Casanovo Using AttentionSmithy

[CCranney]

For those who are not familiar with Casanovo, it's a revolutionary approach to de novo peptide sequencing using a transformer model architecture. I'm including the figure they use for context, but I encourage you to visit the repository and read the papers yourself.

In a nutshell, Casanovo takes in data-dependent acquisition mass spectrometry data (DDA-MS) and assembles a plausible amino acid (peptide) sequence it represents without reference to a library file. For those not familiar with this field, this is an incredibly valuable and powerful tool. Biology gets more complicated the deeper in you go, and so being able to make sense of biological data without needing to know what you're looking for has so many potential applications that the concept is mind-boggling. Read the papers for specifics, and I think they actually have a page in their documentation for ideas on future applications. The Noble Lab has created an amazing model.

In studying their model, I kept coming up with ideas I wanted to try. What happens if you use rotary position strategies for m/z values, for instance? Are there ways you can apply this to data-independent acquisition (DIA-MS) data instead? What changes or updates would make that possible? And how easy would it be to experiment with those options? Some papers have been written that have tried this, but each single experiment is effectively a unique, total rewrite of the Casanovo project.

Situations like this were part of the inspiration for creating AttentionSmithy in the first place. For models that use it as their foundation, it becomes easier to experiment with various alternate strategies without changing the base code too much.

So I'm listing here the idea of writing a version of Casanovo that uses AttentionSmithy as its foundation, such that architectural experimentation becomes easier.