GAG-Former: An LC-MS/MS data-processing workflow with a physics-informed dual-stream transformer backbone for co-elution deconvolution and simultaneous structural identification of glycosaminoglycan disaccharides in biological samples.

Glycosaminoglycan (GAG) disaccharide profiling by LC-MS/MS underpins heparin quality control and glycomics biomarker research. Existing workflows are constrained by co-elution of sulfation-positional isomers and C5 epimers, disconnection between structural identification and peak-area quantification, and matrix-driven transfer loss between standards and biological samples. This study develops GAG-Former, an LC-MS/MS data-processing workflow built on a physics-informed dual-stream Transformer backbone that encodes fragment spectra and chromatographic retention times in parallel. The sulfation biosynthetic order and heparin-lyase cleavage preference are embedded as non-learnable attention biases. A set-prediction triplet decoder outputs disaccharide class, retention window and concentration in a single forward pass, and a gradient-reversal layer with LayerNorm-only test-time entropy minimisation enables cross-matrix self-calibration. The workflow was evaluated on 17 GAG disaccharide standards (510 injections), 32 Latin-square gravimetric mixtures (256 injections), three biological matrices (plasma n=120, urine n=98, mouse liver/kidney/aorta n=144), an OSCS adulteration gradient (210 injections) and four animal-source heparins (96 injections). On Std-17 the workflow reached 89.7% Top-1 identification accuracy and 9.6% MAPE in concentration on the Latin-square mixtures under 5-fold cross-validation, outperforming CandyCrunch, GlycoBERT, QuanFormer and commercial automatic integration; the co-elution-pair RMSE dropped from 0.231 (QuanFormer) to 0.142, and the F1 on rare 3-O-sulfated disaccharides rose from 0.713 (bias-disabled control) to 0.864. Zero-shot identification accuracy across the three biological matrices was 86.3% (plasma), 84.1% (urine) and 80.7% (tissue). OSCS was stably detected at 0.05% (28/30 replicates, P = 0.92) and animal-source discrimination reached 95.8% accuracy. All findings were obtained in a single laboratory on one LC-MS/MS platform using anonymised archival or commercial samples; multicenter validation, blinded external testing, cross-laboratory robustness assessment and head-to-head comparison with pharmacopoeial 1H NMR, SAX-HPLC and MRM methods have not been performed and remain prerequisites for any regulatory or clinical use. The workflow is therefore positioned as a high-throughput pre-screening approach that complements rather than replaces 1H NMR, SAX-HPLC and MRM confirmation.