MSpectraAI

MSpectraAI analyzes raw LC-MS2-based proteomics and metabolomics data using deep neural networks to extract, classify, and visualize spectral-feature swaths for proteome profiling and molecular characterization.

Key Features:

• Deep Neural Network Integration: Utilizes deep neural networks (DNNs) for spectral-feature swath extraction, classification, and visualization beyond traditional protein identification methods.
• Raw Data Mining and Classification: Performs mining and classification of raw LC-MS2-based proteomics and metabolomics datasets.
• User-Customizable Models: Supports construction of deep neural network models implemented with the Keras library.
• Comprehensive Data Handling: Processes large-scale proteomics datasets, demonstrated on ProteomeXchange data from six tumor types comprising 7,997,805 mass spectra.
• High Predictive Accuracy: Achieved an average prediction accuracy of 0.967 for classifying samples based on fingerprint spectrum profiles at the MS1 level, outperforming classical machine learning approaches.

Scientific Applications:

• Proteome Profiling: Deciphers proteome profiles directly from raw LC-MS2 proteomics data.
• Metabolomics Analysis: Applies deep learning-based spectral extraction and classification to LC-MS2 metabolomics datasets.
• Cancer Multi-Tumor Classification: Enables classification and prediction of proteomic information across multi-tumor samples, supporting cancer research and molecular characterization studies.

Methodology:

Implements deep neural networks (DNNs) built with Keras for spectral-feature swath extraction, classification, and visualization from raw LC-MS2 data; evaluated at the MS1 level on a ProteomeXchange proteomics dataset of six tumor types comprising 7,997,805 mass spectra with comparisons to classical machine learning approaches.