This paper presents an interdisciplinary investigation of statistical information retrieval (IR) techniques for protein identification from tandem mass spectra, a challenging problem in proteomic data analysis. We formulate the task as an IR problem, by constructing a “query vector” whose elements are system-predicted peptides with confidence scores based on spectrum analysis of the input sample, and by defining the vector space of “documents” with protein profiles, each of which is constructed based on the theoretical spectrum of a protein. This formulation establishes a new connection from the protein identification problem to a probabilistic language modeling approach as well as the vector space models in IR, and enables us to compare fundamental differences in the IR models and common approaches in protein identification. Our experiments on benchmark spectrometry query sets and large protein databases demonstrate that the IR models significantly outperform well-established methods in protein identification, by enhancing precision in high-recall regions in particular, where the conventional approaches are weak.