Pretrained graph representation of peptides with canonical and non-canonical amino acids, for permeability, solubility and binding prediction.
No providers recorded yet. Browse all providers
Therapeutic peptides are rarely made from the twenty natural amino acids alone. Medicinal chemists N-methylate backbones, swap in D-amino acids, cyclize chains and attach bulky synthetic side chains to buy stability against proteases and a route across cell membranes. A protein language model such as ESM-2 reads a fixed amino acid alphabet and has no token for any of these residues, while small-molecule models see a peptide as one large, repetitive molecule and have no notion of where one residue ends and the next begins. PepLand is a pretrained representation model built for the space between the two: it treats every peptide as a molecular graph, so any residue that can be written as a SMILES string is representable, and it adds a fragment level so the model still sees residue-sized chemical units.
PepLand was developed at Jilin University in Fengfeng Zhou's lab with the Guangzhou peptide-AI company Syneron, the King Abdullah University of Science and Technology group of Xin Gao, and the Beijing Life Science Academy. It was posted to arXiv in November 2023 and published in Briefings in Bioinformatics in 2025.
The encoder is a five-layer heterogeneous graph transformer with 300-dimensional atom and fragment embeddings; the released checkpoint holds roughly 6.5 million parameters. Stage one pretrains on 7,924,509 UniProt-derived sequences of at most 30 residues, converted to SMILES. Stage two continues on 8,977 peptides containing non-canonical amino acids, drawn from the CycPeptMPDB cyclic-peptide permeability database and from the Protein Data Bank. The authors compared fragmentation schemes directly: a coarser operator that preserves only amide bonds produced a 410-entry vocabulary in which 37% of fragments occur five times or fewer and only 88.6% of molecules can be reassembled, against 6.1% rare fragments and 97.5% recovery for AdaFrag.
Evaluation attaches two linear layers to each frozen representation. On non-canonical peptides PepLand reached a Spearman correlation of 0.628 for cyclic-peptide membrane permeability, against 0.539 for Uni-Mol, 0.503 for ChemBERTa-2 (ChemBERTa) and 0.461 for MolCLR, and 0.768 for protein–peptide binding affinity against 0.721 for ChemBERTa-2. On canonical-only tasks it outperformed the small-molecule models but trailed the 15-billion-parameter ESM-2 on cell penetration (AUC 0.838 versus 0.885) and solubility (AUC 0.662 versus 0.725); concatenating the two representations recovered ESM-2's cell-penetration AUC and raised solubility to 0.730.
The model is used as a front end for peptide-drug triage: a chemist encodes candidate sequences written as SMILES, including macrocycles and chemically modified residues, and fits a small head on whatever assay data exists. The study covers four such readouts — cell penetration, membrane permeability, aqueous solubility and protein–peptide binding affinity — and adds two case studies. Ranking cyclic-peptide binders curated from SKEMPI 2.0 gave Spearman correlations of 0.829 and 0.718 on the two protein targets, ahead of docking and free-energy baselines on both. Replacing ECFP fingerprints with PepLand features in a solid-phase synthesis difficulty predictor cut mean absolute error from 0.160 to 0.138 and raised R² from 0.644 to 0.746, which matters for deciding whether a designed sequence is worth committing to the synthesizer.
PepLand established that graph-level pretraining, rather than sequence tokenization, is the workable route to representing peptides whose chemistry falls outside the standard alphabet, and later peptide models reuse it in that role: both SinCAA and PeptiVerse benchmark against it as a pretrained comparator. Code, the pretrained checkpoint and the five evaluation datasets are released under Apache-2.0, so the frozen embeddings are cheap to try on a new assay; the two pretraining corpora themselves are not released, only small example files standing in for them. The scope is deliberately narrow: pretraining used sequences of at most 30 residues, so the representation targets peptides rather than whole proteins, and the model learns no evolutionary signal of the kind that keeps protein language models ahead on canonical-only tasks.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.