CELL-E: Biological Zero-Shot Text-to-Image Synthesis for Protein Localization Prediction

Abstract

Predicting the cellular activities of proteins from their primary amino acid sequences is a highly desirable capability that could greatly enhance our functional understanding of the proteome. Here, we demonstrate CELL-E, a text-to-image transformer architecture, which given a protein sequence and a reference image for cell (or nucleus) morphology, can generate a 2D probability density map of the protein distribution within cells. Unlike previous in silico methods, which rely on existing, discrete class annotation of protein localization to predefined subcellular compartments, CELL-E uses imaging data directly, thus relying on a native description of protein localization relative to the cellular context.