Prediction of Single-Mutation Effects for Fluorescent Immunosensor Engineering with an End-to-End Trained Protein Language Model

Abstract

A quenchbody (Q-body) is a fluorophore-labeled homogeneous immunosensor in which the fluorophore is quenched by tryptophan (Trp) residues in the vicinity of the antigen-binding paratope and dequenched in response to antigen binding. Developing Q-bodies against targets on demand remains challenging due to the large sequence space of the complementarity-determining regions (CDRs) related to antigen binding and fluorophore quenching. In this study, we pioneered a strategy using high-throughput screening and a protein language model (pLM) to predict the effects of mutations on fluorophore quenching with single amino acid resolution, thereby enhancing the performance of Q-bodies. We collected yeasts displaying nanobodies with high- and low-quenching properties for the TAMRA fluorophore from a modified large synthetic nanobody library followed by next-generation sequencing. The pretrained pLM, connected to a single-layer perceptron, was trained end-to-end on the enriched CDR sequences. The achieved quenching prediction model that focused on CDR1 + 3 performed best in the evaluation with precision-recall curves. Using this model, we predicted and validated the effective mutations in two anti-SARS-CoV-2 nanobodies, RBD1i13 and RBD10i14, which converted them into Q-bodies. For RBD1i13, three Trp mutants were predicted to have high probability scores for quenching through in silico Trp scanning. These mutants were verified via yeast surface display, and all showed enhanced quenching. For RBD10i14, mutations at four positions close to an existing Trp gave high scores through in silico saturation mutagenesis scanning. Six of eight high-score mutants, derived from two mutants at each of the four positions, exhibited deeper quenching on the yeast surface. Next, combined with the investigation of antigen binding of the mutants, we successfully achieved Q-bodies with enhanced responses. Overall, our strategy allows the prediction of fluorescence responses solely on the basis of the antibody sequence and will be essential for the rational selection and design of antibodies to achieve immunosensors with larger responses.