Data-independent acquisition (DIA) proteomics allows systematic and unbiased measurement of protein samples and enables fast quantitative analysis of large cohorts of samples. However, sample-specific spectral libraries are usually required prior to perform DIA experiments. The libraries are built by data-dependent acquisition (DDA) proteomic analysis on the same samples normally with pre-fractionation, which is time-consuming and limits the identification/quantification by DIA to the peptides identified by DDA. Herein, we propose DeepDIA, a deep learning based approach to generate in silico spectral libraries for DIA analysis. We benchmarked DeepDIA with HeLa and mixed proteome data sets, showing that the quality of in silico spectral libraries is comparable to that of experimental spectral libraries. We further demonstrated that DeepDIA can be performed on human cell lines and human serum without pre-knowledge of peptides lists by DDA experiments. Compared to the state-of-the-art protocol using DDA-based spectral library with high abundance protein depletion and pre-fractionation, the number of identified and quantified proteins from human serum was increased by >100% using DeepDIA. Accuracy of the identification results was validated using a standard mixture containing >800 stable isotope labelled reference peptides from >500 proteins in human plasma. We expect this work contributing to the studies of quantitative proteomics and especially blood proteomics, whereas expanding the toolbox for DIA proteomics.