cNMF - user experience

[sjfleming]

Notes from meeting with Jim Nemesh.

Improve the user experience for running cNMF.

Usability:

• Streamline the config file
  • don’t require total genes to be input manually (anything that can be inferred from the data should be)
    • or get stuff from a saved checkpoint if applicable
  • require the absolute minimum necessary input in the YAML for functionality
• Ensure HVG finding is streamlined (outside of nmf)
  • this may require writing a separate tool, not sure
  • somehow we need a command-line way for users to use CellariumML to come up with the HVG list for use in a Filter
• Enable users to input a range of k values
  • Do all the runs
  • How do we want to do this? In series? In parallel? Could it be done in parallel...?
    • Maybe we could set things up so that the model itself sequentially fits a lot of k models (in each minibatch)
• Compute and output metrics for all runs in a way that’s readable and plottable (maybe even make the plot)
  • Output “stability” and “error” (use the same terminology as cNMF paper)
  • Have a CSV output file that can be used to make the plot very simply
• Have code to make the necessary stability and consensus clustermap/hist plots in jupyter notebook
  • Simple API for interactive analysis
  • The “consensus” part need not be a command line tool really… they want it to be interactive in a notebook

Note:

• If you have multiple command line steps… then you better do input validation
• Make sure that a user cannot change the input file between the initial fit and the final refit part, for example