Generative model for orientation bias

[davidbenjamin]

The telltale sign of a substitution error occurring on a single strand of DNA is that supporting evidence is all on forward strand read 1 and reverse strand read 2, or vice versa. This lends itself to a graphical model, the hyperparameters of which can be learned from the data.

Further down the road, we might use a neural network to learn the context-specific risk of such artifacts and attach it to the Bayesian model for forward/reverse and read 1/read 2. This would be our first experience with a deep generative model.

[droazen]

Issue moved to broadinstitute/gatk #3016 via ZenHub