TL;DR
This paper tackles the crucial challenge of identifying P-glycoprotein (P-gp) substrates early in drug discovery by developing robust in silico prediction models. They achieved this by compiling the largest P-gp substrate/non-substrate dataset to date and employing a naive Bayesian classification approach, resulting in a highly accurate model for virtual screening.
What Problem Does This Paper Address?
P-glycoprotein (P-gp) is a key transporter protein that actively pumps a wide range of compounds out of cells. This efflux activity is a major contributor to multidrug resistance (MDR) in cancer chemotherapy, leading to treatment failure. It also plays a significant role in drug absorption, distribution, and elimination, influencing a drug's overall ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) profile. Therefore, accurately identifying potential P-gp substrates early in the drug discovery pipeline is critical for developing effective and safe therapeutics. Traditional experimental methods for P-gp substrate identification can be time-consuming and expensive, necessitating the development of reliable computational tools.
Methods
The researchers employed a naive Bayesian classification technique to build predictive models. This statistical method is well-suited for binary classification tasks (substrate vs. non-substrate) and can handle a large number of features. They explored three types of input features for their models:
- Simple physicochemical properties: These are basic molecular characteristics like molecular weight, solubility, and lipophilicity (AlogP, logD).
- Topological descriptors: These are mathematical representations of molecular structure, capturing aspects like connectivity and branching.
- Molecular fingerprints: These are bit strings that encode the presence or absence of specific structural features or substructures within a molecule. The study utilized extended-connectivity fingerprints (ECFC) and path-based fingerprints (FCFC, EPFC, LPFC) with varying maximum distances.
The models were trained and validated using a 5-fold cross-validation approach on the training set and then further evaluated on an independent external test set.
Datasets
The study utilized a newly compiled, extensive dataset for P-gp substrates and non-substrates.
- Training Set: 822 non-duplicated molecules, comprising 423 P-gp substrates and 399 non-substrates. This dataset was split into a training set of 622 compounds (313 substrates, 309 non-substrates) and an independent test set of 200 compounds (110 substrates, 90 non-substrates). The test set was specifically curated to maximize molecular diversity using the Find Diverse Molecules protocol.
- Inhibitor Data: An additional dataset of 735 P-gp inhibitors was used for comparative analysis with substrates.
The authors emphasize that their substrate/non-substrate dataset is the largest published to date, which is crucial for building reliable predictive models.
Key Insights and Ideas
- Molecular Weight and Solubility are Key Differentiators: The study found that molecular weight (MW) and molecular solubility (logS) were the most informative physicochemical properties for distinguishing P-gp substrates from non-substrates. Substrates tended to have higher molecular weights and lower solubility.
- Hydrophobicity and H-bonding Differences between Substrates and Inhibitors: When comparing P-gp substrates to P-gp inhibitors, a clear distinction emerged: inhibitors were significantly more hydrophobic (higher AlogP) than substrates. Conversely, substrates tended to possess more hydrogen bond donors (NHBD) than inhibitors.
- Spatial Arrangement of H-bond Acceptors Matters: While overall H-bond acceptor counts (nHBA) didn't strongly differentiate substrates from non-substrates, the analysis of important structural fragments revealed that the spatial arrangement of H-bond acceptors is critical for P-gp substrate recognition.
- Flexibility Enhances Substrate-likeness: Flexible molecular fragments, which can adjust their conformation, were found to be more favorable for P-gp substrate interactions.
- Naive Bayesian Classifiers are Powerful: The naive Bayesian approach, especially when combined with molecular fingerprints (specifically ECFC_10), proved to be a highly effective method for predicting P-gp substrate status.
Key Results
- Dataset Size: The compiled dataset of 423 P-gp substrates and 399 non-substrates is the largest reported in the literature, providing a robust foundation for model development.
- Model Performance (Training Set): The best naive Bayesian classifier, utilizing molecular properties and ECFC_10 fingerprints, achieved a Matthews correlation coefficient (C) of 0.824 and a global accuracy (GA) of 91.2% on the training set via 5-fold cross-validation.
- Model Performance (Test Set): On the independent external test set of 200 molecules, the best classifier achieved a Matthews correlation coefficient (C) of 0.667 and a global accuracy (GA) of 83.5%. This demonstrates good generalization ability.
- Important Fragments Identified: The analysis of fragments contributing to substrate-likeness highlighted the importance of oxygen-containing groups arranged in specific spatial patterns (type I and type II units) and molecular flexibility. Conversely, negatively charged carboxyl groups and rigid ring systems were associated with non-substrate behavior.
Limitations and Future Work
- Data Quality and Experimental Variability: The authors acknowledge that misclassifications can arise from the inherent variability and potential inconsistencies in experimental data collected from diverse literature sources. Different assays for P-gp substrate identification can yield conflicting results for the same molecule.
- Conformational Requirements: While fingerprint-based models can identify important substructures, they don't fully capture the precise conformational requirements for optimal interaction with P-gp.
- Undetected Important Fragments: It's possible that some crucial fragments for P-gp substrate-likeness are not represented in the training set molecules, leading to mispredictions in the test set.
The authors suggest that their developed classifier can serve as a reliable virtual screening tool for identifying potential P-gp substrates. Future work could focus on refining data curation, incorporating more sophisticated conformational analysis, and exploring other machine learning algorithms to further improve prediction accuracy and address the "uncertain zone" of predictions.
Summary updated