“Data are dumb” [Judea Pearl]. So, how do we model single cell data over time to extract biology? In NeuroVelo we designed a physics-informed Neural ODE architecture to identify gene regulatory networks using interpretable latent spaces in scRNA-seq. https://t.co/XVjhTrFmjd
@GorinGennady 3- That is true to some extent. The method doesn't focus on directions that can be ambigious, but the rest of vector is leanable as we focuses more on giving interpretations to these directions as discribed in the paper. 2/3
@GorinGennady 1- we gain flexibility in approximating nonlinear dynamics, and important stochastic effects might be missing, but as a first approx seems a plausible approach.
2- where do we do it? We use dimensionality reduction precisely to avoid doing data imputation in sparse data sets.