The use of deep neural networks for modelling system dynamics is increasingly popular, but long-term prediction accuracy and out-of-distribution generalization still present challenges. In this study, we address these challenges by considering the parameters of dynamical systems as factors of variation of the data and leverage their ground-truth values to disentangle the representations learned by generative models. Our experimental results in phase-space and observation-space dynamics, demonstrate the effectiveness of latent-space supervision in producing disentangled representations, leading to improved long-term prediction accuracy and out-of-distribution robustness.

MOTIVATION: Present research on gene expression using live cell imaging and fluorescent proteins or tagged RNA requires accurate automated methods of quantification of these molecules from the images. Here, we propose a novel automated method for classifying pixel intensities of fluorescent spots to RNA numbers. RESULTS: The method relies on a new model of intensity distributions of tagged RNAs, for which we estimated parameter values in maximum likelihood sense from measurement data, and constructed a maximum a posteriori classifier to estimate RNA numbers in fluorescent RNA spots. We applied the method to estimate the number of tagged RNAs in individual live Escherichia coli cells containing a gene coding for an RNA with MS2-GFP binding sites. We tested the method using two constructs, coding for either 96 or 48 binding sites, and obtained similar distributions of RNA numbers, showing that the method is adaptive. We further show that the results agree with a method that uses time series data and with quantitative polymerase chain reaction measurements. Lastly, using simulated data, we show that the method is accurate in realistic parameter ranges. This method should, in general, be applicable to live single-cell measurements of low-copy number fluorescence-tagged molecules. AVAILABILITY AND IMPLEMENTATION: MATLAB extensions written in C for parameter estimation and finding decision boundaries are available under Mozilla public license at http://www.cs.tut.fi/%7ehakkin22/estrna/ CONTACT: [email protected].