Abstract
Improving annotation accuracy, coverage, speed and depth of lipid profiles remains a significant challenge in traditional lipid annotation. We introduce LipidIN, an advanced framework designed for flash platform-independent annotation. LipidIN features a 168.5-million lipid fragmentation hierarchical library that encompasses all potential chain compositions and carbon-carbon double bond locations. The expeditious querying module achieves speeds exceeding one hundred billion queries per second across all mass spectral libraries. The lipid categories intelligence model is developed using three relative retention time rules, reducing false positive annotations and predicting unannotated lipids with a 5.7% estimated false discovery rate, covering 8923 lipids cross various species. More importantly, LipidIN integrates a Wide-spectrum Modeling Yield network for regenerating lipid fragment fingerprints to further improve accuracy and coverage with a 20% estimated recall boosting. We further demonstrate the utility of LipidIN in multiple tasks for lipid annotation and biomarker discovery in clinical cohorts.