The number one thing that most biologists need (that most scientists need) is a basic understanding of statistics when doing experimental design. Sure, you can get this by having a computational biologist on staff, always timidly raising their voice to get you to add replicates, for fuck's sake. But - how about learning this stuff yourself, so your experiments (which rely deeply on statistical reasoning) aren't shit to begin with?
My goal as a computational biologist isn't to do work for or correct biologists who don't understand statistics, it's to get biologists tools to do this kind of work easily on their own and bridge the gap in understanding. At the end of the day, computational biologists can't have the important domain-specific insights; biologists need to start understanding the datasets better themselves so they can apply their domain knowledge to it, and computational biologists need to be making tools to help them get there. Interdisciplinary work succeeds best by constructing interdisciplinarians who have all of the skills.