- Python 98.5%
- Jupyter Notebook 0.9%
- Shell 0.6%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| data | ||
| src | ||
| .gitignore | ||
| __init__.py | ||
| active_learning.py | ||
| CITATION.cff | ||
| LICENSE | ||
| predict_binary_gnn.py | ||
| predict_binary_xgb.py | ||
| predict_minisci_gnn.py | ||
| predict_regio.py | ||
| README.md | ||
| setup_env.sh | ||
| train_binary_xgb.py | ||
| train_gnn_ensemble.sh | ||
| train_gnn_workflow.py | ||
Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning
This repository contains a reference implementation detailing the workflows for inference and training as well active learning and the experiments on toy datasets introduced in the paper by Mason Minot, Yannick Stenzhorn, Jens Wolfard, Sebastian Strobel, Philippe Jablonski, Daniel Zimmerli, Martin Binder, Uwe Grether, Rainer E. Martin, Alex T. Müller, David F. Nippa, Kenneth Atz, & Gisbert Schneider.
Dataset preprocessing is embedded into each of the respective workflows.
Table of contents
- Prepare working environment
- Activate working environment
- Downloading weights and data
- Binary prediction XGBoost
- Binary prediction GNN
- Regioselectivity prediction
- Minisci prediction
- Binary training XGBoost
- Binary training GNN
- Regioselectivity training
- Minisci training
- Active learning
- Toy experiments
- Citing This Work
Prepare working environment
To setup the Conda environment, run ./setup_env.sh in the cmd line
Activate working environment
After the initial environment setup and to use any of the scripts below, ensure the environment is activate via
cd active-drug-discovery
conda activate active_disc
Downloading weights and data
Pretrained model weights are distributed via the GitHub Releases page (they are not stored in the git repository). Download the assets from the latest release and place them in src/model_weights/:
mkdir -p src/model_weights/binary_xgb
# Regioselectivity GNN (EquiformerV2) and binary reactivity GNN (PONITA)
curl -L -o src/model_weights/equiformer_regio.pth \
https://github.com/minotm/active-drug-discovery/releases/download/v1.0.0/equiformer_regio.pth
curl -L -o src/model_weights/ponita_binary.pth \
https://github.com/minotm/active-drug-discovery/releases/download/v1.0.0/ponita_binary.pth
# Binary reactivity XGBoost ensemble (+ preprocessing) -> unpack into binary_xgb/
curl -L -o /tmp/binary_xgb.zip \
https://github.com/minotm/active-drug-discovery/releases/download/v1.0.0/binary_xgb.zip
unzip -o /tmp/binary_xgb.zip -d src/model_weights/binary_xgb
No Minisci weights are shipped; train your own (see Minisci training).
Training data
The training data used in the paper ships with this repository under data/training_data/:
| File | Task | Delimiter | Description |
|---|---|---|---|
regioselectivity.csv |
Regioselectivity | tab | Substrate SMILES and borylation product_smiles pairs used to train the regioselectivity model. |
pre_AL_binary_and_yield.csv |
Binary + yield | semicolon | Initial (pre active learning) C–H borylation reactions with reaction conditions, the binary reactivity label, and sum_mono_bo_conv yield/conversion. |
AL_1_binary_and_yield.csv |
Binary + yield | tab | Reactions acquired in the first active learning round (conditions, binary label, sum_mono_bo_conv yield). |
AL_2_binary_and_yield.csv |
Binary + yield | tab | Reactions acquired in the second active learning round. |
AL_3_binary_and_yield.csv |
Binary + yield | tab | Reactions acquired in the third active learning round. |
minisci_data.tsv |
Minisci alkylation | tab | Full Minisci-type C–H alkylation dataset (13,490 reactions, 37 columns) with reaction conditions, mono_alkylation/di_alkylation/non_alkylation yields, and a binary reactivity label. |
The borylation binary/yield files contain full reaction context (substrate, reagent, catalyst, ligand, solvent SMILES plus equivalents and conditions). The binary column is the reactivity label and sum_mono_bo_conv is the measured yield/conversion target.
Minisci dataset source
The Minisci-type C–H alkylation dataset (minisci_data.tsv) is reproduced from the dataset published with Nippa et al., "Expediting hit-to-lead progression in drug discovery through reaction prediction and multi-dimensional optimization", Nature Communications (2025), doi:10.1038/s41467-025-66324-4. The raw data is available on figshare: Minisci-type C–H alkylation dataset.
Inference example data
Small example inputs for the inference / scoring workflows live under data/inference_examples/:
| File | Used by | Description |
|---|---|---|
example_regio.csv |
predict_regio.py |
Example substrates (SMILES, PRODUCT_SMILES, ID) for regioselectivity prediction. |
example_borylation_candidates.csv |
active_learning.py |
Example candidate substrates (startingmat_1_smiles) to score for active learning. |
example_minisci.csv |
minisci workflows | Example Minisci-type reactions (conditions + SMILES) matching the minisci_data.tsv schema. |
Binary prediction XGBoost
Model: XGBoost ensemble
Input: csv SURF file containing the column startingmat_1_smiles
Example:
To run python script with example data
conda activate active_disc
python predict_binary_xgb.py data/binary_substrates.csv binary_prediction_output.csv
This will store binary_prediction_output.csv in the output/ folder, adding the output prediction columns binary_prediction, logits, and uncertainty to the input csv file.
Note: uncertainty is calculated as the standard deviation of ensemble member predictions (logits).
Help options for binary prediction script
predict_binary_xgb.py options:
positional arguments:
input_file Path to the input file with substrates to predict
output_file Name of the output file to save predictions
optional arguments:
--train_data_path PATH Path to training data (default: data/training_data/pre_AL_binary_and_yield.csv)
--ensemble_size N Number of XGBoost classifiers in ensemble
--enumerate_conditions Enumerate conditions (True/False)
--encoding_strategy S Features to include (fp, fp_oh_num, fp_oh)
--balance_strategy S Class balancing strategy (class_weights, none, scale_pos_weight)
--acquisition_function F Acquisition function for scoring
--num_output_mols N Number of molecules to score and output in csv file
--hyperopt_evals N Number of hyperparameter optimization iterations
--run_metrics Run metrics on held-out test set
--num_bins N Number of bins for model calibration error assessment
Binary prediction GNN
Model: GNN
Input: csv SURF file containing the column startingmat_1_smiles
Example:
To run python script with example data
conda activate active_disc
python predict_binary_gnn.py data/binary_substrates.csv binary_prediction_output.csv
This will store binary_prediction_output.csv in the output/ folder, adding the output prediction columns binary_prediction and logits to the input csv file.
Help options for binary prediction script
predict_binary_gnn.py options:
positional arguments:
input_file Path to the input file with substrates to predict
output_file Name of the output file to save predictions
optional arguments:
--train_data_path PATH Path to training data (default: data/training_data/pre_AL_binary_and_yield.csv)
--enumerate_conditions Enumerate conditions (True/False)
--model Model for prediction (default: ponita)
--checkpoint_folder Location of pretrained model checkpoints (default: None)
Predicting with a model you've trained yourself
To run predictions using a model that you've trained yourself, the --checkpoint_folder argument must be specified as in the example below:
conda activate active_disc
python predict_binary_gnn.py data/binary_substrates.csv binary_prediction_output.csv --checkpoint_folder src/model_weights/your_binary_run
Prediction ensembles
Binary prediction occurs via ensembling. If only a single checkpoint is in the folder, only one model will be used.
Passing this flag will cause the ensemble all checkpoints in the folder with 'final' in their name to make predictions. This restricts running predictions to only models that have been trained to completion.
All models in the ensemble make predictions on 5 conformers per molecule. The predictions are then averaged across conformers and across all available models in the ensemble.
Regioselectivity prediction
Model: GNN
Input: csv file containing the column SMILES and the column ID.
Note: if borylation product smiles are avialable, include them in a column titled PRODUCT_SMILES
Example data is provided in data/inference_examples/example_regio.csv
Example
To run the python script with example data:
conda activate active_disc
python predict_regio.py data/inference_examples/example_regio.csv output/preds_regio.xlsx --smiles_col "SMILES" --id_col "ID"
This will store a predictions.xlsx in the output/ folder containing the SMILES, ID, and images of the input molecules annotated with predicted sites of borylation.
Mean and standard deviation of predicted borylation sites are overlayed on the atoms predicted > 10% likely to be borylated. Mean and standard deviation result from model predictions on 5 different conformers of the input molecule.
Predicting with a model you've trained yourself
To run predictions using a model that you've trained yourself, the --checkpoint_folder argument must be specified as in the example below:
conda activate active_disc
python predict_regio.py data/inference_examples/example_regio.csv output/preds_regio.xlsx --smiles_col "SMILES" --id_col "ID" --checkpoint_folder src/model_weights/your_regio_run
Prediction ensembles
Regio prediction occurs via ensembling. If only a single checkpoint is in the folder, only one model will be used.
Passing this flag will cause the ensemble all checkpoints in the folder with 'final' in their name to make predictions. This restricts running predictions to only models that have been trained to completion.
All models in the ensemble make predictions on 5 conformers per molecule. The predictions are then averaged across conformers and across all available models in the ensemble.
Regio prediction model description:
Regioselectivity prediciton uses the Equiformer v2 model.
Training was performed via multi-task learning including
- The main regioselectivity prediction task
- Coordinate denoising auxiliary task.
Coordinate denoising is deactivated during inference.
Help options for regioselectivity prediction script
positional arguments:
input_file Path to the input file.
output_filename Name of the output file.
options:
-h, --help show this help message and exit
-d DELIMITER, --delimiter DELIMITER
Column delimiter of input file.
-c SMILES_COL, --smiles_col SMILES_COL
Name of column that contains SMILES. Defaults to "SMILES"
-i ID_COL, --id_col ID_COL
Name of column that contains compound IDs. Defaults to "ID"
Minisci prediction
Model: GNN (PONITA)
Predicts Minisci-type C–H alkylation outcomes for a reaction. Each reaction has two substrates (startingmat_1_smiles + startingmat_2_smiles) which are concatenated into a single graph; the model outputs one value per reaction.
Two targets are available via --minisci_target:
binary(default): reactivity classification. Output columnslogits(sigmoid probability) andbinary_prediction(thresholded at 0.5).total_yield(alsomono_yield/di_yield): yield regression. Output columnyield_prediction(raw value).
Input: csv (comma) or tsv (tab) file containing rxn_id, startingmat_1_smiles, startingmat_2_smiles, reagent_1_smiles, catalyst_1_smiles, additive_1_smiles, solvent_1_smiles, and solvent_2_smiles. Example data is provided in data/inference_examples/example_minisci.csv.
No pretrained Minisci weights are shipped with this repository.
--checkpoint_folderis required and must point to a directory of checkpoint(s) you trained yourself (see Minisci training). If the folder contains multiple checkpoints, their predictions are averaged.--use_rxn_conditionsmust match how the checkpoint was trained, since it changes the model architecture.
Example
To run binary reactivity prediction with a self-trained checkpoint folder:
conda activate active_disc
python predict_minisci_gnn.py data/inference_examples/example_minisci.csv minisci_predictions.csv --minisci_target binary --checkpoint_folder src/model_weights/your_minisci_run --use_rxn_conditions
For yield regression, set --minisci_target total_yield:
python predict_minisci_gnn.py data/inference_examples/example_minisci.csv minisci_predictions.csv --minisci_target total_yield --checkpoint_folder src/model_weights/your_minisci_run --use_rxn_conditions
This stores minisci_predictions.csv in the output/ folder, joining the prediction column(s) back onto the input rows by rxn_id. Reactions whose substrates fail to parse are skipped during graph mapping and receive an empty prediction.
Binary training XGBoost
Model: XGBoost ensemble
Input: path to csv containing training data with reaction conditions.
Example:
To run python script directly with example data:
conda activate active_disc
python train_binary_xgb.py --train_data_path=data/training_data/pre_AL_binary_and_yield.csv
This will train an ensemble of XGBoost classifiers and preprocessing functions StandardScalar, and OneHotEncoder. The model weights will be saved to model_weights/binary_xgb as .pkl files.
Help options for binary training script
train_binary_xgb.py options:
optional arguments:
--train_data_path PATH Path to training data (default: data/training_data/pre_AL_binary_and_yield.csv)
--model_output_dir PATH Path to save model weights (default: src/model_weights/binary_xgb)
--ensemble_size N Number of XGBoost classifiers in ensemble
--enumerate_conditions Enumerate conditions (True/False)
--encoding_strategy S Features to include (fp, fp_oh_num, fp_oh)
--balance_strategy S Class balancing strategy (class_weights, none, scale_pos_weight)
--acquisition_function F Acquisition function for scoring
--num_output_mols N Number of molecules to score and output in csv file
--hyperopt_evals N Number of hyperparameter optimization iterations
--run_metrics Run metrics on held-out test set
--num_bins N Number of bins for model calibration error assessment
Binary training GNN
Model: GNN
To train a single GNN for binary prediction via python:
conda activate active_disc
python train_gnn_workflow.py --input_csv data/training_data/pre_AL_binary_and_yield.csv --prediction_type binary --epochs 1000
Training arguments
required arguments:
--input_csv PATH Path to the input CSV file with training data
optional arguments:
--prediction_type Type of prediction model to train: binary or regio (default: regio)
--model Model architecture to use (default: ponita)
--output_data_dir Directory to save outputs (default: output)
--checkpoint_path Directory to save model weights (default: src/model_weights/)
--epochs Number of training epochs (default: 100)
--batch_size Training batch size (default: 16)
--learn_rate Learning rate (default: 0.0004)
--patience Patience for early stopping (default: 20)
--wandb_logging Enable Weights & Biases logging (flag)
--seed Random seed for reproducibility (default: 42)
--run_id Run ID for ensemble training (default: 0)
Regioselectivity training
Model: GNN
To train a single GNN for regioselectivity prediction via python:
conda activate active_disc
python train_gnn_workflow.py --input_csv data/training_data/regioselectivity.csv --epochs 1000
To train an ensemble of GNNs for regioselectivity prediction via LSF:
conda activate active_disc
./train_gnn_ensemble.sh --input_csv data/training_data/regioselectivity.csv
Training arguments
required arguments:
--input_csv PATH Path to the input CSV file with training data
optional arguments:
--prediction_type Type of prediction model to train: binary or regio (default: regio)
--model Model architecture to use (default: ponita)
--output_data_dir Directory to save outputs (default: output)
--checkpoint_path Directory to save model weights (default: src/model_weights/)
--epochs Number of training epochs (default: 100)
--batch_size Training batch size (default: 16)
--learn_rate Learning rate (default: 0.0004)
--patience Patience for early stopping (default: 20)
--wandb_logging Enable Weights & Biases logging (flag)
--seed Random seed for reproducibility (default: 42)
--run_id Run ID for ensemble training (default: 0)
Minisci training
Model: GNN (PONITA)
Trains a GNN to predict Minisci-type C–H alkylation outcomes from the full dataset in data/training_data/minisci_data.tsv (see Training data).
Prerequisite: build the reaction-condition vocabulary
The graph mapper encodes categorical reaction conditions from JSON vocab maps in data/minisci_rxn_emb_maps/. These ship with the repository, but to (re)build them from the dataset run:
conda activate active_disc
python -m src.minisci_utils.build_vocab
This reads data/training_data/minisci_data.tsv and writes the vocab maps to data/minisci_rxn_emb_maps/.
Example
To train a single GNN for Minisci binary reactivity prediction, writing checkpoints to a dedicated folder:
conda activate active_disc
python train_gnn_workflow.py --input_csv data/training_data/minisci_data.tsv --prediction_type minisci --minisci_target binary --use_rxn_conditions --checkpoint_path src/model_weights/your_minisci_run --epochs 1000
To train the yield-regression model instead, set --minisci_target total_yield:
python train_gnn_workflow.py --input_csv data/training_data/minisci_data.tsv --prediction_type minisci --minisci_target total_yield --use_rxn_conditions --checkpoint_path src/model_weights/your_minisci_run --epochs 1000
Point predict_minisci_gnn.py --checkpoint_folder at that folder to run Minisci prediction. Use a dedicated --checkpoint_path per run (the prediction script ensembles all .pth files in the folder, so keep only one target's checkpoints together), and use the same --use_rxn_conditions setting at prediction time as during training.
Training arguments
required arguments:
--input_csv PATH Path to the input CSV/TSV file with training data
optional arguments:
--prediction_type Type of prediction model to train: binary, regio, or minisci (default: regio)
--minisci_target Minisci target: binary, total_yield, mono_yield, di_yield (default: binary)
--use_rxn_conditions Include reaction-condition features (minisci) (flag)
--model Model architecture to use (default: ponita)
--output_data_dir Directory to save outputs (default: output)
--checkpoint_path Directory to save model weights (default: src/model_weights/)
--epochs Number of training epochs (default: 100)
--batch_size Training batch size (default: 16)
--learn_rate Learning rate (default: 0.0004)
--patience Patience for early stopping (default: 20)
--wandb_logging Enable Weights & Biases logging (flag)
--seed Random seed for reproducibility (default: 42)
--run_id Run ID for ensemble training (default: 0)
Active learning
Model: XGBoost ensemble
Input: csv or tsv file containing molecules to score for experimental testing. Must contain the column startingmat_1_smiles.
Example data is provided in data/inference_examples/example_borylation_candidates.csv
This will store a al_output.csv in the output/ folder with an added column acquisition_function_value containing the model scores for the molecule.
The output file will be sorted by score in descending order.
Scores are calculated according to Bayesian Active Learning by Disagreement (BALD) via an ensemble of XGBoost classifiers.
Example
To run the python script with example data:
conda activate active_disc
python active_learning.py data/inference_examples/example_borylation_candidates.csv al_output.csv
Help options for active learning script
active_learning.py Options:
positional arguments:
input_file Path to the input file.
output_filename Name of the output file.
optional arguments:
--train_data_path PATH Path to training data (default: data/training_data/pre_AL_binary_and_yield.csv)
--ensemble_size N Number of XGBoost classifiers in ensemble
--acquisition_function F Acquisition function for scoring (default: bald)
--num_output_mols N Number of molecules to score and output in csv file
Toy experiments
Experiments on toy datasets can be found in src/toydataset/
Citing this Work
Please cite our work when referencing this repository.
@article{minot2025active,
title={Active and geometric deep learning advances chemical reaction prediction in data-scarce drug discovery},
author={Minot, Mason and Stenzhorn, Yannick and Wolfard, Jens and Strobel, Sebastian and Jablonski, Philippe and Zimmerli, Daniel and Binder, Martin and Grether, Uwe and Martin, Rainer E and M{\"u}ller, Alex T and Nippa, David F and Atz, Kenneth and Schneider, Gisbert},
year={2025},
publisher={ChemRxiv}
}
To cite this software specifically, use the archived Zenodo release (10.5281/zenodo.20783136), or use the "Cite this repository" button on GitHub (powered by CITATION.cff).
License
This repository is licensed under the GNU General Public License v3.0 (GPL-3.0). See LICENSE for the full text.
Copyright (c) 2026 Mason Minot (Roche & ETH Zurich) and Gisbert Schneider (ETH Zurich).
Third-party components
Several model implementations under src/models/ are derived from third-party
projects and remain governed by their original licenses, not by the GPL-3.0
covering the rest of this repository. Each retains its upstream copyright notice
and a LICENSE_* file in its directory, including:
src/models/fairchem_utils/,src/models/escn/,src/models/schnet.py,src/models/dimenet_plus_plus*.py— Meta (FAIR) / fairchem (MIT)src/models/faenet/— FAENet (MIT)src/models/equiformer_v2/— EquiformerV2src/models/segnn/— SEGNNsrc/models/ponita/— PONITAsrc/models/tfn/— Tensor Field Networks
Refer to the respective LICENSE_* files for the exact terms.