5 ms·

This is an interesting observation: > With AI, they could visualize the three-dimensional structure of the PHGDH protein. Within that structure, they discovere

This is an interesting observation: > With AI, they could visualize the three-dimensional structure of the PHGDH protein. Within that structure, they discovered that the protein has a substructure that is very similar to a known DNA-binding domain in a class of known transcription factors. The similarity is solely in the structure and not in the protein sequence.> Reminds me of: if you come across a dataset you have no idea of what it is representing, graph it.

kylehotchkiss 1y ago
Pardon my poor bio education but couldn’t this same outcome have been reached if the protein was xray crystallographyed?
- falcor84 1y ago
  Are you arguing that an in-silico result could have alternatively been achieved in-vitro? Yes, I suppose it could, but it sounds like that joke "A month in the laboratory can often save an hour in the library".
- colingauvin 1y ago
  I worked for a while on extremeophile Archaeal viruses - the type that infect organisms that manage to live in volcanic hot springs, for instance. These are ecological niches that are old, and extremely divergent. There's little genetic exchange between life around the hot springs, and life within them. The typical route of discovering those viruses was first genetic. When you get a genome (especially back when this work was initiated), you'd BLAST all the gene sequences against all known organisms to look for homologs. That's how you'd annotate what the gene does. Much more often than not, you'd get back zero results - these genes had absolutely no sequence similarity to anything else known. My PI would go through and clone every gene of the virus into bacteria to express the protein. If the protein was soluble, we'd crystallize it. And basically every time, once the structure was solved, if you did a 3D search (using Dali Server or PDBe Fold), there would be a number of near identical hits. In other words, these genes had diverged entirely at the sequence level, but without changing anything at the structural (and thus functional) level. Presumably, if AlphaFold is finding the relationship, there's some information preserved at the sequence level - but that could potentially be indirect, such as co-evolution. Either way, it's finding things no human-guided algorithm has been able to find.
  - im3w1l 1y ago
    What about convergent evolution? Are you ruling that out because you reason that there are many possible structures that could do the same job so it's too much of a coincidence how close it matches?
  - Centigonal 1y ago
    > Presumably, if AlphaFold is finding the relationship, there's some information preserved at the sequence level This is not my area of expertise, and maybe I'm misunderstanding this, but I thought that what AlphaFold does is extrapolate a structure from the sequence. The actual relationship with the other existing proteins would have been found by the investigators through other, more traditional means (like the 3D search you mentioned).
    - IX-103 1y ago
      I'm not sure about that. The way AlphaFold works involves transforming the protein from a vector space representing the sequence to a different vector space representing the folded structure and back again as it performs iterative refinement. Presumably you could perform a comparison in the structure space to find homologs that have completely different sequences - they would just have a high cosine similarity. Checking sub-regions of the structure would be more difficult, but depending on how the structural representation works it could just be computationally intensive.
    - colingauvin 1y ago
      This is a very big misconception about AlphaFold. It's not generating a structure totally de novo from sequencing. Instead it's primarily finding relationships on the sequence level to other solved structures. If those structure/sequence relationships didn't exist somewhere, AF wouldn't work because it doesn't really have much information about protein folding from first principles. There are some small de novo elements, but nothing really groundbreaking. Where AF's true strength lies is in it's ability to detect relationships we have been unable to detect with any other method.
      - 1y ago
- voxic11 1y ago
  Yes but that is a big if. xray crystallography is very hard and expensive and its not even always possible to create crystals of proteins.
davidrupp 1y ago
Whenever I see the term "AI" or similar, I mentally substitute the phrase "a lot of math, done very quickly", which is more concrete, and typically helps me sort out the stuff that still seems plausible, as in the sentence you quoted.
- JensRantil 1y ago
  I always read "ML", because that's what 98% of all AI really is, but rebranded.
  - davidrupp 1y ago
    I do the same substitution for “ML”. :-)

kylehotchkiss 1y ago

Pardon my poor bio education but couldn’t this same outcome have been reached if the protein was xray crystallographyed?

falcor84 1y ago
Are you arguing that an in-silico result could have alternatively been achieved in-vitro? Yes, I suppose it could, but it sounds like that joke "A month in the laboratory can often save an hour in the library".
colingauvin 1y ago
I worked for a while on extremeophile Archaeal viruses - the type that infect organisms that manage to live in volcanic hot springs, for instance. These are ecological niches that are old, and extremely divergent. There's little genetic exchange between life around the hot springs, and life within them. The typical route of discovering those viruses was first genetic. When you get a genome (especially back when this work was initiated), you'd BLAST all the gene sequences against all known organisms to look for homologs. That's how you'd annotate what the gene does. Much more often than not, you'd get back zero results - these genes had absolutely no sequence similarity to anything else known. My PI would go through and clone every gene of the virus into bacteria to express the protein. If the protein was soluble, we'd crystallize it. And basically every time, once the structure was solved, if you did a 3D search (using Dali Server or PDBe Fold), there would be a number of near identical hits. In other words, these genes had diverged entirely at the sequence level, but without changing anything at the structural (and thus functional) level. Presumably, if AlphaFold is finding the relationship, there's some information preserved at the sequence level - but that could potentially be indirect, such as co-evolution. Either way, it's finding things no human-guided algorithm has been able to find.
- im3w1l 1y ago
  What about convergent evolution? Are you ruling that out because you reason that there are many possible structures that could do the same job so it's too much of a coincidence how close it matches?
- Centigonal 1y ago
  > Presumably, if AlphaFold is finding the relationship, there's some information preserved at the sequence level This is not my area of expertise, and maybe I'm misunderstanding this, but I thought that what AlphaFold does is extrapolate a structure from the sequence. The actual relationship with the other existing proteins would have been found by the investigators through other, more traditional means (like the 3D search you mentioned).
  - IX-103 1y ago
    I'm not sure about that. The way AlphaFold works involves transforming the protein from a vector space representing the sequence to a different vector space representing the folded structure and back again as it performs iterative refinement. Presumably you could perform a comparison in the structure space to find homologs that have completely different sequences - they would just have a high cosine similarity. Checking sub-regions of the structure would be more difficult, but depending on how the structural representation works it could just be computationally intensive.
  - colingauvin 1y ago
    This is a very big misconception about AlphaFold. It's not generating a structure totally de novo from sequencing. Instead it's primarily finding relationships on the sequence level to other solved structures. If those structure/sequence relationships didn't exist somewhere, AF wouldn't work because it doesn't really have much information about protein folding from first principles. There are some small de novo elements, but nothing really groundbreaking. Where AF's true strength lies is in it's ability to detect relationships we have been unable to detect with any other method.
    - 1y ago
voxic11 1y ago
Yes but that is a big if. xray crystallography is very hard and expensive and its not even always possible to create crystals of proteins.

davidrupp 1y ago

Whenever I see the term "AI" or similar, I mentally substitute the phrase "a lot of math, done very quickly", which is more concrete, and typically helps me sort out the stuff that still seems plausible, as in the sentence you quoted.

JensRantil 1y ago
I always read "ML", because that's what 98% of all AI really is, but rebranded.
- davidrupp 1y ago
  I do the same substitution for “ML”. :-)