Whole-Virus Screening Coupled with Uncertainty-Aware Deep Learning Enables Discovery of Virus-Binding scFvs

Abstract

Discovery of virus-binding ligands against emerging viral pathogens is of broad interest for diagnostics and therapeutic development. Here, we present an integrated framework that combines whole-virus screening using yeast surface display with uncertainty-aware deep learning and in silico directed evolution for the discovery and optimization of virus-binding single-chain variable fragments (scFvs). Using dengue virus (DENV) as a model system, millions of scFv sequences obtained from iterative whole-virus selections were analyzed by an uncertainty-aware deep learning model to predict virus binding while quantifying prediction uncertainty. Guided by these predictions, an in silico directed evolution algorithm introduced targeted mutations within complementarity-determining regions (CDRs) to improve predicted binding together with developability-related properties, including solubility and humanness. Experimental validation demonstrated measurable DENV binding by two independently AI-designed scFvs, providing proof of concept for the proposed framework.