SurfProp: A surface-based property prediction framework for antibody developability and screening

Therapeutic antibodies are an important class of drugs, increasingly used to treat a variety of conditions. Developing antibodies is challenging in part due to issues with viscosity, aggregation, solubility and more \emph{developability} properties that pose issues for manufacturing and delivery. As these properties require large-scale experiments to measure, \textit{in silico} biophysical molecular descriptors are often considered by proxy, and serve as a basis for screening and optimization. In this work, we introduce SurfProp, a surface-based, differentiable property prediction framework aimed at improving the antibody developability workflow in two ways. (1) The \textit{insilico} arm of SurfProp predicts electrostatics and computes hydrophobicity with a significant speedup over traditional methods, facilitating higher throughput \textit{in silico} property screening. (2) the \textit{experimental} arm of SurfProp uses the pre-trained model from the \textit{in silico} task to predict experimental developability properties; here we demonstrate the ability of the pre-trained model to learn hydrophobic interaction chromatography (HIC) more effectively than a model trained from scratch.