Required inputs

1. Protein complex structure: a mmCIF (.cif / .mmcif) or PDB (.pdb) file for the protein complex of interest.
2. Mutated protein ID (optional): label for the protein your variants apply to. Can be a UniProt accession, gene name, or any alphanumeric label. If left blank it is taken from your variant list when that has ID columns, and otherwise from the selected chain.
3. Partner protein ID (optional): label for its interaction partner.
4. Variants: a list of missense variants.

Variant format

Either a plain list of variants, one per line or separated by spaces or commas, which is taken to mean variants on the mutated protein:

R80G
R174Q
D66V

…or three tab-separated columns per line, no header row, which additionally lets you mutate the partner protein by listing it first:

MUTATED_PROTEIN_ID	VARIANT	PARTNER_ID

• Column 1: ID of the protein carrying the mutation.
• Column 2: Missense variant in single-letter notation, [WT residue][1-based position][MT residue]. Example: V123A (Val→Ala at position 123). Only the standard 20 amino acids are supported.
• Column 3: ID of the interaction partner (the other protein).

Only missense variants are supported. Nonsense, frameshift and indel notation, and synonymous changes where the WT and mutant residues are identical, are rejected. If a list contains unusable entries, all of them are reported together so it can be cleaned in a single pass.

Example

O00548	A653T	P46531
O00548	R661S	P46531
O00548	N34I	P46531
P46531	W89R	O00548
P46531	E100K	O00548

The first three rows specify mutations on O00548. The last two specify mutations on P46531 Note that the mutated protein is listed first.

Obtaining a structure file

• AlphaFold3 Server: alphafoldserver.com generates predicted complex structures in mmCIF format.
  AlphaFold3 outputs are subject to Google DeepMind’s Output Terms of Use (non-commercial use only).
• Protein Data Bank: experimental structures from rcsb.org in mmCIF or PDB format.
• Other prediction tools: AlphaFold-Multimer, RoseTTAFold, or any tool producing mmCIF or PDB output.

The structure must contain chains labeled A and B.

Output

A tab-separated file with one row per variant:

interactor	partner	mutation	score
O00548	P46531	A653T	0.0852
O00548	P46531	R661S	0.0680
O00548	P46531	N34I	0.3541

• interactor: the mutated protein
• partner: its interaction partner. These are separate columns because an accession can itself contain an underscore, which makes a single joined identifier ambiguous.
• mutation: variant notation as provided
• score: predicted probability of interaction disruption (0–1). Scores ≥ 0.5 indicate predicted disruption.

Variant position numbering

Two conventions are accepted, and by default the server works out which one your positions use:

• Sequential: the first residue present in the file is position 1. AlphaFold3 models are naturally in this form.
• Structure numbering: the residue numbers written in the PDB/mmCIF file, which need not start at 1 and may skip positions. Experimental PDB entries are often in this form.

When residue numbers already start at 1 and run consecutively the two are identical. Where they differ, auto-detection picks whichever interprets every one of your variants consistently with the structure sequence; if neither does, the error lists what failed under each convention along with each chain’s actual residue range. Results are reported back in the convention you submitted, and the downloaded file carries both that and the sequential index the model actually scored.

Chain selection

MutPred-PPI scores a two-chain complex. If your structure contains more than two chains, such as extra copies of the complex, DNA, ligands or waters, pick which chain is the mutated protein and which is its partner. Only those two are scored; everything else is discarded, and non-protein chains cannot be selected.

Model version

Jobs run on v1.2 by default: the current model, retrained and released 15 September 2026.

v1.0 is the published model: the ten-model ensemble described in the RECOMB 2026 paper and in bioRxiv preprint versions 1 and 2. It remains selectable on the submission form so the results reported in those papers can be reproduced exactly. If you are checking or extending published numbers, select v1.0.

The two versions use different architectures and training data, so their scores are not directly comparable and should not be pooled or compared across versions. Every results page records the version that produced it.