Model labels which protein residues are active sites and what job they do

The work

Technical · from the record

M3Site is a model that labels each residue of a protein as a non-active site or as one of several active site functional categories, using three inputs at once: a sequence embedding from a protein language model, a graph of the predicted 3D structure encoded by an equivariant graph neural network, and a text description of the protein's function encoded by a biomedical language model. The authors assembled a dataset of 25 883 proteins from Swiss-Prot and the AlphaFold database, and used OpenAI's o1 model plus expert review to consolidate raw active site annotations into six functional categories. On the split clustered at 10% sequence similarity, the paper reports higher scores for M3Site than for the protein representation learning baselines it retrained, across the seven metrics it tabulates; on 33 proteins deposited after the dataset cut-off it reports an F1 of 0.912, and on a held-out Glycosyl-Hydrolase family split an F1 of 0.884. A Gradio application for prediction and visualisation accompanies the code and dataset.

How AI was used

An LLM (OpenAI o1, accessed via API) was prompted to cluster raw UniProt active site function annotations into 5–10 biologically meaningful groups and flag ambiguous cases; domain experts then refined these into six categories used as class labels, and MMSeqs2 sequence clustering at thresholds of 10% to 90% plus an 8:1:1 split produced the training, validation and test sets. Each protein was represented three ways: sequence embeddings from a pretrained protein language model (ESM-3 in the main configuration), a residue graph with edges between Cα atoms closer than 8 Å and 62-dimensional node features, encoded by a two-layer equivariant graph neural network, and a function description embedded with a biomedical language model (PubMedBERT-abs) and globally average pooled. A symmetric cross-attention module (FunICross) fuses the modalities in two branches, and an adaptive weighted fusion mechanism combines the fused and original features with a learned sigmoid coefficient. The classifier was trained end to end for 100 epochs with Adam at a learning rate of 5e-5, warm-up over 10% of steps and cosine annealing, under a loss combining weighted cross entropy with a class-centre loss and an inter-centre margin loss. Baselines, and variants substituting other protein and biomedical language models, were run by the authors with additional transformer layers and a classification head appended for comparable trainable parameter counts. The trained model was then run on the held-out split, on proteins deposited after the dataset cut-off date, on a Glycosyl-Hydrolase family hold-out, and inside a Gradio interface that takes a user's structure and function description.