This code repository contains example code for the paper Topological Regression in Quantitative Structure-Activity Relationship Modeling. Quantitative structure-activity relationship (QSAR) modeling is a powerful tool used in drug discovery, yet the lack of interpretability of commonly used QSAR models hinders their application in molecular design. We proposed a similarity-based regression framework, topological regression (TR), that offers a statistically grounded, computationally fast, and interpretable technique. TR provides predictive performance that competes with deep-learning methods, such as Transformer-CNN, as well as informative visualizations, which can help guide lead optimization in drug discovery.

In this package, we provide Python files for TR, Ensemble TR, k-Nearest-Neighbor Graph Visualization, and Lead optimization pathway visualization, as well as sample ChEMBL datasets and the code to extract all datasets used in the paper.

The code has been tested in Windows 10 and CentOS 8.1 with Python 3.8 or greater and the following packages:

• NumPy >= 1.23.5
• SciPy >= 1.10.1
• pandas >= 1.5.0
• scikit-learn >= 1.2.2
• NetworkX >= 3.1
• metric-learn >= 0.6.2
• Matplotlib >= 3.6.1
• RDKit >= 2020.09.5
• fastparquet >= 2023.7.0

The dependencies can be installed using the provided environment.yml file and activated using the following commands:

conda env create --file environment.yml
conda activate TopoReg

This typically takes around 2-4 minutes depending on various factors.

To train and evaluate TR on a requested dataset, simply run the file 'TR_pipeline.py' with the path argument set to the path to the data folder

python TR_pipeline.py --path 'PATH TO DATA'

The script will output the model performance with metrics NRMSE and Spearman correlation.

One can select the data splitting method, anchor point percentage, RBF gamma, number of CV folds, the distance metric, and the random seed with the following input arguments:

path: str = 'SampleDatasets/ChEMBL/CHEMBL278/'  # the working folder that contains the data files
metric: str='tanimoto'  # the default distance metric to use on FPs
split: str='scaffold'  # which data split to evaluate, either 'scaffold' (default) or 'cv'
seed: int=2021  # random seed 
cv_fold: int=5  # number of folds in cross-validation
anchor_percentage: float=0.5    # Anchor point percentage
rbf_gamma: float=0.5    # gamma for RBF reconstruction

Running the TR_pipeline.py with default parameters will output the following:

Scaffold TR Performance
Spearman: 0.8345442589034985
R2: 0.8614466743805971
RMSE: 0.6902920473802862
NRMSE: 0.3722275186218812

Expected runtime: 554 ms ± 9.41 ms per loop (mean ± std. dev. of 7 runs)

Similarly to above, Ensemble TR can be trained and evaluated by running the 'Ensemble_TR_pipeline.py' file

One can select the data splitting method, number of TR models, average anchor point percentage, standard deviation of anchor point percentages, RBF gamma, number of CV folds, the distance metric, and the random seed with the following input arguments:

path: str = 'SampleDatasets/ChEMBL/CHEMBL278/'  # the working folder that contains the data files
metric: str='tanimoto'  # the default distance metric to use on FPs
split: str='scaffold'  # which data split to evaluate, either 'scaffold' (default) or 'cv'
seed: int=2021  # random seed 
cv_fold: int=5  # number of folds in cross-validation
mean_anchor_percentage: float=0.6   # mean anchor point percentage for TR ensemble
std_anchor_percentage: float=0.2     # std of anchor point percentage for TR ensemble
num_TR_models = 30   # Number of TR models included in the ensemble
rbf_gamma: float=0.5    # gamma for RBF reconstruction

Running the TR_pipeline.py with default parameters will output the following:

Scaffold Ensemble TR Performance
Spearman: 0.8436241650085896
R2: 0.8859742016491053
RMSE: 0.6262185866794664
NRMSE: 0.33767706222202104

Expected runtime: 957 ms ± 14 ms per loop (mean ± std. dev. of 7 runs)

In addition to the TR pipelines, we have provided visualization scripts to regenerate the figures in the main manuscript of the TR paper.

To visualize the predictions made by TR compared to KNN and MLKR, run the 'visualize_NN_test_predictions.py' file with the following default input arguments

# Defaults for regenerating figures from manuscript
path: str = 'SampleDatasets/ChEMBL/CHEMBL2734/'  # the working folder that contains the data files
metric: str='tanimoto'  # the default distance metric to use on FPs
split: str='cv'  # which data split to evaluate, either 'scaffold' (default) or 'cv'
seed: int=2021  # random seed 
cv_fold: int=5  # number of folds in cross-validation
anchor_percentage: float=0.8    # Anchor point percentage
k: int=5             # k for k-NN prediction visualization

For example, running the following from the visualization folder:

python visualize_NN_test_predictions.py

will generate and save the following image to the visualization folder: Expected runtime: 1min 24s ± 548 ms per loop (mean ± std. dev. of 7 runs) (Takes longer due to MLKR predictions)

To visualize lead optimization pathways in the training predictions made by TR, run the 'visualize_TR_leadopt_pathways.py' file with the following default input arguments

  # Defaults for regenerating figures from manuscript
  path: str = 'SampleDatasets/ChEMBL/CHEMBL278/'   # the working folder that contains the data files
  metric: str='tanimoto'  # the default distance metric to use on FPs (ends with FP or FP4)
  split: str='cv'  # which data split to evaluate, either 'scaffold' (default) or 'cv'
  seed: int=2021  # random seed 
  cv_fold: int=5  # number of folds in cross-validation
  anchor_percentage: float=0.9   # Anchor point percentage
  k: int=5             # k for k-NN prediction visualization

For example, running the following from the visualization folder:

python visualize_TR_leadopt_pathways.py

Will generate and save the following image to the visualization folder: Expected runtime: 3.08 s ± 8.17 ms per loop (mean ± std. dev. of 7 runs)

To extract all ChEMBL data [1] used in the manuscript, please see the DataExtraction folder.

[1] Mendez, D., Gaulton, A., Bento, A. P., Chambers, J., De Veij, M., Félix, E., Magariños, M. P., Mosquera, J. F., Mutowo, P., Nowotka, M., Gordillo-Marañón, M., Hunter, F., Junco, L., Mugumbate, G., Rodriguez-Lopez, M., Atkinson, F., Bosc, N., Radoux, C. J., Segura-Cabrera, A., Hersey, A., … Leach, A. R. (2019). ChEMBL: towards direct deposition of bioassay data. Nucleic acids research, 47(D1), D930–D940. https://doi.org/10.1093/nar/gky1075

Once extracted, the results across all 530 datasets can be calculated and averaged to reproduce the provided results. The Transformer-CNN code was downloaded from [1], and the ChemProp code was downloaded from [2]. All other models used the sklearn and metric-learn packages.