PRISM · Physics‑aware generative AI for Double-Blind Spectral Unmixing

Fluorescence is a powerful technique that uses molecular tags, each emitting a unique spectral fingerprint of light, to highlight specific cells or proteins. This enables researchers to map cellular interactions and mechanisms. However, as our research needs grow and we use more tags to create a complete picture, their fingerprints inevitably overlap, creating a mixed signal that obscures critical biological information.Current state-of-the-art unmixing methods operate with a critical restriction: they require the exact number of fluorescent components present and/or the precise spectral fingerprint of each one. This forces time-consuming pre-calibration and prevents the system from identifying unexpected signals, such as tissue autofluorescence, which are often part of the scientific question itself.The PRISM project will overcome this barrier by developing the first-ever ""double-blind"" unmixing framework, which requires neither the spectra nor the number of components to be known in advance. The core of this innovation is a novel, phasor-guided deep learning architecture that integrates physically-constrained generative models (Transformers, Diffusion Models) to analyze data from spectral microscopy and spectral flow cytometry.To ensure maximum impact, the project will deliver three key outcomes: 1) A curated, open-access benchmark dataset of synthetic and real-world data