Use AF2BIND Online
Commercially Available AF2BIND No-Code Web Server
AF2BIND: Predicting Small-Molecule Binding Residues with the AlphaFold2 Pair Representation
AF2BIND (AlphaFold2 Bait-Informed Neural Descriptor) is a method for predicting which residues of a protein bind small molecules, using only the target protein as input. It was developed by Artem Gazizov, Anna Lian, Casper Alexander Goverde, Sergey Ovchinnikov and Nicholas F. Polizzi. It is a logistic regression model trained on AlphaFold2 internal pair representations, and its code and dataset are available on GitHub under the MIT License.
Finding ligand-binding sites is a first step in function annotation and drug discovery. The authors present AF2BIND as a way to predict binding residues without a multiple sequence alignment, a homology model or knowledge of the true ligand. They report that the AlphaFold2 pair representation is expressive enough for this task and that it outperforms single-sequence ESM2 and ESM1-IF representations.
Property | Detail |
|---|---|
Task | Per-residue probability of contacting a small-molecule ligand, P(bind) |
Inputs | Target sequence, target backbone as template, and 20 bait amino acids |
Model | Logistic regression on 5,120 features per residue (20 x 2 x 128) |
Training set | 1,902 proteins and 2,110 ligands curated from the PDB (accessed March 2023) |
Binding label | Residue with any heavy atom within 5 angstroms of the ligand |
License | MIT (code repository) |
How AF2BIND Finds Binding Residues Using Bait Amino Acids
Bait amino acids: one of each of the 20 canonical amino acids is appended to the target as disconnected chains with large residue-index offsets.
Single AlphaFold2 pass: features are extracted in one pass with no recycling, and sidechain dihedrals beyond the beta carbon are masked.
Pair features: pair embeddings between each target residue and the 20 baits are concatenated into 5,120 features.
Classifier: logistic regression converts the features into a binding probability for each residue.
Interpretability
Because each bait contributes separately, the model's output can be attributed to individual bait amino acids. The authors report that hydrophobic baits such as W and F track nonpolar ligands, while polar baits such as Q and N rise for polar ligands, which gives a rough estimate of ligand polarity compatible with a pocket.
AF2BIND Results: Binding-Residue Recovery
Evaluation | Result |
|---|---|
AF2 pair representation alone | 66% average binding-residue recovery |
Combined with ESM2 and ESM1-IF embeddings | About 69% recovery, with the best AUC-ROC |
Backbone robustness (four mu opioid receptor structures) | Mean per-residue spread in P(bind) of 0.02 |
Recovery is measured as the fraction of true binding residues found among the top-N predictions, where N is the number of true binding residues. The authors illustrate the method on the human mu opioid receptor and the BRD4 bromodomain, both from classes held out from training.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with AF2BIND on Tamarind Bio
Ligand-free pocket identification: find likely small-molecule binding residues on a protein with no known ligand.
Predicted and unbound structures: the method is insensitive to sidechain rotamers, which the authors note helps with unbound or predicted structures.
Pocket chemistry: use bait contributions to judge whether a pocket favors polar or nonpolar ligands.
Checking ligand placement: the authors suggest the predictions could flag poorly supported ligand placements in crystal structures.
How to Use AF2BIND on Tamarind Bio
Log in: sign in at tamarind.bio and open the AF2BIND tool.
Provide the target: supply a protein structure, such as a PDB ID and chain or an uploaded structure, if exposed.
Select the chain: choose the single chain to analyze; training covered single chains up to 500 residues.
Set masking options: choose sidechain or sequence masking, if exposed.
Run the job: submit it and wait for the predictions.
Download outputs: retrieve the per-residue binding probabilities.
Interpret: rank residues by P(bind) and map the top residues onto the structure to locate the likely pocket.
Follow up: use the predicted site to guide docking or design.
Things to Keep in Mind
The logistic regression model is intentionally simple, and the authors note room for improvement.
The ceiling for recovery is unclear but likely well below 100%, since the true ligand is not used.
A fixed set of 20 baits may dilute the signal as the number of binding sites grows.
Training covered only single-chain proteins up to 500 residues.
The test set is small, at about 70 complexes.
Some training labels come from poor-quality crystallographic data, despite filtering.
Source: Gazizov A, Lian A, Goverde CA, Ovchinnikov S, Polizzi NF. AF2BIND: Predicting ligand-binding sites using the pair representation of AlphaFold2 (bioRxiv preprint). DOI: 10.1101/2023.10.15.562410. Code and dataset: github.com/sokrypton/af2bind.