【摘要】 为应对质谱图在化合物结构多样性和实验环境差异下所表现出的异质性问题，提出一种提升质谱图间可比性的表征方法。该方法命名为SpecEmbedding，融合正弦嵌入与监督对比学习策略，旨在将高维、复杂的质谱图转化为低维向量表示。在GNPS公共数据集上对该方法进行训练与评估，并将其与主流方法进行对比。实验结果显示，SpecEmbedding在测试集上Top-1命中率指标上达到84.38%，相较目前最优方法CLERMS提高6.3%。该方法能显著增强质谱图间的可比性，有效提升化合物鉴定任务中的准确性与鲁棒性。更多还原

【Abstract】 To address the heterogeneity of mass spectra caused by the structural diversity of compounds and variations in experimental conditions, a novel representation method named SpecEmbedding is proposed to enhance the comparability between mass spectra. SpecEmbedding integrates sinusoidal embedding and supervised contrastive learning strategy, aiming to transform high-dimensional and complex mass spectra into low-dimensional vector representations. This approach is trained and evaluated on the public GNPS dataset, with comparion performed against mainstream methods. Experimental results show that SpecEmbedding achieves a Top-1 hit rate of 84.38% on the test set, representing a 6.3% improvement over CLERMS, the current state-of-the-art method. These findings demonstrate that SpecEmbedding significantly improves the comparability between mass spectra while effectively enhancing accuracy and robustness of compound identification tasks.更多还原