PepOpt — Peptide Lead Optimization

PepOpt is an interpretable machine learning framework for peptide lead optimization in ultralarge design spaces.

Authors: Yixiang Mao (MSD) · Ruibo Zhang (MSD)
Contact: ruibo.zhang@msd.com

---

Table of Contents

• PepOpt — Peptide Lead Optimization
  • Table of Contents
  • Overview
  • Requirements
  • Input Data
    • 1. Peptide Dataset
    • 2. Monomer Lookup Tables
  • Quick Start
  • Full Argument Reference
  • Output Files

---

Overview

PepOpt implements an FP-based Free Wilson Analysis (FP-FWA) pipeline that:

• Learns interpretable per-monomer contributions to a target activity using ridge regression on molecular fingerprints
• Visualizes the most important monomers and their structural substructures
• Enumerates and ranks novel peptide candidates from an ultralarge design space

---

Requirements

Install dependencies via:

pip install -r requirements.txt

---

Input Data

Two types of input files are required to run PepOpt.

1. Peptide Dataset

The peptide input must be pre-aligned, with each position represented as a separate column following the naming convention:

PEPTIDE1_0, PEPTIDE1_1, PEPTIDE1_2, ...

Each cell should contain the HELM monomer name for that position.

The CSV must also include:

Column	Description	Example
ID column	Unique compound identifier in the format letters + ("-" or "_") + numbers	PEP_00123, ABC-45
Target column	The numerical property to be predicted (e.g., pIC50)	pIC50
HELM column	HELM of peptides. HELMs are required for peptide enumeration even though all monomers are provided in prealigned data table

📄 Example file: test_data/example_data_aligned.csv

---

2. Monomer Lookup Tables

Monomer lookup tables map HELM monomer names to their structural SMILES fragments. You must provide at least two tables:

• One for monomers present in the input dataset
• One for monomers available for enumeration / prediction

Table paths are configured in src/config.py.

Each table must contain the following columns:

Column	Description
symbol	HELM monomer name
Partial_SMILES	Clipped SMILES fragment (reaction handles removed)

📄 Example file: test_data/example_monomer_db.csv

---

Quick Start

Run the FP-FWA pipeline from the Pepopt_Public/ directory:

python local_run_fpfw.py test_data/example_data_aligned.csv pIC50 ID M-0001 \
    --model ridge \
    --desc_type ecfp --desc_norm False \
    --fp_len 2048 --fp_radius 2 --fp_sel_n 500 --fp_min_freq 0 \
    --fpbit_plot True --fpbit_top_n 5 \
    --evaluation True \
    --max_enum_out 1000 \
    --enum_positions 0 1 2 3 4 5

The four positional arguments are:

Position	Argument	Description
1	incsv	Path to the aligned input CSV file
2	actcol	Name of the activity/target column
3	idcol	Name of the compound ID column
4	refid	ID of the reference peptide

---

Full Argument Reference

For a complete list of options, run:

python local_run_fpfw.py -h

---

Output Files

All outputs are saved to ./results/ (or the directory specified by --results_dir):

results/
├── fpfwa-coefficients.csv               # Per-monomer coefficients from the model
├── fpfwa-monomer_db_coefficients.csv    # Extrapolated coefficients for monomers in the DB
├── fpfwa_top10_monomers.png             # Heatmap of top monomer coefficients (training set)
├── fpfwa_top10_monomers_freq.png        # Same heatmap, weighted by monomer frequency
├── fpfwa_top10_monomer_db.png           # Heatmap of extrapolated DB monomer coefficients
├── fpfwa_best_1000.csv                  # Top enumerated peptides (training set monomers)
├── fpfwa_best_1000-monomer_db.csv       # Top enumerated peptides (monomer DB)
└── substructure_analysis/              # Per-position Morgan bit substructure figures