AF2BIND

Predict protein ligand-binding residues from AlphaFold2 pair representations

Input

Target protein

Protein structure to score for ligand-binding residues.

Target chain

Mask sidechains

Mask sequence

Rescale by maximum p(bind)

60 credits

Output

Configure inputs to begin

Set options on the left, then click “Submit job”.

What is AF2BIND?

AF2BIND finds the residues on a protein most likely to contact a small-molecule ligand. It reads a single protein structure and returns a per-residue p(bind) score, so the binding pocket shows up as a cluster of high-scoring residues even when no ligand is present in the input.

The method works by borrowing AlphaFold2's learned sense of which residues sit at interfaces. Instead of docking an actual molecule, it presents the protein with a 20-residue bait made from single amino acids acting as stand-in ligands, then reads how strongly AlphaFold2's internal representation couples each target residue to those probes. Residues that the network treats as interface-like score high. This makes AF2BIND a fast way to localize a pocket before committing to docking or pocket geometry analysis.

AF2BIND scores one chain or compact domain at a time. ProteinIQ accepts proteins up to 300 residues; for anything larger, trim to the domain or interface region of interest before scoring.

How to use AF2BIND online

Upload a protein structure as a PDB file or fetch one from RCSB by ID, pick the chain to score, and AF2BIND returns a ranked table of p(bind) values plus a copy of the structure colored by those scores. The prediction runs on GPU infrastructure with no AlphaFold2 setup, no MSA database, and no local install. The top-ranked residues mark the most probable ligand-binding site.

Inputs

Input	Description
`Target protein`	A single PDB file or an RCSB PDB ID. Must contain protein atoms, up to 300 residues.

Settings

Setting	Description
`Target chain`	Chain to score, by default `A`. Only one chain is scored per run.
`Mask sidechains`	Hides target sidechain atoms from the model so scoring leans on backbone geometry. On by default, matching the AF2BIND default `rm_target_sc=True`.
`Mask sequence`	Hides the target sequence identity so scoring leans on structure alone. Off by default.
`Rescale by maximum p(bind)`	Normalizes the `p(bind)` values written into the structure so the highest-scoring residue maps to the top of the color range. Off by default, which preserves raw scores for coloring.

Masking sidechains and keeping the sequence visible is the configuration AF2BIND was validated with. Turning on sequence masking is useful when testing how much a prediction depends on structure versus sequence, but it is not the recommended starting point.

Results

Output	Contents
`results.csv`	One row per residue with rank, chain, residue number, amino acid, and `p(bind)`. Sorted from highest to lowest score.
`output.pdb`	The input structure with `p(bind)` written into the B-factor column, ready to color in a structure viewer.
`output.zip`	The full prediction bundle for download.

How AF2BIND works

AlphaFold2 builds a pair representation, a tensor that encodes the relationship between every pair of residues across all chains in a complex. When two chains form an interface, the pair representation between their residues carries a distinctive signal that the network learned from real protein structures.

AF2BIND exploits this without needing a real ligand. It adds 20 separate single-residue "bait" chains to the prediction, one for each amino acid type, using the bait sequence ACDEFGHIKLMNPQRSTVWY. Each bait acts as a minimal pseudo-ligand. AlphaFold2 then produces a pair representation linking every target residue to every bait. A small trained logistic-regression layer reads those cross-chain features and outputs p(bind), the probability that a given target residue lies in a ligand-binding pocket.

Because the baits are individual amino acids rather than a specific compound, p(bind) flags where ligands tend to bind in general, not the affinity of any particular molecule.

Interpreting p(bind)

p(bind) is a ranking signal, not a calibrated probability with a universal cutoff. The pocket reveals itself as a tight group of high-scoring residues that sit close together in 3D, not as isolated high values scattered across the sequence.

A practical reading approach:

Sort by p(bind) and inspect the top 10 to 15 residues. In the AF2BIND benchmarks, this top slice captures the binding site for most single-pocket proteins.
Map those residues onto the structure using output.pdb. Residues that cluster spatially are the predicted pocket; lone high scorers far from the cluster are usually noise.
Treat the absolute number with care. A top score near 0.9 on one protein and 0.5 on another can both correctly mark the strongest pocket, since the distribution shifts between targets.

For proteins with more than one pocket, expect more than one spatial cluster. The ranking alone will not separate them, so the structural view matters.

When to use AF2BIND vs alternatives

AF2BIND is strongest when a structure exists and the goal is to localize a small-molecule pocket quickly, especially for a protein where no bound ligand has been resolved. It needs no MSA and no docking box.

For pockets defined by surface geometry and cavity detection rather than a learned interface signal, use fpocket, which is faster and gives explicit pocket volumes and druggability scores.
For predicting protein-protein or protein-DNA binding residues rather than small-molecule sites, use ScanNet.
Once a pocket is identified, estimate how a specific compound binds it with AutoDock Vina.

If the structure itself is uncertain, predict it first with AlphaFold 2, then score the model with AF2BIND.

Related tools

SPRINT

SPRINT (Structure-aware Protein-ligand Interaction) predicts drug-target interactions using co-embedded protein and ligand representations. Screen thousands of compounds against a protein target in seconds.

SMRTnet

Deep learning framework for predicting small molecule-RNA interactions using RNA secondary structure. Combines language models, CNNs, and graph attention networks for binding prediction.

ADMET-AI

Predict ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties from SMILES strings using machine learning models trained on Therapeutics Data Commons datasets.

Admetica

Predict 22 ADMET properties from SMILES strings with the native Admetica Chemprop models from Datagrok.

Brenk filter

Identify toxic, reactive, and pharmacokinetically problematic molecular fragments using structural alert patterns

eToxPred

Predict toxicity and synthetic accessibility of small molecules using machine learning. eToxPred combines toxicity risk assessment with synthetic accessibility scoring to help prioritize drug candidates.

Lead-likeness filter

Screen for lead-like compounds using stricter molecular descriptor criteria than Lipinski or Veber rules for early-stage drug discovery

PAINS filter

Screen compounds for Pan-Assay Interference patterns that cause false positives in biological assays

QEPPI

Quantitative estimate for protein-protein interaction inhibitor potential. Evaluates drug-likeness for compounds targeting PPIs.

ToxPred 2.0 (Toxicity prediction)

Screen compounds for structural toxicity alerts using PAINS, Brenk, and NIH filters. For focused screening, see PAINS Filter, Brenk Filter, or Veber's Rule.