Get the top HN stories in your inbox every day.
l33tman
andy_ppp
Does the actual structure of the original protein change to "fit" into the multimer or is it just the protein is fixed and the orientation/tessellation is what is being calculated? Or Both?
cowsandmilk
One of the main comparisons in the paper is to ClusPro docking of Alphafold-generated monomers. ClusPro uses FFT-based sampling of rotations and translations of rigid monomers followed by minimization to allow the monomers to relax and adapt. This paper allows significantly more adaptation in the formation of multimers than ClusPro does.
Historically, most multimer prediction software has required well structured subunits. And the field was largely built around benchmarks that had structures of subunits [0]. Over the last decade, there has been a increased recognition that many heterodimers are made up of a well structured protein and an intrinsically disordered protein [1] (or intrinsically disordered region of another protein). This lead to a lot of efforts to predict what residues are in the interface and then generating ensembles of "peptide" structures (where peptide is <~15 residues) to model the predicted interface without having to predict larger structure.
So, to answer, the structures of the original protein change. An increasingly important area is proteins that don't have much structure as a monomer, but you'll see changes in ones that have well-structured monomers as well. If you dig into some of the difficult cases of the zlab benchmark (using pymol for example), you'll find that in many of those cases, the proteins undergo large conformational changes in forming the complex.
[0] https://zlab.umassmed.edu/benchmark/
[1] https://en.wikipedia.org/wiki/Intrinsically_disordered_prote...
gilleain
So the simple possibility is rigid docking to get a surface between proteins. As with everything in biology, things can be more complicated.
For example, from here :
https://elifesciences.org/articles/22889
it starts with:
>Many eukaryotic regulatory proteins adopt distinct bound and unbound conformations, and use this structural flexibility to bind specifically to multiple partners
In other words, binding of a protein to another can change its shape (conformation).
andy_ppp
It really does seem so incredibly unlikely that biology works!
cowsandmilk
> since they have to be (relatively) stable as a monomer before locking in place
What is your basis for this claim?
l33tman
If the monomer is completely unstable by itself, and wouldn't fold, I'm pretty sure 5 of them are not going to fold simultaneously into a multimer. So it would be a good guess that the monomers are going to fold first into something that at least looks roughly like the final product and then assemble and change configuration.
But no I haven't read any articles studying this in detail. I do recall seeing studies on the evolution of oligomers, where the monomer exists by itself but evolves to polymerize at the expense of its monomeric stability.
awestroke
DeepMind must be an incredible place to work. Fast pace, state of the art.
mach1ne
Employees have said that its amazing to work with very smart people there.
That being said, I'd be a bit worried that the ultimate owner of the AI work done there is Google.
THAT being said, Google has been very open-sourcy about their research, unlike some... ahem Open ahem AI companies
crakenzak
Its amazing having a company with the technology & innovation focus such as Google funding, pushing, and publishing AI research.
Everyone wins, and will continue to do so until the research labs at Google, Facebook, Apple etc... abandon their openness culture -- which I don't see happening anytime soon mostly because of the attitudes and culture of openness of the engineers & researches at said labs.
Great times we live in...
wiz21c
Vast and important domain of knowledge are now under IP protection. It means that society will evolve only at the speed the owner want it to.
You could argue that it is better than nothing. But this you will only know if 50 years when others will have caught up and we'll realize (or not) that things could have been much faster if stuff was more open.
Google opens up its articles, codes, but sure not its infrastructure.
Now does public funded efforts do better, or would do better if they had the financial power of Google ? That's the question. I believe so, but that's rather an idealist point of view.
rcMgD2BwE72F
>which I don't see happening anytime soon mostly because of the attitudes and culture of openness of the engineers & researches at said labs.
Why, this time, would it be different?
amelius
Google's text-to-speech and speech-to-text models are still closed source ...
tziki
Makes me wonder how much DeepMind could make just by keeping their protein folding proprietary and charging per use.
jcims
It seems like one potential endpoint of this type of research, beyond the medical applications, would be to explore the effects of synthetic amino acids that allow the incorporation of other materials into the protein output.
This could open the door to materials synthesis and other types of fabrication.
cowsandmilk
AlphaFold is heavily based on sequence alignments, which means incorporating synthetic amino acids would be extremely difficult as there is no input data for them. There would need to be a different leap forward for this use case.
mrestko
Whatever you're imaging, it's likely it could be accomplished with the standard compliment of amino acids.
Get the top HN stories in your inbox every day.
As an example, the majority of current drug targets are membrane proteins, of which most are multimers, which makes stuff like this a key part of predicting important protein complexes. Membrane proteins are also notoriously difficult to get structures of which makes it even more interesting.
BTW, I'm not sure how well AlphaFold predicts structures that are part of multimers in the first place? But since they have to be (relatively) stable as a monomer before locking in place, I guess they can be predicted without considering their final part of a complex. Still, feels like some kind of symbiosis between this and the structure prediction itself could exist.