An end-to-end machine learning workflow for MS-based proteomics
Date: Monday June 6th
Topic area: Informatics: Workflow and Data Management
In proteomics, machine learning augments various steps in the data analysis, from predicting peptide properties that serve as priors for experimental data, to training target/decoy classifiers for error estimation. How a model is integrated into production systems determines its requirements. However, generating, evaluating and integrating such models remains manual labor. Here we present a workflow that automates all steps from raw data to production-ready model. First, it imports identified spectra and transforms them to training datasets for various peptide properties. Second, a set of model architectures are trained, evaluated, and their hyperparameters optimized. Hyperparameter optimization automatically balances conflicting requirements such as speed and accuracy and is use-case specific. Third, the models are optimized and exported for different deployment scenarios.
