AF2BIND

AF2BIND webserver overview

Many drug targets have a structure but no ligand bound to it, so the pocket has to be found before anything can be docked into it. AF2BIND finds it from the protein alone. It scores every residue with p(bind), the probability that the residue lines a small-molecule binding site, and a pocket shows up as a cluster of high-scoring residues that sit together in the fold.

ProteinIQ runs AF2BIND 1.1 on ColabDesign 1.1.1, following the standard AF2BIND notebook. One job scores one structure and returns a ranked residue table, the structure with p(bind) written into its B-factor column, an amino-acid bait activation heatmap and the native AF2BIND files. The method is described by Gazizov et al. in Nature Methods (2026).

Pricing

Each AF2BIND job costs a flat 60 credits. The selected chains, the number of residues scored and every setting leave the price unchanged. The exact quote is shown before submission.

Inputs

Input	Accepted formats	Limits and requirements
Target protein	.pdb file, or a four-character RCSB PDB ID to fetch	One structure, up to 50 MiB. For multi-model files, AF2BIND reads the first model. The selected chains can hold up to 300 residues in total.

AF2BIND scores protein residues only. Ligands, waters and other HETATM records in the file are ignored, so a structure solved with its ligand is scored as if the pocket were empty. Common modified amino acids such as selenomethionine (MSE) and phosphoserine (SEP), and any residue mapped to a standard amino acid by a MODRES record, count as protein.

Every residue AF2BIND processes needs its N, CA and C backbone atoms. AF2BIND keeps any residue that has an N atom, and without CA or C it would write that residue's atoms at the origin. ProteinIQ therefore rejects the structure and lists the incomplete residues. PDBFixer can rebuild missing atoms, or the residues can be removed.

The 300-residue limit counts only the selected chains, so a large complex is fine when the scored chain is small. For a bigger protein, trim the structure to the domain of interest first.

API submissions can also give a UniProt accession instead of a file. ProteinIQ then downloads the matching predicted structure from the AlphaFold Database. The browser fetch dialog uses RCSB only.

Settings

Parameter	Type	Default	Description
Target chain	string	A	One chain ID or a comma-separated list such as A,B, each listed once. Selected chains are scored together as one target.
Mask sidechains	boolean	true	Removes target sidechains from the model input and uses the AF2BIND classifier trained without sidechains.
Mask sequence	boolean	false	Hides the target's amino-acid identities so the prediction relies on backbone structure.
Activation heatmap cutoff	enum	0.5	0, 0.28, 0.5 or 0.9. Only residues with p(bind) above this value appear in the activation heatmap and activation.csv.
Rescale by maximum p(bind)	boolean	false	Divides the B-factor scores in the output structure by the highest p(bind), so the top residue gets 100.

The defaults match the AF2BIND notebook: sidechains masked and sequence visible. This is the configuration to start with. Masking the sequence as well is a way to test how much a prediction depends on residue identity rather than shape.

The heatmap cutoff and rescaling change only how results are presented. p(bind) scores and results.csv are the same whichever values are chosen.

Outputs

The result opens on four tabs. Structure shows the scored structure, Results lists every residue ranked by p(bind), Activation holds the bait heatmap and the top-residue PyMOL selection, and Files lists the downloads.

File	Contents
results.csv	AF2BIND's residue table in input order, with chain, resi, resn (one-letter amino acid) and p(bind) after a leading row-index column.
viewer.pdb	The structure displayed in ProteinIQ and passed to workflow steps. It keeps AF2BIND's coordinates and B-factor scores and leaves out atom records that do not match the uploaded structure.
output.pdb	The structure as AF2BIND wrote it: input coordinates for the scored residues, with p(bind) × 100 (or the rescaled value) in the B-factor column.
output.zip	AF2BIND's download bundle with output.pdb and results.csv.
activation.csv	Bait contributions for each residue above the heatmap cutoff: one row per residue, one column per amino-acid bait.
source.log	Text output from the AF2BIND run, including its own PyMOL selection command.

In workflows, results.csv is available as binding-site scores and viewer.pdb as the scored protein structure.

Understanding results

p(bind)

p(bind) runs from 0 to 1. The AF2BIND paper recommends p(bind) ≥ 0.28 as the general cutoff for calling a residue part of a binding site. At that threshold its benchmark reached 67% sensitivity, a 4.3% false-positive rate and 63% precision. The AF2BIND notebook uses 0.5 as its default heatmap cutoff and draws sidechains for residues above 0.5 in its structure view.

A single high score is weak evidence. A real pocket appears as several high-scoring residues that sit next to each other in 3D, even when they are far apart in sequence. Lone high scorers far from any cluster are usually noise. A protein with two pockets gives two clusters, and the ranked table mixes them together, so check where the top residues sit in the structure before deciding how many sites there are.

p(bind) marks where small molecules tend to bind. It is not an affinity and says nothing about which compound fits. A low score across the whole protein can mean there is no pocket, or that the pocket only forms when a ligand binds and is closed in the input structure. Repeat runs of the same structure can give slightly different scores, so compare which residues rank highly rather than their exact values.

Ranked table and structure

Results ranks every scored residue from highest to lowest p(bind). Residue numbers follow the uploaded file. Start with the residues at or above 0.28, then read down the top 10 to 15 and look at how they cluster.

The Structure tab opens colored by chain. To color by score, download viewer.pdb and color it by B-factor from 0 to 100, for example with spectrum b, red_white_blue, minimum=0, maximum=100 in PyMOL. This matches the AF2BIND notebook view, where red is low p(bind) and blue is high. With rescaling on, the B-factors are relative to the top residue, so read raw scores from results.csv.

The Activation tab gives a PyMOL command that selects the 15 top-ranked residues. For one chain it has the same form as the notebook's own command, select chA, resi 45 + resi 48 + .... When the top residues span several chains, it names each chain, as in select af2bind_top, (chain A and resi 45+48) or (chain B and resi 12).

Activation heatmap

AF2BIND appends a 20-residue bait, one of each amino acid, and reads how AlphaFold2 couples every target residue to each bait. For each residue, the 20 bait values add up to the score that the logistic function turns into p(bind). A positive value pushes that residue toward binding and a negative value pushes it away.

The heatmap has one column per residue above the cutoff, labeled chain_residue, and one row per bait. Rows follow the AF2BIND order CSTAGPDEQNHRKMILVWYF, which groups similar amino acids. Blue cells favor binding, red cells oppose it, and the color saturates at ±1. These values explain the model's signal for each residue. They do not identify a preferred real ligand. An empty heatmap means no residue passed the cutoff, so lower the cutoff to see more.

How AF2BIND works

AlphaFold2 builds a pair representation, a tensor that encodes the relationship between every pair of residues in a complex. Where two chains form an interface, the pair features between their residues carry a signal that the network learned from real protein structures.

AF2BIND uses that signal without a real ligand. It adds the bait sequence ACDEFGHIKLMNPQRSTVWY as a second chain and spaces its residues 50 positions apart, so each amino acid behaves like a separate single-residue probe. AlphaFold2 then predicts the target together with the baits, using the input structure as a template, and AF2BIND passes the pair features between each target residue and the 20 baits to a trained logistic regression that outputs p(bind). No multiple sequence alignment is used.

When to use AF2BIND vs alternatives

AF2BIND works best when a structure exists and the goal is to find a small-molecule pocket quickly, especially on a protein with no resolved ligand. It needs no MSA and no docking box, which makes it a useful first step for blind docking.

fpocket finds pockets from surface geometry instead of a learned interface signal and reports explicit pocket volumes and druggability scores, so the two make a useful cross-check. For protein-protein or protein-DNA binding residues rather than small-molecule sites, use ScanNet. Once a pocket is located, AutoDock Vina can dock a specific compound into it, with the top AF2BIND residues defining the box.

When no experimental structure exists, predict one with AlphaFold 2 or ESMFold and score the model with AF2BIND. Consider trimming long low-pLDDT stretches first. AF2BIND's own large-protein notebook removes stretches of 7 or more residues with pLDDT below 70 before scoring.

Sources

• Gazizov A, Lian A, Goverde C, Mou J, Ovchinnikov S, Polizzi NF. AF2BIND: predicting small-molecule binding sites using the pair representation of AlphaFold2. Nature Methods (2026).
• AF2BIND source code on GitHub.