Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Alphafold

github.com

Post reply on HN
Print view View original

111–120 of 170 posts

Re: Alphafold

#111

Post by perl4ever »

gopalv wrote: post #107
nikhilsimha wrote: post #41

Earlier quoted context omitted.

The implicit assumption you are making is that the demand increases in lock step with productivity gains. 100x faster drug discovery, 100x more drugs need to be discovered => same number of people employed. These correlations do hold for technical fields, but logically there should be a point beyond which productivity gains outpace, demand growth / demand could even stop growing. One should either retool to solve a n…

> The implicit assumption you are making is that the demand increases in lock step with productivity gains. This is basically the theory around modern industrial revolution - https://en.wikipedia.org/wiki/Jevons_paradox Efficiency increases demand more than the efficiency saves. You could argue that labour does not follow that, but it is more about the technology improving rather than disrupting.

>Efficiency increases demand more than the efficiency saves.

That's not some sort of natural law.

People latched on to this hypothetical situation and keep elevating its universality.

It's just a thing that can happen and may or may not in a given instance.

Re: Alphafold

#112

Post by perl4ever »

sbierwagen wrote: post #97
nikhilsimha wrote: post #41

Earlier quoted context omitted.

The implicit assumption you are making is that the demand increases in lock step with productivity gains. 100x faster drug discovery, 100x more drugs need to be discovered => same number of people employed. These correlations do hold for technical fields, but logically there should be a point beyond which productivity gains outpace, demand growth / demand could even stop growing. One should either retool to solve a n…

>but logically there should be a point beyond which productivity gains outpace The limiting factor on drug approval is clinical trials. Once every living person is enrolled in a clinical trial, we will have hit the maximum rate at which humanity can produce new drugs. That might be more than 10x the current rate, but probably less than 1000x.

The number of people needed for a trial depends on the size of the effect.

I think we have a warped idea of that, because of the practice of looking for a barely "statistically significant" effect from a substance that isn't really understood.

If you have something that just eradicates a disease immediately, because you really understand what you are doing, you don't need very many tests to confirm it works.

Re: Alphafold

#113

Post by lukeplato »

nextos wrote: post #16

Alphafold 2 is very very cool, but we need a little dose of reality. It's still a bit away from really solving protein folding as it was marketed. For example, multi-complex proteins are not well predicted yet and these are really important in many biological processes and drug design: https://occamstypewriter.org/scurry/2020/12/02/no-deepmind-h... A disturbing thing is that the architecture is much less novel than I…

I haven't read the full paper but there is certainly some new/exciting developments i'm seeing just from scanning. The “Invariant Point Attention" which is described as a novel, geometry-aware and equivariant attention operation is pretty huge.

Something along these lines was speculated to be to be used by Fabian Fuchs [0] soon after the original CASP competition. Basically, it's a huge win for the geometric deep learning people, and indicates an exciting direction for mainstream academia to move in.

https://fabianfuchsml.github.io/alphafold2/

Re: Alphafold

#114

Post by Duller-Finite »

slownews45 wrote: post #68
nick238 wrote: post #36

Earlier quoted context omitted.

I haven't worked on the drug-side of things, but here my bio perspective: It's kind of out-of-vogue, but consider the "lock and key" model of proteins and small molecules (drugs). For drug design, what you want to do is get a key that fits just one lock (to pull whatever lever) and not others (to avoid side-effects). It's relatively easy to find a molecule that fits a protein, because that protein is what you might s…

First seems reasonable. I've not heard of anything on the later coming even close credibly - though is an obvious holy grail.

David Baker's group (author of the RoseTTAFold paper out today in science) has multiple exciting examples of de novo design of proteins.

For example, see [1] or [2], and [2] was spun off into a company (Neoleukin Therapeutics).

[1] https://science.sciencemag.org/content/371/6531/eabc8182 [2] https://www.nature.com/articles/s41586-018-0830-7

Re: Alphafold

#115

Post by jefftk »

sbierwagen wrote: post #96
nextos wrote: post #16

Alphafold 2 is very very cool, but we need a little dose of reality. It's still a bit away from really solving protein folding as it was marketed. For example, multi-complex proteins are not well predicted yet and these are really important in many biological processes and drug design: https://occamstypewriter.org/scurry/2020/12/02/no-deepmind-h... A disturbing thing is that the architecture is much less novel than I…

>A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. A similar concern has sparked some worries about "AI overhang" https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-... Most of the compute in ML research seems to be g…

> Once Waymo can replace some human drivers some of the time, things will start changing very quickly.

But that happened 1y+ ago [1][2] without much changing since?

[1] https://www.theverge.com/2019/12/9/21000085/waymo-fully-driv...

[2] https://blog.waymo.com/2020/10/waymo-is-opening-its-fully-dr...

Re: Alphafold

#116

Post by dzhiurgis »

AnimalMuppet wrote: post #73
COGlory wrote: post #24

I am a structural biologist. This is one of the handful of topics that overlaps with my field here. I'm very excited to play with this, although it might eventually put me out of a job.

Here's where I think we need to be going: You go to a doctor's office, sick. 1) They take a blood sample. 2) They find the malignant bacteria and DNA sequence it. 3) If it's a known strain, they know what antibiotics to use on it. 4) If not, they solve protein folding on the genes. 5) From that, they see which existing antibiotics would kill it. 6) If none will, then given the proteins, they have to derive a new anti…

Re step 1 and 2 - here's interesting podcast on how they detect rare infections: https://www.youtube.com/watch?v=MzzD2F73iGU

Basically sequence everything what's in your blood and look for what doesn't match your genome === infection. The problem is this is orders of magnitude more compute intense than whole genome sequencing. Basically increased demand for sequencing far outmatches available compute!

Re: Alphafold

#117

Post by neolog »

nl wrote: post #110
sbierwagen wrote: post #96

Earlier quoted context omitted.

>A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. A similar concern has sparked some worries about "AI overhang" https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-... Most of the compute in ML research seems to be g…

> Most of the compute in ML research seems to be going into architecture search. No it's not. Only Google spends significant time with automatic architecture search, and many people think this is really to try to sell cloud capacity. > Once the architecture is found, training and net finetuning/transfer learning is comparatively cheap Training isn't cheap for significant problems. Getting the data is very expensive,…

parent's EfficientNet graph seems really dramatic. Is it misleading somehow?

Re: Alphafold

#118

Post by ehsankia »

devindotcom wrote: post #76

Also announced today was RoseTTAFold from UW's Baker Lab, which claims nearly the same accuracy at much higher efficiencies. There's a public server and paper in Science. More info here and here: https://www.bakerlab.org/index.php/2021/07/15/accurate-prote... https://techcrunch.com/2021/07/15/researchers-match-deepmind...

Could it be that AlphaFold 2 was open sourced in response to this?

Re: Alphafold

#119

Post by exikyut »

dekhn wrote: post #95
astro-codes wrote: post #87

Earlier quoted context omitted.

Why wouldn't this have much value to pharma? Is it because its application is actually really limited in scope?

there are research groups this would be useful for but structures are not on the critical path to drug discovery or approval.

Out of (probably overoptimistic :) ) curiosity, what do you see are the critical paths?

Re: Alphafold

#120

Post by Tenoke »

nl wrote: post #110
sbierwagen wrote: post #96

Earlier quoted context omitted.

>A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. A similar concern has sparked some worries about "AI overhang" https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-... Most of the compute in ML research seems to be g…

> Most of the compute in ML research seems to be going into architecture search. No it's not. Only Google spends significant time with automatic architecture search, and many people think this is really to try to sell cloud capacity. > Once the architecture is found, training and net finetuning/transfer learning is comparatively cheap Training isn't cheap for significant problems. Getting the data is very expensive,…

>No it's not. Only Google spends significant time with automatic architecture search, and many people think this is really to try to sell cloud capacity.

Maybe not automatic architecture search but a lot does go into testing different architectures and changes to them in a more manual manner. Though yes, those tests are run on a smaller scale so for those huge models, training will be a bigger portion compared to architecture search than for smaller models.

Post reply on HN