Deep Learning-Aided High-Throughput Screening of Time-Resolved Protein Crystallization on Programmable Microliter-Droplet Systems

Abstract

Crystallization of proteins is a critical step in structural biology, biopharmaceutical industry and materials science. Microfluidic technology has emerged as a promising tool for screening the crystallization conditions with advantages of increased throughput, reduced consumption of reagents and lower cost. However, current mirofluidic approches generally lack the module for high-speed data analysis, ignore the time-resolved changes of protein crystalline states or morphologies in the crystallization process and suffer from inconsistency after scaling up due to the subnano-/nano-liter-scaled volume. To address the issues, we propose a deep learning-aided programmable microliter-droplet system, which allows the high-throughput screening of time-resolved protein crystallization in microliter scale. Based on the system, a series of temporal phase diagrams are acquired, which reveal the time-resolved crystallization of target proteins under different crystallization conditions. They provide precise guidance on the scale-up experiment (∼93 % in consistency), and help gain insight into the kinetic characteristics of protein crystallization.