PDZ-binding peptide design moves into a shared latent space: one trained model searches off-target and dual-target objectives without retraining

One trained model can search PBM sequences under different binding objectives, and switching objectives requires no retraining. Earlier off-target-aware design relied on sequence models that approximate structure-derived scores, so adding a PDZ domain meant new structure calculations and a new approximation model for that domain. In four-position redesign for DLG4-PDZ3 starting from the parent sequence STKNYKQTSV, under the same budget of 1,024 evaluated sequences, 84 of the method's top 100 predictions fell in the top 1% of measured BindingPCA fitness in a published library, against 64 for SemiSVR-guided search; the top 0.1% counts were 52 and 33.

Affinity is represented as the cosine similarity between a PDZ vector and a PBM vector, so design objectives become directions in latent space. The model has no decoder that turns a target point into a sequence, so the search runs beam search over real sequences on the fixed encoders, generating single-residue substitutions and keeping the top-scoring candidates at each step. The single-target direction is the target PDZ vector itself, the off-target-aware direction is the normalized difference of target and off-target PDZ vectors, and the dual-target direction is the normalized sum of the two PDZ vectors; on PTPN3 and SCRIB-PDZ3, off-target-aware design widened between-group median score differences to 6.14 and 5.97 kcal/mol but lowered target-domain medians from 25.37 to 18.68 and from 32.14 to 25.74 kcal/mol, while dual-target design gave 24.57 and 30.30 kcal/mol, close to the respective single-target groups.

The demonstrated scope is the PDZ-PBM system: training used 57,581 observations across 133 PDZ domains from the DAPF dataset, external ranking used 448 DLG1-PDZ2 peptides, single-target redesign was checked retrospectively against a published mutant library for DLG4-PDZ3, and unrestricted 10-residue design was compared with three methods on DLG4-PDZ3, DLG1-PDZ2, SYNJ2BP and SNX27. The intended readers are computational and experimental researchers working on peptide design and control of protein interactions.

The binding properties of newly designed candidates have not been validated experimentally, and AlphaFold 3 structure prediction with single-frame MM/PBSA scoring remains an approximation of true binding free energies. Censored weak-binding observations give only upper bounds on pKd, leaving limited quantitative information for learning affinity differences among weak binders. Mean cosine similarity between retained sequences and the objective direction plateaus below 1, and these curves alone cannot distinguish a limit of the search procedure from a limit of the representations attainable by real sequences.