Genie 2.0

GPU-Accelerated MCMC Sampling for Protein Sequences with Codon-Level Mutations

Genie 2.0 is a high-performance tool for generating protein sequences using Direct Coupling Analysis (DCA) models combined with biologically realistic codon substitution dynamics. It implements efficient MCMC sampling on GPUs with two variants:

• Genie: DNA codon-aware evolution with Metropolis-Gibbs sampling
• Genie-AA: Amino acid-only evolution with standard Gibbs sampling

---

Table of Contents

• Features
• Installation
• Quick Start
• Usage
• Command-Line Arguments
• Output Files
• Algorithm Overview
• Performance
• Requirements
• Examples
• Citation
• License

---

Features

Core Capabilities

• Codon-Aware Sampling: Biologically realistic single-nucleotide mutations at DNA level
• Hybrid MCMC: Combined Metropolis-Hastings and Gibbs sampling for better mixing
• Reference-Based: Optional convergence tracking against real sequence data
• GPU-Accelerated: Full CUDA support with PyTorch JIT compilation (2-3x speedup)
• Flexible Input: Start from existing sequences or random initialization

Technical Highlights

• Fully vectorized GPU kernels with zero CPU loops
• Pre-computed codon mutation networks for O(1) neighbor lookups
• Batched random number generation for improved GPU efficiency
• Real-time Pearson correlation tracking for convergence monitoring

---

Installation

From PyPI (Recommended)

pip install genie-dca

From Source

git clone https://github.com/spqb/Genie.py.git
cd Genie.py
pip install .

This installs two command-line tools:

• genie - Codon-aware evolution
• genie-aa - Amino acid evolution

---

Quick Start

Codon-Aware Evolution

genie -p params.dat -n 1000 --num_iterations 50000 -o output_folder

Amino Acid Evolution

genie-aa -p params.dat -n 1000 --num_iterations 50000 -o output_folder

---

Usage

Genie (Codon-Aware Evolution)

# Generate sequences from scratch
genie -p params.dat -n 1000 --num_iterations 50000 -o results/

# Start from existing sequences
genie -c init_sequences.fasta -p params.dat --num_iterations 50000 -o results/

Genie-AA (Amino Acid Evolution)

# Generate sequences from scratch
genie-aa -p params.dat -n 1000 --num_iterations 50000 -o results/

# Start from existing sequences
genie-aa -c init_sequences.fasta -p params.dat --num_iterations 50000 -o results/

Reconstruction Tools

# Reconstruct final sequences from mutation log
reconstruct_chains results/

# Reconstruct sequences at specific timesteps
reconstruct_at_timesteps results/ --timesteps "0,100,500,1000"

Python API

from Genie import reconstruct_at_timesteps, reconstruct_chains_from_log
from adabmDCA.fasta import get_tokens

# Reconstruct sequences at specific timesteps
sequences = reconstruct_at_timesteps(
    initial_chains_file="results/initial_chains.fasta",
    mutation_log_file="results/mutation_log.csv",
    timesteps=[0, 100, 500, 1000],
    alphabet="protein"
)
# Returns: torch.Tensor of shape (len(timesteps), n_chains, L)

# Reconstruct and validate final sequences
tokens = get_tokens(alphabet="protein")
reconstructed_seqs, headers = reconstruct_chains_from_log(
    initial_chains_file="results/initial_chains.fasta",
    mutation_log_file="results/mutation_log.csv",
    tokens=tokens
)

---

Command-Line Arguments

Required Arguments

Argument	Short	Description
--path_params	-p	DCA model parameters file (.dat)
--num_iterations		Number of MCMC iterations

Optional Arguments

Argument	Short	Default	Description
--output	-o	DCA_evolution	Output directory
--num_chains	-n	None	Number of sequences (required if not using -c)
--path_chains	-c	None	Initial sequences (FASTA format)
--seq_index		None	Replicate single sequence from -c file
--save_steps		100	Checkpoint interval or comma-separated list (e.g., "100,500,1000")
--device		auto	Device: 'cuda' or 'cpu'
--dtype		float32	Data type: float32 or float64

Genie-Specific Arguments

Argument	Default	Description
--p_metropolis	0.5	Metropolis vs Gibbs ratio (0.0-1.0)

Genie-AA Specific Arguments

Argument	Default	Description
--alphabet	protein	Alphabet type: 'protein', 'rna', 'dna', or custom

Reconstruction Tool Arguments

reconstruct_chains: Takes output folder as positional argument

reconstruct_at_timesteps:

• folder - Output folder (positional)
• --timesteps - Comma-separated list (e.g., "0,100,500,1000")

---

Output Files

All files are saved in the output directory specified by -o.

Generated Files

File	Description
initial_chains.fasta	Starting sequences (before evolution)
final_chains.fasta	Final sequences (after all iterations)
mutation_log.csv	Log of all mutations at checkpoints

Mutation Log Format

File: mutation_log.csv

CSV file tracking mutations at checkpoints:

Column	Description
iteration	Checkpoint iteration number
chain_id	Sequence identifier
position	Position in sequence (0-indexed)
new_aa	New amino acid at this position

Example:

iteration,chain_id,position,new_aa
100,seq_0,15,A
100,seq_0,42,G
100,seq_1,23,L
200,seq_0,15,V
...

Console Output

Real-time progress showing:

• Iteration number and speed (iter/sec)
• Elapsed time
• Compilation status (first iteration)

---

Algorithm Overview

Genie (Codon Evolution)

1. Initialization: Load DCA model, build codon mutation network
2. Sequence Translation: Convert amino acids to codons
3. MCMC Sampling: Hybrid Metropolis-Gibbs with codon mutations
4. Convergence Tracking: Optional Pearson correlation monitoring

Genie-AA (Amino Acid Only)

1. Initialization: Load DCA model
2. Gibbs Sampling: Standard position-wise sampling
3. Convergence Tracking: Optional correlation monitoring

---

Performance

Hardware: NVIDIA RTX 4090, 1000 sequences, L=100

Mode	Iterations/sec	Speedup
Genie (compiled)	~45-50	2.5x
Genie (eager)	~18-20	1.0x
Genie-AA (compiled)	~120-140	6.5x

Note: First iteration includes ~10-30s JIT compilation overhead

---

Requirements

torch>=2.0.0
numpy>=1.20.0
pandas>=1.3.0
adabmDCA>=1.0.0

Hardware:

• Minimum: CPU with 4GB RAM
• Recommended: NVIDIA GPU (8GB+ VRAM) with CUDA 11.7+

---

Examples

Basic Evolution

# Generate 1000 sequences with codon awareness
genie -p example_data/pf76/params.dat -n 1000 --num_iterations 50000 -o results/pf76

# Generate amino acid sequences only
genie-aa -p example_data/pf76/params.dat -n 1000 --num_iterations 50000 -o results/pf76_aa

Custom Checkpoints

# Save mutations at specific iterations
genie -p params.dat -n 1000 --num_iterations 10000 --save_steps "100,500,1000,5000,10000" -o results/

# Reconstruct sequences at those timesteps
reconstruct_at_timesteps results/ --timesteps "0,100,500,1000,5000,10000"

---

Citation

This software is based on the following article:

@article{
doi:10.1073/pnas.2406807121,
author = {Leonardo Di Bari  and Matteo Bisardi  and Sabrina Cotogno  and Martin Weigt  and Francesco Zamponi },
title = {Emergent time scales of epistasis in protein evolution},
journal = {Proceedings of the National Academy of Sciences},
volume = {121},
number = {40},
pages = {e2406807121},
year = {2024},
doi = {10.1073/pnas.2406807121},
URL = {https://www.pnas.org/doi/abs/10.1073/pnas.2406807121},
eprint = {https://www.pnas.org/doi/pdf/10.1073/pnas.2406807121},
}

A Julia version of Genie is also available: Genie.jl

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Acknowledgments

• Built on the adabmDCA library
• PyTorch team for excellent GPU optimization tools