Liquid–Liquid Phase Separation Prediction of Proteins in Salt Solution by Deep Neural Network

Abstract

Liquid–liquid phase separation (LLPS) underlies the formation of membrane-free organelles in eukaryotic cells and plays an important role in the development of some diseases. The phase boundary of metastable liquid–liquid phase separation as well as the cloud point temperature of some globular proteins characterize the phase behavior of proteins and have been widely studied theoretically and experimentally. In the present study, we used a regression and classification neural network to deal with the phase behavior of lysozyme and bovine serum albumin (BSA). We predicted the cloud point temperature and solubility of a lysozyme solution containing sodium chloride by regression and the reentrant phase behavior of BSA in YCl3 solution containing a surfactant dodecyl dimethyl amine oxide (DDAO) by classification. Specifically, our network model is capable of predicting (a) the solubility of lysozyme in the range: pH 4.0–5.4, temperature 0–25 °C, and NaCl concentration 2–7% (w/v); (b) the cloud point temperature of lysozyme in the range: pH 4.0–4.8, NaCl concentration 2–7%, and lysozyme concentration 0–400 mg/mL; and (c) the phase behavior of BSA in the range: DDAO 1–60 mM, BSA 30–100 mg/mL, and YCl3 1–20 mM. We experimentally tested the model at some prediction points with a high accuracy, which means that deep neural networks can be applicable in qualitative and quantitive analysis of liquid–liquid phase separation.