Neutral mutations don't give neutral results

[mpdunne]

Hi there,

I'm trying out mutations on a DeepSequence model trained on BLAST_ECOLX.

What should be neutral mutations do not seem to give neutral results. For example:

data_helper.delta_elbo(ds_model,[], N_pred_iterations=500) gives 0.25998583984375045

I would expect this to be zero, and it is, approximately. But:

data_helper.delta_elbo(ds_model,[(126,"T","T")], N_pred_iterations=500) gives -4.7841062011718805

I would also expect this to be zero. But it is not! Am I right in thinking this?

data_helper.delta_elbo(ds_model,[(126,"T","Q")], N_pred_iterations=500) gives -9.021644348144548

which looks reasonable.

Also just to double-check, a higher score is better? Thanks!