The article compares ESMFold2, Protenix-v2, and OpenDDE with AlphaFold3 using independent antibody-antigen benchmarks. On the reported FoldBench set, all three open models exceed the previously published AlphaFold3 result; on a stricter out-of-distribution split, the relative margins are smaller and Protenix-v2 does not exceed AF3.
The authors argue that better complex structures can improve downstream design and sequence optimization, because these workflows often use structure predictors to evaluate candidate binders. They emphasize matched inputs and independent evaluation as necessary context for model-performance claims.
The post presents the results as evidence that open models can move the field's frontier, while leaving benchmark choice, dataset leakage, and the specific task central to interpreting a ranking.