Affinity-guided design of PDZ-binding peptides in a shared latent space

Abstract

PDZ-binding motifs (PBMs) are short C-terminal sequences that bind PDZ domains and contribute to protein localization and complex assembly. Designing PBMs could modulate these interactions while leaving most of the original protein sequence unchanged. Because PDZ recognition preferences overlap, modulating intended interactions requires considering off-target binding alongside target affinity. Computational PBM design methods generally use sequence-based affinity prediction and structure-based evaluation, but it remains challenging to guide sequence searches toward different binding objectives with a single trained model. Here, we propose a deep learning-based method to design PBMs by analytically formulating single-target, off-target-aware, and dual-target design objectives as directions in a PDZ-PBM shared latent space learned through quantitative PDZ-PBM affinity prediction. In single-target PBM redesign for DLG4-PDZ3, our method identified 100 top-ranked candidates, 84 of which were in the top 1% for experimentally measured interaction fitness in a published mutant library, compared with 64 using an existing method. In single-target de novo design of 10-residue PBMs, our method achieved the highest median in silico binding scores among the compared methods for three of the four PDZ domains tested. In off-target-aware and dual-target PBM design for PTPN3 and SCRIB-PDZ3, off-target-aware design increased between-group score differences within each domain at the cost of lower target-domain scores, whereas dual-target design yielded scores close to those of the respective single-target designs for both domains. Together, these results show that shared latent representations learned from quantitative affinities allow a single trained model to guide PBM design across different binding objectives without retraining.