Experimental quantification of model identifiability and information loss due to distortions in fluorescence microscopy and image processing

Single-molecule imaging (e.g., using smFISH or MS2-MCP labeling) and fluorescence microscopy can observe and count mRNA molecules within single cells. By quantifying mRNA expression distributions over many cells in different environmental conditions, we can infer discrete stochastic models and discover precise insight into the mechanisms and parameters of gene regulation. However, recent theoretical advances combining Finite State Projection analyses and the Fisher Information Matrix (FSP-FIM) show that inference results depend heavily on experiment design (Fox 2020) and on measurement distortions associated with the microscopy (Vo 2022). In this presentation, we experimentally analyze single-cell preparations using smFISH and MS2-MCP labeling and for many different microscopy and image processing settings. We then use these data to learn probabilistic models for how labeling and imaging settings distort the observation of mRNA spot counts in real cells. Finally, we show how empirically determined distortion operators can be combined with the FSP-FIM to design optimal smFISH experiments that mitigate distortion effects and improve the identification of gene regulation models.