GPU-accelerated Reinforcement learning-enabled Integrated Combination of Automated Transformer-based Approaches with Ligand binding and 3D prediction for Enzyme Evolution

G-ReInCATALiZE is an advanced computational pipeline for in-silico enzyme evolution, developed by the CCBIO team at the University of Applied Sciences (ZHAW) in Wädenswil. The system combines cutting-edge machine learning approaches to optimize enzyme mutations for improved catalytic activity with specific target substrates.

Key Capabilities:

• Find optimal enzyme mutants from wildtype enzymes for targeted ligand transformations
• GPU-accelerated molecular docking and structure prediction
• Reinforcement learning-guided mutation optimization
• Transformer-based mutation effect prediction using ESM2 models

The pipeline integrates four main computational components as shown in the overview diagram:

1. DeepMut - ESM2 transformer-based semi-rational multi-site mutation prediction
2. RESIDORA - Deep reinforcement learning using Proximal Policy Optimization (PPO) for residue substitution
3. Pyroprolex - PyRosetta-based local mutant structure relaxation
4. GAESP - GPU-accelerated enzyme-substrate docking pipeline using Vina-GPU

The system uses an actor-critic reinforcement learning architecture to iteratively improve enzyme candidates through mutation and structural evaluation cycles.

• Docker with GPU support
• NVIDIA Container Toolkit for GPU acceleration
• NVIDIA GPU with CUDA support

1. Build the container:

   docker build --platform linux/amd64 -t gaesp .

2. Run with GPU support:

   docker run -d --gpus all --name gaesp-container -p 80:80 gaesp

3. Execute the pipeline:

   # Standard execution
   PYTHONPATH=. python src/main.py --config src/CONFIG/config_default.yaml
   
   # Debug mode
   PYTHONPATH=. python src/main.py --config src/CONFIG/config_debug.yaml

The pipeline is configuration-driven using YAML files in src/CONFIG/. Key configuration sections:

• globalConfig - Transformer models, GPU settings, file paths
• gaespConfig - Wildtype sequence, structure paths, docking parameters
• pyroprolexConfig - PyRosetta mutation settings

config1:
  globalConfig:
    transformerName: facebook/esm2_t6_8M_UR50D
    transformerDevice: cuda:0
    gpu_vina: true

  gaespConfig:
    wildTypeAASeq: "MSTETLRLQKARATEEGLAFETPGGLT..."
    wildTypeStructurePath: data/raw/enzyme.pdb
    reference_ligand: data/raw/reference/ligand.pdb
    boxSize: 15
    num_modes: 5

pytest tests/

wget https://boostorg.jfrog.io/artifactory/main/release/1.82.0/source/boost_1_82_0.tar.bz2
tar --bzip2 -xf boost_1_82_0.tar.bz2
cd boost_1_82_0
sudo ./bootstrap.sh
sudo ./b2 install

git clone https://github.com/DeltaGroupNJUPT/Vina-GPU-2.0
# Configure Makefile with appropriate paths
cd Vina-GPU-2.0/Vina-GPU+
make clean && make source

conda install -c conda-forge openbabel

# Install dependencies
poetry install

# Add new dependency
poetry add <package_name>

# Code formatting (Black configured for 79 character lines)
black src/

Distributed under the MIT License.

CCBIO Team - University of Applied Sciences (ZHAW), Wädenswil

Developed for computational enzyme engineering and in-silico directed evolution research.