5 ms·

Then we are in the same club. I've published research on global optimization algorithms for protein folding. I'm not a biochemist, but I understood that protein

Then we are in the same club. I've published research on global optimization algorithms for protein folding. I'm not a biochemist, but I understood that protein folding actually has immediate application in understanding disease and drug design. As a computer scientist, I also understood that compiler optimization is a mature field with most of the low-hanging fruit already picked. So, I guess I'm confused and will ask respectfully what problems in compiler optimization make it a thousand times more useful than protein folding and associated medical problems?

timr 18y ago
"I'm not a biochemist, but I understood that protein folding actually has immediate application in understanding disease and drug design....what problems in compiler optimization make it a thousand times more useful than protein folding and associated medical problems?" Short answer: Compilers are used for real work, every day. Nobody is using protein structure prediction for anything practical, and they likely won't be for decades more. At this point, it's blue-sky research. Long answer: "Immediate application" is one of those bits of academic-speak that really means "is related to", but sounds better to grant review boards. While it's true that protein folding is important (after all, most biological processes are mediated by folded proteins), it's not true that protein structure prediction is important. It would be great if we could predict protein structures accurately, but we can't, and until we can, it's not a practically useful discipline. Even the very best, crystallographically determined protein structures are barely sufficient to do rational drug design, and predicted structures don't come close to that level of quality. For example: we can sometimes (very rarely) predict very small (<150 residue) protein structures to within 1 angstrom RMSD of their experimentally determined shapes (i.e. >2 angstrom resolution, in the best case). However, the interactions important to drug binding, protein design, etc., don't start until a tenth of that (scales of ~0.1 angstrom). Throw in the fact that the vast majority of proteins are much larger than 150 angstroms, and that we keep creating cheaper, faster, more automated ways of getting actual experimental information on structure, and the role of protein structure prediction looks increasingly marginalized. It's definitely a cool, fun problem -- just not a very practical one. For whatever it's worth, my first papers were on applying the state-of-the-art method (you've heard of it...I think you're paraphrasing the lab's PR) for protein structure prediction to genome annotation. To call the approach useful was/is a stretch, and that's for a much easier application than drug design (in fact, we were trying to find a practical application for protein structure prediction, and it was the most likely thing we could think of!)
- aswanson 18y ago
  Wow. This is really enlightening; I thought that the computational method was going to open a new phase in disease treatment, but you seem to say here that the empirical method is on its way to making it useless. So the Pande group at Stanford is wasting their time. Interesting.
  - timr 18y ago
    "So the Pande group at Stanford is wasting their time." I wouldn't go quite that far. The research is definitely speculative, but lots of interesting things can come from speculative research. My point is that you don't do research into protein structure prediction with the intent of finding anything useful. It's basic science. We can (and occasionally do) learn things from computer models of proteins. But the PR in this field has been seriously exaggerating the results of a few of the more prominent researchers. We're a long way from curing diseases or designing drugs with this stuff.
    - aswanson 18y ago
      Is it the weakness of the modeling or the lack of computational horsepower that limits the research in this area? And would you mind linking to your papers?
      - timr 18y ago
        That's a matter of debate. Some people think that the problem is search limited, others think that the current models are bad. In my opinion, the bulk of the evidence supports the latter conclusion. Backchannel me, and I'll be happy to provide you with references to the papers I wrote/helped write. Most of them aren't open access, unfortunately.
        aswanson 18y ago
        Will do. And thanks for the info, you may have spared me a decade or so of wasted effort.

timr 18y ago

"I'm not a biochemist, but I understood that protein folding actually has immediate application in understanding disease and drug design....what problems in compiler optimization make it a thousand times more useful than protein folding and associated medical problems?" Short answer: Compilers are used for real work, every day. Nobody is using protein structure prediction for anything practical, and they likely won't be for decades more. At this point, it's blue-sky research. Long answer: "Immediate application" is one of those bits of academic-speak that really means "is related to", but sounds better to grant review boards. While it's true that protein folding is important (after all, most biological processes are mediated by folded proteins), it's not true that protein structure prediction is important. It would be great if we could predict protein structures accurately, but we can't, and until we can, it's not a practically useful discipline. Even the very best, crystallographically determined protein structures are barely sufficient to do rational drug design, and predicted structures don't come close to that level of quality. For example: we can sometimes (very rarely) predict very small (<150 residue) protein structures to within 1 angstrom RMSD of their experimentally determined shapes (i.e. >2 angstrom resolution, in the best case). However, the interactions important to drug binding, protein design, etc., don't start until a tenth of that (scales of ~0.1 angstrom). Throw in the fact that the vast majority of proteins are much larger than 150 angstroms, and that we keep creating cheaper, faster, more automated ways of getting actual experimental information on structure, and the role of protein structure prediction looks increasingly marginalized. It's definitely a cool, fun problem -- just not a very practical one. For whatever it's worth, my first papers were on applying the state-of-the-art method (you've heard of it...I think you're paraphrasing the lab's PR) for protein structure prediction to genome annotation. To call the approach useful was/is a stretch, and that's for a much easier application than drug design (in fact, we were trying to find a practical application for protein structure prediction, and it was the most likely thing we could think of!)

aswanson 18y ago
Wow. This is really enlightening; I thought that the computational method was going to open a new phase in disease treatment, but you seem to say here that the empirical method is on its way to making it useless. So the Pande group at Stanford is wasting their time. Interesting.
- timr 18y ago
  "So the Pande group at Stanford is wasting their time." I wouldn't go quite that far. The research is definitely speculative, but lots of interesting things can come from speculative research. My point is that you don't do research into protein structure prediction with the intent of finding anything useful. It's basic science. We can (and occasionally do) learn things from computer models of proteins. But the PR in this field has been seriously exaggerating the results of a few of the more prominent researchers. We're a long way from curing diseases or designing drugs with this stuff.
  - aswanson 18y ago
    Is it the weakness of the modeling or the lack of computational horsepower that limits the research in this area? And would you mind linking to your papers?
    - timr 18y ago
      That's a matter of debate. Some people think that the problem is search limited, others think that the current models are bad. In my opinion, the bulk of the evidence supports the latter conclusion. Backchannel me, and I'll be happy to provide you with references to the papers I wrote/helped write. Most of them aren't open access, unfortunately.
      - aswanson 18y ago
        Will do. And thanks for the info, you may have spared me a decade or so of wasted effort.