A second new preprint supplies a useful warning for any automated discovery loop. Researchers combined AlphaFold 3 modeling with microscale thermophoresis—a physical binding assay—to study glucan-binding protein C and a biofilm-inhibiting aptamer plus five derivative sequences. The models proposed a binding motif and a possible mechanism, but experiment only partly supported the predictions. More pointedly, AlphaFold's confidence measures did not correlate directly with measured binding affinity.
That does not mean the structure model was useless. Confidence can help nominate plausible interfaces for follow-up. It does mean that structural plausibility and binding strength are different targets. Ranking candidates by a model's internal confidence would answer a different question from the laboratory assay. In an autonomous lab, the objective must say which property governs the next selection, and the loop must preserve a physical measurement for that property rather than quietly substituting a convenient proxy. This is one protein/aptamer system, not a universal verdict on AlphaFold 3, and the work has not yet been peer reviewed. Preprint