This thesis covers the issue of classification of proteins into protein families. With the help of the description of amino acids by numerical vectors, the aim is to improve the method of iterative proteome search. By studying the distribution of proteins in the vector space, we try to observe the geometric structure between them. After stating the mathematical and biological concepts necessary to understand this paper and introducing the data structure on which the analysis is performed, an algorithm that finds a sphere in the vector space that groups proteins, which are biologically proved to belong to a particular protein family, is presented. The procedure was performed on proteomes of thale cress, potato and tomato. Algorithm performance was measured on both small and large data sets. F1-score was the primary metric used for model success. Results obtained show that the algorithm improves iterative search - it significantly increases the percentage of biologically correct proteins in the response, by removing the wrong ones. The model proved to be robust to changes in the data size, with extremely fast execution time.