SteroidXtract: Deep Learning-Based Pattern Recognition Enables Comprehensive and Rapid Extraction of Steroid-Like...

Despite the vast amount of metabolic information that can be captured in untargeted metabolomics, many biological applications are looking for a biology-driven metabolomics platform that targets a set of metabolites that are relevant to the given biological question. Steroids are a class of important molecules that play critical roles in many physiological systems and diseases. Besides known steroids, there are a large number of unknown steroids that have not been reported in the literature. The ability to rapidly detect and quantify both known and unknown steroid molecules in a biological sample can greatly accelerate a broad range of steroid-focused life science research. This work describes the development and application of SteroidXtract, a convolutional neural network (CNN)-based bioinformatics tool that can recognize steroid molecules in mass spectrometry (MS)-based untargeted metabolomics using their unique tandem MS (MS²) spectral patterns. SteroidXtract was trained using a comprehensive set of standard MS² spectra from MassBank of North America (MoNA) and an in-house steroid library. Data augmentation strategies, including intensity thresholding and Gaussian noise addition, were created and applied to minimize data overfitting caused by the limited number of standard steroid MS² spectra. The CNN model embedded in SteroidXtract was further compared with random forest and XGBoost using nested cross-validations to demonstrate its performance. Finally, SteroidXtract was applied in several metabolomics studies to demonstrate its sensitivity, specificity, and robustness. Compared to conventional statistics-driven metabolomics data interpretation, our work offers a novel automated biology-driven approach to interpreting untargeted metabolomics data, prioritizing biologically important molecules with high throughput and sensitivity.