MPBind

Proteins do their work at interfaces. The same chain may present one surface to a partner protein, another to DNA, a metal-coordinating pocket somewhere else, and a hydrophobic patch where a lipid inserts. Computational tools have mostly attacked these one at a time: a DNA-binding predictor here, a metal-site predictor there, each trained on its own curated set with its own installation and output format. That fragmentation is inconvenient, and it throws away signal, because the structural cues marking one kind of interface are partly shared with the others.

MPBind, from Jianlin Cheng's lab at the University of Missouri, predicts all five classes in a single pass. Given a 3D protein structure it labels every residue with the probability of contacting a protein, a nucleic acid, an ion, a small-molecule ligand, or a lipid, trained jointly across the five tasks so that each benefits from the others. The architecture pairs frozen protein language model embeddings — which carry evolutionary and functional context that coordinates alone do not — with an equivariant graph neural network that reads the geometry directly. It was published in Bioinformatics in 2025, following an April 2025 preprint.

The design descends from PeSTo, the parameter-free geometric transformer whose data splits MPBind reuses exactly, and adds the language model channel that PeSTo omits.

#Key Features

• Five partner classes, one model: Protein, DNA/RNA, ion, ligand, and lipid binding sites come from one multitask head, and users select which to compute per run instead of installing five separate tools.
• Language model plus geometry: Residue features combine embeddings from ProtT5-XL-UniRef50 in the ProtTrans suite and from ProstT5, whose structure-aware training gives a second, complementary view of the same sequence.
• Equivariant by construction: The graph network updates node features from relative positions and distances alone, so predictions are unchanged by rotating or translating the input structure and no augmentation is needed to teach that invariance.
• Works on predicted structures: The authors annotate 23,391 human proteins from the AlphaFold Protein Structure Database, and the shipped examples run on AlphaFold 3 models, so experimental coordinates are not a prerequisite.
• Checkpoints in the repository: Five trained .pt files sit in the repo's weight/ directory under GPL-3.0, with version 2 recommended for inference.

#Technical Details

Each residue is described by ProtTrans embeddings, ProstT5 amino-acid embeddings, DSSP secondary structure, and atomic and geometric residue features; an ablation showed the ProstT5 amino-acid channel raised protein-protein ROC AUC to 0.84 and PR AUC to 0.62, while the 3Di structural-alphabet channel did not help. Four EGNN blocks update node features, and a multitask MLP head emits a score between 0 and 1 per binding type. Training follows the PeSTo split of PDB chains clustered at 30% sequence identity: 376,216 training chains, 101,700 for validation, and 97,424 held out, from which 40,651 chains under 8,192 atoms form the Test1 benchmark.

On Test1, MPBind reaches ROC AUC 0.83 and PR AUC 0.54 for protein-protein interfaces against 0.76 and 0.38 for PeSTo trained on identical data. A stricter benchmark, Test2, holds 1,452 chains released between 2022 and mid-2024 with at most 30% identity to any training protein; on its 1,208 protein-interacting chains MPBind scores 0.84 ROC AUC and 0.62 PR AUC, ahead of PeSTo at 0.78 and 0.48 and of the task-specific ScanNet by 0.01 and 0.03. Across the other partner classes on Test2, ROC AUC is 0.8116 for DNA/RNA, 0.7579 for ions, 0.7366 for ligands, and 0.7409 for lipids, exceeding both PeSTo and the task-specific CLAPE, LMetalSite, and GraphBind.

#Applications

MPBind suits any workflow that needs interface annotation before an experiment: choosing residues to mutate in an interaction study, mapping the nucleic-acid contact surface of a putative transcription factor, flagging metal sites in a structural genomics target, or scanning a predicted proteome for druggable ligand pockets. Predictions are written into the B-factor column of an output PDB file, so results color directly in PyMOL or ChimeraX. The authors' human-proteome run, deposited on Zenodo alongside the model, is a ready-made resource for cross-referencing predicted interfaces with UniProt annotations.

#Impact

MPBind makes a clean case that multitask training plus language model features can beat specialists on their own tasks, the pattern that has reshaped protein prediction more broadly. Its release is practical for reuse: code, trained checkpoints, dataset construction scripts, and the human-proteome predictions are all public. Two limits are worth stating. The model needs 3D coordinates, so proteins without experimental structures must be modeled first, and prediction quality inherits any error in that model. And accuracy is uneven across partner classes — protein and nucleic-acid interfaces are well resolved, while ligand and lipid sites sit near 0.74 ROC AUC, reflecting how few labeled lipid-binding chains exist. The GPL-3.0 license is copyleft, which constrains embedding the code in closed pipelines.