Predictive Minisci and P450 Late Stage Functionalization with Transfer Learning

Code for the ¹³C NMR pretraining and LSF finetuning as described in King-Smith et al.

Getting Your Bearings

set_lsf:

• The modules used to train the best model(s).
• The scoring function noted in the Supporting Information.
• The module to predict LSF regioselectivity on new molecules.
• Generating the Random Forest Baseline.
• The pickle file containing the predictions of the prospective molecules from the best model, MPNN_LSF.

data:

• The molecules used in the prospective validation, pre- and post-Glasgow Subgraph Solver processing.
• The open-source ¹³C NMR data used in the pretraining.

neural_nets & trained_models:

• The Message Passing Neural Networks (MPNNs) used for pretraining, finetuning, and for running new molecules.
• The best trained models for each of the neural networks.

P450:

• The P450-specific setup utils and finetuning train/test module.

utils:

• The modules for preparing the data for running in the MPNNs.
• The modules for automatically extracting the reaction sites from input excel data.
• The module used to find and cluster maximally different molecules used in prospective validation.

Fukui_Additions:

• The module to set up the MPNN with Fukui indices as part of the atom featurization.

Jensen_comparison:

• The modules used to setup the training / testing data for comparison to Jensen et al.'s ml-QM-GNN.

How to Use

To run predictions on new molecules:

1. Prepare an excel file in the format of prospective_with_product_smiles.xlsx.
2. Run reacting_centres.py -d {PATH to your excel file} -s {PATH to save file}.
3. Run Run_New_Molecules.py -d {PATH TO gss_to_react_centers.py OUTPUT} -m neural_nets/trained_models/best_retrospective_model -s {PATH TO SAVE FILE}.

Dependencies

Run on python 3.7.

• numpy==1.21.2
• pandas==1.3.4
• rdkit==2020.09.1
• torch==1.8.1+cu111
• logging==0.5.1.2
• argparse==1.1
• matplotlib==3.3.4
• networkx==1.11
• tdqm==4.62.3
• sklearn==1.0.1
• os
• pathlib
• collections