Untargeted mass spectrometry (MS) workflows are more suitable than targeted workflows for high throughput characterization of complex biological samples. However, analysis workflows for untargeted methods are inadequate for characterization of… Click to show full abstract
Untargeted mass spectrometry (MS) workflows are more suitable than targeted workflows for high throughput characterization of complex biological samples. However, analysis workflows for untargeted methods are inadequate for characterization of complex samples that contain multiple classes of compounds as each chemical class might require a different type of data processing approach. To increase the feasibility of analyzing MS data for multi-class/component complex mixtures (i.e., mixtures containing more than one major class of biomolecules), we developed a neural network-based approach for classification of MS data. In our in silico fractionation (iSF) approach, we utilize a neural decision tree to sequentially classify biomolecules based on their MS-detected isotopic patterns. In the presented demonstration, the neural decision tree consisted of two supervised binary classifiers to positively classify polypeptides and lipids, respectively, and a third supervised network was trained to classify lipids into the eight main sub-categories of lipids. The two binary classifiers assigned polypeptide and lipid experimental components with 100% sensitivity and 100% specificity; however, the 8-target classifier assigned lipids into their respective subclasses with 95% sensitivity and 99% specificity. Here, we discuss important relationships between class-specific chemical properties and MS isotopic envelopes that enable analyte classification. Moreover, we evaluate the performance characteristics of the utilized networks.
               
Click one of the above tabs to view related content.