AI Pool Ranking Improves Antibody Structure Prediction
A preprint introduces AbPoolRank, a selector that ranks structures from several predictors without relying on model-specific confidence scores.
Updated on Nov. 19, 2026 in Artificial Intelligence

Researchers introduced AbPoolRank, a predictor-agnostic structural selector for antibody-antigen structure prediction. Its 15-structure mixed pool reached 58.0% medium-or-better Success@1 on 103 antibody-antigen complexes.
Why it matters
Antibody-antigen structure prediction has typically relied on deep sampling of one co-folding predictor and choosing its top output by model-specific confidence scores. The approach described here combines shallow sampling across several predictors with a confidence-free ranking method.
The 15-structure pool combined three structures each from Boltz-2, Promera, Protenix-v2, ESMFold2-Fast and OpenDDE, and used 1.8-fold less aggregate GPU time than the most accurate single-predictor setting. Removing any one predictor during training lowered accuracy by at most 3.0 percentage points.
The details
AbPoolRank ranks structures using cross-pose agreement, within-source reproducibility and interface geometry, and operates without native confidence scores, predictor identities or predictor-specific parameters. The frozen model accepted full MSA-conditioned ESMFold2 without retraining or loss of accuracy.
Timeline
2021: Cutoff year for the post-2021 complexes in the benchmark.
November 17, 2026: Preprint identifier date for the bioRxiv submission.
Under the Hood
The method departs from the established practice of deep sampling a single co-folding predictor and selecting its highest-confidence output. Instead it draws on several predictors at once and ranks their outputs with a confidence-free selector.
For researchers and drug developers, the approach reduces GPU time while avoiding reliance on any single predictor's confidence scores. The reported gains are based on a benchmark of 103 complexes and have not yet been validated beyond that set.
The takeaway
Combining structures from several predictors and ranking them without confidence scores may offer a more reliable route to accurate antibody models than deep sampling a single predictor.
Further reading
Explore more on Artificial Intelligence developments shaping scientific modeling.
Source note: This article includes information reported by Biorxiv.



