I definetely have something different in mind from Ockham's and PhilC's exploding methods, but I don't want to distract you from MetaBalls :)

Wish I could say I was a great scripter/programmer or mathematician, as I would certainly help if I could in that area.  If I'm understanding correctly, I see what you mean about the speed when not using a compiled extension.  Hmm. 

If you can get around the speed problem and IF you want go for animation/dynamics , it might be better to start off by looking at some 2D pixel based particle systems, I think many of the principals still apply to 3D if you're using spheres. 
  Let me know if any of my ideas are getting too far fetched, but for animation in a particle system, think of a sphere as just one of those pixels/points, where as the point is the center of the sphere. You could probably generate spheres from scratch based on that center point with varying size options, etc.  So when you go to animate, instead animating geometry directly, you're just animating those points(dummy objects?).  Generating the spheres and isosurface for them would still present speed issues though I suppose.

Forgive me, I realize this is probably much more complicated than what you had in mind, but if you pulled it off, I'd pay for it for sure