional coupling. The result is a residue-level predictor whose accuracy and recall sit alongside the leading structure-based tools while requiring no protein structure at inference.

#Key Features

• Sequence-only inference: Predictions are made from the amino acid sequence plus an orthosteric site annotation, avoiding the high-resolution structure that every prior allosteric predictor requires.
• Orthosteric conditioning: A binary orthosteric-site mask is linearly projected to the embedding dimension and added element-wise to the token embeddings before the forward pass, injecting functional context without an extra loss term.
• Residue-level output: The model labels individual residues rather than ranking whole pockets, giving finer-grained hypotheses than pocket-detection pipelines.
• Parameter-efficient fine-tuning: LoRA adapters with focal loss handle a training set in which only about 6% of residue labels are positive, making full fine-tuning of 1.5–3B parameter backbones unnecessary.
• High-sensitivity operating point: The conditioned model is tuned toward recall, surfacing candidate sites for downstream filtering rather than returning a short, conservative list.

#Technical Details

All three backbones are loaded from public HuggingFace checkpoints and paired with the ConvBert token-classification head from the Ankh repository, which combines convolutions with multi-head self-attention. Training data comes from the Allosteric Database (ASD, release 202306_XF): roughly 3,000 binding sites, with residues within 7 Å of a modulator atom labelled allosteric. Splits are 0.8/0.1/0.1 and enforced with MMseqs2 so that no held-out sequence exceeds 30% identity with training sequences. Method 1, a classifier over frozen embeddings, reaches an average precision score (APS) of 0.448–0.502, improving slightly when the head is pre-trained on ~14,000 PDBbind ligand-binding sites. Method 2, LoRA fine-tuning, raises APS to 0.553 and recall to 0.651 (ProtT5 XL, focal loss).

Method 3 adds orthosteric conditioning on the ~1,900-sequence ASD subset that carries both site annotations, initialised from the best Method 2 checkpoint. It reaches accuracy 0.924–0.953, recall 0.766–0.836, AUC-ROC 0.932–0.948 and APS 0.587–0.614, with precision falling to 0.329–0.455. ESM-2 3B is strongest on accuracy (95.3%, recall 78.8%) and Ankh Large on recall (83.6%, accuracy 92.4%), against PASSer's reported 97.4% accuracy and 84.7% recall. That comparison is drawn from a published survey table and sets residue-level predictions against pocket-level ones.

#Applications

The model suits early-stage allosteric drug discovery, where a target's active site is known but no allosteric pocket has been characterised — a common situation for proteins whose orthosteric site is considered undruggable. Because it needs only sequence, it extends to targets without a ligand-bound high-resolution structure, which is precisely where structure-based predictors are weakest. Its recall-oriented behaviour makes it a triage step: generate a broad set of candidate allosteric residues, then narrow them with docking, molecular dynamics, or mutagenesis.

#Impact

The work is a bioRxiv preprint released under a CC BY-NC licence and has not been peer reviewed; the repository named in its code and data availability statement is not publicly accessible, and no trained checkpoints have been released. Its contribution is methodological — evidence that sequence-only protein language models, which tend to capture local sequence patterns rather than the global dynamics underlying allostery, can approach structure-based allosteric predictors once given minimal functional context. The authors are explicit about the limits: the method applies only where the orthosteric site is known or confidently predicted, precision remains low, and the ASD training set is small and severely imbalanced. Adding dynamic descriptors such as normal modes is the next step they identify.