Deep learning-assisted concentration gradient generation for the study of 3D cell cultures in hydrogel beads of varying stiffness.

Abstract

The study of dose-response relationships underpins analytical biosciences. Droplet microfluidics platforms can automate the generation of microreactors encapsulating varying concentrations of an assay component, providing datasets across a large chemical space in a single experiment. A classical method consists in varying the flow rate of multiple solutions co-flowing into a single microchannel (producing different volume fractions) before encapsulating the contents into water-in-oil droplets. This process can be automated through controlling the pumping elements but lacks the ability to adapt to unpredictable experimental scenarios, often requiring constant human supervision. In this paper, we introduce an image-based, closed-loop control system for assessing and adjusting volume fractions, thereby generating unsupervised, uniform concentration gradients. We trained a shallow convolutional neural network to assess the position of the laminar flow interface between two co-flowing fluids and used this model to adjust flow rates in real-time. We apply the method to generate alginate microbeads in which HEK293FT cells could grow in three dimensions. The stiffnesses ranged from 50Pa to close to 1kPa in Young modulus and were encoded with a fluorescent marker. We trained deep learning models based on the YOLOv4 object detector to efficiently detect both microbeads and multicellular spheroids from high-content screening images. This allowed us to map relationships between hydrogel stiffness and multicellular spheroid growth.