Mass Spectrometry Prediction Based on Few-Shot Neural Network and Chemical Prior Knowledge for Regulatory Screening of Cosmetic Glucocorticoid

Abstract

Glucocorticoids are prohibited in cosmetics due to health risks, yet illegal additions persist through structural modifications that evade conventional screening. Current detection methods face critical limitations, including scarce reference standards, restricted coverage, and high training costs, presenting a global regulatory challenge. Here, we integrate message-passing neural networks with high-resolution mass spectrometry to develop GCsMSPredictor, an AI-driven screening platform trained on only 107 reference compounds and chemical prior knowledge. Using scaffold matching, we curated 32420 potential glucocorticoid analogues from PubChem and generated a predicted mass spectra library for comprehensive screening. This approach outperforms generalized models such as FIORA in accuracy while requiring substantially fewer training data and computational resources. Application to 300 cosmetic samples detected six known glucocorticoids and identified five novel structural variants that were not covered by existing standards. This framework enables cost-effective surveillance of prohibited substances in consumer products and provides a scalable paradigm for screening other compound classes that share common scaffolds, advancing intelligent chemical safety monitoring.