TL;DR
This review examines the current state of in silico models for predicting key ADME (Absorption, Distribution, Metabolism, Excretion) properties: solubility, human intestinal absorption (HIA), and oral bioavailability. While progress has been made, especially for simpler properties like solubility, complex phenomena like oral bioavailability remain challenging to predict accurately, largely due to data limitations and inherent biological complexity.
What Problem Does This Paper Address?
The pharmaceutical industry faces a high attrition rate for drug candidates, with poor ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties being a major cause of failure. Traditionally, ADMET properties were evaluated late in the drug discovery process, leading to costly late-stage failures. To mitigate this, there's a growing need for reliable in silico models that can predict ADMET properties early on, allowing for better prioritization of compounds and a more efficient drug discovery pipeline. This paper specifically addresses the accuracy and limitations of current computational models for predicting three crucial ADME properties: solubility, human intestinal absorption, and oral bioavailability.
Methods
The paper reviews existing in silico prediction models, primarily focusing on Quantitative Structure-Property Relationship (QSPR) approaches. Instead of detailing the models themselves, the authors emphasize the discussion of their prediction accuracies. They categorize solubility prediction models into three types:
- Experiment-related models: Utilize experimental physicochemical properties.
- QSPR-based models: Correlate molecular descriptors with experimental data using statistical techniques.
- Atom/group contribution models: Sum up contributions of atoms or functional groups.
For human intestinal absorption (HIA) and oral bioavailability, the review discusses models that utilize various molecular descriptors, including 1-D, 2-D, and 3-D descriptors, such as polar surface area (PSA), partition coefficients (logP/logD), and hydrogen bonding capabilities. The authors also touch upon the use of techniques like genetic function approximation (GFA) and support vector machines (SVM) for developing these models.
Datasets
The paper references several datasets used for training and evaluating ADME prediction models:
- Huuskonen dataset: A widely used dataset of 1297 organic compounds for solubility prediction, split into training and test sets.
- Kühne and co-workers' dataset: An in-house dataset of 1876 compounds with validated experimental water solubility values, used for model comparison.
- Hou and co-workers' HIA dataset: A large dataset of 647 drug and drug-like molecules, including 578 believed to be transported by passive diffusion.
- Hou and co-workers' oral bioavailability dataset: A database of 768 chemical compounds.
- GlaxoSmithKline and Glaxo Wellcome internal databases: Mentioned as examples of large, consistently measured compound property datasets used by pharmaceutical companies, though not publicly available.
The authors note that many models are developed using small, historical datasets from the literature, which limits their reliability. They highlight the recent availability of larger datasets for HIA and oral bioavailability as a positive development for future model development.
Key Insights and Ideas
- Complexity is Key: The predictive power and robustness of ADME models are directly tied to the complexity of the ADME property being predicted. Simpler properties are more amenable to accurate modeling.
- Bioavailability is Tricky: Predicting oral bioavailability is significantly more challenging due to its multifactorial nature, involving dissolution, intestinal permeation, first-pass metabolism, and even formulation effects.
- Data is King (and Scarce): The lack of large, high-quality, and publicly available experimental datasets is a major bottleneck hindering the development of reliable ADME prediction models, especially for in vivo properties.
- HIA is More Predictable: Passive human intestinal absorption can be predicted with relatively good accuracy compared to oral bioavailability.
- Rule-of-Five Limitations: While useful for initial filtering, Lipinski's "rule-of-five" provides only a rough classification and is insufficient for precise prediction.
- Descriptor Choice Matters: Descriptors like polar surface area (PSA) and distribution coefficient (logD) are crucial for HIA prediction, but no single descriptor or simple set appears sufficient.
- Consensus Models Show Promise: Combining multiple models (consensus modeling) can lead to greater accuracy than any single model, as demonstrated in solubility prediction.
Key Results
- Solubility: The best solubility prediction models (e.g., Meylan's or Hou's) can achieve a standard error of less than 0.9 log units. Atom/group contribution methods are considered practical as they don't require experimental data and are time-saving.
- Human Intestinal Absorption (HIA): Models can achieve good classification performance. For example, a support vector machine (SVM) classifier demonstrated high accuracy in discriminating well-absorbed from poorly absorbed compounds, with a total misclassification rate of 15 out of 113 compounds in one test. Regression models can achieve an average absolute error of less than 10% for fractional absorption.
- Oral Bioavailability: Prediction models currently struggle to provide reliable estimations. Regression models often show a significant drop in correlation (q²) during cross-validation compared to training (e.g., from r²=0.71 to q²=0.58). Root mean square errors (RMSE) for prediction are often greater than 20%.
Limitations and Future Work
- Data Availability: The primary limitation acknowledged is the scarcity of extensive and reliable experimental data, particularly for in vivo oral bioavailability and human intestinal absorption. The authors emphasize the need for more publicly available ADME data.
- Model Progress: Despite numerous models developed, the authors note that the prediction accuracy for some properties, like oral bioavailability, hasn't shown significant progress in recent years, even with larger datasets.
- Ionization and Crystal Effects: Many solubility models work best for non-charged compounds and don't explicitly account for ionization or crystal size, which significantly impact solubility. Future work should integrate these aspects.
- First-Pass Metabolism: Predicting oral bioavailability is hampered by the difficulty in accurately modeling first-pass metabolism. Developing new rules or descriptors specifically for this process is crucial.
- Integrated Models: The authors suggest developing separated prediction models for different components of oral bioavailability (passive transport, carrier-mediated transport, first-pass metabolism) and then integrating them, though this is currently challenging due to the unreliability of some individual component predictions.
- Software Performance: There's a need for more comparisons of the performance of various commercial ADME prediction software systems on specific ADME properties to guide users.
Summary updated