Multi-Modal Foundation · ChatterjeeLab
PeptiVerse
A suite of machine learning models predicting biophysical and therapeutic properties of natural and modified peptides.
What PeptiVerse is
PeptiVerse is a collection of machine learning predictors developed by ChatterjeeLab for evaluating key therapeutic and biophysical characteristics of peptides. It accepts inputs either as standard amino acid sequences for canonical peptides or as SMILES strings, enabling analysis of non-canonical and chemically modified structures.
Designing therapeutic peptides requires optimising across multiple competing profiles, including membrane permeability, stability, safety, and target affinity. PeptiVerse unifies property prediction into a single system, helping researchers screen candidate molecules for hemolysis risk, aqueous solubility, non-fouling properties, cellular penetrance, toxicity, serum half-life, and peptide-target binding affinity.
The framework provides uncertainty estimates alongside predictions, using deep ensemble entropy for neural network classifiers and adaptive conformal prediction intervals for regression tasks to assist experimental triage.
How it works
PeptiVerse builds downstream predictive models on top of frozen representations from pre-trained foundation models. It derives amino acid sequence representations using ESM-2 (650M), and SMILES molecular representations using PeptideCLM or ChemBERTa (77M). Downstream heads operate on either mean-pooled fixed-length vectors or unpooled token-level representations.
Depending on the property and input mode, the best predictive heads utilise gradient-boosted decision trees (XGBoost), one-dimensional Convolutional Neural Networks (CNNs), multilayer perceptrons, self-attention Transformers, Support Vector Machines/Regression (SVM/SVR), or ElasticNet models. Peptide-protein binding affinity is evaluated using a transformer architecture with cross-attention between target and peptide embeddings.
Key facts
- Underlying foundation models
- ESM-2 (650M), PeptideCLM (23M), ChemBERTa (77M)
- Prediction heads
- CNN, Transformer, MLP, XGBoost, SVM, SVR, ElasticNet
- Input formats
- Amino acid sequences (canonical) and SMILES strings (canonical and non-canonical)
- Properties modelled
- Hemolysis, non-fouling, solubility, penetrance, toxicity, PAMPA, Caco-2, half-life, binding affinity
What you can use it for
- Hemolysis and Toxicity Screening
- Assess the probability of red blood cell disruption (hemolysis based on HC50 threshold) and general toxicity risks for therapeutic candidates using sequence or SMILES input.
- Membrane Permeability Profiling
- Predict qualitative penetrance as well as quantitative log P permeability across PAMPA and Caco-2 cellular assays.
- Peptide-Target Binding Affinity Estimation
- Score interaction strength between a target protein sequence and candidate peptide to identify binders falling into nanomolar, picomolar, or micromolar affinity bands.
- Stability and Aqueous Formulation Evaluation
- Evaluate relative serum persistence in hours alongside aqueous solubility to filter designs likely to suffer from rapid degradation or aggregation.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Supports both amino acid sequences and SMILES representations, allowing the characterisation of non-canonical residues and chemical modifications.
- Incorporates principled uncertainty quantification via predictive entropy for classifiers and adaptive conformal prediction intervals with 90% marginal coverage for regressors.
- Offers flexible deployment tiers, including a lightweight CPU/GPU setup and a GPU-accelerated suite using RAPIDS cuML for SVM and ElasticNet models.
Watch out for
- Certain prediction tasks, such as toxicity, PAMPA, and Caco-2 permeability, are available solely using SMILES representations.
- SVM and ElasticNet predictors depend on RAPIDS cuML, which requires a CUDA-capable GPU environment.
- Conformal prediction intervals provide marginal rather than conditional coverage guarantees, meaning intervals may not guarantee individual correctness for severe out-of-distribution molecules.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use PeptiVerse commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
Can PeptiVerse evaluate peptides containing non-canonical amino acids?
Yes. PeptiVerse supports SMILES inputs modelled via PeptideCLM and ChemBERTa embeddings, allowing it to represent and predict properties for chemically modified and non-canonical peptides.
How should the binding affinity score be interpreted?
Scores of 9 or greater indicate tight binders in the nanomolar to picomolar range (Kd ≤ 10⁻⁹ M), scores between 7 and 9 reflect medium binders (10⁻⁷ to 10⁻⁹ M), and scores under 7 indicate weak binders (Kd ≥ 10⁻⁶ M). A single-unit shift corresponds to roughly a tenfold difference in affinity.
What information does the uncertainty output convey?
For classification tasks, uncertainty is reported as predictive entropy where values near 0 imply high confidence and values near 0.693 denote maximum ambiguity. For regression tasks, it provides an adaptive conformal interval (lo, hi) delivering a 90% marginal coverage guarantee.
Are target protein structures required for binding affinity predictions?
No. The binding affinity model uses sequence inputs for both the target protein and the peptide, applying cross-attention over foundational embeddings.
How to cite it
Zhang, Y., Tang, S., Chen, T., Mahood, E., Vincoff, S., Chatterjee, P. PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction. bioRxiv, 2026. doi: 10.64898/2025.12.31.697180.
Written from PeptiVerse's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.