To discover frequently occurring DNA patterns related to inherent diseases or gene regulation associated diseases, we must clarify which sequences interact with transcription factors in genome. A probabilistic neural network model was introduced to represent variable length DNA sequence patterns. This model, combined with an EM algorithm, was used to discover conserved sequence patterns from some DNA sequences, and was successfully tested on two datasets, one containing simulated sequences and the other containing upstream sequences of genes in E.coli. Both fixed length and variable length patterns were discovered from the two datasets. The sensitivity of this method was higher than two compared methods, and regulatory sequences of genes were discovered from real DNA sequences of gene clusters. This method could also be used for discovering patterns of protein sequences.