Drug-Excipient Nanoparticle Design Using Yoked Deep Learning Molecular Simulation

Introduction: In recent years, it has been demonstrated that certain small molecule excipients possess the capability to co-formulate nanoparticles via self-assembly alongside specific drugs. These excipients small molecules, such as Congo red, IR783, and glycyrrhizin, can rapidly create ultra highly loaded nanoparticles (>95%) for some drugs. However, only few such excipients are known, these known excipients are yet unable to encapsulate several important drugs, and several excipients pose risks including allergenic or oncogenic potentials. Therefore, there is an urgent need to identify novel excipients for drug-excipient nanoparticles to expand our material armamentarium, encapsulate drugs that are inaccessible through current excipients, and prevent excipient-triggered side-effects.
In this study, we present a novel, data-free computational approach for predicting nanoparticles formation. Specifically, we conducted molecular dynamics (MD) simulations of drug-excipient pairs in an aqueous environment, coupled with enhanced sampling techniques to compute the binding free energy (BFE) between them. Compared to machine learning methods that rely on training on experimental data, our approach is purely computational. Therefore, it enables us to explore novel excipient molecules with a broader scope, offering promising avenues for innovation and discovery. Based on our predictions, we synthesized several new nanoparticles using our robotic-assisted liquid handling platform and characterized nanoparticle size and stability via dynamic light scattering (DLS). Leveraging our new computational methodology, we uncovered many novel drug-excipient nanoparticles that use new excipients with large potential to further expand the design space of drug-excipient nanoparticles.

Materials and

Methods: On-the-fly Probability Enhanced Sampling (OPES) was employed to calculate the potential of mean force between drug and excipient molecules. The resulting free energy difference between the bound and unbound states quantifies the BFE of the drug-excipient pair.
To calculate the BFE between a drug and an excipient pair, approximately 15 GPU hours are required using RTX 2080. Based on our computational resources, the BFE calculations are limited to a few thousand pairs, far less than the vast chemical space we aimed to explore. To address this challenge, we developed a yoked deep learning molecular simulation (YoDeL-MS) framework. Specifically, we employed Gaussian Process Regression-Marginalized Graph Kernel-RDKit features (GPR-MGK-RDKit) as the teacher model to select only the most relevant molecule pairs for simulation, thereby reducing the computational burden to a manageable level. Then we trained a graph transformer neural network (GraphGPS) as the student model on the select simulations to predict BFE of new pairs that have not yet been simulated.
We then utilized molecules from the Enamine screening library as the chemical space for potential excipient exploration. Directed Message-Passing Neural Networks (DMPNNs) were trained for predicting various toxicity and only molecules predicted to exhibit no toxicity were taken into consideration in the YoDeL-MS framework. We finally purchased top new excipient candidates based on highest average binding BFE to our drug library. Excipients were purchased and tested for nanoparticle formation using our robotic-assisted liquid handling station synthesis platform. Nanoparticle formation was assessed through DLS.

Results, Conclusions, and Discussions: At the beginning, we employed high-throughput molecular simulations to calculate the BFE for all drug-excipient pairs within a dataset of 1440 drug-excipient combinations published previously (Nature Nanotechnology 2021), aiming to elucidate the relationship between the calculated BFE between a drug and an excipient and their ability to form nanoparticles. Our analysis revealed that pairs exhibiting lower BFE (strong binding) displayed a propensity to form nanoparticles, contrasting with those exhibiting higher BFE (weak binding), with a ROC-AUC of 0.63. Furthermore, this relationship was found to hold significance solely for negatively charged excipients, yielding a ROC-AUC of 0.74. We speculate that the negative charge of these excipients facilitates the development of a negatively charged surface on the nanoparticles, thereby enhancing their stability. Additionally, in comparison to the random forest (RF) platform that was published previously (with a similar ROC-AUC of 0.72), BFE demonstrated the ability to identify novel nanoparticles overlooked by RF.
Based on the high-throughput BFE dataset, we conducted a thorough evaluation of various active learning and yoke learning algorithms to enable us to explore new excipient spaces. Our findings highlight that GPR-MGK-RDKit yoked GraphGPS emerged as the most effective algorithm combination. Subsequently, the YoDeL-MS was deployed for large-scale calculations. Following computational non-toxicity screening, we identified approximately 1600 potential excipients from the Enamine database. Leveraging the YoDeLMS framework, we computed BFE values for approximately 160k drugs-excipient pairs. Remarkably, a large proportion of drug-excipient pairs have low BFE values, yet the RF model predicts them as unable to form nanoparticles. After experimental verification, we found several drug-excipient pairs that successfully form nanoparticles, showcasing the power of our new computational platform to identify drug-excipient pairs that cannot be identified with machine learning. Furthermore, this data highlights the ability of our approach to identify novel excipients for drug-excipient nanoparticles.
In summary, we have presented a MD-based computational framework designed for predicting drug-excipient nanoparticle formation. Through this innovative methodology, we discovered several new excipients capable of formulating nanoparticles with a variety of drugs. Our platform holds the potential to accelerate and de-risk the identification of nanomedicines across multiple indications, thereby fostering advancements in therapeutic delivery.

Acknowledgements (Optional):