The invention, which belongs to the technical field of biological information, discloses a protein-protein interaction prediction method based on a deep forest. According to the method, pseudo amino acid composition, a mutual information descriptor, composition, and distribution, a conversion descriptor, an amino acid composition position specificity score matrix and a dipeptide composition position specificity score matrix are fused to convert a protein sequence into a numerical vector; sequence information, physicochemical property information and evolution information of the protein pair are fused as initial characteristics of a sample; an elastic network is used for feature selection, and redundant and irrelevant features are removed; and a fused optimal feature vector is inputted intoa constructed multi-granularity cascade depth forest to predict protein-protein interaction. The method is simple and effective, the deep forest can represent the high-level feature information of the protein pair, the results of the training set and the test set are obviously superior to those of other prediction methods, and a certain reference can be provided for drug target prediction and human disease treatment.