Skip to content

Latest commit

 

History

History
35 lines (27 loc) · 3.14 KB

Liu et al (2021).md

File metadata and controls

35 lines (27 loc) · 3.14 KB
Raw

  • https://doi.org/10.1021/acs.chemrestox.0c00343

  • Aim:

    • Developed ensemble models for predicting the BBB permeability of compounds using models trained based on different molecular fingerprints as base classifiers; compared with previous developed models; exemplified the substructures of two chemicals

  • Data:

    • Data from three studies (one dataset for external validation)
    • Molecular fingerprints from the PaDEL-Descriptor software

  • Data curation:

    • Remove redundant data

  • Feature selection:

    • This paper used two feature selection methods to filter the data set:
      • (1) the nearZeroVar function in the R package caret was used to filter redundant or very similar features in all of the samples
      • (2) the f indcorrelation function caret was used to filter out highly correlated features
    • Used the Tanimoto coefficient to assess whether a compound has one or more highly correlated feature (high-correlation feature filtering with a Tanimoto correlation coefficient threshold of 0.95)

  • Data imbalance:

    • The resampling method Synthetic Minority Oversampling Technique (SMOTE) to effectively solve the problem of imbalanced data sets
    • SMOTE generates extra examples from the minority class in an attempt to have its data set size match that of the majority class to combat the existing imbalance

  • Approach:

    • Built qualitative models (Qualitative models can binary classify the BBB permeability of chemicals with high accuracy)
    • Three machine-learning methods, an SVM, an RF, and an extreme gradient boosting (XGBoost), were used to construct 27 base classifier models
    • The base classifiers were added to the ensemble machine-learning model based on the AUC value of each classifier from highest to lowest
    • RF: To better understand the contribution of several substructures related to blood−brain barrier permeability prediction, the mean decrease in the Gini coefficient (Mean Decrease Gini) was used to assess the importance of these substructures with the RF algorithm; the higher the mean decrease in the Gini coefficient, the closer the feature is to blood−brain barrier permeability prediction

  • Model performance:

    • Internal validation involved 5-fold cross-validation with 100 repeats, and external validation was performed using an external validation data set
    • 5-fold cross-validation randomly divided the training set into 5 parts: 4 were used as a training set to build the model, and the remaining part was used to test the performance of the model. This process was repeated 5 times so that each part could be used as a test set. Following 5-fold cross-validation, a series of models with different performance results were obtained. All 5-fold cross-validations were based on 100 repeated experiments, and the average value of the 100 runs was used to evaluate the predictive performance of the model
    • Four performance indicators to evaluate the performance of the model: the ACC, AUC (area under the receiver-operating characteristic (ROC) curve), SPE (specificity), and SEN (sensitivity)

  • Key learnings:

    • The ensemble model was superior to various base classifiers in terms of accuracy and AUC.