Machine Learning Prediction of Antibody-mediated Effector Function Using Structural Features of Fc-Glycans and Binding Affinity to an Fc-receptor

Fc-glycosylation critically modulates antibody effector functions including antibody-dependent cellular cytotoxicity (ADCC) and antibody-dependent cellular phagocytosis (ADCP), yet computational tools for predicting Fc-receptor binding affinity remain limited. Here, we present a systematic machine learning benchmark evaluating 10 machine learning methods for predicting Fc binding affinity to FcγRIIIa receptor which is responsible for the ADCC activity. Using a curated dataset of 22 structurally diverse N-glycans on antibodies with experimentally measured binding affinities of antibody-glycoforms FcγRIIIa-receptor, we compared traditional machine learning approaches using glycan structural motif features against deep learning methods employing pretrained graph neural network embeddings. Support vector regression with a linear kernel achieved the best performance (Spearman ρ = 0.85), followed closely by ElasticNet (ρ = 0.82) and Ridge regression (ρ = 0.81). Notably, simpler linear models consistently outperformed more complex ensemble methods and deep learning approaches, with the pretrained SweetNet graph neural network yielding substantially lower correlation (ρ = 0.58). The superior performance of linear models suggests that glycan structural motifs capture essential features to predict receptor binding affinity. Our rigorous nested leave-one-out cross-validation protocol with bootstrap confidence intervals provides robust performance estimates suitable for small datasets. This study demonstrated that by decoding the impact of Fc-glycosylation on Fc-receptor binding, we can build the dataset for AI-driven prediction of ADCC, and this approach is extensible to other Fc receptors and the corresponding effector functions