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

A shared latent-space deep learning model redesigned PDZ-binding peptides for DLG4-PDZ3, yielding 100 top candidates with 84 in the top 1% of measured interaction fitness.

• Why it matters: PDZ-binding motif design must balance desired affinity with off-target binding, but existing models struggle to steer sequence searches toward different objectives using one trained system.
• What they did: The method learns a PDZ-PBM shared latent space from quantitative PDZ-PBM affinity prediction, then formulates single-target, off-target-aware, and dual-target objectives as latent directions.
• The result: For DLG4-PDZ3, it outperformed an existing method (84 vs 64 in the top 1%); for de novo 10-residue PBMs it had the highest median in silico binding scores for three of four PDZ domains, and for PTPN3/SCRIB-PDZ3 it improved off-target separation or matched single-target scores without retraining.

AI summary of the abstract · may contain errors · not medical advice · read the paper.