Protein language model and two classifiers predict where ATP binds on proteins

The work

Technical · from the record

The study builds a sequence-only predictor of ATP-binding residues in proteins. Each residue is encoded as a 1280-dimensional vector by the pre-trained ESM-2 protein language model, and two classifiers trained on those vectors — a Transformer with combined local-window and global self-attention plus a contrastive learning branch, and a deep Q-network that treats each residue call as a reinforcement-learning action — are combined by a weighted sum whose weights were set by five-fold cross-validation to 0.6 and 0.4. On the ATP-41 independent testing set the authors report an AUC of 0.9188 and an MCC of 0.6564, and on ATP-17 an AUC of 0.9353 and an MCC of 0.6625, which they state are higher than the sequence-based methods they compared against on both sets. Residue-level and Pfam-family breakdowns report the highest recognition rates for glycine and lysine and the lowest for cysteine and proline, and the highest family-level MCC for AAA+ ATPases.

How AI was used

Protein chains with annotated ATP-binding residues were taken from four published benchmark sets (ATP-388/ATP-41 and ATP-227/ATP-17). The pre-trained ESM-2 model esm2_t33_650M_UR50D was run over each FASTA sequence to produce a 1280-dimensional embedding per residue, replacing handcrafted sequence features. Two classifiers were then fitted on these embeddings: a Transformer with learnable position embeddings, an eight-head global self-attention module, a local attention module over windows of fixed size, a learnable weight fusing the two attention outputs, a contrastive projection head with temperature-scaled loss, and a hybrid Focal Loss plus MCC-penalty objective optimised with Adam and cosine annealing; and a deep Q-network whose feedforward policy and target networks (512 and 256 hidden units) map a residue state to Q-values for the binary binding/non-binding action, trained with experience replay, an epsilon-greedy policy decaying from 1.0 to 0.01, discount factor 0.95 and target updates every ten episodes. Five-fold cross-validation on the training sets selected the local window size among six candidates and the ensemble weight split; the two classifiers' probability scores were then weighted and summed per residue for the independent testing sets, with the decision threshold chosen by maximising MCC. Ablation variants (traditional, local-only and global-only Transformers, and models with and without the contrastive branch) were trained for comparison, and t-SNE was used to visualise the learned features.