Loading…
Peptide-protein interaction prediction that scores a pair and maps the interface residues on both the peptide and the protein from sequence alone.
No providers recorded yet. Browse all providers
A peptide binds through a short motif, a handful of residues doing most of the work, while its protein partner presents a groove shaped by hundreds of residues of context. Most sequence-based predictors push both sides through the same encoder and concatenate the results, asking one representation to serve two very different length scales and discarding the fact that recognition is conditional: the same peptide is a binder against one protein and noise against another. ConGA-PepPI, from Lanzhou University with collaborators at City University of Hong Kong and the Shanghai Innovation Institute, treats that asymmetry as architecture rather than as a nuisance.
Peptide and protein sequences are embedded by a shared ESMC protein language model and then routed into branch-specific extractors: a text-CNN pathway for the peptide's local motifs, a dilated-convolution pathway for the protein's longer-range context, each paired with a self-attention pathway and a gated fusion step weighting local against global cues. Two rounds of bidirectional cross-attention follow, so peptide features refine the protein representation and protein context updates the peptide representation, and the model scores a conditional pair rather than two independently embedded strings.
The second idea addresses a supervision gap. Sequence-level "do these interact?" labels are far more plentiful than residue-level interface annotations, so ConGA-PepPI is trained first on the larger pair-level task and then transferred to residue localization on the smaller annotated set. The April 2026 preprint introduces it alongside TC-PepGen, a target-conditioned peptide generator that shares its protein encoder; the current version of that preprint is a September 2026 revision.
Training pairs come from RCSB PDB peptide-protein complexes deposited before 2023, with interface residues annotated by PDB-BRE and chains aligned to UniProt. Pairs with more than 20% unknown or non-standard residues were removed, peptides padded or truncated to 50 residues and proteins to 800. Because experimentally confirmed non-binders are scarce, negatives were approximated by randomly pairing peptides with non-cognate proteins at a 1:1 ratio. A supervised contrastive head regularizes the pair embedding during sequence-level training. Under five-fold cross-validation the model reaches 0.839 accuracy, 0.921 AUROC and 0.915 AUPR, ahead of CAMP, IIDL-PepPI, DeepPepPI, BridgeDPI and classical baselines on the same split; an ablation series that adds cross-attention, branch specialization, gated fusion and contrastive regularization in turn lifts accuracy from 0.808 to 0.839. On external benchmarks it records AUROC 0.843 on the temporally separated low-homology Test167, 0.936 on LEADS-PEP and 0.940 on Test251. For residue identification it reaches peptide-side AUPR 0.950 and AUROC 0.922, and protein-side AUPR 0.601, the harder direction where long chains and sparse interfaces make class imbalance severe; AlphaFold 3 reaches 0.847 and 0.544 on the same two tasks. SHAP attribution over the 1024-dimensional fused representation puts attention features ahead of convolutional ones, 667.03 against 389.39.
The intended setting is early-stage triage, narrowing a large candidate pool before anything reaches a docking run or a synthesis order. Against GSK-3 beta, 1,000 generated candidates scored a mean predicted binding probability of 0.5927 versus 0.1910 for a 10,000-peptide background drawn from UniProt, a Cliff's delta of 0.778. Under a stringent MDM2 setting built by removing every training sample above 60% similarity on either side, the known ligand pMI and four analogs from the same motif family still scored between 0.942 and 1.000. For HLA-B27 in PDB structure 1FG2, a target excluded from training, an alanine scan of an 11-mer candidate ranked positions by predicted sensitivity, and the most sensitive ones mapped near the modeled interface.
The argument here is that peptide-protein recognition benefits more from modeling partner asymmetry and pair conditioning than from adding capacity, and the ablation series is the evidence: each architectural piece moves the metric, and the largest single step is the one that makes the prediction conditional. The residue-localization result separately supports reusing abundant pair labels as pretraining for annotation-poor tasks. The limits are worth stating plainly. Negatives are computational rather than experimental, so rates on real screening libraries may differ from the reported benchmarks; every result is in silico, with no prospective binding assay; the work remains a preprint. The authors released MIT-licensed inference code for binding and binding-site prediction, but the trained checkpoints it 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.