Complementary peaks in spectrum vectors

[bittremieux]

Currently spectra are converted to vectors by binning their peaks, after which they are transformed to lower-dimensional vectors.

We should explore whether appending complementary peaks or neutral losses to the vector improves performance of open searches. Specifically, the high-dimensional spectrum vector would consist of the binned peaks and the binned neutral losses, after which the low-dimensional transformation is performed on the concatenated vector.

The hypothesis is that this will improve the candidate selection step from the ANN index. Candidate selection is performed using the cosine similarity, without considering shifted peaks. Instead, complementary peaks might still capture some peak shifts.