PeptiVerse

PeptiVerse is a unified machine-learning platform for predicting the biophysical and therapeutic properties of peptides. Therapeutic peptides occupy the space between small molecules and biologics, and their viability as drugs depends on a cluster of developability properties — solubility, membrane permeability, hemolysis, toxicity, serum half-life, non-fouling behavior, and target binding — that must be assessed early in a design campaign. Existing computational tools address this fragmentarily: sequence models such as PeptideBERT are confined to the 20 canonical amino acids and cannot represent chemical modifications, while small-molecule ADMET predictors are trained on drug-like chemical space that transfers poorly to peptides.

PeptiVerse addresses that gap by predicting a broad panel of properties from either of the two representations peptide chemists actually use: canonical amino-acid sequences and chemically modified peptide SMILES that encode non-natural residues, cyclization, and staples. Rather than train property predictors from scratch, it builds task-specific models on top of frozen foundation-model embeddings, inheriting representations learned from large protein and peptide corpora. This lets a single framework cover both natural and non-canonical peptides within one interface.

PeptiVerse was introduced in a January 2026 bioRxiv preprint by Yinuo Zhang, Sophia Tang, Tong Chen, Elizabeth Mahood, Sophia Vincoff, and Pranam Chatterjee of the Programmable Biology Group at the University of Pennsylvania, with a co-affiliation at Duke-NUS Medical School. The preprint awaits peer review.

#Key Features

• Dual-representation input: Every property can be predicted from either an amino-acid sequence or a peptide SMILES string, so the platform serves both natural peptides and chemically modified, macrocyclic, or stapled designs.
• Foundation-model embeddings: Predictors are built on ESM-2 (650M parameters) for amino-acid sequences and PeptideCLM (23M) for SMILES, whose 586-token vocabulary captures non-canonical peptide chemistry.
• Broad property panel: A single interface covers hemolysis, non-fouling, solubility, permeability (penetrance, PAMPA, and Caco-2), toxicity, half-life, and peptide-protein binding affinity.
• Best-model-per-task selection: For each property, candidate architectures (XGBoost, elastic net, SVM, MLP, CNN, and Transformer) are tuned with Optuna and the strongest predictor is retained, rather than forcing one architecture across all tasks.
• Calibrated uncertainty: Regressors report adaptive conformal prediction intervals with a 90% coverage target, and classifiers report predictive entropy, giving each prediction an interpretable confidence estimate.
• Open release: Model weights, processed datasets, and an interactive web application are released publicly for direct use and reproduction.

#Technical Details

PeptiVerse encodes each peptide with a frozen foundation model — ESM-2 (esm2_t33_650M_UR50D) for amino-acid sequences and PeptideCLM-23M for SMILES — and trains lightweight predictors on the resulting embeddings, selecting the best architecture per property via 200-trial Optuna searches. Binding-affinity models add cross-attention over paired peptide and target-protein representations. Training data are curated per property from public resources including DBAASP, PROSO II, CycPeptMPDB, ToxinPred, THPdb2, PepLand, and the PDB, ranging from a few hundred labeled examples (half-life, Caco-2) to over 17,000 (non-fouling). To measure out-of-distribution generalization rather than memorization, canonical sequences are split with MMseqs2 (min-seq-id 0.3, coverage 0.8) and non-canonical peptides by Tanimoto similarity of Morgan fingerprints. On these stringent splits PeptiVerse reports gains over prior tools — for example, penetrance-permeability F1 of 0.929 versus 0.838 for PepLand and solubility F1 of 0.754 versus 0.662 — with binding-affinity Spearman correlations near 0.56.

#Applications

PeptiVerse is aimed at researchers running early-stage peptide therapeutic campaigns, where triaging candidates on developability before synthesis saves substantial experimental effort. Because it scores both natural sequences and chemically modified SMILES, it fits generative design workflows as a property filter: candidates from a peptide generator can be ranked on solubility, permeability, toxicity, half-life, and target affinity in one pass, with uncertainty estimates flagging predictions that warrant experimental follow-up. An interactive web application makes the predictors usable without local setup.

#Impact

PeptiVerse consolidates peptide developability prediction — previously spread across tools that each handled a narrow property set or a single input representation — into one framework spanning natural and non-canonical peptides. Its release of open model weights, processed datasets, and a public web application under permissive licensing lowers the barrier to property-aware peptide design and makes the reported benchmarks reproducible. As a recent preprint, its results await peer review, and the reported metrics are in-silico benchmarks on similarity-based splits rather than prospective wet-lab outcomes; predictive performance also varies by property, with data-limited tasks such as half-life and Caco-2 permeability resting on only a few hundred labeled examples.