We present YASS -- a new tool for computing local similarities in DNA sequences. Similar to many existent algorithms (e.g. BLAST), YASS first searches for small patterns (``seeds'') shared by input sequences and then extends some of those (``hits'') into larger alignments. New features of YASS include, on the one hand, a new transition-constrained seed model and, on the other hand, a flexible group criterion of forming hits out of individual seeds. Computer experiments confirm that the program achieves a good sensitivity/selectivity trade-off compared to existing programs, which makes of YASS an efficient tool for comparing long DNA sequences, such as eukaryotic chromosomes.