In nearly every case, it comes down to what the GPU is good at. Architecturally modern GPUs consist of a massive number of small cores. They are as such good at embarassingly parallel problems and happen to be good at vector calculations (as opposed to floating point, where CPUs often do better). Rather than using a GPU as well, a GPU, all these tasks use it as a GPGPU
It so happens that protein folding, signal analysis, and bitcoin mining are embarassingly parallel, and distribute well amongst many nodes, which are what folding@home, seti and bitcoin do!
Also, bitcoin mining involves solving equations and that somehow has a random chance of you getting a bitcoin, which is an arbitrary, cyrptographically secure, non legal tender currency - see this closed question for more details