Picking the top AlphaFold3 model per antibody-antigen target with a confidence score

Stable ID: structural-20261009-protocol-fromm2026-abag-model-selection

areas

proteins-complexes

contexts

research

protocol

AlphaFold3 is run with 40 seeds and 5 diffusion samples per seed, giving 200 models per target. Each ranking score selects the model it rates highest; the score is the mean DockQ of the selected models over 110 targets. Two correlations with DockQ are also printed: across all models of all targets (R) and the mean of per-target correlations (<R>). DockQ is computed for the antibody-antigen interface with multi-chain antigens merged into one chain.

version

Fromm et al. 2026, sections 2.2 to 2.7 and 3.5 to 3.6; Table 1

metric

dockq

metric direction

higher

unit

unitless

metric definition

<DockQ>: mean over targets of the DockQ of the model each score ranks first. R: Spearman correlation between the score and DockQ over all models. <R>: mean over targets of the per-target Spearman correlation.

selection

Highest value of the named score among the 200 models of each target

limitations

Antibody-antigen complexes only; results do not transfer to other complex classes.; Structural accuracy against the deposited complex, not binding, affinity or the outcome of an experiment.; Only AlphaFold3 models are ranked in Table 1; the source does not print the same table for AlphaFold2.3, Boltz-1 or Chai-1.; The DockQ, aeTM, aeiTM and aeRankConf rows use the experimental structure, so they are upper bounds, not usable selection rules.; Figure 6 and the text give a per-target ranking-confidence correlation of 0.28, while Table 1 prints 0.214 for RankConf under a Spearman caption; the figure's correlation type is not stated.; No uncertainty is printed.; Table 1 does not state how many models it ranks, but Methods 2.2.2 and section 2.4 state 200 models per complex for AlphaFold3, and the table's DockQ row (0.544) matches the best-of-200 value in section 3.6 (0.54).; Section 3.5 gives the mean per-target correlation for ranking confidence as 0.28 while Table 1 prints 0.214; the text names no correlation type. The table value is stored and carries the conflict as a source warning.

source locator

Sections 2.2.2, 2.4, 2.7, 3.5 and 3.6; Table 1

• uses data: Fromm et al. 2026 antibody-antigen benchmark: 110 complexes released after 30 September 2021
• subject: headline_finding: structural-20261009-protocol-fromm2026-abag-model-selection
• assessment: AlphaFold3 top model by aeiTM on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by aeRankConf on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by aeTM on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by DockQ on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by ipSAE on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by ipTM on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by pDockQ2 on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by pTM on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessment: AlphaFold3 top model by RankConf on 110 post-cutoff antibody-antigen complexes (Fromm et al. 2026)
• assessed by: Choose structural hypotheses to guide experiments