Target-conditioned peptide binder generation that re-injects the encoded target protein through cross-attention at every autoregressive step.
No providers recorded yet. Browse all providers
A peptide generator told about its target only once, at the start of decoding, is in an awkward position: every residue it emits pushes that instruction further into the past, until the model is mostly continuing its own output rather than answering the protein it was asked about. Conditioning dilutes, and generation drifts back toward a generic peptide language model — fluent, plausible, and specific to nothing. TC-PepGen's response is to stop treating the target as an initial condition at all.
Built at Lanzhou University with collaborators at City University of Hong Kong and the Shanghai Innovation Institute, the model pairs a frozen ESMC encoder, which reads the target protein sequence, with an autoregressive ProtGPT2 decoder that writes the peptide. The encoded target is projected once and then supplied as keys and values to cross-attention inside every decoder block, so the target representation is re-consulted at each step of generation rather than consumed at the first one. The design goal the authors state is not peptide language modeling but target-conditioned design, and the architecture is what enforces the distinction.
TC-PepGen is the generative half of an April 2026 preprint whose other half, ConGA-PepPI, scores and localizes peptide–protein interactions. The two share the frozen protein encoder and are meant to run in sequence — generate a candidate pool against a target, then rank it. The current version of the preprint is a September 2026 revision.
The frozen ESMC encoder emits a 960-dimensional target representation that a linear projection with dropout maps to the decoder's 1280-dimensional width, where it is injected as cross-attention keys and values in each ProtGPT2 block. Training data merges the peptide–target set released with PepNN with Propedia, deduplicated; target proteins were clustered with MMseqs2 at 0.8 sequence identity and entries whose targets exceeded 80% similarity were removed, yielding 10,203 unique pairs — 10,000 for training and 203 held out — with peptides capped at 50 residues and targets at 500. Evaluation counts a generated peptide as a hit when its predicted interface score exceeds that of the native test peptide for the same target. Under the length-conditioned benchmark, 45.81% of TC-PepGen peptides beat the native template on Chai-1 ipTM, 53.69% on ESMFold relative ipTM and 40.39% on AlphaFold 3 ipTM, ahead of PepMLM and RFdiffusion on all three evaluators. Without a supplied length, 27.09% still exceeded the native template on AlphaFold 3 ipTM, and generated minus reference length peaks between 0 and +5 residues. On held-out complex 1ZKY, final-layer decoder attention concentrated near residue 55, in the region of the predicted binding groove.
The model is aimed at the front of a peptide-therapeutic screening funnel, where the task is to produce a candidate pool for a specific target cheaply enough that a predictor can triage it. Three settings show the shape of that workflow: 1,000 candidates generated against GSK-3 beta and ranked by ConGA-PepPI, whose predicted binding probabilities averaged 0.5927 against 0.1910 for a length-matched natural background from UniProt; an 11-mer for HLA-B27 in PDB structure 1FG2, a target excluded from training; and 100 candidates for MDM2 after deleting every training sample above 60% similarity on either side, whose best-ranked novel 14-mer reached AlphaFold 3 ipTM 0.84 against 0.87 for the known ligand pMI despite near-zero sequence identity to it.
The contribution is narrow and testable: persistent conditioning beats one-shot conditioning for target-specific generation, and the unconstrained-length result is the part that carries the argument, since it removes the external hint that makes controlled comparisons flattering. The cross-attention analysis supports the mechanistic reading rather than only the score. The limits are real. Every evaluation is in silico, scored by structure predictors against native templates, with no synthesis and no binding assay; the authors name prospective experimental assessment as future work. Unconstrained generation is markedly weaker than the length-conditioned setting, and the work remains a preprint. The authors released MIT-licensed inference code together with the generation training and test splits, but the trained checkpoints the code loads are not distributed with it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.