Poster M5 - Machine Learning Guided Combination Nanoparticle Design.

Introduction: All but one currently approved nanoparticles deliver only a single drug payload. However, this method of treatment can lead to implications with single drug toxicity and drug resistance over time. We harness text-mining, machine learning, and robotic-assisted synthesis to develop multi-drug loaded nanoparticles to accelerate and de-risk the design of combination therapy nanoparticles.

Materials and

Methods: We first performed large-scale text-mining of the literature using biomedical databases such as PubMed and Web of Science and extracted information on drug combinations. We then leveraged the use of a liquid handling robot to synthesize a library of combination therapy nanoparticles with our top candidates through nano precipitation experiments with drugs and excipients. Nanoparticles were characterized with dynamic light scattering (DLS) and high-performance liquid chromatography (HPLC) with a UV-vis detector to determine nanoparticle size, stability, and drug loading. We then used this data of successful and unsuccessful nanoparticles to train machine learning models to predict successful nanoparticle formation. Top performing nanoparticles that encapsulate synergistic drug combinations were then evaluated in vitro in multiple cell lines to assess treatment efficacy of the nanoparticles compared to unformulated drug mixtures.

Results, Conclusions, and Discussions: Results, Conclusions+ Discussions
Preliminary text-mining literature searches led to the development of a small screening library of 84 drug combinations (seven drugs and four excipients) that have previously been reported to be synergistic. The robotic protocols successfully produced nanoparticles with two drugs and one excipient. Additionally, all single drug-excipient and drug-drug pairs were synthesized to compare to the drug-drug-excipient nanoparticles. DLS characterizations of these multi-drug nanoparticles showed high success rates of nanoparticle formation with the selected excipients (approximately 40%), suggesting the feasibility of using our nanoparticle synthesis platform to encapsulate several drugs. We then trained a random forest model on the data to assess the ability of machine learning to triage experiments and identify suitable drug-drug-excipient combinations for successful nanoparticle formation. We found high accuracy in retrospective, stratified 10-fold cross validations of our model, suggesting its utility for successful material selections. Therefore, we used our best model to predict and selected excipients to use with specific pairs of drugs for downstream testing. We successfully synthesized two nanoparticles based on our predictions. HPLC showed high drug loading (over 80%) in the combination nanoparticles. In the in vitro testing, we observed better efficacy of our combination therapy nanoparticles in comparison to other tested conditions such as single drug-excipient nanoparticles, single drugs, and non-nanoparticle combinations. These results were reproducible in various cell lines. Overall, our study was successful at utilizing machine learning for finding pairs of drugs to test in combination for nanoparticle fabrication, drug loading characterization, and in vitro efficacy studies. We aim to use our pipeline to develop more combination nanoparticles for broader indications and test out particles in vivo in mouse models.

Acknowledgements (Optional):