Meta-Learning for Binding Affinity Prediction in Fixed-Target DARPin Engineering

Abstract

Most protein engineering campaigns are aimed at improving protein properties, such as thermostability and binding affinity, a difficult task due to the vast combinatorial sequence space and challenges associated with modeling the sequence-function relationship. Various computational methods have been developed, based on physics principles as well as more recent data-driven machine learning methods for sequencestructure-property mappings, to aid in both in-silico protein design and engineering. In this work, we develop a forward prediction model for learning the sequence-function relationship of Designed Ankyrin Repeat Proteins (DARPins), an antibodysurrogate binder scaffold with highly promising therapeutic potential. To facilitate better learning of DARPins' representations for property prediction, we explore various amino acid representation schemes, leveraging both knowledge-driven and data-driven descriptors. Using a real-world DARPin affinity maturation dataset, we design a meta-learning framework to maximize the utility of available experimental data and achieve accurate and reliable property prediction. Our computational model has demonstrated great potential to guide DARPin design campaigns.