A Machine Learning Workflow to Enhance Microfluidic Development of Nanomedicines

Artificial intelligence (AI) is being implemented in almost every facet of modern-day life, and machine learning, a subfield of AI, has the potential to greatly streamline the process of developing various nanomedicines using microfluidic fabrication techniques. The availability of open-source machine learning frameworks makes these powerful tools accessible to scientists, facilitating the integration of wet-bench activities with computational analysis. Here, we present a machine learning workflow aimed at optimizing the microfluidic formulation development of nanomedicines. A database of almost 200 unique nanomedicine formulations with over 550 total measurements was curated by producing liposomes, lipid nanoparticles, and PLGA nanoparticles using a benchtop microfluidic system. Microfluidic and materials input features, including the total flow rate, aqueous:organic flow rate ratio, and reagent concentrations, were systematically varied, and the resulting particles were characterized for their hydrodynamic diameter (dH), polydispersity index (PdI), and encapsulation efficiency (EE) for a model therapeutic agent, curcumin (CURC). These data were used to train, test, and validate 13 different machine learning models with the task of returning the most accurate prediction of the nanomedicine attributes – dH, PdI, EE. The most accurate machine learning models, based on random forest regression, were implemented to provide the optimal formulation to yield particles with user-specified attributes. Finally, this system, dubbed MicrofluidicML, was compared against generative large language models as represented by the Open AI ChatGPT and Google’s Gemini platforms. The application of machine learning in the field of nanomedicine is inevitable, and MicrofluidicML represents a step towards implementing a machine learning framework towards accelerating formulation development.