In the following document, we provide steps and code snippets to reproduce the models and results for the surrogate models presented in "Surrogate Machine Learning Models for ADME Prediction: From Synthetic Labels to Experimental Data" by Peteani et al. (Parrondo-Pizarro, Gerebtzoff, and Rodríguez-Pérez)
Corresponding author: raquel.rodriguez_perez@novartis.com
open_source_data/
data/ # Raw and processed public ADMET datasets
files/ # Intermediate aggregated files produced by data_preparation
finetune_data/ # Per-source train/test splits used for fine-tuning
all/ # Union of the three sources
all_unfreeze/ # Same, padded with extra endpoint columns
surrogate_dataset/ # 5-fold splits for surrogate pretraining
5fold_split/fold{1..5}_data.csv
src/
data_preparation/ # Notebooks to aggregate/pivot/split public data
public_ADME_data_aggregation.ipynb
process_public_data.ipynb
generate_5folds.ipynb
surrogate_pretraining/
architectures_benchmark/ # Benchmark of pretraining architectures
ST_ML_models/ # Single-task XGBoost / MLP
ST_GNN/ # Single-task Chemprop GNN (5 endpoints)
MT_GNN_5/ # Multi-task Chemprop GNN (5 endpoints)
MT_GNN_property/ # Multi-task Chemprop GNN per property group
MT_GNN_all/ # Multi-task Chemprop GNN, all 25 endpoints
overall_evaluation.py # Aggregate per-endpoint metrics across architectures
surrogate/ # Final MT GNN property related surrogate models used downstream
finetuning/ # Fine-tuning the pretrained surrogates
surrogate_feat_extraction/ # Frozen backbone + new initialized head
surrogate_unfreeze/ # Full model unfrozen
evaluation/ # Predict on public benchmark test sets
envs/ # Conda environment files
chemprop_env.yml # For GNN
surrogate_env.yml # For XGBoost + MLP baselines
Two conda environments cover the full pipeline. Create them with:
conda env create -f envs/chemprop_env.yml # for GNN
conda env create -f envs/surrogate_env.yml # for XGBoost + MLP baselineschemprop_env— Chemprop v2 + Ray Tune + hyperopt. Used for all GNN training, final surrogate training, zero-shot evaluation, fine-tuning, and the data-preparation notebooks.surrogate_env— PyTorch + XGBoost + Optuna + scikit-learn. Used only for theST_ML_modelsbaselines.
Download the public dataset at the following link:
| Path | Content | Source |
|---|---|---|
open_source_data/surrogate_dataset/protacdb2.0_zinc_chembl_dataset.csv |
Surrogate dataset | https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41467-024-49979-3/MediaObjects/41467_2024_49979_MOESM4_ESM.zip |
open_source_data/data/ADME_public_set_3521.csv |
Fang et al. ADME dataset | https://pubs.acs.org/mpohbp/article/22/4/1892/3749317/Multitask-Deep-Learning-Models-of-Combined |
open_source_data/data/ADMET.csv |
OpenADMET compounds | https://openadmet.org/datasetsmodels/ |
The end-to-end pipeline has five stages:
- 5-fold pretraining data preparation - generate the fold splits used by the surrogate benchmark.
- Surrogate architecture benchmark + 5-fold evaluation - compare ST/MT architectures on the fold splits and pick the best per property.
- Public data preparation - aggregate the three public ADME sources and produce per-source/union train/test splits used for evaluation and fine-tuning.
- Zero-shot evaluation of the pretrained surrogate on public data - use the final surrogate directly on the public test sets, no fine-tuning.
- Fine-tuning on public data + evaluation - fine-tune the surrogate on the public training sets (two strategies) and predict on the public test sets.
jupyter notebook src/data_preparation/generate_5folds.ipynbReads open_source_data/surrogate_dataset/protacdb2.0_zinc_chembl_dataset.csv
and writes:
open_source_data/surrogate_dataset/5fold_split/fold{1..5}_data.csv
Each fold file contains a smiles column, a split column with values
train/val/test, and one pred(<endpoint>) column per target endpoint.
Each subfolder of src/surrogate_pretraining/architectures_benchmark/ follows
the same three-step Chemprop v2 pipeline:
cd src/surrogate_pretraining/architectures_benchmark/<architecture>/
bash hyper_opt.sh # Ray Tune HP search (or hyper_opt_<scope>.sh)
bash model_train.sh # Train tuned model as a 10-model ensemble
bash predict.sh # 5-fold ensemble prediction on the test splitsFor the classical ML baselines under ST_ML_models/:
cd src/surrogate_pretraining/architectures_benchmark/ST_ML_models
bash train_xgboost.sh # or train_mlp.sh
bash evaluate_xgboost.sh # or evaluate_MLP.shRun the two notebooks in order:
jupyter notebook src/data_preparation/public_ADME_data_aggregation.ipynb
jupyter notebook src/data_preparation/process_public_data.ipynbpublic_ADME_data_aggregation.ipynbmerges the three public sources (ASAP x OpenADMET, ExpansionRx x OpenADMET, Fang et al. 2023) into a single long-format table underopen_source_data/files/.process_public_data.ipynbproduces the unions of the three sources train sets and test sets..
Apply the final surrogate directly to the public test sets without any fine-tuning:
bash src/evaluation/clint/chemprop_predict.sh
bash src/evaluation/permeability/chemprop_predict.sh
bash src/evaluation/ppb/chemprop_predict.shEach script points at the surrogate checkpoint produced at the end of stage 2 and predicts on the corresponding public test set from stage 3.
Two transfer-learning strategies are provided under src/finetuning/:
surrogate_feat_extraction/- frozen backbone, only a newly-initialized head is trained.surrogate_unfreeze/- the full model is fine-tuned.
For each property, first fine-tune, then predict on the public test set:
cd src/finetuning/surrogate_unfreeze/clint
bash finetune_surrogate_clint.sh
bash CLint_chemprop_predict.sh
# ... same pattern for permeability/ and ppb/, and for surrogate_feat_extraction/For any further question or inquiry about data and code availability/reproducibility, please get in touch with the paper corresponding author.