TL;DR
This paper tackles the urgent need for new treatments for Opioid Use Disorder (OUD) by using machine learning to screen existing drugs for repurposing. They developed models to predict how well DrugBank compounds bind to key opioid receptors and identified promising candidates with favorable ADMET properties, offering a faster and cheaper drug discovery pathway.
What Problem Does This Paper Address?
Opioid Use Disorder (OUD) is a chronic and debilitating condition with significant societal impact. While existing medications like methadone, buprenorphine, and naltrexone are available, they have limitations, including side effects, adherence issues, and relapse rates. Developing new, more effective treatments is a major challenge, as traditional drug discovery is time-consuming and expensive. Drug repurposing, the process of finding new uses for existing drugs, offers a promising alternative by leveraging existing safety and efficacy data, potentially accelerating the development of new OUD therapies. However, identifying suitable candidates from vast drug libraries requires efficient screening methods.
Methods
The researchers employed a machine learning approach to predict the binding affinity of DrugBank compounds to four opioid receptors: MOR, KOR, DOR, and NOR. They utilized a gradient boosting decision tree (GBDT) algorithm, known for its robustness and ease of implementation. To represent drug molecules, they generated molecular fingerprints using three methods:
- Transformer (TF-FP) and Autoencoder (AE-FP): These are novel NLP-based techniques that learn latent representations from SMILES strings.
- Extended-Connectivity Fingerprints (ECFP): A traditional 2D topological fingerprint.
They built individual models for each fingerprint type and then created consensus models by averaging predictions to improve accuracy. The models were trained on inhibitor datasets from the ChEMBL database. Additionally, they incorporated a screening step for hERG channel inhibition to avoid potential cardiotoxicity. Promising candidates were further evaluated for ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties using ADMETlab 2.0. Molecular docking was also performed on select compounds to visualize their interactions with opioid receptors.
Datasets
The study utilized the following datasets:
- ChEMBL database: Provided inhibitor data for MOR, KOR, DOR, NOR, and hERG.
- DrugBank database (version 5.1.10): A comprehensive database containing 8865 compounds, including 1806 FDA-approved drugs and 7059 investigational or off-market drugs, which were screened for OUD treatment potential.
Key Insights and Ideas
- NLP-based fingerprints show promise: The study highlights the effectiveness of transformer and autoencoder-based molecular fingerprints, demonstrating their ability to capture crucial molecular information for predictive modeling, sometimes outperforming traditional methods.
- Consensus models enhance prediction: Combining predictions from multiple models (using different fingerprints) significantly improved predictive performance, underscoring the benefit of ensemble approaches in drug discovery.
- Selectivity is key: The researchers emphasized the importance of drug selectivity for specific opioid receptors to achieve desired therapeutic effects while minimizing off-target side effects. Their models allowed for the discrimination of compounds with varying selectivity profiles.
- ADMET is crucial for repurposing: Beyond binding affinity, evaluating ADMET properties is essential for identifying viable drug candidates for OUD treatment, as it directly impacts a drug's safety and efficacy in vivo.
Key Results
- The machine learning models demonstrated strong predictive power for binding affinity, with Pearson correlation coefficients (R) ranging from 0.783 to 0.866 and RMSE values between 0.944 and 0.969 kcal/mol for the opioid receptors. The hERG model achieved an R of 0.786 and an RMSE of 0.773 kcal/mol.
- The consensus model employing TF, AE, and ECFP fingerprints showed the best overall performance across the datasets.
- The study identified 15 FDA-approved drugs and 19 investigational drugs predicted to have high potency on opioid receptors, with some showing selectivity.
- A subset of these drugs also exhibited favorable ADMET profiles, suggesting their potential for OUD treatment.
- Molecular docking provided insights into the specific interactions between promising drug candidates and opioid receptors, such as hydrogen bonds and hydrophobic interactions.
Limitations and Future Work
- NOR dataset bias: The NOR dataset exhibited a label distribution with a high percentage of high-affinity values, leading to potential inaccuracies in predictions for compounds with low similarity to the training set. The authors acknowledge this bias and suggest it's a challenge to address.
- Further experimental validation needed: The identified drug candidates require further experimental studies, including animal experiments and clinical trials, to confirm their safety and efficacy for OUD treatment.
- Broader ADMET and pharmacological profiling: While ADMET properties were screened, a comprehensive evaluation of other factors like blood-brain barrier permeability and specific antagonist/agonist effects is necessary for a complete assessment of therapeutic potential.
- Generative models for novel drug design: The authors mention the potential of generative network modules to assist in designing new candidate compounds based on existing ones.
Summary updated