MS-FGNet: An Attention-Guided Deep Learning Framework for Fine-Grained Pathological Subtype Prediction from DIA...

Proteomics plays a vital role in precision medicine, enabling biomarker discovery and pathological subtype prediction of complex diseases. Among analytical platforms, data-independent acquisition mass spectrometry (DIA-MS) offers high reproducibility and comprehensive proteomic coverage for rapid disease diagnosis. However, it is challenged by redundant signals and high similarity between different subtypes, making data interpretation more complicated. To address these challenges, we propose MS-FGNet, an attention-guided deep learning framework for fine-grained pathological subtype prediction based on DIA-MS pseudo imaging. Built upon the state-of-the-art DINOv3-pretrained ConvNeXt backbone, MS-FGNet introduces two innovations: (i) an attention-guided recursive feature encoding (ARE) strategy that refines DINOv3-derived representations to progressively locate disease-related spectral regions from noisy DIA-MS images, and capture multiscale feature representation through a global-to-local attention mechanism, (ii) a multibranch bilinear fusion (MBF) strategy to aggregate multibranch features and construct high-order, fine-grained fusion representations, thereby addressing the challenge of high inter-subtype similarity and enabling accurate pathological subtype prediction. When applied to 2006 human formalin-fixed paraffin-embedded samples, including follicular adenoma, multinodular goiter, follicular thyroid cancer, papillary thyroid cancer, and healthy control tissues, MS-FGNet accurately classified thyroid nodules into different pathological subtypes, achieving a mean accuracy of 90.01 ± 0.81% and a macro-averaged F1 score of 89.60 ± 0.46%. Importantly, identified spectral regions by the model effectively avoid redundant signals and exhibit variability among different phenotypes, enhancing clinical interpretability. Experimental results confirm that MS-FGNet not only performs well on thyroid DIA data but also generalizes effectively to kidney DDA data, demonstrating its potential adaptability across different data acquisition strategies.