When I was a postdoc many years ago, we had a patient who was missing an exon in a tumor suppressor gene that coded for the domain that interacted with a kinase to regulate it. My PI asked if I could translate the mRNA sequence into amino acids and then "use some software tool" to get the 3D structure for the protein. I said, "Um, that doesn't exist and if it did, we would need a supercomputer to run it." The best I could do was highlight the deletion using an existing crystal structure to show what was missing.
I feel like I'm kind of a smart person. I feel like I kind of understand protein folding. I feel like I kind of understand AI. That was completely Greek to me.
At a high level it just showed the AI a bunch of before and afters, so the RNA before and the folded protein it produced after. The AI had to draw conclusions on the rules surrounded the folding in N Dimensional space. Then they had it go back and try to figure out what the protein looks like with only the RNA and the accuracy of the prediction was based on how closely it matched the actual protein. If there were errors it just changed the weights on the neural network or added layers or whatever needed to happen until the decision surface for protein folding had been reached. Now that they know the rules encapsulated in the neural net of the AI they can fold any RNA sequence into a protein with a reasonable degree of accuracy. Hopefully that was more clear on how it works, I of course left a ton of the technical details out but they are only relevant if you intend on using AlphaFold2.
I think the narrator glossed over the notion of 'residues' a bit. The way I understand it, the large protein molecule, which is very complex to analyze, is chopped into smaller pieces. Like a puzzle that gets pulled apart. The properties of these smaller pieces, the residues, are easier to analyze, and this information is fed into the system. Then, step by step, the system tries to find which piece would match with which other piece. Once the puzzle is complete, a 3D result is hypothesized, tested and optimized. If it meets the test criteria, the system has produced a result. Some please correct me if I have this wrong.
According to my opinion that’s kind of analysis couldn’t replace the old Spectrogotometric analysis of entire protein structure of a given protein, as we should understand the in vivo function and molecular interactions. Unfortunately the human and generally living cells don’t have so strict rules to follow before inserting a mutation into their structure. The only rule is the evolutionistic survival rule and the changing structure on life proteins require time and errors. If we cannot understand this natural process we could find and define abnormal a protein sequence that is present and functioning in Nature, but not in our computer.
That’s the diffference between Homo sapiens and Homo fabbricans . That’s why it was builded Titanic , it was so big and so perfect , but an iceberg was enough to destroy the ship 😇
So, if I'm understanding this correctly, the pairwise relationships are intercorrelated within a hyperspace in a multi-dimensional system and is then reduced to a scalar across seven delirious elves who wrote the last seasons of Game of Thrones?
Some proteins can be very hard-to-impossible to image nicely with methods like X-ray crystallography or cryoelectron microscopy (cryo-EM). Smaller proteins are easier to fold computationally (less computer memory used and fewer possible structures to rule out), while larger proteins might be easier to see with imaging. Some methods also require crystallizing many copies of proteins together, which can be difficult. Stuff like this can limit the resolution you can image up to. If you get stuck imaging at 5-10 Angstrom resolution, you might only get 3D electron densities that are very blobby/low resolution, which don't tell you a ton about the protein's structure. Nice experimental structures are the gold standard (they have more biological/realistic info), but if the experiments aren't working out, then computer-generated structures can be used. Good experimental structure > good computed structure > bad experimental structure.
When I was a postdoc many years ago, we had a patient who was missing an exon in a tumor suppressor gene that coded for the domain that interacted with a kinase to regulate it. My PI asked if I could translate the mRNA sequence into amino acids and then "use some software tool" to get the 3D structure for the protein. I said, "Um, that doesn't exist and if it did, we would need a supercomputer to run it." The best I could do was highlight the deletion using an existing crystal structure to show what was missing.
There are gaps in this explanation of the structure.
I feel like I'm kind of a smart person. I feel like I kind of understand protein folding. I feel like I kind of understand AI. That was completely Greek to me.
same but this video took me like 25 minutes bc i had to skip back to make notes or try to understand it somehow
At a high level it just showed the AI a bunch of before and afters, so the RNA before and the folded protein it produced after. The AI had to draw conclusions on the rules surrounded the folding in N Dimensional space. Then they had it go back and try to figure out what the protein looks like with only the RNA and the accuracy of the prediction was based on how closely it matched the actual protein. If there were errors it just changed the weights on the neural network or added layers or whatever needed to happen until the decision surface for protein folding had been reached. Now that they know the rules encapsulated in the neural net of the AI they can fold any RNA sequence into a protein with a reasonable degree of accuracy. Hopefully that was more clear on how it works, I of course left a ton of the technical details out but they are only relevant if you intend on using AlphaFold2.
I think the narrator glossed over the notion of 'residues' a bit. The way I understand it, the large protein molecule, which is very complex to analyze, is chopped into smaller pieces. Like a puzzle that gets pulled apart. The properties of these smaller pieces, the residues, are easier to analyze, and this information is fed into the system. Then, step by step, the system tries to find which piece would match with which other piece. Once the puzzle is complete, a 3D result is hypothesized, tested and optimized. If it meets the test criteria, the system has produced a result.
Some please correct me if I have this wrong.
According to my opinion that’s kind of analysis couldn’t replace the old Spectrogotometric analysis of entire protein structure of a given protein, as we should understand the in vivo function and molecular interactions. Unfortunately the human and generally living cells don’t have so strict rules to follow before inserting a mutation into their structure. The only rule is the evolutionistic survival rule and the changing structure on life proteins require time and errors. If we cannot understand this natural process we could find and define abnormal a protein sequence that is present and functioning in Nature, but not in our computer.
That’s the diffference between Homo sapiens and Homo fabbricans . That’s why it was builded Titanic , it was so big and so perfect , but an iceberg was enough to destroy the ship 😇
So, if I'm understanding this correctly, the pairwise relationships are intercorrelated within a hyperspace in a multi-dimensional system and is then reduced to a scalar across seven delirious elves who wrote the last seasons of Game of Thrones?
Technical Jargon Overload
You know I don't speak Spanish... in English, please!
"almost always" ???
Some proteins can be very hard-to-impossible to image nicely with methods like X-ray crystallography or cryoelectron microscopy (cryo-EM). Smaller proteins are easier to fold computationally (less computer memory used and fewer possible structures to rule out), while larger proteins might be easier to see with imaging. Some methods also require crystallizing many copies of proteins together, which can be difficult. Stuff like this can limit the resolution you can image up to. If you get stuck imaging at 5-10 Angstrom resolution, you might only get 3D electron densities that are very blobby/low resolution, which don't tell you a ton about the protein's structure.
Nice experimental structures are the gold standard (they have more biological/realistic info), but if the experiments aren't working out, then computer-generated structures can be used.
Good experimental structure > good computed structure > bad experimental structure.
Thanks 😂❤
Didn’t understand sorry
i try this with price d ata on assets
Ok so all this for what? Again please!
I thought of fentanyl
❤
Even if I am a scientist, you have badly explained how it works, you should consider that a YT channel is not a scientific conference between pairs.
Hmmmm
jj
Not a great explanation
Any reco for a good explanation please?
🐇