AlphaFold-Multimer Tutorial: Predict Protein-Protein Interaction Structures with ColabFold
Protein-protein interactions drive virtually every cellular process — but their 3D structures are notoriously hard to determine experimentally. AlphaFold-Multimer changed that. This tutorial covers what it does, when it’s the right tool, how to run it free with ColabFold, and — critically — how to tell whether the predicted interface is trustworthy.
What AlphaFold-Multimer is and how it differs from AF2
AlphaFold-Multimer is an extension of AlphaFold2 specifically trained to predict the structures of protein complexes — two or more chains that interact to form a single functional assembly. Released in 2021 and updated in subsequent versions, it processes multiple protein sequences simultaneously and predicts how they fold together, including the binding interface between chains.
- One protein sequence → one 3D structure
- Confidence: pLDDT per residue + PAE within chain
- Cannot predict binding interfaces between proteins
- Treats each chain as independent
- Pre-computed models in AFDB for 200M+ proteins
- Multiple sequences → complex 3D structure
- Confidence: ipTM (interface), pTM (overall) + inter-chain PAE
- Predicts which residues form the binding interface
- Generates multiple stoichiometry configurations
- Available via ColabFold — same notebook, different input format
The key architectural difference is in training: AlphaFold-Multimer was trained on experimentally determined protein complex structures from the PDB, learning the co-evolutionary signals that distinguish true binding interfaces from non-interacting surfaces. It jointly processes the MSAs for all input chains, allowing co-evolutionary information between the chains to inform the interface prediction.
When to use it — and when not to
| Scenario | AF-Multimer? | Notes |
|---|---|---|
| Predicting a known stable protein-protein complex | Yes — ideal | Homodimers, heterodimers, enzyme-substrate complexes, receptor-ligand binding proteins |
| Mapping a binding interface for mutagenesis design | Yes | Identify interface residues, design mutations to disrupt or enhance binding |
| Characterizing a transient signaling interaction | Possible | Accuracy drops for weak/transient interactions — check ipTM carefully |
| Large assemblies (>5 chains or >3,000 residues total) | Challenging | Accuracy degrades; computational cost increases; consider AlphaFold3 for large assemblies |
| Confirming whether two proteins interact at all | Partial | ipTM > 0.75 suggests interaction; not reliable as sole evidence — use with biochemical data |
| Protein-DNA or protein-small molecule complexes | No | Use AlphaFold3 server — AF-Multimer only handles protein chains |
Step 1 — Prepare your sequences
ColabFold uses a simple colon-separated format to specify multiple chains. Paste the sequences of all chains into the query sequence field, separated by colons. Each sequence should be the mature protein sequence — remove signal peptides and propeptides first.
# Chain A sequence : Chain B sequence
MKTAYIAKQRQISFVKSHFSRQLEERLGL...KSTVEAI:MGSSHHHHHHSSGENLYFQGHM...RRFVSSFor a homodimer (two copies of the same protein), repeat the sequence with a colon between copies. For a heterotrimer, separate three sequences with colons.
Name the job descriptively — include both protein names and the stoichiometry. Use underscores, not spaces: EGFR_HER2_heterodimer rather than EGFR HER2.
Step 2 — Run with ColabFold
The ColabFold notebook automatically detects multimer input when it sees a colon in the sequence field and switches to AF-Multimer mode. Most settings remain the same as monomer prediction — but check these multimer-specific options:
- num_models = 5 — keep at 5. For complexes, model agreement across the 5 predictions is even more informative than for monomers. Five models that all agree on the interface geometry is a strong confidence signal.
- num_recycles = 3 — the default. Increasing to 6 can improve accuracy for difficult complexes at the cost of longer runtime.
- use_dropout = False — leave unchecked for standard predictions. Enabling it generates conformationally diverse models, useful for exploring interface flexibility.
- MSA mode: mmseqs2_uniref_env — the recommended default. For each chain, ColabFold searches for homologs independently and also performs paired MSA search to capture inter-species co-evolution between the two chains — a key signal for genuine binding partners.
Runtime for a typical heterodimer (500 + 300 residues) is 20–40 minutes. The notebook outputs the same files as monomer prediction — PDBs ranked by confidence score — plus additional complex-specific confidence metrics in the JSON files.
Step 3 — Interpret ipTM and pTM scores
AlphaFold-Multimer outputs two composite confidence scores alongside the per-residue pLDDT:
pTM (predicted TM-score) — estimates the overall structural accuracy of the entire complex, including all chains. Ranges from 0 to 1. Analogous to pLDDT for the whole complex. A pTM above 0.5 suggests the overall fold is predicted reliably.
ipTM (interface predicted TM-score) — estimates the accuracy of the interface between chains specifically. This is the key number for protein-protein interaction predictions. Ranges from 0 to 1. Higher ipTM means higher confidence that the predicted binding interface is correct.
ColabFold ranks the five models by a combined score: 0.8 × ipTM + 0.2 × pTM. The rank_1 model maximizes this combined score and is your primary result.
_scores_rank_*.json files. To read them: import json; d = json.load(open('scores_rank_001.json')); print(d['iptm'], d['ptm']). The JSON also contains the full PAE matrix, which you need for the next step.
Step 4 — Read the inter-chain PAE plot
For complexes, the PAE matrix is larger — it covers all residues of all chains combined. The matrix is divided into quadrants: the diagonal blocks represent intra-chain confidence (same as monomer PAE), and the off-diagonal blocks represent inter-chain confidence — AlphaFold’s certainty about the relative position of residues in Chain A versus Chain B.
The off-diagonal PAE blocks are the key diagnostic for interface confidence. Dark blue off-diagonal blocks mean AlphaFold is confident about the relative orientation of the two chains — they are predicted to interact in a specific, well-defined geometry. Light or white off-diagonal blocks mean the chains’ relative positions are uncertain — they may or may not interact, or they may interact in multiple ways.
Step 5 — Analyze the predicted interface
With the top-ranked complex PDB downloaded, load it into PyMOL to analyze the interface. Color the two chains differently to immediately see the binding interface:
# Load and visualize the complex
load jobname_relaxed_rank_001.pdb
show cartoon
color slate, chain A
color tv_orange, chain B
set cartoon_fancy_helices, 1
# Select interface residues (within 5 Å of the other chain)
select chainA_interface, chain A and (byres chain B around 5)
select chainB_interface, chain B and (byres chain A around 5)
# Show interface residues as sticks
show sticks, chainA_interface or chainB_interface
color cyan, chainA_interface
color salmon, chainB_interface
# Measure H-bonds at the interface
distance hbonds, chain A, chain B, mode=2 # mode=2 = H-bonds onlyRecord which residues form the interface and what interactions they make. Cross-reference with literature — if published mutagenesis data identifies key interface residues and your prediction places them at the interface, this is strong validation. If important known interface residues are absent from the predicted contact set, treat the model with more caution.
Check model agreement across the 5 predictions
Load all five ranked models and superimpose them on Chain A. Check whether the predicted position of Chain B is consistent across models:
# Load all 5 models
load jobname_relaxed_rank_001.pdb, model1
load jobname_relaxed_rank_002.pdb, model2
load jobname_relaxed_rank_003.pdb, model3
# Align on Chain A of model 1, check where Chain B lands
align model2 and chain A, model1 and chain A
align model3 and chain A, model1 and chain AIf all five models place Chain B in the same general position (low RMSD between Chain B positions after aligning on Chain A), the interface prediction is robust. If Chain B is positioned differently across the five models, the interface is genuinely uncertain and the ipTM may be inflated.
Limitations to report
AlphaFold-Multimer is powerful but has known limitations that must be acknowledged when publishing results:
- Cannot distinguish binding modes. If two proteins have multiple possible binding interfaces — allosteric sites, secondary binding sites, pH-dependent rearrangements — Multimer typically predicts one. Biochemical data is required to determine which mode is biologically relevant.
- Lower accuracy for large complexes. Accuracy decreases as the number of chains and total residues increases. Predictions for large assemblies (ribosomes, proteasomes, large signaling complexes) are less reliable than for simple dimers.
- Does not account for conformational change upon binding. Some proteins undergo significant structural rearrangement when they bind a partner. Multimer predicts a single conformation of the complex; it may not capture the conformational dynamics of the binding event.
- Paired MSA quality matters. For distantly related proteins without many shared homologs, the paired MSA that drives interface prediction is sparse, and accuracy suffers. The coverage plot in the output shows how well paired MSA coverage is — low paired coverage predicts lower interface accuracy.
AlphaFold-Multimer in one paragraph
AlphaFold-Multimer predicts protein complex structures by jointly processing multiple protein chains and learning from co-evolutionary signals between them. Run it via the standard ColabFold notebook by separating chain sequences with colons. The key confidence metric is ipTM — above 0.75 indicates a reliable interface, below 0.40 means the prediction is not trustworthy. Always check the inter-chain PAE plot: dark off-diagonal blocks confirm the interface is confidently placed; light blocks mean the relative chain orientation is uncertain. Cross-validate interface contacts with mutagenesis data, check all five model predictions for consistency, and acknowledge the limitations when publishing. A high-confidence Multimer prediction is a strong hypothesis generator — not a replacement for experimental validation of the interface.