TL;DR
This paper develops and validates machine learning models, specifically Naive Bayesian classifiers, to predict the "drug-likeness" of molecules. By incorporating structural fingerprints alongside physicochemical properties, they achieve high prediction accuracy and demonstrate that Traditional Chinese Medicine compounds are a rich source of potentially drug-like molecules.
What Problem Does This Paper Address?
The drug discovery pipeline is notoriously long, expensive, and inefficient. A major bottleneck is identifying promising "lead compounds" from the vast chemical space. "Drug-likeness" is a concept used to filter molecules that are more likely to possess favorable pharmacokinetic and pharmacodynamic properties (like absorption, distribution, metabolism, excretion, and toxicity - ADMET) and thus have a higher chance of becoming successful drugs. While simple rules like Lipinski's "rule-of-five" exist, they are often too simplistic and don't capture the full complexity of drug-likeness. More sophisticated models are needed to accurately assess this property, especially for novel compound libraries like those found in Traditional Chinese Medicine (TCM).
Methods
The researchers employed two main machine learning approaches:
- Naive Bayesian Classification (NBC): This probabilistic classifier is known for its speed and ability to handle large datasets. They trained NBC models using various combinations of molecular physicochemical properties and structural fingerprints.
- Recursive Partitioning (RP): This technique builds decision trees, which can offer more interpretable models. They used RP to create classification trees and compared their performance to the NBC models.
A key aspect of their methodology was systematically investigating the impact of:
- Molecular Descriptors: Comparing the predictive power of 21 physicochemical properties versus various types of structural fingerprints (ECFP, FCFP, LCFP, EPFP, FPFP, LPFP).
- Training Set Size and Balance: Evaluating how the number of molecules and the ratio of drug-like to non-drug-like compounds in the training data affected model accuracy.
Datasets
The study utilized the following datasets:
- MDDR (MDL Drug Data Report): Used as a source of known drug-like molecules.
- ACD (Available Chemicals Directory): Used as a source of non-drug-like molecules.
- TCMCD (Traditional Chinese Medicine Compound Database): The primary target dataset for drug-likeness evaluation.
They preprocessed these datasets, notably filtering for molecules with a molecular weight less than 600 to ensure comparability. They also created balanced training and test sets to mitigate bias.
Key Insights and Ideas
- Structural Fingerprints are Crucial: The study strongly suggests that structural fingerprints (like ECFP and LCFP) are more essential for accurate drug-likeness prediction than physicochemical properties alone. This highlights the importance of capturing detailed molecular substructures.
- Balanced Training is Key: The accuracy of the predictive models is highly sensitive to the balance of drug-like and non-drug-like molecules in the training set. Unbalanced datasets led to significantly lower prediction accuracy.
- TCM as a Drug-Like Reservoir: A substantial portion of TCM compounds were predicted to be drug-like, suggesting TCM is a valuable and underexplored source for drug discovery.
- Fingerprints as Pharmacological Signatures: The identified "good" molecular fragments (those associated with drug-likeness) often correlated with known pharmacological activities in the MDDR dataset. This opens up the possibility of inferring potential biological effects of TCM compounds based on their structural features.
- Bayesian Models Outperform Decision Trees: The Naive Bayesian classifiers demonstrated superior prediction accuracy compared to the Recursive Partitioning models.
Key Results
- High Prediction Accuracy: The best Naive Bayesian classifier, using 21 physicochemical properties and the LCFP_6 fingerprint set, achieved a leave-one-out (LOO) cross-validated accuracy of 91.4% on the training set and 90.9% on the test set.
- Impact of Training Set Size: Model performance improved with increasing training set size, but plateaued around 140,000 molecules, suggesting a good balance between computational efficiency and accuracy.
- TCMCD Drug-Likeness: Approximately 59.37% of the 33,961 TCMCD molecules (with MW < 600) were predicted to be drug-like.
- Fragment Importance: The study identified specific molecular fragments that were highly predictive of drug-likeness, providing valuable insights for medicinal chemists. For example, certain fragments were linked to bronchodilator, antiarthritic, and antihypertensive activities.
- RP Model Performance: The best RP model achieved a Matthews correlation coefficient (C) of 0.525 on the test set, significantly lower than the best Bayesian classifier.
Limitations and Future Work
- "Uncertain Zone": The Bayesian score distribution revealed an "uncertain zone" (around -25 to 25) where predictions are less reliable, suggesting potential for further refinement.
- ACD as Non-Drug-like: While ACD is a common source for non-drug-like molecules, it's acknowledged that some ACD compounds are used in high-throughput screenings, meaning it's not a perfectly pure non-drug-like dataset.
- Inferring Pharmacology: The authors propose that the identified important molecular fragments can be used to annotate the pharmacological activities of TCM compounds, which is a promising avenue for future research.
- Development Stage of Misclassified Molecules: Some misclassified molecules in MDDR were found to be in early stages of drug discovery, suggesting that the models might be identifying potential issues early in the pipeline.
- Stimulating TCM Development: The study aims to provide clues to accelerate the development of TCMs by highlighting their potential as a source of drug-like molecules and by offering a method to infer their biological activities.
Summary updated