Machine Learning-Guided Engineering of Protein Phase Separation Properties in Immune Regulation

Phase separation (PS) underpins compartmentalization in living cells, facilitating the formation of membraneless organelles and the regulation of cellular processes. Despite the increasingly pivotal role of engineering protein PS properties in the study and regulation of cellular physiological processes, manipulating PS ability through single amino acid alterations remains a challenge. Here, we develop phase separation scalpel (PScalpel), a machine learning-based tool identifying and recommends protein engineering strategies for directed changes in PS ability. Based on our biological experimental data, we apply transfer learning to achieve the feedback-driven optimization of specific protein prediction accuracy--markedly enhancing the predictive performance for TDP43, a neurodegenerative disease-associated protein. Furthermore, by engineering the crucial nucleic acid sensor cGAS as a model application, we successfully modulate its PS ability in the anticipated direction by altering a single amino acid, which subsequently optimizes its immune function and impacts the activity of engineered macrophages. Transcriptomic analysis of these cGAS-engineered macrophages further demonstrated that the immune function of macrophages can be altered by the manipulation of cGAS PS ability. In summary, PScalpel is an effective tool for guiding PS ability engineering, enabling targeted molecular and cellular modifications and providing nuanced methods for precise biomolecular engineering in future research.