Regularized Deep Neural Networks for Combining Heterogeneous Features of Peptides in Data Independent Acquisition Mass Spectrometry

Data-independent acquisition (DIA) has gained much attention in mass spectrometry (MS)-based proteomics for its improved reproducibility and unbiased data acquisition. In DIA-MS, the spectral library is crucial in peptide identification. However, this method is limited to peptides previously identified via data-dependent acquisition (DDA) MS experiments. This study proposes a deep learning approach for generating spectral libraries, even for previously unseen peptides. While most deep learning-based methods rely on one-hot encoding representation for peptides, the proposed method incorporates physicochemical features, including atomic composition, hydrophobicity, flexibility, fractional surface probability, and aromaticity. We introduce sparsity regu-larized neural network layers to facilitate the selection and combination of important high-dimensional physicochemical features and improve prediction performance. Fur-thermore, we suggest a transfer learning strategy for training the proposed deep neural networks having multiple heterogeneous input channels. Numerical experiments using benchmark DDA-MS data demonstrated that the proposed deep learning model out-performed existing benchmark models, such as Prosit and DeepDIA, particularly in predicting retention times. And it was demonstrated that the proposed models with sparsity regularization identified more peptides from HeLa cell DIA data compared to the other deep learning models.