what kind of improvement is it most reasonable to expect within our lifetime?

Perfluorocarbons

It's hard to say anything about the potential developments within decades from now without sounding like a lunatic, so lets keep it at the obvious for now. The near term effects of computational protein folding will be 1) structure-based drug design has a wealth of data to work on now, so a lot of drug candidates for novel targets will be designed, though the experimental pipeline still remains the main bottleneck in the drug discovery process for now, but organ-on-a-chip and similar technologies are challenging that, and 2) protein design will be accelerated significantly and with it synthetic biology. Given that proteins are the main machinery of life and most drugs in current use try to target specific proteins to intervene in disease mechanisms, the potential for what can be built and cured is enormous. Beyond biology, the most exciting aspect could be that advanced AI will increasingly be able to tackle scientific questions that seemed unsolvable for long time, or might even be totally out of human reach. I think mathematics and physics are good candidates for major events, such as AI systems becoming able to deduce relevant, original theorems on their own in exceptional speed, or discovering new relationships between observables that humans did not recognize at all. Also, control of various physical processes will improve significantly, where a clear application that's actively being worked on is plasma control in nuclear fusion. The more speculative take would be about AI becoming an increasingly capable general-purpose problem solver, which can transfer abstract knowledge learned in certain domains well to other domains, which can reliably generate new knowledge by advanced machine reasoning, and which at some point might design better algorithms for doing so, effectively entering a recursive self-improvement loop of unknown limitations. So within our lifetimes, a system that can read the entire scientific literature and inspect petabytes of harvested experimental data to come up with new hypotheses to test doesn't sound too outlandish to me, which would have unprecedented effects on science and technology.