Chapter 7: Protein Complex Modeling & AlphaFold-Multimer
Chapter 7: Protein Complex Modeling & AlphaFold-Multimer
Most biological functions are executed not by isolated monomeric proteins, but by multi-protein quaternary complexes (dimers, trimers, large macromolecular assemblies). AlphaFold-Multimer extends the AlphaFold architecture to predict multi-chain complexes and protein-protein interactions (PPIs).
In multimer prediction, how does the model know which organism sequence in Chain A's MSA corresponds to which sequence in Chain B's MSA? AlphaFold-Multimer solves this by matching UniProt taxonomic IDs and genetic operon locus distances to pair interacting orthologs.
1. The Interface pTM (ipTM) Metric
In addition to global TM-score, AlphaFold-Multimer computes ipTM (interface predicted Template Modeling score), evaluating confidence specifically at the binding interface between chains:
# Multimer Model Scoring Metric:
# Model Score = 0.8 * ipTM + 0.2 * pTM
#
# ipTM > 0.85: Extremely confident binding interface; highly reliable complex model
# ipTM 0.6 - 0.85: Confident interface fold; exact side-chain contact poses may vary
# ipTM < 0.6: Unconfident interface; likely non-interacting transient proteins or false binding2. Predicting Complexes with ColabFold Multimer
# Fast multimer prediction by separating FASTA sequences with a colon ':'
cat << 'FASTA_EOF' > complex.fasta
>CDK2_CyclinA
MENFQKVEKIGEGTYGVVYKARNKLTGEVVALKKIRL:MRAILVDWLVEVGEEYKLQNETLHLAVNYIDRFLSSMSVPR
FASTA_EOF
colabfold_batch complex.fasta ./colabfold_multimer_out/ --model-type alphafold2_multimer_v3