Probabilistic-Guided Dynamic Fusion Multitask (PDFM) Framework for Mass Spectrometry Classification.

Area of Science:

• Biomedical data analysis
• Computational biology
• Machine learning for mass spectrometry

Background:

• Traditional deep learning models struggle with mass spectrometry (MS) data due to deterministic features and fixed fusion.
• Limitations include single-task optimization, hindering accurate MS data classification.

Purpose of the Study:

• To propose PDFM, a progressive deep learning framework for enhanced MS data classification.
• To improve accuracy, robustness, and identification of rare categories in MS datasets.

Main Methods:

• PDFM framework implemented using a Variational Autoencoder (VAE) and Transformer with Dynamic Weights (TDW) architecture.
• VAE guides peak attention; dynamic weights fuse global (VAE) and local (Transformer) features.
• Multiobjective loss function integrates classification, reconstruction, and distribution alignment for enhanced robustness.

Main Results:

• Achieved a 4.73% accuracy gain on batch-effect-free datasets.
• Demonstrated 3.49%-4.66% cross-batch improvement in classification accuracy.
• Showcased up to a 44.07% F1-score boost for rare categories in small sample sizes.

Conclusions:

• PDFM offers a novel and precise approach for mass spectrometry data analysis.
• The framework shows significant potential for advancing translational applications in biomedicine and clinical diagnostics.