how bonito normalize raw signals when save-ctc

[xieyy46]

Hi bonito team!
I want to know how bonito normalize raw signals! I noticed that in the chunks.npy, the signal seem to be normalized to (-1, 1). Does bonito normalize raw signals in per chunk level? or per level? or per run level?

[iiSeymour]

Hi @xieyy46 most commonly, normalization happens per read using med_mad or later quantile strategies (reference implementation can be found in the code). However, in the very latest models we have switched to a global strategy.

[joshtburdick]

I was thinking about normalization a bit.

In particular, I was thinking about a very naive model of the five bases in a pore, as five resistors connected in series. I'm sure the pore is more complicated than this, but it's a simple model. With that (admittedly way oversimplified) model, the resistance of all five is just the sum, and so a linear function.

So, you could use the assumed voltage, and convert from pA to pico-ohms or nano-ohms.

I'm wondering if this might be slightly easier for the NN to parse? I realize that the NN can model all sorts of nonlinear things, and I wouldn't expect a large improvement. But it seems like a (relatively) small software-only thing to try (although no doubt complicated).