Machine learning-guided design and optimization of red fluorescent protein-based genetically encoded calcium indicators

Abstract

Calcium signaling is one of the most widely used mechanisms by which cells convert external stimuli, electrical activity, and biochemical cues into intracellular responses. Because calcium dynamics regulate processes ranging from neurotransmission and muscle contraction to epithelial differentiation, development, and cancer-associated signaling, the ability to monitor calcium in living cells has become essential across biology and bioengineering. Genetically encoded calcium indicators (GECIs) have transformed this measurement problem by enabling optical, genetically targetable, and repeated recording of calcium dynamics in defined cell types and subcellular compartments. Despite major advances in GECI engineering, red fluorescent calcium indicators remain limited by trade-offs among brightness, dynamic range, response kinetics, photostability, intracellular localization, and compatibility with optogenetic stimulation. These trade-offs are especially important for experiments that require spectral separation from green reporters or simultaneous use of blue-light-activated actuators. Red GECIs therefore represent both a powerful imaging technology and an unresolved protein engineering challenge. This thesis addresses that challenge using machine learning-guided protein engineering. Rather than relying solely on iterative random mutagenesis and experimental screening, this work uses existing sequence–function data to predict beneficial mutations, experimentally validates those mutations across multiple cellular contexts, and investigates whether the resulting sequence–function insights can be transferred across related red GECI scaffolds. Together, these approaches aim to improve red calcium indicators while revealing the mutational and biophysical trade-offs that shape sensor performance.