2026-07-29
On 439 dimers, shuffling MSA pairings leaves AlphaFold3 accuracy nearly unchanged (0.612 vs 0.613); depth, not pairing, is what drives complex prediction.
AlphaFold-Multimer and AlphaFold3 pushed protein complex structure prediction to striking accuracy. Their standard recipe is paired MSA: aligning sequences from different species so that co-evolving partners are matched, letting the model read off inter-chain co-evolutionary signals. A wave of follow-up work (ESMPair, DeepSCFold) has tried to pair sequences more precisely.
What nobody had measured systematically is how much that pairing step actually buys you. Maybe the depth of the MSA, meaning how many homologs you have, is what matters, and pairing is decorative. This Shandong University paper builds a clean benchmark to settle it.
The authors assemble HD439: 439 heterodimers (two different chains bound together), with sequences similar to the AlphaFold3 training set removed so the model cannot memorize answers. Quality is scored with DockQ (0 to 1, higher is better).
The core experiment compares four MSA strategies: mMSA is each chain's monomer MSA alone with no cross-chain pairing (baseline); pMSA is the standard cross-chain paired MSA, AF3's default; sMSA takes the paired sequences from pMSA and randomly shuffles them, preserving depth and composition while destroying only the pairing; uMSA skips pairing but piles in every UniProt homolog that fits.
sMSA is the crux: it isolates pairing from depth. If shuffling pairings leaves accuracy intact, pairing itself does nothing.
Pairing barely helps. pMSA averages DockQ 0.613 against mMSA's 0.602, a 0.011 gap.
Shuffling pairings is essentially free: sMSA scores 0.612, statistically indistinguishable from pMSA (P = 0.96). For cross-species complexes, sMSA even beats pMSA (0.561 vs 0.545), because species-based matching injects spurious co-evolutionary noise. One case (PDB 8JLE): unpaired mMSA reaches DockQ > 0.8, the standard paired pMSA is almost entirely wrong (DockQ about 0), and shuffling the pairings rescues it to 0.800.
The strongest strategy is uMSA: skip pairing, maximize homologs, average DockQ 0.623, the highest, and consistently ahead for both intra- and inter-species complexes.
| MSA strategy | Mean DockQ | Note |
| mMSA (unpaired baseline) | 0.602 | per-chain only |
| pMSA (standard pairing) | 0.613 | AF3 default |
| sMSA (shuffled pairing) | 0.612 | only pairing disrupted |
| uMSA (unpaired, max homologs) | 0.623 | best |
The conclusion is counterintuitive: the cross-chain pairing algorithms everyone has been polishing may be oversold. What drives accuracy is MSA depth, meaning how many homologs you have. For anyone doing structure prediction, this says stop agonizing over pairing and find ways to pack in more useful homologs (the recent omicMSA strategy confirms the same). Mechanistically, AlphaFold3 leans on physicochemical complementarity between chains (shape and charge matching) plus its iterative update module to dock correctly without co-evolutionary priors.
Systematic testing covers only AlphaFold3 and dimers; the extension to higher-order multimers is mentioned in passing. Antibody-antigen and host-pathogen interfaces, where co-evolutionary signal is weak, remain hard. The authors admit AF3's internal ranking score is unstable: in several cases a high-accuracy model was generated but ranked lower. The 439-sample benchmark, though controlled for training-set leakage, is still a curated set and may not match real-world distributions.