cfanalysis

The repository provides code and data to implement the CFA analysis introduced in the paper "Enhancing ADMET Property Models Performance through Combinatorial Fusion Analysis", categorized by drug encoding schemes: Morgan Circular Fingerprints, RDKit 2D descriptors, and MCFP.

Link to the manuscript

⚠️ Note: Previously, this repository was hosted at github.com/FLIDM/CFA4DD, which is no longer an active URL.

Python package: cfanalysis

Library to perform Combinatorial Fusion Analysis

⚠️ Introduction

A Combinatorial Fusion Analysis (CFA) to enhance Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) performance of ML models. CFA models show superior performance compared to most of the individual ML models. The CFA model architecture and the performance of CFA models on TDC and other internal datasets is presented in the paper. Significant enhancement suggests that CFA is a viable tool for improving ADMET ML model performance, offering promise for faster and more cost-effective drug development pipelines.

👩‍💻 Installation

You can install this package using pip. Run the following commands to install it and then import it:

pip install cfanalysis==0.1.13
import cfanalysis
from cfanalysis import cfafunctions

⚡ Usage

An example using Therapeutic Data Common's (TDC) Clearance_Microsome_AZ dataset.

Import the required libraries

import numpy as np
import pandas as pd

from tdc.single_pred import ADME
from tdc.benchmark_group import admet_group
from tdc import BenchmarkGroup

Import the TDC dataset and input our predictions: ⚠️ you may use your own dataset and predictions

Clearance_Microsome_AZ has a continuous target variable (regression)

group = admet_group(path = 'data/')
name = 'Clearance_Microsome_AZ' #you need to change dataset name to get the model fusion result
benchmark = group.get(name)
train_val, test = benchmark['train_val'], benchmark['test']
y_test = np.array(test.Y) # store the target variable for the test set

y_valid = {}  # create a dictionary to store target variable for the validation set for each seed

for seed in [1, 2, 3, 4, 5]:
    train, valid = group.get_train_valid_split(benchmark=name, split_type='default', seed=seed)
    y_valid[seed] = valid.Y # store the target variable for the validation set

# read in the dataframes with predicted values for the validation sets and the test sets, these datasets are available in the data dir
predictions_val_xgb = pd.read_csv('%s_predictions_val_xgb.csv' % name)
predictions_val_rf = pd.read_csv('%s_predictions_val_rf.csv' % name)
predictions_val_svm = pd.read_csv('%s_predictions_val_svm.csv' % name)


predictions_test_xgb = pd.read_csv('%s_predictions_test_xgb.csv' % name)
predictions_test_rf = pd.read_csv('%s_predictions_test_rf.csv' % name)
predictions_test_svm = pd.read_csv('%s_predictions_test_svm.csv' % name)

📝 Input model data

model_names = ['xgb', 'rf', 'svm'] # mention model names 
preds = cfafunctions.model_predictions(
    len(model_names),
    model_names,
    val_dfs_list=[predictions_val_xgb, predictions_val_rf, predictions_val_svm],
    test_dfs_list=[predictions_test_xgb, predictions_test_rf, predictions_test_svm]
)

cfafunctions.model_predictions takes 4 arguments:

len(model_names): Number of models to perfrom combinatorial fusion analysis

model_names: Vector of model names for identification

val_dfs_list: Predictions from the validation set for each ML model

test_dfs_list: Predictions from the test set for each ML model

⚠️ Predictions, val_dfs_list and test_dfs_list, need to be probabilities for the positive class for classification ML models and predictions from 5 seeds are required for each model.

✅ Obtain the results from the CFA analysis

cfafunctions.calculate_spearman_corr(model_names, preds, y_test, y_valid) # use Spearman-Rank correlation (regression) 

Other options for metrics for CFA

>> cfafunctions.calculate_MAE(model_names, preds, y_test, y_valid) # use Mean Absolute Error (regression) 
>> cfafunctions.calculate_auroc(model_names, preds, y_test, y_valid) # use Area Under the Receiver Operating Characteristics (classification)
>> cfafunctions.calculate_auprc(model_names, preds, y_test, y_valid) # use Area Under the Precision-Recall Curve (classification)

📈 Interpreting results

modelcombination_ds: implies diversity strength
modelcombination_r: implies rank combination
modelcombination_ps: implies performance strength
modelcombination: implies average score strength
modelname: implies raw model metric

📎 Citation

Jiang, Nan, et al. "Enhancing ADMET Property Models Performance through Combinatorial Fusion Analysis." arXiv preprint arXiv:#### (2023).

@article{Jiang2023cfa,
      title={Enhancing ADMET Property Models Performance through Combinatorial Fusion Analysis}, 
      author={Nan Jiang and Mohammed Quazi and Christina Schweikert and D. Frank Hsu and Tudor I. Oprea and Suman Sirimulla},
      year={2023},
      eprint={####},
      archivePrefix={arXiv},
      primaryClass={####},
      publisher={arXiv}
}