PepCCD

Peptides sit between small molecules and antibodies as a therapeutic modality, and they are one of the few practical routes to the roughly 80% of disease-relevant proteins that lack the stable pockets small-molecule inhibitors require. Most computational methods for designing a peptide binder are structure-based: they dock, hallucinate, or diffuse against a high-resolution model of the target interface. That is a real constraint, because experimental structures are expensive and slow to obtain and intrinsically disordered or conformationally dynamic targets may have no single structure to design against.

PepCCD (Contrastive Conditioned Diffusion) drops that dependency: it takes the target protein's amino acid sequence as its only input and generates binder sequences directly. The framework was developed at the School of Artificial Intelligence, Shenzhen University, posted as a bioRxiv preprint in September 2025 and published in the Proceedings of AAAI-2026. Its central idea is to learn a shared embedding space where a target protein and its binding peptides sit close together, then use the target's embedding as the conditioning vector for a diffusion model that denoises Gaussian noise into a peptide representation.

Earlier sequence-only methods mutate or extend a known template peptide and so explore a narrow neighborhood of sequence space. Sampling from noise instead, PepCCD draws fresh candidates for any target with no seed sequence, candidate library, or post hoc screening pass.

#Key Features

• Sequence-only target conditioning: A target is given by its amino acid sequence alone, so the method applies to intrinsically disordered proteins and others without usable interface structures.
• Contrastive protein–peptide alignment: Two ESM-2 encoders, one for proteins and one for peptides, are fine-tuned with an InfoNCE objective so that matched binding pairs sit close in embedding space and mismatched pairs far apart.
• Latent diffusion over peptide embeddings: A Transformer denoiser of multi-head self-attention and MLP blocks reverses a 500-step Gaussian noising process in embedding space, and the result is decoded into a sequence.
• Template-free sampling: Candidates come from noise conditioned on the target, giving lower sequence and structural similarity to native binders than either baseline while retaining near-native amino acid composition.
• Sub-second inference: One peptide takes roughly 1 second on a single NVIDIA 4090 GPU, against 4 seconds for the sequence-based baseline and 5 minutes for the structure-based pipeline.

#Technical Details

Training proceeds in three stages. Stage 1 aligns paired protein and peptide ESM-2 encoders (the released code uses the 150M-parameter esm2_t30_150M_UR50D checkpoint) with InfoNCE on protein–peptide pairs from RCSB PDB co-crystal structures, keeping interfaces with buried surface area of at least 50 Ų and clustering proteins with CD-HIT at 90% identity — 15,110 pairs for training, 5,480 for testing. Stage 2 pre-trains the diffusion model unconditionally on 68,958,049 peptide fragments under 30 residues, cut by sliding a 30-residue window with stride 15 over UniProt sequences, under a mean-squared-error reconstruction loss on the embeddings. Stage 3 adds the target condition vector and fine-tunes with a joint MSE plus cross-entropy objective over decoded tokens.

Evaluation covered 209 target proteins with 10 peptides generated per target, against RFdiffusion (with ProteinMPNN for sequence design) as the structure-guided baseline and PepPrCLIP as the sequence-guided one. PepCCD reached a best-case interface pTM of 0.7418 and an average of 0.5557, above PepPrCLIP (0.7141 / 0.4901) and below RFdiffusion (0.7838 / 0.5960), while achieving the lowest sequence similarity to native binders (0.1895), the lowest structural similarity (0.2701), the highest predicted bioactivity (0.4419), and the lowest instability index (37.66). On hit rate — the share of peptides matching or beating their template's ipTM — PepCCD exceeded RFdiffusion. Ablations attribute specificity to the alignment stage and diversity to pre-training. In 100 ns all-atom GROMACS simulations against the SARS-CoV-2 spike receptor-binding domain (PDB 6M0J), PepCCD peptides had the lowest MM/GBSA and MM/PBSA binding free energies and the smallest variance.

#Applications

PepCCD targets early-stage peptide discovery where structure-based design is impractical: intrinsically disordered proteins, conformationally heterogeneous targets, and interaction interfaces that structure predictors model poorly. Because a run needs only a target sequence and about a second per candidate, it is cheap enough to sweep whole target panels and pass the output to downstream filtering, docking, or synthesis. Code and checkpoints for all three stages are released under Apache-2.0, with weights and the train/test splits on Hugging Face.

#Impact

PepCCD is the first published framework to condition a diffusion model on a target protein sequence for de novo peptide design, and it shows that a sequence-only method can approach structure-guided design on interface quality while producing far more diverse candidates orders of magnitude faster. The caveats matter: every reported result is in silico, resting on predicted structures, Rosetta energies, and simulation rather than binding assays. The benchmark targets all come from the PDB and so have solved structures, which the authors note may favor the structure-based baseline. The permissive release of code, weights, and data makes the framework straightforward to reproduce and build on.