Assessing Structural Prediction Accuracy for Nanobody-Small Molecule Complexes

Generative models have advanced profusely in recent years, generating great impact in computational modelling and structural bioinformatics. A part of computational protein design focuses on antibody and nanobody engineering, targeting proteic epitopes. By contrast, the development of nanobodies for small-molecule sensing remains a largely experimental field. In this work, we evaluate the performance of several state-of-the-art prediction softwares (AlfaFold3, Chai-1, Boltz-1, RosettaFold-AllAtom, FlowDock and OmegaFold), on nanobody-small molecule complexes, in an effort to pave the way of computational studies in this area. We tested the general nanobody and CDR accuracy, as well as ligand placement and orientation. We also explored the correlation of the results with intrinsic metrics of the models, such as pLDDT, and with the number of samples and recycles. Results show that most predictors perform well at predicting nanobody structures, but some struggle at ligand placement. AlphaFold3 outperformed the other softwares in all these tasks. Co-folding increased the accuracy of the predictions, modelling better CDR1. Contact analysis revealed that CDR1 was mostly involved in ligand binding instead of CDR3. This could point to a memorization tendency in the models, as most nanobody-antigen complexes target other proteins. Results also pointed to pLDDT as a good score to indicate CDR accuracy. Accuracy did slightly improve when increasing the number of samples but did not with the number of recycles. These findings highlight both the advantages and limitations of structure prediction methods for nanobody-small molecule complexes.