Physics-informed self-supervised learning enables spectra-free multiplexed imaging on standard fluorescence microscopes

Abstract

Multiplexed fluorescence imaging is limited by spectral overlap and the small number of excitation or emission channels available on standard microscopes, restricting most laboratories to low-plex imaging. Here we introduce physics-informed spectra-free multiplexed imaging (PhySMI), a self-supervised framework for underdetermined spectral unmixing that enables highly multiplexed imaging without dense spectral measurements after training. By embedding the spectral forward-mixing process into a self-consistent architecture, PhySMI recovers physically plausible source decompositions from unlabeled data without paired ground-truth labels while suppressing stochastic acquisition noise. PhySMI resolves five subcellular structures from only three excitation channels, overcoming the conventional channel-number limit while preserving spectral fidelity and minimizing crosstalk (<0.5%). The framework also generalizes across imaging systems, enabling zero-shot deployment on standard fluorescence microscopes. In live cells, PhySMI enables fast five-color imaging of dynamic multi-organelle interactions with improved temporal resolution and reduced photobleaching and phototoxicity relative to conventional spectral imaging. These results establish a general strategy for physics-informed learning in underdetermined imaging inverse problems and represent a step toward a general-purpose framework for highly multiplexed fluorescence imaging on standard microscopy platforms.