Pattern recognition approach to classifying CYP 2C19 isoform

Bartosz KrawczykOpen 2011 · DOI

Abstract

In this paper a pattern recognition approach to classifying quantitative structure-property relationships (QSPR) of the CYP2C19 isoform is presented. QSPR is a correlative computer modelling of the properties of chemical molecules and is widely used in cheminformatics and the pharmaceutical industry. Predicting whether or not a particular chemical will be metabolized by 2C19 is of primary importance to the pharmaceutical industry. This task poses certain challenges. First of all analyzed data are characterized by a significant biological noise. Additionally the training set is unbalanced, with objects from negative class outnumbering the positives four times. Presented solution deals with those problems, additionally incorporating a throughout feature selection for improving the stability of received results. A strong emphasis is put on the outlier detection and proper model validation to achieve the best predictive power.

Introduction

In the pharmaceutical industry, the development of new drugs is a high-stakes, decade-long process. A critical component of this journey is pharmacokinetics—understanding how human enzymes, particularly the Cytochrome P450 (CYP) family, metabolize compounds. The CYP 2C19 isoform is especially significant; identifying whether a drug is a substrate for this enzyme is vital for predicting its efficacy and potential toxicity.

However, biological datasets for CYP 2C19 are notoriously difficult to work with. They are often plagued by significant biological noise from experimental variability, a severe class imbalance where non-substrates (negative class) outnumber substrates (positive class) by a 4:1 ratio, and high dimensionality due to the hundreds of chemical descriptors used to characterize molecules.

Methodology

To address these challenges, a robust four-stage pattern recognition pipeline was developed to refine the data and optimize predictive accuracy.

The four-stage pattern recognition pipeline used for CYP 2C19 classification.
  1. Outlier Detection: Utilizing Hampel’s Test, a median-based approach, 78 corrupted objects were identified and removed from the initial 837-molecule dataset, providing a cleaner foundation than standard mean-based filters.
  2. Feature Selection: The Fast Correlation-Based Feature Filter (FCBF) was employed to reduce the dimensionality from 242 molecular descriptors to the 98 most relevant ones, mitigating noise and improving computational efficiency.
  3. Balancing the Training Set: The Synthetic Minority Over-sampling Technique (SMOTE) was used to correct the 4:1 class imbalance. SMOTE generates synthetic substrate examples by interpolating between existing positive instances in the feature space.
  4. Classification: The study evaluated several models, including Support Vector Machines (SVM) with Linear and Gaussian (RBF) kernels, Neural Networks (NN), and Gaussian Processes (GP).

Results and Discussion

The experimental results demonstrate that the integration of outlier detection, feature selection, and SMOTE significantly enhances model performance across all classifier types.

Detailed performance metrics across various classifiers and preprocessing configurations.

The SVM with a Gaussian kernel (SVMg) emerged as the top performer when utilizing the full pipeline, achieving an accuracy of 88.76%. Crucially, the pipeline dramatically improved sensitivity—the ability to correctly identify substrates—which rose from 63.56% in the unprocessed model to 85.95% in the optimized version.

A key insight from this research is that rigorous preprocessing is more impactful than the choice of classifier alone. For instance, the raw SVMg model only achieved 72.23% accuracy. Furthermore, the synergy between feature selection and SMOTE proved vital; reducing dimensionality before oversampling ensured that synthetic data points were generated based on truly relevant chemical properties. Interestingly, while SMOTE increases the training set size, the preceding feature selection step often resulted in faster execution times compared to raw datasets.

Conclusion

This study presents a highly effective machine learning pipeline for predicting CYP 2C19 metabolism. By systematically addressing noise, imbalance, and dimensionality, the proposed method turns challenging biological data into reliable predictive models. While the computational cost of balancing and the "black box" nature of some descriptors remain areas for future exploration, the methodology provides a strong framework that is likely applicable to other P450 isoforms.

Summary updated

Abstract

Abstract In this paper a pattern recognition approach to classifying quantitative structure-property relationships (QSPR) of the CYP2C19 isoform is presented. QSPR is a correlative computer modelling of the properties of chemical molecules and is widely used in cheminformatics and the pharmaceutical industry. Predicting whether or not a particular chemical will be metabolized by 2C19 is of primary importance to the pharmaceutical industry. This task poses certain challenges. First of all analyzed data are characterized by a significant biological noise. Additionally the training set is unbalanced, with objects from negative class outnumbering the positives four times. Presented solution deals with those problems, additionally incorporating a throughout feature selection for improving the stability of received results. A strong emphasis is put on the outlier detection and proper model validation to achieve the best predictive power.
BibTeX
@inproceedings{krawczyk2011pattern,
  author = {Bartosz Krawczyk},
  title = {{Pattern recognition approach to classifying CYP 2C19 isoform}},
  booktitle = {Open},
  year = {2011},
  doi = {10.2478/s11536-011-0120-3},
}
Figure