Merck/PepEVOLVE-RBP-Benchmark

1. Dataset Overview

• Dataset Name: PepEVOLVE-RBP-Benchmark
• Summary: Benchmark dataset for evaluating macrocyclic peptide optimization methods on the Rev-binding peptide (RBP) task. Contains generated peptide sequences (in both SMILES and CHUCKLES representations) and multi-objective scoring components produced by PepEVOLVE (4 configurations: self-mask single-agent, self-mask multi-agent, neighbor-mask single-agent, neighbor-mask multi-agent) and PepINVENT during reinforcement-learning-driven optimization of the therapeutically relevant Rev-binding macrocycle. The goal is to optimize permeability, lipophilicity, maximum ring size, and structural alerts (SMARTS) for the cyclic RBP peptide derived from the antiviral candidate YPAASYR.
• Dataset Size: Approximately 4.3 M rows total across all configurations. The dataset consists of multiple CSV files (one per model configuration for a single run), with 5 model configurations × 1 random runs = 5 CSV files.
• Developed by: Merck & Co., Inc. — Trieu Nguyen, Peter Zhiping Zhang, Nicolas Boyer, Cheng Fang, Liying Zhang, Sebastian Schneider, Hao-Wei Pang, Shasha Feng.
• License:
• Source Data: The dataset is entirely computationally generated. Peptide sequences were produced by the PepEVOLVE and PepINVENT generative models during reinforcement learning optimization runs. The starting macrocycle is derived from the Rev-binding peptide YPAASYR (with two appended glycine residues and head-to-tail macrocyclization), as described in Wu et al (Wu, H.; Mousseau, G.; Mediouni, S.; Valente, S. T.; Kodadek, T. Cell-Permeable Peptides Containing Cycloalanine Residues. Angew Chem Int Ed Engl 2016, 55 (41), 12637–12642. DOI:10.1002/anie.201605745). The CHUCKLES representation of the starting macrocycle is: N2[C@@H](Cc1ccc(O)cc1)C(=O)|N1[C@@H](CCC1)C(=O)|N[C@@H](C)C(=O)|N[C@@H](C)C(=O)|N[C@@H](Cc1ccc(O)cc1)C(=O)|N[C@@H](CCCNC(=N)N)C(=O)|NCC(=O)|NCC2(=O)
• Dataset Structure: Tabular CSV format. One file per model configuration per run (e.g., pepevolve_SS_run1.csv, pepevolve_SM_run1.csv, pepevolve_NS_run1.csv, pepevolve_NM_run1.csv, pepinvent_run1.csv, etc.). Each row represents a single generated peptide at a given optimization step. Columns include both SMILES and CHUCKLES molecular representations, individual scoring component values (permeability, lipophilicity, maximum ring size, SMARTS alerts), the aggregated total score, and metadata (model name, mode, run number, step number).
• Languages: N/A (chemical/molecular dataset, not NLP).
• Task Categories: Multi-objective molecular optimization, generative molecular design, reinforcement-learning-based lead optimization, macrocyclic peptide design benchmarking.

2. Intended Use

• Direct Use: This dataset can be used to:
  • Benchmark generative peptide optimization methods against PepEVOLVE and PepINVENT on the Rev-binding peptide task.
  • Analyze optimization dynamics (score trajectories, convergence behavior, diversity-quality trade-offs) of different model configurations.
  • Study chemical space exploration patterns during reinforcement learning for macrocyclic peptide design.
  • Evaluate yield-quality trade-offs and sample efficiency of generative optimization strategies.
• Out-of-Scope Use: This dataset should not be used as a training set for generative models without understanding that it contains optimization trajectories (early-step peptides are low-quality by design). It is not suitable for direct wet-lab synthesis decisions without additional experimental validation.

3. Bias, Risks, and Limitations

• Known Limitations:
  • Scoring is performed entirely in silico using computational predictors for permeability, lipophilicity, ring size, and SMARTS alerts. The solubility component used in the original PepINVENT study was not publicly available and is therefore excluded.
  • The benchmark is limited to a single starting macrocycle (RBP-derived), so conclusions about model performance may not generalize to all peptide scaffolds.
• Biases: The scoring function weights permeability more heavily (exponent = 3 in the geometric mean), which biases the generated peptides toward high-permeability solutions. Lipophilicity is targeted around -4.0, which constrains the explored chemical space. No demographic or social biases are applicable to this molecular dataset.
• Risk Areas: None. The dataset contains only computationally generated macrocyclic peptide sequences and associated scores. No PII (personal identifiable information), toxic content, or ethically sensitive material is present.
• Recommendations: NONE — no biases or risks requiring mitigation were identified beyond standard caveats about in silico scoring accuracy.

4. Dataset Creation Details

• Data Collection: Data was generated through reinforcement learning optimization runs using two generative frameworks:

  • PepEVOLVE (4 configurations): self-mask single-agent (SS), self-mask multi-agent (SM), neighbor-mask single-agent (NS), and neighbor-mask multi-agent (NM). Each configuration was run for 250 steps × 4 input contexts = 1000 total update steps, with evolving phase parameters K=16 (top seeds) and G=8 (candidates per seed).
  • PepINVENT: Run for 1000 steps to equalize total update steps across methods.

  Each configuration was repeated 3 times with different random seeds for robustness. The scoring function is a permeability-weighted geometric mean: Score = (S_perm^3 × S_ring × S_lipophilicity × S_SMARTS)^(1/6).

• Data Preprocessing/Cleaning: Generated SMILES/CHUCKLES strings are recorded as-is from the model outputs. Scoring is computed using the multi-objective scoring function defined in the paper.

• Annotations: No manual annotations. All scores are computed programmatically using the scoring components.

• Personal and Sensitive Information: The dataset contains no personally identifiable information (PII) or sensitive content. All data is computationally generated molecular structures and associated numerical scores.

5. Evaluation and Splits

• Splits: No train/validation/test splits. The dataset contains the raw generation output from reinforcement learning optimization runs. Each file represents one complete optimization trajectory (all steps from a single run of a single configuration). Users may define their own splits if needed for downstream analysis.
• Evaluation Metrics:
  • Total multi-objective score (geometric mean, higher is better)
  • Individual component scores: permeability, lipophilicity, ring size, SMARTS
  • Unique SMILES count per step (sample efficiency)
  • Chemical diversity (1 − mean Tanimoto similarity over generated set)
  • Score distribution and yield in score bins (0.40–0.60, 0.60–0.80, 0.80–0.90, 0.90–0.94, 0.94–1.00)

6. Usage

• Sample:

from datasets import load_dataset 

# Load the full dataset 
dataset = load_dataset("merck/PepEVOLVE-RBP-Benchmark") 

# Or load a specific configuration file 
import pandas as pd 
df = pd.read_csv("pepevolve_SS_run1.csv") 

# Iterate through generated peptides 
for idx, row in df.iterrows():
    smiles = row["SMILES"]
    chuckles = row["CHUCKLES"]
    total_score = row["total_score"]
    step = row["step"]
    print(f"Step {step}: {smiles} -> Score: {total_score:.3f}") 

# Filter high-scoring peptides 
high_score = df[df["total_score"] >= 0.9] 
print(f"Number of peptides with score >= 0.9: {len(high_score)}") 

# Compare models 
import glob 
all_files = glob.glob("pepevolve_*.csv") + glob.glob("pepinvent_*.csv") 
combined = pd.concat([pd.read_csv(f) for f in all_files]) 
print(combined.groupby("model")["total_score"].describe())

• Data Format: Each CSV file contains the following columns:

  • SMILES: SMILES string of the generated peptide
  • CHUCKLES: CHUCKLES representation of the generated peptide
  • total_score: Aggregated multi-objective score (geometric mean)
  • predictive_model: Permeability component score
  • lipophilicity: Lipophilicity component score
  • maximum_ring_size: Maximum ring size component score
  • custom_alerts: Structural alerts (SMARTS) component score
  • step: Optimization step number

• GitHub Links (Optional): https://github.com/MSDLLCpapers/PepEvolve/

7. 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

8. Contributors

• Names and Roles: N/A — see Developed by in Section 1.

9. Contact Information

• Support Channels: For questions or feedback, please open an issue on the HuggingFace dataset repository or contact the corresponding authors (Hao-Wei Pang, Shasha Feng) at Merck & Co., Inc.

10. Acknowledgements

This work was conducted at Merck & Co., Inc., Boston, MA, USA, with a contribution from the University of South Florida, FL, USA.

11. Disclaimer

• Legal Disclaimer (Mandatory): This dataset is provided "as is" without warranty of any kind. The generated peptide sequences and associated scores are computational predictions and have not been experimentally validated. Users should independently verify any findings before making experimental or clinical decisions. Merck & Co., Inc. makes no representations regarding the suitability of this data for any particular purpose. Use of this dataset is subject to the CC-BY-4.0 license terms.