Post AbMvnSX0RI5RP26C36 by PathIntegral@social.quodverum.com
(DIR) More posts by PathIntegral@social.quodverum.com
(DIR) Post #AbMewtwcBmWvoAzg6i by darulharb@social.quodverum.com
2023-11-01T13:53:00Z
0 likes, 0 repeats
Having a bit of future shock here...@hughhewitt: "From 'News Items' (news-items.com) from @EllisItems comes an announcement from Google Deep Mind about what is arguably the most important development in synthetic biology yet: [screenshot]"#Google #AI #biologyhttps://nitter.poast.org/hughhewitt/status/1719654539891011931(H/T @RealSaavedra RT )
(DIR) Post #AbMfH2E96U5pHwnIhM by darulharb@social.quodverum.com
2023-11-01T13:56:38Z
0 likes, 0 repeats
https://deepmind.google/discover/blog/a-glimpse-of-the-next-generation-of-alphafold/
(DIR) Post #AbMfs7cm02omk4zvlY by darulharb@social.quodverum.com
2023-11-01T14:03:20Z
0 likes, 0 repeats
Besides the implications for pharma and basic molecular biology research, this capability could also advance biologically-based molecular nanotechnology. The Google blog already mentions the use of enzymes which break down plastic pollution.
(DIR) Post #AbMgkUO7a55Syp4KRM by darulharb@social.quodverum.com
2023-11-01T14:13:09Z
0 likes, 0 repeats
Current AI systems are tremendously good at pattern recognition, on all sorts of data, so it's not surprising that given large databases of protein structures, they're able to predict protein folding for new structures.Solving the "protein folding problem" has been a challenge for molecular biology for decades. While the amino acid sequence of a protein could be determined from a given gene sequence, figuring out how the protein folded into its functional three-dimensional structure was [...]
(DIR) Post #AbMhYsukfPPlWgMV1M by darulharb@social.quodverum.com
2023-11-01T14:22:16Z
0 likes, 0 repeats
[...] really tough, because the protein molecules are so complex, with so many atoms. Imagine a chain necklace with tiny magnets on parts of the chain. How many different unique snarls can the necklace get into? In the case of the protein molecule, different parts of the amino acid sequence are attracted or repelled by each other, but it reliably self-assembles into a functional shape.
(DIR) Post #AbMjusTnAEOCSNMmjQ by darulharb@social.quodverum.com
2023-11-01T14:48:39Z
0 likes, 0 repeats
Individual protein molecules are, however, usually too small to see much detail, even with an electron microscope. Here, for example, is an electron micrograph of an antibody taken about 10 years ago. Antibodies are proteins that are part of the immune system which stick to foreign substances to neutralize them, or tag them for destruction by immune system cells (the Y-shaped structures sticking out from the mica surface are the antibody molecules).
(DIR) Post #AbMkEgkwSdlPtApox6 by PathIntegral@social.quodverum.com
2023-11-01T14:52:13Z
0 likes, 0 repeats
@darulharb This is probably the most insidious of the machine learning propaganda stories currently blasting around the internet.Experimental results on protein structure require crystallizing the protein, which forces it into an artificial conformation which typically deviates from that it would have naturally. Thus, the primary problem is not the first prediction, but the refinement of the structure, something which was recognized long before AlphaFold came along.
(DIR) Post #AbMkRITy1OpuMyuJ3w by darulharb@social.quodverum.com
2023-11-01T14:54:31Z
0 likes, 0 repeats
The scale bar in the image (10 nanometers), is about the size of the smallest features in a computer chip, or ~1/720th the diameter of a red blood cell.So you can't figure out a protein's structure just by "looking" it, even with a microscope.
(DIR) Post #AbMkTfyXe5dboNLIcy by darulharb@social.quodverum.com
2023-11-01T14:54:56Z
0 likes, 0 repeats
@PathIntegral I'm getting to that part...
(DIR) Post #AbMkdPwI7giwyIw1ya by darulharb@social.quodverum.com
2023-11-01T14:56:42Z
0 likes, 0 repeats
@PathIntegral I'm getting to that part...But a good point that x-ray crystallography has its own problems and artifacts.
(DIR) Post #AbMlonVgvWydDPU2DI by darulharb@social.quodverum.com
2023-11-01T15:09:57Z
0 likes, 0 repeats
Google is claiming to have expanded their predictive AI model in the latest iteration to include other molecules that the proteins would interact with.The "gold standard" for determining protein structure, and the basis for the database used to train the AI, has required isolating the protein, and making a crystal out of it, which can then be hit with X-rays, a process called x-ray crystallography.
(DIR) Post #AbMm2xRSnVHs4YG3WK by darulharb@social.quodverum.com
2023-11-01T15:12:31Z
0 likes, 0 repeats
The way the x-rays are scattered by the crystal tells scientists about the positions of the atoms making up the crystal, and thus the protein the crystal is made from. Getting certain proteins purified and then to form crystals is a "painstaking process" (I.e. a PITA). Not to mention all the calculations involved in figuring out the positions of the atoms from looking at the x-ray scattering data.
(DIR) Post #AbMmp6oNGUFT20w09Y by darulharb@social.quodverum.com
2023-11-01T15:21:14Z
0 likes, 0 repeats
Over decades, scientists have built this hard-won structural database, which is derived from the x-ray crystallography data, and in recent years, IT guys have been training AI systems on databases like these to see if they can figure out things that humans, who can't keep track of all these parameters without help, have overlooked.
(DIR) Post #AbMmsAHIpggmeSvGz2 by darulharb@social.quodverum.com
2023-11-01T15:21:46Z
0 likes, 0 repeats
This is still early days, but the potential for what can be discovered using these AI tools is very exciting (and as with other powerful technologies, a bit frightening as well).--end--Soon to be polished and enhanced for my Substack, I expect...
(DIR) Post #AbMnAlSVKnk3B6201I by Timmerax@social.quodverum.com
2023-11-01T15:25:08Z
0 likes, 0 repeats
@darulharbExcellent thread, lots to digest.I am excited and frightened at the same time.Many discoveries to come.
(DIR) Post #AbMvnSX0RI5RP26C36 by PathIntegral@social.quodverum.com
2023-11-01T17:01:43Z
0 likes, 0 repeats
@darulharb The other unmentioned bottleneck is quantum mechanics.The drug-protein interaction, aside from being dynamic in nature (that is, one must understand the atomic dynamics of how and where the molecule may bind, in addition to how the solvent reorganizes), is also a very complex electronic interaction.A vast literature attempts to glom this into various statistical models, but the colossal complexity renders it intractable to machine learning techniques.
(DIR) Post #AbjXQdJyECVpwSf9UW by darulharb@social.quodverum.com
2023-11-12T14:47:05Z
0 likes, 0 repeats
And here it is...https://darulharb.substack.com/p/parlez-vous-proteines