Not solved, no. Definitely much much better than before. The difference AlphaFold made is significant: we’re talking about getting a decent model in a couple minutes using a PC compared to several months of calculations before.
However we still need experimental data: in many occasions AlphaFold gives an incorrect model. With some experimental data that model can be improved, but we still have no reliable way to know what the structure of a protein is starting from the amino acidic sequence without extensive experimentation.
That’s the big promise of quantum computers, there are however two major problems in my opinion:
There’s still no theoretical framework which explains how once we have a quantum computer we may tackle protein folding
Plenty quantum computing companies closed shortly after AlphaFold was published since they lost all funding because protein folding was “solved”
Agreed, I still do hope they can maintain some of their promises. However until now, I have not really seen any real advances towards making something useful.
I do not have a deep knowledge of quantum computers, but I know plenty people working on them and often get to talk about it.
I know people working on chemical problems who are basically approximating atoms to point charges. And either way those calculations are slower than on a CPU. For the uninformed, in chemistry the interactions between electronic orbitals is fundamental; this is in no way an approximation useful to obtain any kind of information.
This is fine, I understand methodologies take time to develop; however as far as I understand it those techniques they’re using are mathematically limited to using point charges: no matter how much they improve them that’ll be the highest level of accuracy.
I hope someone finds a way to handle such things better: as much as you can make a great machine learning model you’re always depending on available data.
That’s really intereting. Thank you for the informative comment.
I once did a internship where we solved protein structures by using Cristalisation and X-ray fracturing. Is this what you’re referring to with extensive research or has the methodology advanced?
Yes, X-ray is the gold standard. Technology has advanced in the sense that the protein crystallization is now more standardized and automated, as well as the analysis of the results.
It is not the only technique, for example there are cheaper ones based on mass spectrometry which do not resolve the full structure but allow to understand which amino acids are spatially near; such information is useful when developing a protein model and to validate whether a model is plausible.
The other two major techniques for structure resolution are NMR spectra analysis and the fairly novel technique of cryo electro microscopy.
These in general do not resolve the protein structure to the same resolution as X-ray but have other advantages: they allow you to observe the protein structure when in solution, which may be significantly different from the crystallized structure.
I doubt there is a comparable correctness metric between LLMs and protein structure prediction models.
You can measure how many times they correctly predict a thing, but results will greatly change according to what your objective is. Those are only comparable when you’re trying to predict the same thing.
As such my reply would be: sometimes more incorrect sometimes more correct. However, in general, a mishandled incorrect protein structure prediction is way more expensive than an LLM hallucination.
I mean… machine learning extends far beyond the stupid lying machines, and can produce far more trustworthy results than the dipshits at “Open” AI, but results will still always require validation, as all results require a lot of validation in science before they should be trusted.
Its somewhat complex. Protein structures are usually solved with x-ray crystallography. You get a 2-dimensional diffraction pattern. There are different methods to “solve” the diffraction pattern and get the Protein structure. Most are a lot of work, including biochemical lab work. But if you have a model it is relatively easy. The modern computer generated models have revolutionized protein x-ray crystallography.
The interesting part would be to a reliable and computationally accessible way to handle disordered proteins. Seeing how they can move could be quite revolutionary.
I had to work on some disordered proteins and you’re pretty much just guessing, plausibly you’re better off going to a casino blindfolded and play blackjack.
Wait I thought protein folding was “solved”? Is it just significantly improved or was I lied to?
Not solved, no. Definitely much much better than before. The difference AlphaFold made is significant: we’re talking about getting a decent model in a couple minutes using a PC compared to several months of calculations before.
However we still need experimental data: in many occasions AlphaFold gives an incorrect model. With some experimental data that model can be improved, but we still have no reliable way to know what the structure of a protein is starting from the amino acidic sequence without extensive experimentation.
That’s the big promise of quantum computers, there are however two major problems in my opinion:
90% of quantum computing is a hype scam anyway.
Agreed, I still do hope they can maintain some of their promises. However until now, I have not really seen any real advances towards making something useful.
I do not have a deep knowledge of quantum computers, but I know plenty people working on them and often get to talk about it.
I know people working on chemical problems who are basically approximating atoms to point charges. And either way those calculations are slower than on a CPU. For the uninformed, in chemistry the interactions between electronic orbitals is fundamental; this is in no way an approximation useful to obtain any kind of information.
This is fine, I understand methodologies take time to develop; however as far as I understand it those techniques they’re using are mathematically limited to using point charges: no matter how much they improve them that’ll be the highest level of accuracy.
I hope someone finds a way to handle such things better: as much as you can make a great machine learning model you’re always depending on available data.
That’s really intereting. Thank you for the informative comment.
I once did a internship where we solved protein structures by using Cristalisation and X-ray fracturing. Is this what you’re referring to with extensive research or has the methodology advanced?
Yes, X-ray is the gold standard. Technology has advanced in the sense that the protein crystallization is now more standardized and automated, as well as the analysis of the results.
It is not the only technique, for example there are cheaper ones based on mass spectrometry which do not resolve the full structure but allow to understand which amino acids are spatially near; such information is useful when developing a protein model and to validate whether a model is plausible.
The other two major techniques for structure resolution are NMR spectra analysis and the fairly novel technique of cryo electro microscopy.
These in general do not resolve the protein structure to the same resolution as X-ray but have other advantages: they allow you to observe the protein structure when in solution, which may be significantly different from the crystallized structure.
Gotcha, thank you ❤️
We traded it in for something that will hallucinate a solution.
AlphaFold is incorrect in plenty occasions.
More incorrect than a text prediction algorithm?
I doubt there is a comparable correctness metric between LLMs and protein structure prediction models.
You can measure how many times they correctly predict a thing, but results will greatly change according to what your objective is. Those are only comparable when you’re trying to predict the same thing.
As such my reply would be: sometimes more incorrect sometimes more correct. However, in general, a mishandled incorrect protein structure prediction is way more expensive than an LLM hallucination.
I mean… machine learning extends far beyond the stupid lying machines, and can produce far more trustworthy results than the dipshits at “Open” AI, but results will still always require validation, as all results require a lot of validation in science before they should be trusted.
Ah, yes I can imagine that being the case
How do you solve protein folding when there are so many different ways for proteins to arrange?
A reliable way to simulate any kind of protein fold. As long as you have a method you could hypothetically “solve” the issue
You dont solve anything. You generate models. That’s it. Without experimental validation they are just cartoons.
Then why has modelling been so focused on?
Because the models become a database you can search for certain characteristics.
(This is a guess I consider reasonable.)
Its somewhat complex. Protein structures are usually solved with x-ray crystallography. You get a 2-dimensional diffraction pattern. There are different methods to “solve” the diffraction pattern and get the Protein structure. Most are a lot of work, including biochemical lab work. But if you have a model it is relatively easy. The modern computer generated models have revolutionized protein x-ray crystallography.
There actually aren’t. However, some of the most interesting proteins have no structure at all.
The interesting part would be to a reliable and computationally accessible way to handle disordered proteins. Seeing how they can move could be quite revolutionary.
I had to work on some disordered proteins and you’re pretty much just guessing, plausibly you’re better off going to a casino blindfolded and play blackjack.