TL;DR
This paper introduces a novel platform that combines molecular dynamics (MD) simulations, density functional theory (DFT), and explainable machine learning (ML) to accurately predict the binding properties of antiviral drugs (favipiravir and ebselen) on monolayer black phosphorus (BP). The approach demonstrates the potential of functionalized 2D materials as drug delivery systems and offers insights into drug-surface interactions.
What Problem Does This Paper Address?
The development of effective drug delivery systems is a critical area in modern medicine. Two-dimensional (2D) materials, like black phosphorus (BP), have emerged as promising candidates due to their unique properties. However, understanding the intricate interactions between antiviral drugs and these 2D nanomaterials at an atomistic level is crucial for designing efficient drug delivery platforms. Traditional methods for this analysis can be computationally expensive and time-consuming. This paper aims to bridge this gap by developing a fast, accurate, and interpretable computational framework to predict these interactions, specifically focusing on the antiviral drugs favipiravir (FP) and ebselen (EB) interacting with monolayer BP.
Methods
The study employs a multi-pronged approach:
- Molecular Dynamics (MD) Simulations: Used to replicate non-covalent geometries and capture the dynamic interactions between antiviral molecules and the BP surface under various conditions (temperature, pressure). The Forcite module in BIOVIA Material Studio was utilized.
- Density Functional Theory (DFT): Employed for accurate electronic structure calculations and to determine preferred adsorption sites and interaction energies. The Vienna Ab Initio Simulation Package (VASP) was used, incorporating the PBE functional and Grimme's D3-BJ dispersion correction.
- Ensemble Machine Learning (ML): Four powerful ensemble learning models – Random Forest, Gradient Boosting, XGBoost, and CatBoost – were trained to predict key energy parameters (total potential energy, van der Waals energy, and non-bond energy). This approach aims to accelerate the prediction process and approximate DFT results.
- Explainable AI (SHAP): The SHapley Additive exPlanations (SHAP) framework was used to interpret the ML models' predictions, providing insights into the contribution of each feature (e.g., different energy terms) to the overall prediction. This helps understand why the models make certain predictions.
Datasets
The study utilized datasets generated from MD simulations of favipiravir (FP), ebselen (EB), and a combined FP_EB system adsorbed on a monolayer black phosphorus (BP) surface. These datasets consisted of nine input variables (total kinetic energy, bond energy, angle energy, torsion energy, inversion energy, non-bond energy, van der Waals energy, pressure, and temperature) and were used to predict three output energy parameters. The data was split into 80% for training and 20% for testing, with five-fold cross-validation employed during training.
Key Insights and Ideas
- Synergistic Effects in FP_EB: The combined FP_EB system exhibits more stable configurations and lower potential energy compared to individual FP or EB, suggesting a synergistic interaction that enhances stability.
- Dominance of Van der Waals Forces: Van der Waals (vdW) interactions consistently emerge as the most influential factor in predicting total potential energy and van der Waals energy, highlighting their critical role in the adsorption of these drugs on BP.
- Interpretability is Key: The use of SHAP values provides crucial insights into which molecular features (energy terms) are most important for predicting different energy parameters, moving beyond black-box ML models.
- BP as a Promising Nanocarrier: The functionalized 2D BP complex shows robust thermostability and potential for adsorption, indicating its promise as an effective drug delivery entity.
- ML as a DFT Accelerator: The study demonstrates that well-trained ML models can accurately approximate DFT calculations, significantly speeding up the process of predicting molecular interactions.
Key Results
- High Prediction Accuracy: The ensemble ML models achieved high prediction accuracy for total potential energy, van der Waals energy, and non-bond energy, with R² values consistently close to 1 and low RMSE and MAE scores across all tested scenarios (EB, FP, and FP_EB).
- FP_EB Interaction: The binding energy for the FP_EB system (-19.58 kcal/mol) was significantly higher than for FP (-18.91 kcal/mol) and EB (-10.21 kcal/mol), indicating stronger overall interaction.
- Feature Importance: SHAP analysis revealed that van der Waals energy and non-bond energy were the most significant predictors for total potential energy and van der Waals energy, respectively. For non-bond energy prediction, Total Potential Energy was dominant for EB, while van der Waals energy was dominant for FP and FP_EB.
Limitations and Future Work
- Limitations: While the models show high accuracy, the prediction performance for the FP_EB scenario slightly decreased compared to individual drugs, attributed to the increased complexity of the combined system. The study focused on a specific vacuum interface, and real-world conditions might involve more complex environments.
- Future Work: The authors suggest several avenues for future research:
- Advanced ML Models: Incorporating more sophisticated ML algorithms, such as deep learning architectures, for potentially higher accuracy.
- Broader Drug/System Scope: Extending the study to more drugs and molecular systems to gain a more comprehensive understanding.
- Environmental Factors: Investigating the influence of environmental factors like temperature and pH on drug binding to low-dimensional materials.
- Real-World Applications: Exploring how these findings can be translated into more effective and reliable pharmaceutical therapies.
Summary updated