Benchmarking of reaction representations

Installation

BERT+RXNFP environment

• Due to the various dependencies of different packages, it is not possible to have a single environment that satisfies them all. In particular, the rxnfp package relies on python 3.7 (latest python version for which is works). Therefore it is recommended to have a separate environment to run the BERT+RXNFP models:

conda create -n rxnfp python=3.7 -y
conda activate rxnfp
conda install -c rdkit rdkit=2020.03.3 -y
conda install -c tmap tmap -y
git clone https://github.com/rxn4chemistry/rxnfp
pip install -e rxnfp

• Note that the above does not work on Mac computers with the M1 chip onwards, since they do not support python 3.7. This env has been tried and tested on several Linux platforms.
• It is recommended to run BERT+DRFP in a GPU-enabled environment.

Chemprop environment

• This can be installed using conda create -n <environment-name> --file requirements_chemprop.txt
• Note a modified version of chemprop is used (forked from 1.6.1), which allows for hyperparameter optimisation only on the first CV fold and reports timings for training and inference separately.
• It is recommended to run Chemprop in a GPU-enabled environment.

Fingerprints environment

• This can be installed using conda create -n <environment-name> --file environment_fingerprints.yml
• The key packages are qml (for SLATM and B2R2 dependencies), drfp (for DRFP), rdkit (for MFP, careful with the version) for the fingerprints and scikit-learn and numpy for the models
• These fingerprints will not benefit from a GPU environment, run on CPU only.

EquiReact environment

• Consult the dedicated repo for installation instructions: https://github.com/lcmd-epfl/EquiReact
• It is recommended to run EquiReact in a GPU-enabled environment.

Using the repo / reproducing the results

2D/3D fingerprints

• 2D reps here are the DRFP and MFP
• 3D reps are SLATM and $B^2R^2_l$, all in the difference variation for reactions
• To generate/load the reps and perform CV-fold cross-validated predictions with a train fraction of tr, the file src/run_all_fingerprints.py should be used
• Here, the datasets are specified as additional arguments, -c for Cyclo-23-TS, -g for GDB7-22-TS and -p for Proparg-21-TS
• For example, for a train fraction of 0.8, with corresponding test and validation of 0.1 each, and 10-fold CV, the command python src/run_all_fingerprints.py -c -g -p --CV=10 --train=0.8 will generate/load representations and run the models or load the results
• The hyperparameters for each dataset can be found in src/hypers.py

Language models

• A pre-trained BERT model is fine-tuned on the appropriate datasets (cyclo, gdb, or proparg)
• Here, the dataset is provided as an optional argument to src/train_bert.py
• To train on and then predict on the cyclo dataset for example, the command is python src/train_bert.py -t -p --dataset='gdb' --test_size=0.2
• The optional --CV flag determines whether to shuffle the train and predict datasets over the CV iterations
• The prediction MAE of 10 CV iterations are saved to outs/cyclo_bert_pretrained/results.txt (and the equivalent for the other datasets)
• The hyperparameters for each dataset can be found in src/hypers.py

Graph based models

• The CGR model was run using the provided atom mapping, automatic atom mapping from rxnmapper, and no atom mapping
• The data are in the csv files data/gdb7-22-ts/ccsdtf12_dz.csv for GDB7-22-TS, data/cyclo/full_dataset.csv for Cyclo-23-TS and data/proparg/data.csv for Proparg-21-TS
• SMILES with the provided ("true") atom mapping are in the rxn_smiles (Cyclo-23-TS, GDB7-22-TS) or rxn_smiles_mapped (Proparg-21-TS) columns of each csv file
• To run rxnmapper, run src/mapper.py
• Atom mapped SMILES from rxnmapper are in the rxn_smiles_rxnmapper column of each csv file
• Python file specifying how each CGR model was run is src/cgr.py. Note the hypers for each dataset are in data (these are referenced by cgr.py
• Results of the CGR runs can be found in results

EquiReact

• The submission files for each of the EquiReact jobs can be found in the equireact-subfiles directory. Note this will need to be changed somewhat depending on your cluster structure (there are hard-coded paths for scratch for example).
• The results of running EquiReact can be found in the equireact-results directory