Revisiting earlier protease inhibitor studies using modern AI structure modeling tools

Crystallographic studies of proteases and inhibitors by Wolfram Bode and colleagues in the late twentieth century relied on simple graphical modeling, force-field energy functions, and human intuition. These systems now provide interesting tests of modern AI methods. We examine three classes of protease inhibitors and their complexes. For cystatin, originally modeled with papain by manual fitting of protein surfaces, AI methods recover both the inhibitor fold and binding geometry to differing extents, with some approaches favoring a single binding mode and others generating broader ensembles that may explore alternative geometries. De novo design against papain reveals a similar contrast between conservative and exploratory modes of prediction. For the two-headed serine protease inhibitor from mung bean, prediction of its symmetric ternary complex with two trypsin molecules depends strongly on recovery of the small disulfide-rich inhibitor fold. Finally, modeling α1-antitrypsin tests whether AI methods can distinguish known uncleaved and cleaved conformations of a structurally more plastic inhibitor. Local restructuring of the central β-sheet is central to mechanism, while the overall fold remains conserved. These applications show diverse uses of AI modeling tools, from reproducing known structures to exploring assemblies and generating novel interactions, while exposing limitations in flexibility and sampling.