Merck/PepEVOLVE

1. Model Overview

• Model Name: PepEVOLVE
• Summary: PepEVOLVE is a position-aware, dynamic generative framework for linear and macrocyclic peptide lead optimization. It learns both where to edit and how to optimize peptides by combining dynamic masking and CHUCKLES shifting during pretraining, a context-free multi-armed-bandit router that discovers high-value residue positions, and an evolving reinforcement-learning loop with group-relative advantage (GRA). This enables property-biased peptide design when the optimal edit sites are not known a priori.
• Model Specification: Encoder–decoder Transformer. ~20M parameters.
• Developed by: Merck & Co., Inc. (Boston, MA, USA) and University of South Florida (FL, USA).
• License: Apache 2.0
• Base Model: PepINVENT (same transformer generator implementation and parameter configuration).
• Model Type: Transformer
• Languages: CHUCKLES & SMILES (chemical representation)
• Pipeline Tag: text2text-generation for peptide optimization
• Library: Transformers, PyTorch

2. Intended Use

• Direct Use:
  • Generation of chemically valid linear and macrocyclic peptides in CHUCKLES representation.
  • Position-aware peptide lead optimization when the mutable residues are not known a priori.
  • Router-guided identification of high-value residue positions from reward feedback.
• Downstream Use:
  • Integration into computational peptide design and multi-objective optimization pipelines.
• Out-of-Scope Use: Direct wet-lab synthesis or clinical decision-making without independent experimental validation; use as a substitute for medicinal chemistry, safety, IP, or developability review.

3. Bias, Risks, and Limitations

• Known Limitations: The public benchmark centers on a single Rev-binding macrocycle task with predefined mutable positions. The comparative scoring objective excludes solubility because the PepINVENT solubility code was not publicly available. The router is context-free (sequence-agnostic) and may miss sequence-dependent or 3D structural contingencies. Group-relative advantage stabilizes optimization but can reduce diversity through group-wise normalization pressure.
• Biases: Inherits biases from the semi-synthetic PepINVENT training distribution, CHUCKLES encoding choices, and the selected reward/scoring function. In the RBP benchmark the score weights permeability more heavily (exponent = 3) and targets lipophilicity around −4.0.
• Risk Areas: Intended for research use. Generated peptides and scores are computational predictions with no PII or ethically sensitive content, but generated chemical matter should undergo standard experimental, safety, and IP review.
• Recommendations: Validate generated peptides experimentally; interpret rankings as dependent on the chosen scoring objective and input scaffold.

4. Training Details

• Training Data: Semi-synthetic peptide dataset introduced in PepINVENT (~900,000 training and ~50,000 validation peptides), assembled from canonical and non-canonical alpha-amino acids, with stereochemical randomization, N-methylations, and both linear and cyclic designs.
• Training Data Preprocessing:
  • Peptides represented in CHUCKLES, a SMILES-derived encoding preserving atom-level detail while retaining monomer boundaries.
  • Residues concatenated in N-to-C order and separated by the "|" delimiter.
  • Dataset spans 6–18 residues; ~40% linear and ~60% macrocyclic; ~30% non-canonical residues.
• Pre-Training: Conditional text-infilling / masked language modeling over CHUCKLES. The number of masked positions is sampled from a triangular distribution biased toward single-residue masking; masked positions are dynamically resampled each epoch; CHUCKLES shifting is applied for robustness to cyclic re-indexing.
• Training Procedure:
  • Supervised sequence-to-sequence pretraining (masked-residue reconstruction).
  • Context-free multi-armed-bandit router learns residue-position selection via policy gradient.
  • Evolving optimization loop retains top-K candidates as seeds for subsequent steps.
  • Group-relative advantage (GRA) normalizes rewards within seed-derived groups to stabilize RL updates.
• Fine-Tuning: Task-specific reinforcement-learning optimization on objectives such as permeability, lipophilicity, maximum ring size, and SMARTS alerts.
• Environmental impact: Not applicable.
• Societal Impact Assessment: Intended for research use in computational peptide design.

5. Evaluation

• Metrics:
  • Validation loss on standard and CHUCKLES-shifted validation sets
  • Router selection probability and convergence to known target positions
  • Total multi-objective score and component scores (permeability, lipophilicity, maximum ring size, SMARTS)
  • Unique SMILES count, chemical diversity, and score-stratified yield
• Benchmarks:
  • Pretraining generalization on standard and CHUCKLES-shifted validation sets
  • Router tasks: hydrogen-bond donor reduction, LogP decrease, and adversarial LogP increase
  • Rev-binding peptide macrocycle optimization vs. PepINVENT
• Testing Data: Companion PepEVOLVE-RBP-Benchmark dataset of computationally generated peptides and scoring components from PepEVOLVE and PepINVENT optimization runs.

6. Model Architecture

• Architecture Details: PepEVOLVE uses the PepINVENT transformer generator as its generative engine — a conditional masked language model over CHUCKLES that predicts masked residue tokens autoregressively until an [END] token or a maximum length of 500 characters. PepEVOLVE augments this with dynamic pretraining, a router policy for automatic residue selection, and an evolving optimization loop with group-relative advantage.
• Diagram: See the PepEVOLVE manuscript figures (workflow, pretraining strategy, router analysis, and Rev-binding macrocycle optimization).

7. Usage

• Sample Inference Code: Described conceptually in the publication; inference conditions the model on a CHUCKLES peptide with one or more residue positions replaced by the mask token "?" and samples completions (e.g., with beam search).
• Sample Fine-Tuning Code: To be provided with the released code/model repository.
• Prompt Format: A CHUCKLES peptide string with one or more residue positions replaced by "?"; masking context depends on the configuration (self-mask or neighbor-mask).
• Hardware Requirements: Not specified.
• GitHub Links: https://github.com/MSDLLCpapers/PepEvolve

8. Citation

• BibTeX:

@article{nguyen2025pepevolve,
  title={PepEVOLVE: Position-Aware Dynamic Peptide Optimization via Group-Relative Advantage},
  author={Nguyen, Trieu and Zhang, Peter Zhiping and Boyer, Nicolas and Fang, Cheng and Zhang, Liying and Schneider, Sebastian and Pang, Hao-Wei and Feng, Shasha},
  journal={arXiv preprint arXiv:2511.16912},
  year={2025},
  doi={10.48550/arXiv.2511.16912}
}

• DOI: https://doi.org/10.48550/arXiv.2511.16912

9. Contributors

• Names and Roles: See Developed by in Section 1.

10. Contact Information

• Support Channels: Please contact the corresponding authors, Hao-Wei Pang and Shasha Feng, at Merck & Co., Inc.

11. Acknowledgements

• Acknowledgements: This work was conducted at Merck & Co., Inc., Boston, MA, USA, with a contribution from the University of South Florida, FL, USA. The authors thank James P. Jewell, Edward Miller, Ryan Chau, Ruheng Wang, Ruibo Zhang, Yixiang Mao, Alec Glisman, Haote Li, Kenneth López Pérez, and Song Yin for discussions and insights.

12. Disclaimer

• Legal Disclaimer: This model is provided "as is" without warranty of any kind. Generated peptide sequences and associated scores are computational predictions and have not been experimentally validated. Users should independently verify any generated candidates before making experimental or clinical decisions. Merck & Co., Inc. makes no representations regarding suitability for any particular purpose. Use is subject to the Apache 2.0 license terms.