Evaluation of AI-Predicted GH11 Xylanase Models Against a Previously Unreported Experimental Structure: Implications for Conformational Accuracy and Ligand Binding

Artificial intelligence (AI)-based structure prediction tools have emerged as powerful methods for understanding previously unsolved structures. AI-predicted models are widely used for protein function identification, drug development, and protein engineering. Although AI-predicted structures offer significant opportunities to advance research, their inaccuracies can lead to misinterpretations of molecular mechanisms. Thus, evaluating the structural differences between AI-predicted and experimental structures is crucial for accurately understanding molecular mechanisms and guiding the design of subsequent experiments. In this study, the previously unreported crystal structure of xylanase from Hypocrea virens (HviGH11) was compared with the structures predicted by ESMFold, AlphaFold2, AlphaFold3, and RoseTTAFold. The overall fold of HviGH11 was highly similar between the experimental and AI-predicted models; however, the conformation of the thumb domain of the protein varied across the models. The substrate-binding cleft of experimental HviGH11 was similar to that in the model structures generated by ESMFold, AlphaFold2, and AlphaFold3, but significantly different from those in the model structures generated by RoseTTAFold. The substrate docking study illustrated that the binding mode of xylohexaose in the substrate-binding cleft differed between the experimental and AI-predicted HviGH11 structures. These findings provide insights into the applications of AI-predicted models and offer guidance for appropriate application in structural and functional studies and biotechnology.