PMGen: From Structure Prediction to Peptide Generation

PMGen (Peptide MHC Generator) is a comprehensive pipeline for predicting peptide-MHC (pMHC) complex structures and designing optimized peptide sequences.

Key Features

• Fast & accurate structure prediction using AlphaFold with template engineering or initial guess mode
• Peptide sequence design with structure-aware optimization
• MHC pseudo-sequence design for customized allele engineering
• Iterative peptide optimization with binding prediction
• Mutation screening for systematic variant analysis
• Batch processing for multiple pMHC complexes

Installation

Requirements

• Python 3.8+
• Conda or Mamba
• Git Optional
• Modeller (requires a license key - get it here)
• CUDA-enabled GPU (Required for faster Alphafold predictions)

Setup

git clone https://github.com/soedinglab/PMGen.git
cd PMGen
bash -l install.sh
#or, for CPU only support run: bash -l install.sh --cpu
conda activate PMGen

You will be prompted to enter your Modeller license key. The script automatically:

• Creates the PMGen Conda environment
• Installs all dependencies
• Downloads AlphaFold parameters
• Clones PANDORA and ProteinMPNN

Optional: Configure NetMHCpan (Recommended)

Install NetMHCpan and NetMHCIIpan, then edit user_setting.py:

netmhcipan_path = '/path/to/netMHCpan'
netmhciipan_path = '/path/to/netMHCIIpan'

Quick Start

Prepare Input File

Create a tab-separated file (input.tsv) with your pMHC data:

peptide	mhc_seq	mhc_type	anchors	id
GILGFVFTL	GSHSMRYFFTSVSRPGRGEPRFIAVGYVDDTQFVRFDSDAASQRMEPRAPWIEQEGPEYWDGETRKVKAHSQTHRVDLGTLRGYYNQSEAGSHTVQRMYGCDVGSDWRFLRGYHQYAYDGKDYIALKEDLRSWTAADMAAQTTKHKWEAAHVAEQLRAYLEGTCVEWLRRYLENGKETLQRTDAPKTHMTHHAVSDHEATLRCWALSFYPAEITLTWQRDGEDQTQDTELVETRPAGDGTFQKWAAVVVPSGQEQRYTCHVQHEGLPKPLTLRWE	1		sample1
KLGGALQAK	GSHSLKYFHTSVSRPGRGEPRFISVGYVDDTQFVRFDSDAASPRGEPRAPWVEQEGPEYWDRNTQIFKTNTQTYRENLRIALRYYNQSEAGSHIIQRMYGCDLGPDGRLLRGHDQYAYDGKDYIALNEDLRSWTAADTAAQITQRKWEAAREAEQRRAYLEGTCVEWLRRYLKNGNATLLRTDSPKTHMTHHPISDHEATLRCWALGFYPAEITLTWQRDGEDQTQDTELVETRPAGDRTFQKWAAVVVPSGEEQRYTCHVQHEGLPKPLTLRWE	1		sample2

Columns:

• peptide: Peptide sequence
• mhc_seq: MHC sequence (for MHC-II: Alpha/Beta separated by /)
• mhc_type: 1 for MHC-I, 2 for MHC-II
• anchors: Anchor positions (leave empty for prediction)
• id: Unique identifier

Basic Structure Prediction (Recommended)

Single-threaded mode with initial guess (fastest & most accurate):

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess

This is the preferred mode for most users. It uses:

• --mode wrapper: Works for one or more than one prediction per run.
• --run single: Sequential processing (unparallel)
• --initial_guess: Fast and more accurate AlphaFold mode without homology modeling (recommended)

Common Use Cases

1. Structure Prediction with Multiple Models

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess \
  --models model_1_ptm model_2_ptm model_3_ptm

2. Peptide Sequence Design

Generate optimized peptide variants:

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess \
  --peptide_design \
  --num_sequences_peptide 50 \
  --binder_pred

3. MHC Pseudo-Sequence Design

Customize MHC binding groove residues:

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess \
  --only_pseudo_sequence_design \
  --num_sequences_mhc 20

4. Iterative Peptide Optimization

Optimize peptides over multiple rounds:

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess \
  --peptide_design \
  --binder_pred \
  --iterative_peptide_gen 3 \
  --fix_anchors

5. Mutation Screening

Systematically test point mutations:

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess \
  --mutation_screen \
  --n_mutations 1

Key Options

Flag	Description
--mode wrapper	Batch processing mode (recommended)
--run single	Sequential processing (recommended)
--initial_guess	Fast AF mode without templates (recommended)
--peptide_design	Enable peptide sequence generation
--only_pseudo_sequence_design	Design MHC binding groove only
--binder_pred	Predict binding affinity (requires NetMHCpan)
--fix_anchors	Keep anchor positions fixed during design
--iterative_peptide_gen N	Run N rounds of optimization
--mutation_screen	Systematic mutation analysis
--num_templates	Number of structural templates (default: 4)
--num_recycles	AlphaFold recycles (default: 3)

Output Structure

output/
├── pandora/              # Template structures
├── alphafold/            # Predicted pMHC structures
├── proteinmpnn/          # Designed sequences
│   └── {id}/
│       ├── peptide_design/
│       └── only_pseudo_sequence_design/
└── best_structures/      # Top-ranked models (if --best_structures used)

Citation

@article{asgary2025pmgen,
  author = {Asgary, Amir H. and Amirreza and others},
  title = {PMGen: From Peptide-MHC Prediction to Neoantigen Generation},
  journal = {bioRxiv},
  year = {2025},
  month = {11},
  date = {2025-11-14},
  doi = {10.1101/2025.11.14.688404},
  url = {https://doi.org/10.1101/2025.11.14.688404},
  note = {Preprint}
}

PMGen uses the following tools and papers:

• PANDORA: Antunes et al., Front. Immunol. 2022
• AlphaFold: Jumper et al., Nature 2021
• AFfine: Bradley et al., PNAS 2023
• ProteinMPNN: Dauparas et al., Science 2022

Support

For issues or questions, please open an issue on GitHub.