That's been made many times, but it may not be as useful as it sounds. You might learn something about what special characters you use more than the average person, but outside that, generic datasets are good enough. If you have a reasonable variety of thumb keys (instead of a single giant spacebar), you are probably better off putting your symbols on another layer anyway. Keyboard layout improvement is relatively logarithmic: you can only optimize so well for the set of permutations of keys you will type. Eventually, the variety of permutations becomes a ceiling for optimization. Any recent intentionally optimized layout will be drastically better than qwerty, marginally better than Dvorak, and roughly comparable to Colemak.
--
I've been using a Model 01 (the original release of the 100) for a few years now. It's to the point where several switches are wearing out from regular use. I happen to enjoy nearly every design decision, but am most particularly fond of the thumb clusters. I plan to eventually build a dactyl-like from scratch, with very similar thumb clusters. My favorite feature by far is the thumb-knuckle/palm key.
When I switched to the Model 01, I also learned the Workman layout. Changing both together helped isolate the change of muscle memory to its own additive context. Right now, while I'm away from home, I'm typing on a traditional board (a generic laptop) configured with qwerty. I would rather be at home typing on my Model 01, but I can live without it: the traditional qwerty experience is no worse than it was before.