Deep-learning enables proteome-scale identification of phase-separated protein candidates from immunofluorescence images

Intrinsically disordered region (IDR) analysis has been widely used in the screening of phase-separated proteins. However, the precise sequences determining phase separation remain unclear. Furthermore, a large number of phase-separated proteins that exhibit relatively low IDR content remain uncharacterized. Phase-separated proteins appear as spherical droplet structures in immunofluorescence (IF) images, which renders them distinguishable from non-phase-separated proteins. Here, we transformed the problem of phase-separated protein recognition into a binary classification problem of image recognition. In addition, we established a method named IDeepPhase to identify IF images with spherical droplet structures based on convolutional neural networks. Using IDeepPhase on proteome-scale IF images from the Human Protein Atlas database, we generated a comprehensive list of phase-separated candidates which displayed spherical droplet structures in IF images, allowing nomination of proteins, antibodies and cell lines for subsequent phase separation study.