12 ms·

There is an unfortunate lack of technical information in this article, but before the ML naysayers arrive in full force to point out that this isn't real AI or

There is an unfortunate lack of technical information in this article, but before the ML naysayers arrive in full force to point out that this isn't real AI or ML or what have you, it's quite possible they were using deep neural nets to heuristically approximate complex simulations 3-6 orders of magnitude more quickly than typical simulation. We're doing the same in other domains now.

natechols 7y ago
Simulations are not the major bottleneck in drug development. More generally, the naysayers in this case are likely to be people who actually have some knowledge of drug development, and have seen countless companies that promised to revolutionize drug discovery by using computers, most of which are no longer in existence. If an AI-powered company can suddenly start passing Phase III clinical trials at an unprecedented rate, then I'll be genuinely impressed.
ashleyn 7y ago
>We're doing the same in other domains now. One of the first examples I remember reading about this was an antenna designed by an AI at NASA[1]. This was back in the days when "AI" wasn't the buzzword so much as "genetic algorithms" was. It ended up looking like this crazy zigzag thing. 1) https://www.nasa.gov/centers/ames/research/exploringtheuniverse/borg.html https://www.nasa.gov/centers/ames/research/exploringtheunive...
jcranmer 7y ago
Drug development roughly consists of the following steps: 1. Figure out what enzyme or receptor you have to mess with to cause the disease. You can skip this step if you decide just to blast away in a phenotypic screen. 2. Design a molecule that will actually mess with the enzyme/receptor to prevent it from malfunctioning in whatever way is causing the disease. 3. Figure out how to actually synthesize said molecule. 4. Now make sure this molecule can actually get to the sites where it needs to go to hit the target... 5. ... that it lasts there long enough and in enough potency to do its job ... 6. ... that it doesn't screw anything else up along the way (or get converted by your body into something that does)... 7. ... and that there's no compensatory mechanism that still causes the disease after your drug successfully gummed up what it was supposed to. Or maybe you failed at step 1 and you're barking up the wrong tree. Roughly speaking, the last four steps will correspond to the different trials you have to run to get your drug approved. About 90% of drug candidates that make it to step 4 fail to make it all the way through step 7. The difficult things that kill drugs are understanding their toxicity profiles and other off-target effects. We suck at this because we have so little data (and the effects are quite complicated), and modern AI techniques are mostly borne on shoveling piles of data and hoping for the best--not a good match.
- natechols 7y ago
  One thing that depresses me about this entire field is that there are probably enough data points buried in all of the various Big Pharma databases for AI to come up with some truly novel and useful insights. Even though most of the data will be for drugs that failed, however, there's still zero incentive for the companies to share the data. In general I'm extremely skeptical of claims that central planning will be more efficient than capitalism, but this is one case where the competition seems actively harmful.
  - dekhn 7y ago
    I've worked with big pharma databases. Nearly all of them are inappropriate for the techniques used at large internet companies (deep networks trained to predict user behavior to maximize some objective). Pharma is slowly coming to the understanding that they could use their resources to produce high quality datasets that are amenable to machine learning, but it's a slow process and it always has to fight against the status quo.

natechols 7y ago

Simulations are not the major bottleneck in drug development. More generally, the naysayers in this case are likely to be people who actually have some knowledge of drug development, and have seen countless companies that promised to revolutionize drug discovery by using computers, most of which are no longer in existence. If an AI-powered company can suddenly start passing Phase III clinical trials at an unprecedented rate, then I'll be genuinely impressed.

ashleyn 7y ago

>We're doing the same in other domains now. One of the first examples I remember reading about this was an antenna designed by an AI at NASA[1]. This was back in the days when "AI" wasn't the buzzword so much as "genetic algorithms" was. It ended up looking like this crazy zigzag thing. 1) https://www.nasa.gov/centers/ames/research/exploringtheuniverse/borg.html https://www.nasa.gov/centers/ames/research/exploringtheunive...

jcranmer 7y ago

Drug development roughly consists of the following steps: 1. Figure out what enzyme or receptor you have to mess with to cause the disease. You can skip this step if you decide just to blast away in a phenotypic screen. 2. Design a molecule that will actually mess with the enzyme/receptor to prevent it from malfunctioning in whatever way is causing the disease. 3. Figure out how to actually synthesize said molecule. 4. Now make sure this molecule can actually get to the sites where it needs to go to hit the target... 5. ... that it lasts there long enough and in enough potency to do its job ... 6. ... that it doesn't screw anything else up along the way (or get converted by your body into something that does)... 7. ... and that there's no compensatory mechanism that still causes the disease after your drug successfully gummed up what it was supposed to. Or maybe you failed at step 1 and you're barking up the wrong tree. Roughly speaking, the last four steps will correspond to the different trials you have to run to get your drug approved. About 90% of drug candidates that make it to step 4 fail to make it all the way through step 7. The difficult things that kill drugs are understanding their toxicity profiles and other off-target effects. We suck at this because we have so little data (and the effects are quite complicated), and modern AI techniques are mostly borne on shoveling piles of data and hoping for the best--not a good match.

natechols 7y ago
One thing that depresses me about this entire field is that there are probably enough data points buried in all of the various Big Pharma databases for AI to come up with some truly novel and useful insights. Even though most of the data will be for drugs that failed, however, there's still zero incentive for the companies to share the data. In general I'm extremely skeptical of claims that central planning will be more efficient than capitalism, but this is one case where the competition seems actively harmful.
- dekhn 7y ago
  I've worked with big pharma databases. Nearly all of them are inappropriate for the techniques used at large internet companies (deep networks trained to predict user behavior to maximize some objective). Pharma is slowly coming to the understanding that they could use their resources to produce high quality datasets that are amenable to machine learning, but it's a slow process and it always has to fight against the status quo.