Large-scale prospective evaluation of co-folding across hundreds of Mac1-ligand complexes and three virtual screens.

Deep-learning co-folding methods accurately predicted over 50% of Mac1 ligand poses within 2 Å RMSD for 489 SARS-CoV-2 complexes, despite limited structural rearrangement recapitulation.

• Why it matters: Accurate prediction and ranking of ligand-protein complexes are crucial for drug discovery, but evaluating these methods has been limited by small test sets and dependence on training data.
• What they did: The study assessed AlphaFold3, Boltz-2, and Chai-1 on 489 Mac1 complexes determined after training cutoff, and prospectively tested co-folding scores for ligand enrichment among molecules targeting AmpC β-lactamase, dopamine D4, and σ2 receptors.
• The result: Boltz-2 affinity predictions correlated most strongly with experimental potency and outperformed baseline predictors after calibration, while co-folding confidence weakly tracked potency; integrating these approaches may improve hit prioritization.