Google DeepMind
DeepMind's extension of AlphaFold 2 for predicting protein complex structures, using paired MSA processing and ipTM scoring to model multimeric assemblies.
AlphaFold-Multimer is an extension of AlphaFold 2 developed by Richard Evans, Michael O'Neill, Alexander Pritzel, Natasha Antropova, Andrew Senior, Tim Green, Augustin Zidek, Russ Bates, Sam Blackwell, Jason Yim, and colleagues at Google DeepMind, with a preprint deposited on bioRxiv in October 2021. Where AlphaFold 2 was trained and optimized for single-chain protein structure prediction, AlphaFold-Multimer specifically addresses the challenge of predicting how multiple protein chains come together to form functional complexes. Protein complexes — from simple homodimers to large hetero-oligomeric assemblies — underlie virtually all of molecular biology, and accurately predicting their structures poses distinct challenges beyond what single-chain folding requires.
The core problem AlphaFold-Multimer addresses is that the original AlphaFold 2 architecture, while it can be applied to multi-chain inputs in an ad hoc fashion, was not designed to encode cross-chain evolutionary information or to reason about the symmetry and permutational equivalence of identical subunits. Protein complexes have coevolved interfaces: residues on one chain that contact a partner chain tend to vary in correlated ways across evolution, and this co-evolutionary signal encodes the geometry of the interface. Extracting this signal requires carefully pairing sequences from the same organism across the chains of a heteromeric complex in the MSA — a non-trivial alignment task that the standard AlphaFold pipeline does not perform.
AlphaFold-Multimer was released as part of the same open-source AlphaFold GitHub repository used for AlphaFold 2, making it immediately accessible to the structural biology community with the same infrastructure. It went on to become one of the most widely used tools for protein complex structure prediction, serving as the baseline that subsequent multi-chain prediction methods including AF2-Multimer v2 and v3 improvements, AFsample, and AlphaFold 3 were benchmarked against and built upon.
AlphaFold-Multimer shares the same fundamental architecture as AlphaFold 2 — a 93-million-parameter network built around the 48-block Evoformer stack and the Invariant Point Attention-based Structure Module — but introduces several targeted modifications for multi-chain inputs. The most critical is MSA pairing: for heteromeric complexes, sequences from the MSAs of individual chains are paired by matching organisms using UniProt species annotation. When multiple sequences from the same species exist for a chain, candidates are ranked by similarity to the respective target sequence, and pairs of equal rank are concatenated into cross-chain rows of the MSA. Both paired (cross-chain) and unpaired (single-chain) MSA rows are used at training and inference, allowing the model to draw on both intra-chain evolutionary information and cross-chain co-evolution signals simultaneously.
The model also modifies residue cropping during training to preferentially sample interface residues, ensuring that interface contacts are well-represented in training batches rather than diluted by large internal regions. Loss function modifications include terms for inter-chain distances and a permutation-equivariant formulation that allows the model to correctly score symmetric assemblies without being sensitive to arbitrary chain ordering in the input. At inference, the model outputs five candidate complex structures per submission (sampled from different random seeds), ranked by the composite ipTM + pTM confidence score. On benchmark datasets of heterodimers without templates, AlphaFold-Multimer achieves at least medium accuracy (DockQ ≥ 0.49) on approximately 70% of cases and high accuracy (DockQ ≥ 0.8) on approximately 26%. Performance is generally stronger for homomers than heteromers, consistent with the greater availability of co-evolutionary signal within a single chain's MSA for homomeric interfaces.
AlphaFold-Multimer is broadly used in any context where the structure of a protein complex is needed and experimental data are unavailable or insufficient. Drug discovery teams use it to model drug target proteins in their biologically relevant oligomeric state — a membrane receptor dimer, a protease-inhibitor complex, or an enzyme-cofactor assembly — where monomeric models may be misleading. Structural biologists use AlphaFold-Multimer predictions as molecular replacement models for phasing X-ray crystallography data collected on multi-chain assemblies and as initial models for fitting cryo-EM density maps of large complexes. Biologists studying protein-protein interaction networks use it to generate structural hypotheses for binary interactions identified in pull-down or two-hybrid experiments, prioritizing pairs for follow-up experimental validation. In immunology, AlphaFold-Multimer is widely applied to predict antibody-antigen complexes, MHC-peptide-TCR ternary complexes, and cytokine-receptor assemblies. Systems biologists model entire pathway-relevant assemblies — signaling complexes, scaffolding protein networks — to understand how mutations at one chain's interface alter the geometry of adjacent chains.
AlphaFold-Multimer substantially expanded the scope of accurate computational structural biology from individual proteins to multi-chain assemblies, enabling structural hypotheses about protein-protein interactions at proteome scale. Its release alongside the AlphaFold 2 open-source codebase ensured immediate, broad adoption with minimal infrastructure barrier. The ipTM confidence metric introduced by AlphaFold-Multimer has become a standard for evaluating predicted complex quality and is routinely reported alongside multimer predictions in the literature. The model served as the primary reference for CASP15's multimer prediction challenge (2022), where teams built upon and compared against it extensively. Subsequent improvements — AlphaFold-Multimer v2 (2022) and v3 updates incorporated into the AlphaFold GitHub releases — further improved interface accuracy, particularly for difficult cases with sparse MSAs. AlphaFold 3 eventually superseded AlphaFold-Multimer for many use cases by adopting a diffusion-based architecture with broader molecular coverage, but AlphaFold-Multimer remains the most widely cited and deployed tool for protein complex prediction due to its established open-source availability, well-characterized performance, and direct integration with the AlphaFold ecosystem. Key limitations include reduced accuracy for complexes with transient or flexible interfaces, for very large assemblies where chain cropping at training limits coverage of distant inter-chain contacts, and for complexes between proteins with sparse MSAs where the co-evolutionary signal that drives interface prediction is weak.