train deep network with lstm layer with missing (NaN) data input and target (not synced)

Hello,

is it possible to train a lstm network with NaN values on input and target data like we can with timedelay nets (in this case there exists the fixunknowns function?

I tried a simple case adding just 1 NaN value and the training didn't converged....

Is there a "fixunknown function" for deep networks...