# Improving Spectral Similarity and Molecular Network Reliability through Noise Signal Filtering in MS/MS Spectra

## TL;DR

This paper shows how removing noise from MS/MS spectra improves the accuracy of molecular networks and compound identification in metabolomics.

## Contribution

A data-specific workflow and faster denoising method are introduced to enhance spectral similarity and network reliability.

## Key findings

- Noise elimination increased similarity scores for homologous spectra and improved molecular network structure.
- Denoised networks showed denser regions and fewer false-positive connections.
- A tailored intensity-based denoising method performed as well as fixed threshold approaches.

## Abstract

In mass spectrometry, fragmentation spectra play a central
role
in compound identification. However, noise in MS/MS spectra can significantly
impact similarity scores and molecular network (MN) reliability, leading
to inaccurate compound annotation in untargeted metabolomics. This
work investigates the influence of noise on MS/MS similarity scores
and molecular network structure. Noise elimination increased similarity
scores for homologous spectra, enhancing match affordability. In MNs,
effective noise management improved network structure, resulting in
more interpretable networks with fewer edges and enhanced clustering,
decreasing false-positive connections. To quantitatively assess these
improvements, a minimum spanning tree (MST) analysis was performed,
revealing denser regions in the denoised MNs. An increasing cutoff
of noise threshold can lead to an overlay between two or more different
compound spectra. A data-specific workflow was developed to identify
the optimal threshold for denoising, balancing spectra quality and
network integrity during noise elimination, by incorporating statistics
calculated on the distribution of the MST distances and the number
of fragment ions, which could be explained by an in-silico fragmentation
algorithm. Finally, a faster-tailored denoising method, based solely
on the intensity of individual spectral ions, demonstrated performance
comparable to the previously cited fixed threshold approaches.