The interpretation and application of data-independent acquisition (DIA) MS-based proteomic data are centered on identification and quantification of proteins from complex mass spectra. This leads to loss of signals for proteins absent in a prior protein database. We developed a novel tensor-centric analysis strategy as well as a data format, namely, DIA tensor (DIAT) to analyze the complete DIA-MS map without the need for peptide precursor identification. DIAT as a tensor format minimizes the storage space and can be directly fed into a deep neural network on GPUs to predict phenotypes. This prediction framework was validated in a variety of DIA-MS schemes, including ScanningSWATH, diaPASEF and DI-SPA. We have also developed feature interpretating module to identify peptides from a selected region of the DIAT file. DIAT represents a novel analytics for DIA-based proteomic big data.