Explainable Machine Learning for Identifying the Full Heterogeneity of Peptidoforms and Proteoforms

Mass spectrometry driven proteomics allows deep insights into the working of cells.

Still, the vast majority of proteoforms, representing the full heterogeneity of molecular forms of protein products in a sample, currently remain undetected in proteomics experiments.

This lack of information strongly restricts our knowledge of disease progression, possible biomarkers, and therapeutic targets across a large number of diseases.

Several machine learning approaches have been developed for proteomics data, but not being trained end-to-end, they cannot capture the full wealth of proteomic mass spectra and commonly remain unexplained black boxes.

Within explAInProt, my team and I will develop representations of spectra that allow deploying explainable, end-to-end machine learning models on the wealth of proteomic data available, regarding both bottom-up and topdown spectra to identify novel protein variants.

Explanations will allow identifying the origin of predictions and allow reducing bias and building up the trustworthiness of AI systems required for clinical applications.

To verify results, we will pioneer orthogonal real-time strategies based on selective sequencing approaches and calling of amino acids that we will introduce for nanopore sequencing devices as a complementary acquisition method.

All combined, this will allow to drastically increase our knowledge about the current dark matter of mass spectrometry driven proteomics: those proteins and peptides that are non-canonically modified, non-tryptic, have potentially multiple amino acid substation, or no close match in databases or result from structural variants such as fusion proteins that they remain undetected in current analyses.

We will highlight applicability in two areas of particular concern in current approaches: the detection of structural variants in proteomic mass spectra and the characterization of novel microbial organisms without sufficient database information.