Models Matter

This package extends AiZynthFinder to seamlessly integrate various single-step retrosynthesis models, as illustrated in our papers Models Matter: The Impact of Single-Step Models on Synthesis Prediction and Mind the Retrosynthesis Gap: Bridging the divide between Single-step and Multi-step Retrosynthesis Prediction

Overview

Models Matter uses the default AiZynthFinder implementation and introduces the following enhancements:

Smiles-based Expansion Strategy: This feature allows the integration of any retrosynthesis model operating at smiles-level with AiZynthFinder, not depending on the underlying search algorithm.
ModelZoo Integration: The package includes ModelZoo, a dedicated framework that defines the single-step retrosynthesis approach within the smiles-based expansion strategy. Currently supported implementations are Chemformer, MHNreact, and LocalRetro.

Installation Procedure

To install the package, follow the sequential steps below:

# Clone this repository and all submodules
git clone --recursive https://github.com/AlanHassen/modelsmatter

# Create and initialize a conda environment
conda env create -f environments/environment.yml -n ssbenchmark
conda activate ssbenchmark

# Transition to the SSBenchmark directory
cd external/modelsmatter_modelzoo/

# First the installation of the ModelZoo
poetry install

# Navigate back to the models matter directory
cd ../..

# Finalize the installation process
poetry install

# Subsequently, install the appropriate single-step models or the necessary libraries.

Usage Guidelines

Adaptations to AiZynthFinder config are necessary to accommodate the different single-step models (examples provided in config/). The configurable settings include:

gpu_mode: Enable GPU mode for accelerated inference.
module_path: The path of the single-step retrosynthesis model repository.
model_path: The location of the trained model.
Additional parameters can be set, such as specifying the vocabulary for Chemformer via vocab_path.

Utilize the extended functionalities of AiZynthFinder through Models Matter for a comprehensive synthesis prediction experience.

Datasets & Models

Datasets and models that are not proprietary will be accessible upon publication.