6 ms·

AI and Drug Discovery: Attacking the Right Problems

101 points by mhb 6y ago

text70 6y ago
The last time I was working with molecular modeling, proteins were still the main target. With the advent of cryoEM, glycobiology of proteins can now be discerned. Now, AI is starting to be able to predict entire secondary structures from sequence alone. It will take massive computational power, but I can see how toxicology may soon be able to be predicted from in silico experiments.
- ramraj07 6y ago
  The point of this post (and what I realised during my PhD) was that for the most part this doesn’t matter; choice of targets and predicting the human efficacy of a therapy that works in a tube or mice is far more important and typically more ignored. People don’t focus on that because it’s a hard problem especially for math and computer people who think their absence is the reason cancer isn’t cured yet and come over to try and win the Nobel prize (and don’t; cue relevant xkcd).
ta988 6y ago
An advice for the VC among us. Don't get fooled by all these startups promising to revolutionalize drug-discovery with AI. Go talk to scientists that are specialists but have no stakes in those approaches. I've seen so much money diverted from projects (many that turned out to work once sanity came back) just to pay for some over-confident startupers shitty solution... Yes AI will help drug discovery no doubt about that, but those companies telling you they can learn from existing data are lying, most data available is of really low quality, acquired in conditions adequate only for specific types of questions... One I know closely will promise you a knowledge graph based on a specific slice of drug sources that can supposedly answer every question. It will almost certainly end up like many of those companies that hire humans to do the work in the back and sell it to you as AI... And their source data which I know well is of really low quality so we can't expect much more from their models either...
- ramraj07 6y ago
  A lot of times they are not lying, they are genuinely clueless. Part of the problem with academia is that they follow the edict set by George Constanza: “the best liar truly believes in their own lie.” So often they truly believe that their field’s method of choice is how we will use AIDS to cure cancer, because the grants-system ensures that anyone who doesn’t truly believe in their vomit won’t get their grant approved (you need to write them with conviction, not doubt, if you want them awarded). Then mediocre postdocs with an “entrepreneurial flair” decide they’re ready to take their nanobots and cure breast cancer and come to SV to pitch to VCs who are either clueless or are being advised by the same heads that are pouring the Kool-aid in academia to begin with! Exceptions exist of course, and some of these bets turn good purely because of luck (if you take twenty random chemicals and inject, one of them might just cure cancer; incidentally how stuff like paclitaxel & vinblastine was discovered anyway). But this only seems to reinforce the conviction of these institutions to continue their methods. One of the most important differences I feel VCs should acknowledge as a contrast between tech and biotech: with tech you can often choose one method of solving a problem and make it work. With biology however, nature will eventually give you a roundhouse kick if you try to force anything but an actual solution down it’s throat.
  - ta988 6y ago
    You are right I assigned to malice what is probably just lack of skills and knowledge. Your are also right on the fact that what works is grind and find. You can do it at the scale of one lab or the scale of many labs, but in the end what works is to try a lot of things because we are still really shooting in the dark for a lot of targets or diseases we don't know any druggable target yet.
    - ramraj07 6y ago
      I’ll propose that we actually kind of know a lot of things about many (not all) diseases, we really just need an objective look at how to address them. The majority of scientists today are mostly quite narrow minded in terms of scientific thinking, every one of them wants to use the method that they are experts at to solve the problem. When the Manhattan project was deployed all the top physicists came and tackled the problem that needed to be solved with the most suitable methods, but there’s been “cancers Manhattan project” or “moonshot” initiatives in biology numerous times and they all fail because the scientists incharge invariably push their own agendas in these. Take Janelia farms for example- without reading their editorialised website if you try to gander what the hell this institution is trying to do, from just the papers they publish, you would fail miserably. Not saying that institution sucks, well I suppose I am saying that, but only insofar because all of academia seems to suck in the same way.
rscho 6y ago
If you want to see people who understand the why and the how: https://www.uab.edu/medicine/pmi/ https://www.uab.edu/medicine/pmi/ Lead by Matt Might and Will Byrd. As an MD, I must say MediKanren is the only convincing use of medical AI I've seen.
- srckinase123 6y ago
  Quick tangential question: Do you see a benefit in incoming medical students knowing programming? If so, which specialty do you see it benefiting from the most (e.g., radiology)?
  - bionhoward 6y ago
    One hundred percent! I did this, and I'm glad I did. Programming changes your perspective. When you learn to code you learn to value abstraction. That's valuable in medicine because medicine is pathologically reductionist...
  - rscho 6y ago
    For practicing day-to-day medicine, programming is useless. However, for a medical student hoping to become an academic it's a highly important asset. Others and even more fundamental assets are to have solid knowledge of epidemiology and statistics. Essentially, if you want to stay in the academic system you'll be slaving under older academics for quite some time. Being able to hack together a dataset from csv with a bit of R in the mix will make you a godsend to those people. Be aware that this advantage comes at a high price, as those people usually understand nothing and will attempt to pressure you into doing impossible/unethical things (extreme p-hacking, etc.). Now, doing anything more than shitty statistics or code-monkey javascript is almost impossible due to other constraints. In medicine, almost every system is totally locked up. There are of course exceptions as some students will go above and beyond, but the yield in terms of work/achievement will always be poor. It's far easier to just fill in excel spreadsheets and publish lots of shitty papers. Much to my dismay. As for the sibling comment speaking of abstraction I'm sorry to disappoint, but no. Current medicine is far from being hard science. You cannot approach it like a formal system (yet). These days, docs are still far closer to gardeners than they are to programmers or any kind of engineer. That said, you can find use cases for programming in any specialty. Sometimes it will be more obvious than others: Image analysis -> radiology Statistical genomics -> genetics Dsp -> anesthesia ... But anyways, don't get your hopes up too much. Students would usually be much better served by learning to care for people. And that doesn't involve programming.
1cvmask 6y ago
AI is just a tool like Excel and not a panacea. It might create some efficiencies but it is not a short cut to the real grind and the too often overlooked serendipity involved. Serendipity plays such a big role in many “discoveries” but we don’t want to ascribe too much to it because it is just potluck by brute forcing OTHER problems. Perhaps serendipity should be rephrased as “no method to our madness”.
- randcraw 6y ago
  Unlike Excel, AI truly can deliver a competitive advantage when used judiciously. In the right situation, it can deliver that small catalytic boost that lets you see just a little farther than before. And incremental advancement is the basis for almost all scientific or commercial advancement. That said, deep CNNs and LSTMs have delivered only a fairly narrow set of catalytic advancements to date, even fewer of which have changed the status quo in big businesses, where R&D ROI really adds up. Accurately identifying the next use of AI that will meaningfully move the needle is beyond hard. But hasn't that always been true of all forms of innovation? That's why so many are keen on AI today; as hard as it is to wield, it still seems to be the sharpest arrow in our quiver.
  - dumb1224 6y ago
    I agree with you. I transited from computing to drug discovery. A lot of projects attempt to harness the power of ML but there still seem to be a lot of misunderstanding. A lot of the time the valuable ML methods are supporting / assisting hypothesis generation rather than 'doing it all'. The domain experts will judge the value of ML methods by their usual merits in the drug discovery world. If the more sophisticated methods can predict hit compounds no better than the traditional standard deviation based cut off, or if the models can not explain key insight of any indication of mechanism of action, then it may not be a good method after all. But meanwhile if it opens up a new way (e.g a new phenotype based screening) then it is worth pursuing.
hamsterbooster 6y ago
Insitro (led by Daphne Koller) just raised $400M dollar and seems to be making good progress on applying AI to drug discovery (https://insitro.com/ https://insitro.com/).
- thesausageking 6y ago
  What progress have they made (other than raising more money)?

101 points by mhb 6y ago

The last time I was working with molecular modeling, proteins were still the main target. With the advent of cryoEM, glycobiology of proteins can now be discerned. Now, AI is starting to be able to predict entire secondary structures from sequence alone. It will take massive computational power, but I can see how toxicology may soon be able to be predicted from in silico experiments.

ramraj07 6y ago
The point of this post (and what I realised during my PhD) was that for the most part this doesn’t matter; choice of targets and predicting the human efficacy of a therapy that works in a tube or mice is far more important and typically more ignored. People don’t focus on that because it’s a hard problem especially for math and computer people who think their absence is the reason cancer isn’t cured yet and come over to try and win the Nobel prize (and don’t; cue relevant xkcd).

ta988 6y ago

An advice for the VC among us. Don't get fooled by all these startups promising to revolutionalize drug-discovery with AI. Go talk to scientists that are specialists but have no stakes in those approaches. I've seen so much money diverted from projects (many that turned out to work once sanity came back) just to pay for some over-confident startupers shitty solution... Yes AI will help drug discovery no doubt about that, but those companies telling you they can learn from existing data are lying, most data available is of really low quality, acquired in conditions adequate only for specific types of questions... One I know closely will promise you a knowledge graph based on a specific slice of drug sources that can supposedly answer every question. It will almost certainly end up like many of those companies that hire humans to do the work in the back and sell it to you as AI... And their source data which I know well is of really low quality so we can't expect much more from their models either...

ramraj07 6y ago
A lot of times they are not lying, they are genuinely clueless. Part of the problem with academia is that they follow the edict set by George Constanza: “the best liar truly believes in their own lie.” So often they truly believe that their field’s method of choice is how we will use AIDS to cure cancer, because the grants-system ensures that anyone who doesn’t truly believe in their vomit won’t get their grant approved (you need to write them with conviction, not doubt, if you want them awarded). Then mediocre postdocs with an “entrepreneurial flair” decide they’re ready to take their nanobots and cure breast cancer and come to SV to pitch to VCs who are either clueless or are being advised by the same heads that are pouring the Kool-aid in academia to begin with! Exceptions exist of course, and some of these bets turn good purely because of luck (if you take twenty random chemicals and inject, one of them might just cure cancer; incidentally how stuff like paclitaxel & vinblastine was discovered anyway). But this only seems to reinforce the conviction of these institutions to continue their methods. One of the most important differences I feel VCs should acknowledge as a contrast between tech and biotech: with tech you can often choose one method of solving a problem and make it work. With biology however, nature will eventually give you a roundhouse kick if you try to force anything but an actual solution down it’s throat.
- ta988 6y ago
  You are right I assigned to malice what is probably just lack of skills and knowledge. Your are also right on the fact that what works is grind and find. You can do it at the scale of one lab or the scale of many labs, but in the end what works is to try a lot of things because we are still really shooting in the dark for a lot of targets or diseases we don't know any druggable target yet.
  - ramraj07 6y ago
    I’ll propose that we actually kind of know a lot of things about many (not all) diseases, we really just need an objective look at how to address them. The majority of scientists today are mostly quite narrow minded in terms of scientific thinking, every one of them wants to use the method that they are experts at to solve the problem. When the Manhattan project was deployed all the top physicists came and tackled the problem that needed to be solved with the most suitable methods, but there’s been “cancers Manhattan project” or “moonshot” initiatives in biology numerous times and they all fail because the scientists incharge invariably push their own agendas in these. Take Janelia farms for example- without reading their editorialised website if you try to gander what the hell this institution is trying to do, from just the papers they publish, you would fail miserably. Not saying that institution sucks, well I suppose I am saying that, but only insofar because all of academia seems to suck in the same way.

rscho 6y ago

If you want to see people who understand the why and the how: https://www.uab.edu/medicine/pmi/ https://www.uab.edu/medicine/pmi/ Lead by Matt Might and Will Byrd. As an MD, I must say MediKanren is the only convincing use of medical AI I've seen.

srckinase123 6y ago
Quick tangential question: Do you see a benefit in incoming medical students knowing programming? If so, which specialty do you see it benefiting from the most (e.g., radiology)?
- bionhoward 6y ago
  One hundred percent! I did this, and I'm glad I did. Programming changes your perspective. When you learn to code you learn to value abstraction. That's valuable in medicine because medicine is pathologically reductionist...
- rscho 6y ago
  For practicing day-to-day medicine, programming is useless. However, for a medical student hoping to become an academic it's a highly important asset. Others and even more fundamental assets are to have solid knowledge of epidemiology and statistics. Essentially, if you want to stay in the academic system you'll be slaving under older academics for quite some time. Being able to hack together a dataset from csv with a bit of R in the mix will make you a godsend to those people. Be aware that this advantage comes at a high price, as those people usually understand nothing and will attempt to pressure you into doing impossible/unethical things (extreme p-hacking, etc.). Now, doing anything more than shitty statistics or code-monkey javascript is almost impossible due to other constraints. In medicine, almost every system is totally locked up. There are of course exceptions as some students will go above and beyond, but the yield in terms of work/achievement will always be poor. It's far easier to just fill in excel spreadsheets and publish lots of shitty papers. Much to my dismay. As for the sibling comment speaking of abstraction I'm sorry to disappoint, but no. Current medicine is far from being hard science. You cannot approach it like a formal system (yet). These days, docs are still far closer to gardeners than they are to programmers or any kind of engineer. That said, you can find use cases for programming in any specialty. Sometimes it will be more obvious than others: Image analysis -> radiology Statistical genomics -> genetics Dsp -> anesthesia ... But anyways, don't get your hopes up too much. Students would usually be much better served by learning to care for people. And that doesn't involve programming.

1cvmask 6y ago

AI is just a tool like Excel and not a panacea. It might create some efficiencies but it is not a short cut to the real grind and the too often overlooked serendipity involved. Serendipity plays such a big role in many “discoveries” but we don’t want to ascribe too much to it because it is just potluck by brute forcing OTHER problems. Perhaps serendipity should be rephrased as “no method to our madness”.

randcraw 6y ago
Unlike Excel, AI truly can deliver a competitive advantage when used judiciously. In the right situation, it can deliver that small catalytic boost that lets you see just a little farther than before. And incremental advancement is the basis for almost all scientific or commercial advancement. That said, deep CNNs and LSTMs have delivered only a fairly narrow set of catalytic advancements to date, even fewer of which have changed the status quo in big businesses, where R&D ROI really adds up. Accurately identifying the next use of AI that will meaningfully move the needle is beyond hard. But hasn't that always been true of all forms of innovation? That's why so many are keen on AI today; as hard as it is to wield, it still seems to be the sharpest arrow in our quiver.
- dumb1224 6y ago
  I agree with you. I transited from computing to drug discovery. A lot of projects attempt to harness the power of ML but there still seem to be a lot of misunderstanding. A lot of the time the valuable ML methods are supporting / assisting hypothesis generation rather than 'doing it all'. The domain experts will judge the value of ML methods by their usual merits in the drug discovery world. If the more sophisticated methods can predict hit compounds no better than the traditional standard deviation based cut off, or if the models can not explain key insight of any indication of mechanism of action, then it may not be a good method after all. But meanwhile if it opens up a new way (e.g a new phenotype based screening) then it is worth pursuing.

hamsterbooster 6y ago

Insitro (led by Daphne Koller) just raised $400M dollar and seems to be making good progress on applying AI to drug discovery (https://insitro.com/ https://insitro.com/).

thesausageking 6y ago
What progress have they made (other than raising more money)?