4 ms·

The author doesn’t answer the question. If not this, then what will? Because as far as I can tell, we’re nowhere near extracting the full value of AI-generated

by aabhay 4y ago

The author doesn’t answer the question. If not this, then what will? Because as far as I can tell, we’re nowhere near extracting the full value of AI-generated protein structures. Why plant this flag and be wrong later if you have no real idea of what should be done instead?

ChrisRackauckas 4y ago
The questions he asks here are exactly the questions that quantitative systems pharmacology (QSP) seeks to answer (and as a result, it's booming as a field). Just because you can build a drug to inactivate said protein doesn't mean you should. 85% of clinical trials fail as he states, and one of the main reasons why is because the target ends up being incorrect. Targeting some protein because a lot of it seems to exist when a given disease is occurring might end up targeting the symptom instead of the cause. Understanding how the complex systems interact, their feedbacks and their nonlinearities, is essential to knowing what needs to be targeted. We had already been able to quickly create new drug candidates, and with protein folding predictions we can now do that even faster. Those drugs can be tested in a lab to see if they bind to the proteins they're supposed to, and they keep getting quicker at hitting exactly the function they expected. But without making the billion dollar clinical trial more likely to be solving the actual problem, we're still going to be limited by "okay, so what in this pool of possible drugs should we risk trying next"? We can accurately knock out protein function, but we're still fishing in the dark when it comes to how to actually fix and regulate bodies.
echelon 4y ago
Because we have to be honest with ourselves. Don't tell the crystallographers they're no longer necessary for structure determination. If people flee the field and ML doesn't pan out, then we're worse off. Treat this as it is. An exciting approach that may help some now and yield fantastic results in the future. Don't count the chickens before they hatch. Even if the structures were entirely correct - and they're definitely not - there's a massive complex metabolome to figure out. Google is certainly milking the PR as much as they can, and that can be dangerous to the laymen approving research budgets.
- tigershark 4y ago
  Alphafold demonstrated beyond any reasonable doubt that crystallography by itself is useless in certain circumstances. There are plenty of research groups working on crystallography that found the correct solution only combining their data with Alphafold data. In the last competition, if I remember correctly, there was one protein that escaped crystallography for many years until they used Alphafold predicted structure. I’m not really sure how can you simply discount these really groundbreaking results when crystallography provided much less wins in many more years.
  - l33tman 4y ago
    You are aware of that almost all known protein structures come from crystallography?
  - pas 4y ago
    How much of the AlphaFold training data is from crystallography results?
  - evouga 4y ago
    The crystallographically-determined structures are the ground truth! [1] Saying that AlphaFold makes x-ray crystallography useless is like saying DALL-E makes photography useless or Copilot makes GitHub useless. You've got the dependency chain backwards. [1] (Or at least, they're treated as the ground truth---they don't necessarily predict the conformation of proteins in solution, but that's a separate topic for another thread).
freemint 4y ago
Near real time (max 100 times slower then real time) differentiable, stochastic multi organ simulations with chemically accurate time and environment depending dynamic structure changes at all possible binding targets or interactions with body own components and third party drugs. Without machine learning at every atom is dynamic precision we are at 10^-18 L (liters) at 20 micro seconds a week with a specialised super computer https://dl.acm.org/doi/abs/10.1145/3458817.3487397 https://dl.acm.org/doi/abs/10.1145/3458817.3487397 . A solution does not need that precision everywhere. However a machine learning proxy of such precision in every relevant environment including 2d surface along non mixing fluid etc for every likely type of interaction is required so we can be certain of the possible outcomes. That would allow humanity to pre-screen a bunch of edge conditions and check for unintended or previously explained side effects. The derived surrogates for environment dependent reaction rates could be used in a spatially distributed event based simulations with level of precision ranging from atoms with position and electrons in orbits subject to electro-magnetic force interaction, molecules as things with position and rotation and folding state, concentration gradients of those as stochastic 3d PDEs, 2d PDEs, 1d PDEs and ODEs of the number of moles with relevant boundary conditions. If we had those reaction rates down and knew of all the proteins and other structures i am positive that a proxy model of relevant parts of the human body could achieve enough accuracy to be practical at pre-screening drugs with todays super computers.
ramraj07 4y ago
A model system that’s expendable (meaning you can repeat experiments in hundreds or thousands), cheap, and perfectly represents a human organ and also its immune system. Currently mice are used. But mice aren’t human! This didn’t matter too much when the drugs were simple and affected proteins that don’t change much in mammals but now we’re generally worried about diseases very specific to humans. Do you know we have many types of humanized mice? There are now very successful organoids research teams making great progress. These are far more exciting for drug discovery than getting the whole proteome structure catalogued. But that doesn’t involve computers so it’s not as exciting for a large population of technical people who want to help. Unless of course you think you can truly model the entire human body pharmacologically using AI or whatever. That’ll help. https://www.jax.org/news-and-insights/jax-blog/2020/july/why-humanized-mice https://www.jax.org/news-and-insights/jax-blog/2020/july/why...
sytelus 4y ago
Ultimate goal is to create molecules with desired properties. For example, you might say that I want material that is as hard as turtle shell. Or molecules that makes up transparent eye material. So first question is how do you go from property -> protein shape -> amino acid sequence. This is reverse of what we have now, i.e., amino acid sequence -> protein shape. However, it’s step in right direction.

by aabhay 4y ago

ChrisRackauckas 4y ago

The questions he asks here are exactly the questions that quantitative systems pharmacology (QSP) seeks to answer (and as a result, it's booming as a field). Just because you can build a drug to inactivate said protein doesn't mean you should. 85% of clinical trials fail as he states, and one of the main reasons why is because the target ends up being incorrect. Targeting some protein because a lot of it seems to exist when a given disease is occurring might end up targeting the symptom instead of the cause. Understanding how the complex systems interact, their feedbacks and their nonlinearities, is essential to knowing what needs to be targeted. We had already been able to quickly create new drug candidates, and with protein folding predictions we can now do that even faster. Those drugs can be tested in a lab to see if they bind to the proteins they're supposed to, and they keep getting quicker at hitting exactly the function they expected. But without making the billion dollar clinical trial more likely to be solving the actual problem, we're still going to be limited by "okay, so what in this pool of possible drugs should we risk trying next"? We can accurately knock out protein function, but we're still fishing in the dark when it comes to how to actually fix and regulate bodies.

echelon 4y ago

Because we have to be honest with ourselves. Don't tell the crystallographers they're no longer necessary for structure determination. If people flee the field and ML doesn't pan out, then we're worse off. Treat this as it is. An exciting approach that may help some now and yield fantastic results in the future. Don't count the chickens before they hatch. Even if the structures were entirely correct - and they're definitely not - there's a massive complex metabolome to figure out. Google is certainly milking the PR as much as they can, and that can be dangerous to the laymen approving research budgets.

tigershark 4y ago
Alphafold demonstrated beyond any reasonable doubt that crystallography by itself is useless in certain circumstances. There are plenty of research groups working on crystallography that found the correct solution only combining their data with Alphafold data. In the last competition, if I remember correctly, there was one protein that escaped crystallography for many years until they used Alphafold predicted structure. I’m not really sure how can you simply discount these really groundbreaking results when crystallography provided much less wins in many more years.
- l33tman 4y ago
  You are aware of that almost all known protein structures come from crystallography?
- pas 4y ago
  How much of the AlphaFold training data is from crystallography results?
- evouga 4y ago
  The crystallographically-determined structures are the ground truth! [1] Saying that AlphaFold makes x-ray crystallography useless is like saying DALL-E makes photography useless or Copilot makes GitHub useless. You've got the dependency chain backwards. [1] (Or at least, they're treated as the ground truth---they don't necessarily predict the conformation of proteins in solution, but that's a separate topic for another thread).

freemint 4y ago

Near real time (max 100 times slower then real time) differentiable, stochastic multi organ simulations with chemically accurate time and environment depending dynamic structure changes at all possible binding targets or interactions with body own components and third party drugs. Without machine learning at every atom is dynamic precision we are at 10^-18 L (liters) at 20 micro seconds a week with a specialised super computer https://dl.acm.org/doi/abs/10.1145/3458817.3487397 https://dl.acm.org/doi/abs/10.1145/3458817.3487397 . A solution does not need that precision everywhere. However a machine learning proxy of such precision in every relevant environment including 2d surface along non mixing fluid etc for every likely type of interaction is required so we can be certain of the possible outcomes. That would allow humanity to pre-screen a bunch of edge conditions and check for unintended or previously explained side effects. The derived surrogates for environment dependent reaction rates could be used in a spatially distributed event based simulations with level of precision ranging from atoms with position and electrons in orbits subject to electro-magnetic force interaction, molecules as things with position and rotation and folding state, concentration gradients of those as stochastic 3d PDEs, 2d PDEs, 1d PDEs and ODEs of the number of moles with relevant boundary conditions. If we had those reaction rates down and knew of all the proteins and other structures i am positive that a proxy model of relevant parts of the human body could achieve enough accuracy to be practical at pre-screening drugs with todays super computers.

ramraj07 4y ago

A model system that’s expendable (meaning you can repeat experiments in hundreds or thousands), cheap, and perfectly represents a human organ and also its immune system. Currently mice are used. But mice aren’t human! This didn’t matter too much when the drugs were simple and affected proteins that don’t change much in mammals but now we’re generally worried about diseases very specific to humans. Do you know we have many types of humanized mice? There are now very successful organoids research teams making great progress. These are far more exciting for drug discovery than getting the whole proteome structure catalogued. But that doesn’t involve computers so it’s not as exciting for a large population of technical people who want to help. Unless of course you think you can truly model the entire human body pharmacologically using AI or whatever. That’ll help. https://www.jax.org/news-and-insights/jax-blog/2020/july/why-humanized-mice https://www.jax.org/news-and-insights/jax-blog/2020/july/why...

sytelus 4y ago

Ultimate goal is to create molecules with desired properties. For example, you might say that I want material that is as hard as turtle shell. Or molecules that makes up transparent eye material. So first question is how do you go from property -> protein shape -> amino acid sequence. This is reverse of what we have now, i.e., amino acid sequence -> protein shape. However, it’s step in right direction.