The impact of different negative training data on regulatory sequence predictions