With the development of mass spectrometry technologies and bioinformatics analysis algorithms, disease research-driven human proteome project (HPP) is advancing rapidly. Protein biomarkers play critical roles in clinical applications and the biomarker discovery strategies and methods have become one of research hotspots. Feature selection and machine learning methods have good effects on solving the "dimensionality" and "sparsity" problems of proteomics data, which have been widely used in the discovery of protein biomarkers. Here, we systematically review the strategy of protein biomarker discovery and the frequently-used machine learning methods. Also, the review illustrates the prospects and limitations of deep learning in this field. It is aimed at providing a valuable reference for corresponding researchers.