Abstract
A promising technique of discovering disease biomarkers is to measure therelative protein abundance in multiple biofluid samples through liquidchromatography with tandem mass spectrometry (LC-MS/MS) based quantitativeproteomics. The key step involves peptide feature detection in LC-MS map, alongwith its charge and intensity. Existing heuristic algorithms suffer frominaccurate parameters since different settings of the parameters result insignificantly different outcomes. Therefore, we propose PointIso, to serve thenecessity of an automated system for peptide feature detection that is able tofind out the proper parameters itself, and is easily adaptable to differenttypes of datasets. It consists of an attention based scanning step forsegmenting the multi-isotopic pattern of peptide features along with charge anda sequence classification step for grouping those isotopes into potentialpeptide features. PointIso is the first point cloud based, arbitrary-precisiondeep learning network to address the problem and achieves 98% detection of highquality MS/MS identifications in a benchmark dataset, which is higher thanseveral other widely used algorithms. Besides contributing to the proteomicsstudy, we believe our novel segmentation technique should serve the generalimage processing domain as well.