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

Traditional deep learning models (e.g., convolutional neural networks (CNNs) and transformers) often rely on deterministic feature representations, fixed feature fusion mechanisms, and single-task optimization, which limit their effectiveness in accurately classifying mass spectrometry (MS) data. This study proposes PDFM, a progressive framework for MS data classification, implemented via the VAE-TDW architecture. VAE models latent distributions to guide peak attention, while dynamic weights adaptively fuse global (VAE) and local (Transformer) features. A reconstruction branch and adversarial samples enhance robustness. The multiobjective loss integrates classification, reconstruction, and distribution alignment. Evaluations show a 4.73% accuracy gain on six batch-effect-free data sets; 3.49%-4.66% cross-batch improvement; and up to a 44.07% F1-score boost for rare categories in small samples. PDFM represents a novel approach for the precise analysis of mass spectrometry data, demonstrating substantial potential to advance translational applications in biomedicine and clinical diagnostics.