jpvert
commented
7 years ago One possibility would be to run the model with your batch and library size effects captured by covariates, and with "enough" latent factors to infer W and (alpha_mu,alpha_pi). Then Walpha_mu is our best estimate of the log-expression level, and Walpha_pi is our best estimate of the logic-probability of dropout. I would not recommend directly fitting a mode without latent factors and then subtracting the estimated batch and library size effects from the raw counts, since our model really assumes a statistical distribution for observed counts which is not so simple. If the method you want to use afterwards require a matrix of data that "looks" like the original counts but without batch and library size effects, then maybe a possibility is to simulate counts using the ZINB distributions with parameters Walpha_mu and Walpha_pi, although this looks a bit artificial compared to directly work with the matrices Walpha_mu and Walpha_pi. Finally, regarding how many latent factors to use to estimate W and alpha, you can set it to a small number (say k=2) if you believe the real data have two degrees of freedom, or just set it to a larger value (say k=10 or 50) if not; note that we did not really test the later case, the math says that this should not be a problem, but I hope the numerical errors will not accumulate. Let us know!
drisso
JohnReid
Thanks for a nice model with a good implementation. I'm curious if I can use the model just to normalise my single cell counts. I'm not interested in a factor analysis but would like to remove batch effects and library size effects from the counts before applying my own model. What would you recommend to do? I'm thinking just to fit the model and subtract the expected non-zero counts from the counts. This ignores the ZI probability though. I suppose it is difficult to separate the technical effects from the biological effects for the ZI probabilities.