Verily’s DeepMass, a deep-learning application for mass spectrometry, is highly accurate at predicting peptide mass spectra, according to a manuscript published in Nature Methods.
The researchers applied DeepMass and found that it demonstrated a more accurate method to interpret data coming from experimental platforms that analyze proteins and increase the ability to identify and characterize known biomarkers in a sample. The findings can be used in data-dependent and data-independent acquisition computational workflows to improve peptide identification rates and reduce the reliance on spectral libraries.
The study authors wrote that they anticipate that the model’s greatest impact will be through providing in silico-based spectral libraries. Another application can be to supplement an existing experimental library with a small number of hypothesis-driven peptides to expand the range of scientific and clinical questions beyond measuring the levels of proteins in a sample.
“We hope that DeepMass, available on Google Cloud, will enable researchers to characterize disease-relevant protein profiles to build new diagnostic tools and therapeutics,” wrote Peter Cimermancic and Roie Levy, computational biologists at Verily, a sister company to Google.
The research team, which included experts from Verily, Google and Jurgen Cox’s Computational Systems Biology research group at the Max Planck Institute of Biochemistry, found that the using DeepMass, the cross-correlation coefficient between the actual and predicted spectra is 0.944, which is better than the available state-of-the-art (0.871).
Researchers at Verily measure protein profile using mass spectrometry to search for new biomarkers of disease. The company also integrates protein signals with other biomolecular data like genomics and transcriptomics to find out how genetics and behavior affect protein profiles. Data-independent acquisition is used to identify and quantify proteins more accurately and precisely than previous methods. But the method relies on experimentally determined spectral libraries — curated, annotated and non-redundant collection of spectra — for data interpretation, which is time- and resource-intensive.
The researchers developed DeepMass to generate spectral libraries by computation and hypothesized that the tool would help them more quickly develop the necessary reference material to interpret large data sets of protein profiles generated by mass spectrometry.
In its first application to clinical data at Verily, DeepMass helped expand the coverage of known biomarkers more than twofold.
Researchers demonstrated that the use of DeepMass-calculated spectral libraries is equivalent to the experimental ones.
The researchers discovered that the model correctly learned known and new chemical rules that govern a peptide fragmentation, which the team found surprising, according to Cimermancic and Roie.
Verily will continue to apply machine learning to proteins, proteogenomics and other fields to continue its mission of making health data useful so people can live healthier lives.
Send this to a friend