Abstract
Because of their high affinity, specificity, and environmental stability, nanobodies (Nbs) have continuously received attention from the field of biological research. However, it is tough work to obtain high-affinity Nbs using experimental methods. In the current study, 12 machine learning algorithms were compared in parallel to explore the potential patterns between Nb-ligand affinity and eight noncovalent interactions. After model comparison and optimization, four optimized models (SVMrB, RotFB, RFB, and C50B) and two stacked models (StackKNN and StackRF) based on nine uncorrelated (correlation coefficient <0.65) optimized models were selected. All the models showed an accuracy of around 0.70 and high specificity. Compared to the other models, RotFB and RFB were not capable of predicting nonaffinitive Nbs with lower precision (<0.44) but showed higher sensitivity at 0.6761 and 0.3521 and good model robustness (F1 score and MCC values). On the contrary, SVMrB, C50B, and StackKNN were able to effectively predict the future nonaffinitive Nbs (specificity >0.92) and reduce the number of true affinitive Nbs (precision >0.5). On the other hand, StackRF showed intermediate model performance. Furthermore, an in-depth feature analysis indicated that hydrogen bonding and aromatic-associated interactions were the key noncovalent interactions in determining Nb-ligand binding affinity. In summary, the current study provides, for the first time, a tool that can effectively predict whether there is an affinity between nanobodies and their intended ligands and explores the key factors that influence their affinity, which could improve the screening and design process of Nbs and accelerate the development of Nb drugs and applications.