Computational models for predicting substrates or inhibitors of P-glycoprotein

Lei Chen, Youyong Li, Huidong Yu, Liling Zhang, Tingjun HouDrug 2011 · DOI

TL;DR

This review explores the landscape of computational models designed to predict whether a molecule will be a substrate or inhibitor of P-glycoprotein (P-gp), a key player in multidrug resistance and drug pharmacokinetics. The authors highlight the challenges posed by P-gp's polyspecificity and the need for robust in silico tools to accelerate drug discovery, especially in light of recent P-gp crystal structure advancements.

What Problem Does This Paper Address?

P-glycoprotein (P-gp) is a crucial efflux pump found at various physiological barriers, including the blood-brain barrier and the gut epithelium. Its primary role is to transport a wide range of structurally diverse compounds out of cells. This activity significantly impacts drug absorption, distribution, metabolism, and excretion (ADME) properties. Critically, P-gp is a major contributor to multidrug resistance (MDR) in cancer chemotherapy, where cancer cells overexpress P-gp to expel therapeutic agents, rendering treatments ineffective.

Identifying P-gp substrates and inhibitors is vital for drug development. Traditionally, this involves expensive and time-consuming experimental assays. The problem this paper addresses is the development and application of in silico (computational) models that can rapidly and cost-effectively predict P-gp substrate or inhibitor activity. This would allow researchers to screen large compound libraries and prioritize candidates for experimental validation, thereby accelerating the discovery of new drugs and strategies to overcome MDR. The paper specifically reviews the progress made in this area, the available datasets, and the computational approaches employed.

Methods

The paper reviews several computational methodologies used for predicting P-gp substrates and inhibitors:

  • Quantitative Structure-Activity Relationship (QSAR): This involves establishing mathematical relationships between the chemical structure of molecules and their biological activity. The review discusses both 2D-QSAR (using molecular descriptors like fingerprints and electrotopological states) and 3D-QSAR (which considers the spatial arrangement of atoms, such as with Comparative Molecular Field Analysis - CoMFA and Comparative Molecular Similarity Index Analysis - COMSIA).
  • Pharmacophore Modeling: This technique identifies the essential 3D features (e.g., hydrogen bond donors/acceptors, hydrophobic regions) that a molecule must possess to interact with a biological target. Models are built based on known active compounds and then used to screen new molecules.
  • Molecular Docking: With the advent of P-gp crystal structures, molecular docking has become a significant approach. This method simulates the binding of a ligand (potential substrate or inhibitor) into the 3D structure of the P-gp binding site to predict binding affinity and pose.
  • Machine Learning Approaches: Various statistical and machine learning algorithms are employed to build predictive models, including Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), Naive Bayes classifiers, Decision Trees (DT), and k-Nearest Neighbors (kNN).

The paper highlights that while QSAR and pharmacophore modeling were dominant in earlier stages due to the lack of structural information, molecular docking has gained prominence with the availability of P-gp crystal structures.

Datasets

The availability of high-quality and extensive experimental datasets is crucial for developing reliable computational models. The paper discusses several datasets used:

  • Early Datasets: Often contained a limited number of compounds (less than 200). Examples include datasets reported by Penzotti et al. (195 compounds), Xue et al. (201 compounds), and Gombar et al. (98 compounds).
  • Larger Datasets: More recent efforts have focused on compiling larger datasets. Sun et al. reported a dataset of 609 compounds, and Wang et al. a dataset of 332 compounds.
  • Largest Reported Dataset: The authors themselves reported the largest dataset to date, comprising 1273 structurally diverse molecules, with 797 P-gp inhibitors and 476 non-inhibitors. This dataset was derived from experimental data on multidrug resistance reversal (MDRR) activity.
  • Substrate Datasets: The paper notes that while inhibitor datasets are growing, large datasets for P-gp substrates are still needed, with the largest reported containing only 332 compounds.

Key Insights and Ideas

  • P-gp's Polyspecificity is a Major Hurdle: The ability of P-gp to bind a vast array of structurally diverse molecules is a double-edged sword. While it contributes to its protective role, it makes developing highly specific and accurate predictive models extremely challenging.
  • Crystal Structures are Promising but Not a Panacea: The availability of P-gp crystal structures has opened doors for molecular docking. However, the paper points out that even with these structures, distinguishing substrates from non-substrates based solely on docking scores remains difficult due to significant overlap in score distributions.
  • Integrated Models May Offer Higher Confidence: The authors suggest that combining different modeling principles (e.g., pharmacophore models integrated into a classification tree) can lead to more robust predictions, especially given P-gp's complex binding mechanisms.
  • Structural Fragments Matter: Bayesian classifiers, when used with molecular fingerprints, can highlight specific structural fragments that are important for P-gp inhibition, providing valuable insights for medicinal chemists.
  • Data Quality is Paramount: The paper emphasizes that mixing data from different experimental protocols can lead to unreliable models. The availability of large, high-quality, and well-curated datasets is a critical bottleneck.

Key Results

  • Diverse Computational Approaches: The review details numerous studies employing QSAR (2D and 3D), pharmacophore modeling, and molecular docking, often combined with various machine learning algorithms (LDA, SVM, Bayes, DT, etc.).
  • Model Performance Varies: Table 1 summarizes the performance of various models, with reported accuracies ranging from around 63% to over 90% on test sets. For example, Gombar et al. achieved 86.2% accuracy, while Chen et al. reported an 81% accuracy for their Bayesian classifier.
  • Molecular Docking Limitations: Despite using crystal structures, the authors' own molecular docking study showed significant overlap in docking scores between P-gp substrates and non-substrates, indicating that current docking approaches struggle to differentiate them effectively.
  • Fragment Identification: The Bayesian classifier approach identified specific "good" and "bad" fragments for P-gp inhibition, offering potential guidance for drug design.

Limitations and Future Work

  • Dataset Size and Quality: A major limitation acknowledged is the lack of sufficiently large and reliable experimental datasets, particularly for P-gp substrates. This hinders the development of highly accurate and generalizable computational models.
  • P-gp's Complex Binding Site: The large, flexible binding site of P-gp, capable of accommodating multiple molecules and diverse binding modes, presents a significant challenge for developing a single, universal predictive model.
  • Structure-Based Prediction Challenges: While high-resolution P-gp structures are available, translating this information into reliable structure-based prediction models is still an active area of research. The weak and unspecific binding properties of P-gp amplify inherent defects in docking approaches.
  • Future Directions: The authors suggest the need for continued efforts in generating high-quality experimental data. They also propose developing integrated prediction platforms that combine multiple modeling principles and leveraging the available structural information to create more accurate structure-based prediction models, potentially with more co-crystallized ligands.

Summary updated

BibTeX
@inproceedings{chen2011computational,
  author = {Lei Chen and Youyong Li and Huidong Yu and Liling Zhang and Tingjun Hou},
  title = {{Computational models for predicting substrates or inhibitors of P-glycoprotein}},
  booktitle = {Drug},
  year = {2011},
  doi = {10.1016/j.drudis.2011.11.003},
}
Figure