深度神经网络筛选蛋白质组学高置信度定量肽段

A Method for Analyzing DIA-NN Output Peptides Based on Squeeze-and-Excitation Neural Network

摘要: 质谱分析是蛋白质组学的重要研究方法。数据不依赖获取是一种稳定且复现性高的质谱仪数据采集方式，具有质荷比范围宽广，通量高等特点。DIA-NN是处理DIA蛋白质组学数据的主流定量软件之一。由于DIA-NN分析DIA数据后输出的肽段中存在低置信度肽段，生物学家需要根据肽段碎片离子色谱峰组图(XICs)的相似性来人工筛选出高置信度肽段。人工筛选的任务量大、耗时长，并且筛选标准因人而异，这导致结果具有主观性。本文提出了一种名为MSDeepFilter的算法，它基于深度学习技术，能够自动筛选出高置信度的肽段。MSDeepFilter算法结合压缩激励神经网络和残差网络设计深度学习模型，从XICs中提取特征，以此区分高置信度和低置信度肽段。与传统机器学习模型Adaboosting和支持向量机模型相比，MSDeepFilter模型在基准数据集上的多项分类性能指标均表现更优，测试集AUC值达到了98.7%。这表明MSDeepFilter具有优秀性能，可以替代人工筛选的环节。

Abstract: Mass spectrometry is an important analytical method of proteomics. Data-Independent acquisition (DIA) is a stable and highly reproducible data acquisition method of mass spectrometer, which has the characteristics of wide range of mass to charge ratio and high throughput. DIA-NN is one of the mainstream quantitative software based on deep learning in the field of DIA proteomics data pro-cessing. The output of DIA-NN analysis of DIA data contains low confidence peptides, so biologists need to manually filter out high confidence peptides based on the similarity of peptide fragment ion chromatogram peak profiles (XICs). The task of manual filter is time-consuming, and the filter crite-ria vary from person to person, leading to subjective results. In this work, we propose an algorithm MSDeepFilter that can automatically filter out high-confidence peptides based on deep learning. The algorithm extracts the features of XICs by a deep learning model designed based on Squeeze-and-Excitation Networks and Residual networks as a way to distinguish high confidence peptides from low confidence peptides. Compared with the traditional machine learning models Adaboosting and Support Vector Machine models, the MSDeepFilter model performs better in sev-eral classification performance metrics on the benchmark dataset, with a test set AUC value of 98.7%. This indicates that MSDeepFilter has excellent performance and can replace the manual fil-tering process.