Abstract
Cytochrome P450 (CYP) is the most important drug-metabolizing enzyme in human beings. Each CYP isoform is able to metabolize a large number of compounds, and if patients take more than one drug during treatment, it is possible that some drugs would be metabolized by the same CYP isoform, leading to potential drug-drug interactions and side effects. Therefore, it is necessary to investigate the isoform specificity of CYP substrates. In this study, we constructed a data set consisting of 10 major CYP isoforms associated with 776 substrates and used machine learning methods to construct predictive models based on the features of structural and physicochemical properties of substrates. We also proposed a new method called the Improved Bayesian method, which is suitable for small data sets and is able to construct more stable and accurate predictive models compared with other traditional machine learning models. Based on this method, the predictive performance of our method achieved an accuracy of 86% for the independent test, which was significantly better than existing models. We believe that our proposed method will facilitate the understanding of drug metabolisms and help the large-scale analysis of drug-drug interactions.
Introduction
Cytochrome P450 (CYP) enzymes are the primary drivers of human drug metabolism, processing over 90% of clinical drugs. However, these enzymes lack strict specificity; multiple drugs can be metabolized by the same CYP isoform (e.g., CYP3A4 or CYP2D6). When a patient is prescribed multiple drugs metabolized by the same enzyme, competition can lead to dangerous drug-drug interactions (DDIs), toxicity, or reduced therapeutic efficacy.
Predicting which specific CYP isoform will metabolize a new drug candidate is a critical challenge in drug discovery. While experimental assays are the gold standard, they are slow and costly. Computational models offer a faster alternative, but they often struggle with the "small data" limitations of biological datasets or require complex, computationally expensive structural simulations of enzyme active sites.
The Improved Bayesian Method
To address these challenges, this study introduces an Improved Bayesian Method. Unlike the standard "Naïve" Bayes approach, which assumes all molecular features are independent, this method accounts for the correlations between features.
The core of this approach is a two-step training strategy:
- Background Training: The model first analyzes a large "background" dataset of compounds with unknown metabolic activity (e.g., from DrugBank). This step establishes the natural distribution and correlation of molecular features.
- Activity Training: The model is then refined using a smaller set of "activity samples" (known CYP substrates). By leveraging the feature correlations established in the first step, the model remains stable and avoids overfitting, even with limited labeled data.
Technically, the method transforms molecular features into a standard normal distribution. The transformation is defined by Equation (1):
This transformation ensures that the resulting variable $\hat{X}$ follows a standard normal distribution, as shown in Equation (2):
By calculating a correlation coefficient matrix ($R$) from the background data, the model can compute joint probability distributions without the flawed independence assumption of Naïve Bayes.
Data and Feature Selection
The study utilized two primary datasets:
- Primary Dataset: 776 compounds from the Meta-CYP database, covering 10 major CYP isoforms (1A1, 1A2, 2A6, 2B6, 2C8, 2C9, 2C19, 2D6, 2E1, and 3A4).
- Background Dataset: 7,114 compounds from DrugBank, used to stabilize feature distributions.
Initially, 176 molecular descriptors (structural and physicochemical) were calculated. To optimize the models, the top 20 descriptors for each isoform were selected using the mRMR (Minimum Redundancy Maximum Relevance) algorithm.
Key Insights and Results
The Improved Bayesian Method demonstrated significant advantages over traditional machine learning approaches:
- Superior Performance: The method achieved an average accuracy of 86% on an independent test set. In head-to-head comparisons using Area Under the Curve (AUC), it consistently outperformed Naïve Bayes, k-Nearest Neighbors (kNN), and Support Vector Machines (SVM). For example, for CYP1A1, the Improved Bayesian AUC was 0.92, compared to 0.79 for Naïve Bayes and 0.85 for SVM.
- High Reliability: The model showed an average specificity of 0.90, making it particularly effective at identifying which enzymes will not metabolize a drug—a key factor in avoiding DDIs.
- Isoform-Specific Preferences: Feature selection revealed that different isoforms have distinct structural "tastes." CYP3A4 tends to metabolize larger molecules, whereas CYP2A6 and CYP2E1 prefer smaller ones. Interestingly, molecular weight was not a significant factor for CYP2D6.
- The Role of LogP: Surprisingly, LogP (hydrophobicity) was not among the top descriptors for any of the 10 models, suggesting that isoform specificity is driven more by molecular shape and connectivity than by simple solubility.
Conclusion
The Improved Bayesian Method provides a robust framework for predicting CYP isoform specificity, particularly when dealing with small biological datasets. By incorporating background chemical knowledge, the model achieves high accuracy and stability. While the current model focuses on identifying the metabolizing enzyme, future work could extend this framework to predict specific metabolic products or regulatory interactions like enzyme inhibition and induction.
Summary updated