One latent space steers PDZ-binding peptide design toward single, dual or selective targets
- A deep learning model learns a shared latent space of PDZ domains and their binding motifs by predicting quantitative affinity; design goals (single target, off-target-aware, dual target) become directions in that space, so one trained model serves all without retraining.
- Redesigning the DLG4-PDZ3 binding motif, it outperformed an existing method, with a clearly higher share of top candidates landing in the top 1% of measured interaction fitness.
- For de novo 10-residue peptides, it gave the highest median in silico binding scores on three of four PDZ domains. Off-target-aware design widened target/off-target score gaps at some cost to target scores; dual-target designs scored close to single-target ones.
- Validation is computational: scored against existing measured fitness data and in silico predictors, not new wet-lab binding assays.
💡 Points Based on the abstract
No new wet-lab binding data here: scoring against existing fitness datasets and in silico predictors invites circularity, especially if related models score the designs. Still, one latent space where off-target avoidance is just a direction is appealing. Next step: SPR or ITC on a handful of designs, including the off-target-aware ones, to see whether the selectivity gap survives.
📄 Paper · bioRxiv ↗ Affinity-guided design of PDZ-binding peptides in a shared latent space