Robustify non-uniform model estimation for small data

[ahonkela]

A user reported that non-uniform model estimation can yield very bad results when the number of reads is extremely low (of the order of thousands).

[magnusrattray]

Since this is a probabilistic model, can we use a prior to give a sensible behaviour in the case where the data is used to fit the model is sparse? Really, for small datasets I think the model should revert to the uniform read model. A more sophisticated solution would be to revert to a simpler non-uniform model using model selection. A related idea is to use multiple datasets (e.g. replicates or many single-cell datasets) to learn the bias model.