Characterizing ssRNA and dsRNA electrophoretic behavior: empirical insights with neural network-aided predictions

RNA-based therapeutics are currently at the forefront of the biopharmaceutical industry because of their safety, efficacy, and shortened time from disease discovery to therapy development. Microfluidic electrophoresis provides a great analytical platform to analyze nucleic acids in unprecedented detail. However, while DNA has been studied extensively within microfluidic systems, there is limited data available for RNA, particularly of chemically modified molecules, such as those used in the COVID-19 mRNA vaccines, and for long double-stranded RNA molecules, which may accompany, intentionally or as a by-product, RNA therapeutics. To this end, this study focused on the empirical microfluidic electrophoretic analysis of double- and single-stranded RNA, non-modified and pseudouridine-modified, at varying gel concentrations. It then compared the findings to the electrophoretic mobility models in the literature. This work was then complemented with data-driven and physics-informed neural networks that successfully predicted the migration time and length of different RNA molecules with an average error of 12.34% for the data-driven model and 0.77% for the physics-informed model. The low error in the physics-informed neural networks opens the doors to the electrophoretic characterization of molecules, even beyond RNA, without the need for extensive experimental data.