DeepMind AlphaFold: A Gift To Humanity! 🧬
Вставка
- Опубліковано 13 тра 2024
- ❤️ Train a neural network and track your experiments with Weights & Biases here: wandb.me/paperintro
📝 The paper "Highly accurate protein structure prediction with #AlphaFold" is available here:
www.nature.com/articles/s4158...
deepmind.com/blog/article/alp...
Protein database:
alphafold.ebi.ac.uk/
More on this:
deepmind.com/blog/article/alp...
deepmind.com/research/case-st...
deepmind.com/blog/article/put...
❤️ Watch these videos in early access on our Patreon page or join us here on UA-cam:
- / twominutepapers
- / @twominutepapers
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bryan Learn, Christian Ahlin, Eric Martel, Gordon Child, Ivo Galic, Jace O'Brien, Javier Bustamante, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Paul F, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: / twominutepapers
Thumbnail background design: Felícia Zsolnai-Fehér - felicia.hu
Chapters:
00:00 What is protein folding?
01:00 Is it so hard!
04:08 CASP
05:15 AlphaFold's results
06:41 Let's look inside!
07:41 Iterative refinement
08:41 Convolutional neural networks
08:51 Transformers
09:55 Everything matters!
10:58 Adding physics knowledge
13:00 What is this good for?
13:33 A gift to humanity!
14:05 Future works
14:56 AlphaFold's weaknesses
15:16 An important message
Meet and discuss your ideas with other Fellow Scholars on the Two Minute Papers Discord: discordapp.com/invite/hbcTJu2
Károly Zsolnai-Fehér's links:
Instagram: / twominutepapers
Twitter: / twominutepapers
Web: cg.tuwien.ac.at/~zsolnai/ - Наука та технологія
The real gift to humanity is an 18 minute two minute papers video
You are too kind. This really made my day - thank you so much! 🙏
@@TwoMinutePapers brother you've made a massive impact in my life with how you deliver the important details and applications without getting too bogged down in the academics, sparking ideas about how to integrate these idea modules into ai solutions.
You make it so accessible. Instead of "only 3 people understand quantum in the world, IDK who the other two are," your delivery is more like "guys! guys! Check out what we can do now!"
Thank you so much for what you do!
One really great gift :)
@@lexscarlet That is exactly the goal of Two Minute Papers. So happy to hear that you are enjoying it, thank you so much once again!
@@TwoMinutePapers It's called TWO minute papers, not EIGHTEEN minute papers! :D Just kidding! I love the videos.
I'm a PhD student investigating Fe-S cluster biosynthesis, and AlphaFold has been an absolute treat for my work. I think people haven't truly understood how hugely important (and absolutely essential) this is for the Biochemistry field. It can also predict protein complexes, and it is much more accurate than any previous docking software I have used. Thanks for the video, it's great to see you getting everyone interested!
ok
What a time to be alive!
Tbh i still dont really get what implication this is gonna have on our future.. probably not that many but at least research will be accelerated
So youre saying alpha fold will find a bunch of protein that could extend human life indefinitely?
The reason is probably that no one has given a clear answer on how the structure of a protein helps. I don't doubt it will have amazing impacts but it seems to be very unclear what these impacts are.
As a Biophysicist I can assure that this blue bar is indeed jaw-dropping. Accurate protein folding prediction is basically the holy grail of molecular biology. It is insanely powerful!
what are the potential applications of this discovery and what benefits can regular ppl expect ?
@@hdjfjd8 E.g. prion predicting, e.g. to counter Alzheimer's. Building controlled drug delivery (routed) cages.
Hey good to see a fellow Biophysicist!
Key to the gate of the church containing the altar by which is the tabernacle containing a safe within which is the Holy Grail. Decent start, mind. :)
Have you considered the possible negative impacts? Nuclear physics was a big step, too, and it hasn't been all upside. Whether it's been a positive change, on balance, is debatable.
For those wondering. This is truly history in the making! As a molecular biologist myself currently working with aptamers, I shiver of the possibilities this will open not only for my field of research, but for medicine, in better understanding disease, and in the nucleus structures we currently know so little.
This is as ground breaking as the 2013 Nobel Prize work, won by Michael Levitt and others, which helped develop better computational models. Hold on to your papers and hats folks!
Great context, thanks 🙏
Physicists were largely pretty impressed with themselves, too, when they invented the hydrogen bomb. Have you stopped to think about how this technology could be used against The People, and how it will in fact be used be the powers that be?
I've tried to get in contact with them since they won the competition. I want them to know that this is invaluable, but as someone who have studied molecular biology myself as well, it would be even better if they focus on the ligand aspect of this. You give it some sequence you are working with, then it tells you the structure for the protein, and you just highlight an area, or related sequence, and it would generate sequences that fold into structures that would bind to the selected area. That would cause an explosion in the biological engineering field.
@@bumpty9830 With the discovery of metal working we made knives, knifes can help us do lots of things, but can be used to kill people.
With a rock we can kill people, we can kill people with hands.
Don't you dare compare this with a hydrogen bomb.
This can solve so many of our problems.
With more advancements in this field we might be able to finally cute cancer, cure dementia, comined with CRISPR, the standard of living on say a 50 years will be vastly better for future generations.
Want to fearmonger. Do that about Big Oil paying politicians to deny climate change which poses a greater risk to our collective demise than a bio weapon made with AI
Do you have a more specific problem description or a review paper with existing methods for exactly this you mention ? I can take a look I do a bit of ML
This is really big news, and a really big step, but, as you said, it is just one of a thousand step journey. Here's what's next:
Protein folding gets exponentially more complex with each additional aminoacid, meaning large proteins are more difficult to solve. Plus, the problem as it is is mostly based on Wan Der Vaals interactions, but in reality we also have covalent cross bonding between specific aminoacids, ionic interactions with metal atoms, and non-natural folding driven by exterior forces, mainly auxiliar proteins called chaperones. Moreover, the really useful part of this is protein-protein interactions and protein-drug interactions, which are out of the scope of the current algorithm. And last but not least, the solution varies depending on temperature and pH. So all in all, we still have a loooooong way to go, but still, this is an *unbelievably* big step which I cannot understate. Amazing video as always!
Considering how quickly we got to this stage, what do you think is a reasonable expectation for how long it will take us to compete those steps?
@@grimaffiliations3671 Welcome to the world of quantum computing !
@@santoshrrrr Excited about Microsoft's recent developments in topological qubits, we man be able to rationally design novel enzymes in
Thanks for the context, fantastically helpful comment 🙏
P.S. cannot overstate*
@@santoshrrrr Honestly at the rate Google's going, I think it's more a race to see which gets there first - AI or quantum computing.
Of course once they're both working we can put them together... Oh boy...
"Yes, this is the place on the internet where we get unreasonably excited by a large blue bar. Welcome to Two Minute Papers!"
Well said, Ka'role! 😂❤️
Dr. Zsolnai-Fehér, as a complete stranger to the subject I cannot overstate how much these videos lift the weight of ingnorance from my shoulders while adding knowledge to my brain.
I'm a whitness through your tales and, for that, i can't thank you enough
As a computer architecture researcher, I'm now on the task of making this run fast and cheap. Hope myself good luck!
❤️
BEST OF LUCK HOWARD
Run it king, you are the future.
How expensive is it right now?
godspeed to you!
That blue bar was massive, did not see that coming!
"This is the place on the internet where we get unreasonable excited by a large blue bar, Welcome to Two Minute Papers" fits suprisingly well to many vids here
The blue bar was AlphaFold 2 years ago. I wonder what can it do right now…..it give me the chills….
You know a Two Minute Papers video is going to be good when it's almost 18 minutes.
18 Minute Paper
Nice to get some genuinely good news amongst everything going on, this is exciting.
Let's just hope our species lasts long enough to realise the vast potential we can unlock for us all...
amongst
@@turolretar bit suss, eh?
You should read the book (or a summary) '10 global trends every smart person should know'
What a time to be alive :D
Would love to see some more videos on multidisciplinary topics, even if u cannot explain everything properly, it always gets me excited about the future ^^
Man, I love You for this. Reaching to authors of discused paper, respect for all fields of science and comprehensive format. What is Scott Manley for rockets and astronomy, You are for AI and graphics. Have a great day and many thanks.
It’s so nice to have something so positive to put into my “Moments in History” playlist. Thank you for sharing this :)
Whenever the daily news gets too depressing, revert to this: one of the most amazing things ever! I spent 1 year during my diploma thesis purifying proteins for structure elucidation... now this can be done from a sequence in minutes. It allowed anyone with a PC to visualize how similar or different the spike protein is between omicron an delta, for example, helping to make an educated guess on whether vaccines were bound to continue to have a reasonable effect or not. Indeed, what a time to...! 👍👏
We all appreciate your dedication to make sure the information in your videos is as accurate as possible. Great video!
Would love to see everything that we discover because of this
There is already a ton of stuff using this.. and there seems no end to it
@@nulled7888 advances in AI are happening at a truly blistering pace. So exciting
@@grimaffiliations3671 ikr.. and not only ai, every technological, biological and basically every field of research is getting blasted with new results and so on.. its fantastic what digitalization and especially ai enables us to do
Oh boy, a new episode of Twenty Minute Papers!
Thank you for taking your time to make the video accurate and super comprehensible even though it was not your area of expertise! It means a lot
I'm so impressed by you actually insist making this video correct by asking the best person knowing this work!!! Thank you so much!!
Károly you have always been a source of inspiration for everyone, especially people like me from outside research community! You explain these breakthroughs in such simple illustrations, it really is making a dent in the universe, for better :)
I highly appreciate your approach of quality content and the actual interest in the topics you discuss
A few comments are asking about what you can do with knowledge about a protein structure. The answer is "a lot"...and that is why protein biochemists sometimes work multiple years to try to figure out the structure of a single protein.
OK, here's a better explanation of what you can do if you can predict a protein's structure from the sequence of amino acids it is made of:
1) You could understand how a protein works and where the most important sites are.
2) You could screen the structures of small molecules to identify the ones that could interact with the protein in a therapeutic way (and, even better small molecules could be identified if the current best small molecule has some limitations).
3) You could identify small molecules that could interfere with two proteins from interacting with each other to disrupt a cellular process.
4) You could study two different proteins to understand how they might interact with each other.
5) Proteins are often modified (decorated) with small molecule groups and Alphafold could understand the impact of the modifications on how they could change the protein's unmodified structure.
6) You could test ideas about whether changing the amino acid sequence at specific sites in the protein could alter the protein's function in the way your wish it to be changed.
7) You could create proteins that do not occur naturally (designed/bespoke proteins) with the properties you are seeking and check that your design would be correct by having Alphafold analyze the structure of the protein you dreamed up.
here in Nairobi Kenya, i shall use Alphafold to predict protein complexes that interfere with tau tangle mentor synthesis. we already have strong indication that specific proteins can derail the protein complex involved in tau tangle establishment in dementia progression. Alphafold is a mighty blessing for the dementia research community and a hope for patients-a valid real hope at last
I really like the current trend that authors of papers help videos like this be made. Hopefully you can get more such opportunities!
Amazing video, thank you for taking out time to enlighten us on this topic🙏🏻
I'm very happy this video was able to be made. Thank you to this channel and the people who helped this happen.
Thanks you for all the effort you put in to your videos.
Glad you take quality and delay over the views, very respectable characteristic to have.
Really appriaciate your content, thanks again
I am studying protein folding/unfolding simulations for an undergraduate degree; this video is exceptional. Very well done!
You are too kind. Thank you so much!
How do you study this, I’m still in Highschool, but I want to study bio-chemistry and bio-informatics
@@egoxagony4623 Looking at how the amino acids that make up the protein cause structure to the protein (clefts and active sites). Then look for how the structure is lost in mutations of the amino acids using computational simulations of the protein environment.
@@badgerff thank you
Yes, this is a place where we get excited about a big blue bar :D
Thank you for explaining the papers in a way, that even I can follow!
I remember taking part in the crowd sourced training of this. At the time I heard that humans were better at folding so the system would watch how you did it and learn... this system didn't learn anything from me because I was terrible at it.
Don't be so hard on yourself. Even AlphaFold has low confidence in its solutions sometimes. ;u) Also that "game" was super hard!
Based.
@@JACKRAIDEN97 Please elaborate, as you may have gleaned I'm an idiot.
@@newM0nkey Oh that's just how I am, I didn't know anyone who could finish a protein.
I installed it a long time ago but never got any good at it at all, lol. There was this other project to try to map neurons in 3d space too.
I really like the longer videos! I know some papers don't have 20 minutes of stuff to talk about, or at least you like to keep it brief so it doesn't feel drawn out, but still, I like to sit down with breakfast and watch long videos to get my morning going lol
Thank you so much for this video! I have been so confused by the lack of cover for AlphaFold and how even many of my ML peers have not heard of it before I started fosming about it.
Thanks to you, I have one more amazing video to share about the subject, and will be instantly sharing it in our ML some groups
I love this. I really hope to see this develop further!
Great video! Love the longer format . Thanks for explaining this development so with such clarify and enthusiasm.
Wow! incredible. However, note that a AI for scarce food may say eat one another or animals. As a vegan this is very wrong. AI without ethics could be very very disastrous.... (if AI think that humans are doing bad, it may gun them...) ....
This is pretty amazing. One of my honours supervistors' group did a lot of work for protein structure, specifically protein binding. Their method involved fine simulations of predicted active sites (IE based on Van Der Waals, dipole, etc) and simple spring simulations for atomic bonds further away. The future of this will be a game-changer.
Wow! I remember learning about protein folding in a talk many years ago, and the speaker said we had a big undiscovered gap there. i feel that gap is now being filled. This will lead to many good things.
Really great video, Very excited to see a few paper down the road where this goes, and hopefully more two minute papers videos to tell me whats happening haha. :)
One of the best channels on YT. Congratz and ty for your excelent service to science communication.
Absolutely appreciate waiting to make something that is accurate and reviewed. Now that's a gift to humanity ! Kudos !
I have been telling everyone I know about this scientific development since first hearing about it. Thank you, I now have a video to share instead of my unqualified explanations! This seems like one of those "once in a generation" discoveries that we keep having.
That was amazing, thank you Two Minute Papers and DeepMind! It is indeed very exciting to be alive right now👏👏👏
Not just «not bad» but really impressive result this time!
Kudos to developers and congratulations to all humanity!
The fact that he helped with the video shows the long term thinking. These videos bring a lot of new people into the field, and who knows, maybe one who was a viewer becomes one who gets a breakthrough.
Also i really wonder how good they will be in this years competition
RoseTTAFold is one of many projects looking to improve on Alphafold 2
Should rebrand as "17 minute papers"
Gae u r
Great video. Thank you for explaining it in such detail and diving deep into the workings, this is really enlightening. And bravo on verifying your work before publishing it here ! :)
thank you for this video, and connecting all types of tech that you've shown previously, to protein folding
An 18 Minute Paper! YAY! It was very fun and exciting to watch, thank you!
I hope this AI is implemented into solutions as soon as possible
AlphaFold 2 was used to develop and verify some recent mRNA vaccines
This is really going to help with my molecular simulation project !
Wow! incredible. However, note that a AI for scarce food may say eat one another or animals. As a vegan this is very wrong. AI without ethics could be very very disastrous.... (if AI think that humans are doing bad, it may gun them...) ....
The incredible effort that was AlphaFold is already astonishing, but reaching such a high level of accuracy in only two years is absolutely mindblowing. I personally do not own any paper that I could've held on strong enough.
WHAT A TIME TO BE ALIVE ; thank you john & karoly and everyone involved =]
There are already 2 years since it got that good. I can't wait to see how it is now, in 2022!
I am a 1000% novice to all sciences but absolutely love to learn everything I can about them. Your channel is criminally underrated and it seems like every new video you put up something amazing is always happening! Keep up your hard work!
Amazing job researching and making this video! I am sure you spent a lot of time making it.
Exciting start for a bright futur. Thank you for the vid!
Does alpha fold include ribosome activity/topology to adduce protein geometries or is it all inference from post folded protein data? Are there any good examples of physics simulations of peptide catalysis of mRNA read amino acids travelling over the surface of a ribosome and what they generate in terms of protein structure versus the empirical data?
Loved this video, wow!!
I am a casual viewer and love that it is a accessible to me. thank you
I remember watching your older videos and being amazed how in just a few minutes you introduced a topic, explained the problem, and articulated the core basic concept of "how it works".
Here, I think we missed out on the last one, and I would love it if you'd go back to explaining things in a little more thorough way.
As a molecular biologist/bioinformatician deeply interested in protein folding AlphaFold is the worst thing that could happen.
First of all it reinforces shallow thinking about protein structures - that the functional protein structures are static - which they are not! Proteins are dynamic beasts, oscillating around a single or a few different structures - and this dynamics are often crucial for understanding of their biochemical function. AlphaFold will give you one structure - but what about the proteins for which it should give you a few structures?
Second, AlphaFold it is still based on homology (notice it uses MSA's, multiple sequence alignments), which is a crutch on which we were standing for a long time - but we NEED to eventually stop using it. First problem with homology modelling is that a single change in the 50-60 long sequence can COMPLETELY change the final structure, something homology modelling will have super hard time catching - if at all possible. Second problem is that it cannot by definition build structures of rare topologies which did not been yet discovered experimentally.
Finally the AlphaFold doesnt solve the problem at all! We need to understand HOW the protein folding works, not just be able to magically conjure up a structure from the sequence. Protein folding process can and often does fail in real life and sometimes this leads to significant health problems. On top of that, the understanding of this process is absolutely crucial if we ever want to design our own proteins - which we absolutely do.
But now lots of people will think that this is a solved problem and move on, leaving lots of issues swept under the rug.
I've heard some people say that maybe this just proves that there isn't much to discover about protein folding in the way that people were expecting. In the sense that it may just be a problem where the only solution is doing extremely computationally expensive simulations of the chemistry, and there isn't really any general principles or theories that could be applied across all different proteins. Idk tho I'm not an expert at all that's just what I've heard some people say in articles about this stuff
You make excellent points. I think the multiple structures issue is really interesting. As I understand it, changing shape in response to the environment is how many proteins do useful work. I can't agree that AlphaFold is a bad thing though. it is massively useful, but should not be seen as the final word protein structures. I am sure DeepMind will continue to develop it.
@@RealUlrichLeland That cannot possibly be. People saying that do not understand how evolution works, nor do they know about the progress in the field of studying hte protein folding. While our studies of the evolution of protein structures are still in their infancy - and studies of how protein folding evolved is practically an opened question - one does not simply get a folded protein out of thin air. While there probably isnt a SINGLE pathway for any given protein to fold - much less a single pathway for all proteins to fold - there must be some mechanisms or else the proteins:
1. would never fold
2. could not evolve new proteins
3. chaperone proteins would not be a thing
The thing is, it seems from what've seen so far from the results we COULD achieve is that for any given protein there is a labyrinth of possible states it can fold through, but the labyrinth is biased toward the functional structure. By changing the rates of equilibrium constants between the states you can shift your protein and with some work uncover previously unavailable structures.
Now about chaperones - they do understand folding. They dont just randomly refold everything they've touched - a lab next to the one I've made my MSc in made such a chaperone by a few point mutations and it wrecked havoc upon the cells. They somehow target what needs to be refolded. Now, I am evolutionary biologist as much as structural one, I believe that the key to understanding folding is to see how chaperones evolved and understand how do they work. How do they detect what needs to be refolded? What do they *exactly* do to refold the protein? There are lots of details i dont have space to dive in here, pointing that there is a lot of method to this madness.
This attitude 'oooh we can never understand folding, lets not try to do this' is what i am talking about in my initial comment - its the worst that could happen, and like it or not, AlphaFold is boosting it.
@@Paul-rs4gd I specifically did not say that its a bad thing, please dont put words in my mouth. I said that it has a very very bad effect. This is a different and undeniable statement.
Cars, antibiotics, electricity - all of those things are good things overall, but they do have bad effects too, effects we should not forget to tackle.
I am not keen on AI in science. They are good to develop tools to solve specific problems, but they cannot help us *understand* anything because you cannot reverse engineer the AI to tell you WHY it predicted this or that outcome. AlphaFold is a perfect example of that and any other AI will be no different. They are basically a very, very good tea leaf divination machines.
I love large blue bars. 💙
Thanks for whispering the whole time. Great video.
I have seen so many explanations for transformers, yours is the only easy one that makes sense. Thank you.
As a structural biologist myself, I want to say thank you for this excellent summary of why AlphaFold has made such a huge splash and why it is important! I couldn't have put it better myself. Keep up the great work!
This is absolutely mind boggling and I can't wait for future papers to come out!
as always a great content ... I have never been disappointed following this channel. Thanks
I just love watching your videos. They make me feel excited to be alive today
Question: I've been running Rosetta@home on devices for the better part of two decades. Is the distributed computing model of protein folding even relevant anymore? Should I continue to donate cycles to projects that are completely outclassed, rather than dedicating all my compute to projects like Einstein@home or Universe@home?
We’re so close to make so many things that used to be considered miracles nothing but a trivial doctor visit
🤩 the future is coming at us so fast 🤩
we're also currently living with so many things that wouldve been considered miracles. the healing power of antibiotics probably would've been viewed as an act of god just a few generations ago
some still view all medicine as acts of god though so I guess that's not an amazing example
One day we will download medicines
Like what? Can you please elaborate for the less educated in this topic?
Wake me when the doctor's visit itself can go through an AI. Maybe the process of seeing a doctor will be made less painful and costly. (Unfortunately, I know that's not the way capitalism works. They'll charge 90% of what a doctor charges to avoid "leaving money on the table". But one can dream, and it will make it cheaper for developing countries that don't respect intellectual property laws.)
Thank you for your work! Happy to share!
Awesome! Thanks again for another great video. Keep it up!
While I am highly impressed by what Alphafold has achieved, I would like to add my two cents: After reading about how the AI works, I am skeptical about using it to predict structures and dictate what experiments should be done. Just a few months ago we were investigating how a couple of mutations in a protein affected its structure and while working on the project, I decided to see what Alphafold would predict the new structure would be, just for fun. After the experiment was done, we saw that the mutations lead to a substantial loss of structure in the protein, but Alphafold predicted it would remain almost exactly intact. So if we would've relied on the AI from the beginning, we would've been mislead.
To be fair, Alphafold is very clearly advertised in that it is not designed to, or capable of, giving answers about the effect of mutations on the protein fold. Thus you really should of expected zero change to the fold in your experiment. This makes a lot of sense as the strength of the technique is reliant on multiple sequence alignment and detecting co-evolution signals across the protein sequence to help determine whats near and far.
You are correct in that one of Alphafolds weaknesses is that it cannot handle mutations and will not correctly predict the effect of mutations on folding, but its a bit on the nose to attempt to use a technique for a purpose its literally not designed to (and warns you not to try) achieve. In your case, you would never of relied on the AI from the beginning because you should already know that it cannot supply information regarding mutations affecting folding.
Interestingly that doesn't meant just because it wasn't designed to do something it can't do some interesting things - a common technique people have been using is to add large glycine linkers between two domains to help test interaction between two proteins - which is surprisingly valid as the linker is simply poorly predicted and the MSA can still detect the relative positions quite well. However, obviously you couldn't then make mutations to try break that predicted interaction, as it will always be there no matter what (so long as enough info remains for the MSA). This technique of making fusions plays to the strengths of alphafolds workings, which is why it works so well.
@@MichaelFlambe i dont understand how can handling mutations not be within the scope of alpha fold’s capacity? I thought amino acid differences and it’s consequent effect on chemical interaction are its main drivers. This is a genuine question.
@@mau345 Howdy, its definitely an easy assumption to miss, as its its often phrased as "amino acid (AA) sequence goes in, structure comes out", however this is only partly true, the multiple sequence alignment between homologous proteins is a major part of the what allows the predictions to work. In this sense, it may be better to think the AA themselves are not driving the solution in a physics simulation sense, but more the patterns within and across AA that are known when compared to your target sequence are driving a global fold. This may be more obvious when you consider that many proteins can be very different at the primary sequence level (i.e. many different aminoacids) but still share a common fold. It's the relative patterns of AA which matter, not what the aminoacids are per se.
Bit of a rambly answer sorry, but hopefully that helps.
Does it predict the possible combinations, under all conditions (pH, temperature, aa et al. availability)?
Are all those combinations stable? Are they possible in the natural world? Are they safe (not toxic)?
We need to address these questions as it becomes easier to "create" these molecules.
It is a great advancement and will make many scientists happy 👍
Alphafold outputs multiple putative structures with different confidence. One could interpret these multiple structures as potential alternative conformations of the protein. However this needs to be validated on multiple protein families before we can accept it completely
It really is amazing, I am so privileged to have worked with deep mind and google over the past week
Thank you so much for bringing this amazing topic to our attention in such a good way.
be interesting to see how they deal with proteins folding other proteins
i was thinking of the same thing! wouldn't the presence of folding chaperones at least slightly change the shape of the proteins, how would AF2 account for these small things? Then i guess the next step would be predicting information regarding interaction like formation of polymers or substrate interaction, which should be within reach. Its amazing how it developed that much in only 2 years, curious what crazy things we will have in the next decade!
They abandoned the AlphaStar project, most likely because they couldn't come up with a definitively super human level AI. The AlphaStar we got was only grand master level, it couldn't even beat the top human players.
It's interesting to hear some criticism of this tech, after all the breathless praise. Their results with simple Atari games were also impressive but as I understand it the AI also struggled to learn more complex games.
@@chrisf1600 Clearly StarCraft was harder than they thought, the bot had incredible intuition but it's lacking something that stops it from becoming superhuman. I think it's the fact that it's entirely based on a sort of intuition and zero analytical thought process, the opposite of hand crafted algorithms. Simple things like a patch change that nerfs some units would mean the network has to be re-trained where as a human player you can just tell him oh that unit's nerfed, use it less often and he'll understand without having to play another 20,000 games.
@@chengong388 yep. They also had extra info by being allowed to access the whole map simultaneously. This is a huge advantage and wouldn’t be possible for a human. Even with that, the ai mostly relied on superhuman micro used in extreme bursts, as opposed to playing the game. IMO they failed to make the ai play under the constrains of a human and so they eased the constrains to make it win more. They also constrained apm but only to a average. So they allowed it to do crazy bursts of inhuman levels of apm, tens of thousands of commands per second, which is just impossible do with hands and a mouse. So eh it really wasn’t that impressive. I was disappointed that they declared victory without really succeeding. It would have been more interesting to admit they could not defeat humans at StarCraft.
Love your videos where we get excited by a big blue bar. Amazing. Thanks 🙏
Thank you so much for this video, it was truly inspiring!
What's next? Reversing the process. I.e., given a structural protein, solve for the amino acid sequence. Why? to construct enzymes that will recognize and act on a variety of molecules in useful ways. Once you have sequence you can build DNA and incorporate it into genes to make useful "worker cells". Make your own bio-factories to extract carbon from CO2 and make graphite or nano - tubes or even diamonds.
Inverting the process is surely the next challenge. There are standard algorithms that can use successive approximation and search to invert some functions, given a fast function in the forward direction - they may not work on something this complex, but it is worth trying. Alternatively maybe another AI can be trained end-to-end to do the reverse process (using the existing datasets !).
As someone in the field, the protein folding problem has not yet been solved. Alpha fold 2 is incredible but as mentioned in the video there are still areas it struggles with. We need more papers!
Quantum Computers will solve it I hope. 😄
@@kamilkaya5367 Delve into Quantum Computing a little bit more, and you realize that Quantum Computers can't magically accelerate the progress of AI training. It's a bit less straightforward than that. (Maybe some day.)
Thanks for the video, Károly! You rock!
thank you DeepMind!! yall are truly amazing
DeepMind: *creates and keeps improving AlphaFold*
Cancer: *sweating profusely*
Some cures to cancers people don't want to take. The HPV vaccines effectively cures some cancers in men and woman, but people avoid taking it or having their children take it because of some weird religious association with a fear of sex. Also, 40% of the US population is not vaccinated against COVID. This all makes me think that we might find "cures to every cancer", but a large majority of the population will still be afraid to take it. (And the HPV vaccine is like more than a decade in, and there is still fear.)
I guess it's like the old saying about leading a horse to water but you can't make it drink.
@@fitybux4664, Your point being? Vaccination against SARS-CoV-2 does practically nothing to prevent it from spreading, and most likely barely minimizes its effects at all. The worldwide percentage of vaccinated people is about 50%, so I don't see how any of it surprise or even remotely related to religion.
I live in one of the most heavily atheistic countries in the Northern Hemisphere, and we had a huge trouble getting over 50% of the population vaccinated, and that was at a time when the vaccines still had some effect, before the latest mutations.
The "fear" is very much justified. We have absolutely no idea how any of this will come to haunt us in the future. Countless time we've seen governments approve extremely dangerous "treatments" in the name of advancement, so how could you blame anything from doubting this extremely suspicious situations where governments forced citizens to get vaccinated when their safety couldn't be guaranteed and the effects were already diminishing around the same time...
There are countless idiots, but scepticism and cautiousness are never wrong.
Everyone I know has all three doses, and every single one of them had been infected at least once, even though they have been barely in contact with anyone.
Medicine is always a gamble, and you have to consider your odds, and decide for yourself. It's easier to die of cancer than to undergo some magical treatment no one has ever heard of.
In three months next CASP will begin and alphafold 3.0 will be released
Thank you for the detailed explanation. And also thanks to Deepmind to make this as an open resource!!
My roommate was part of this project. Keep up the great work!
This combined with CRISPR gives us almost omniscient control over biology. We're nearly at the point where we can design an organism to do anything you can imagine: nanobots that clean wastewater, crops that grow in any environment, super effective CO2 sequestering, drugs tailored to any disease, all genetic disorders curable, evolution itself directly controllable.
super effective co2 sequestration might already be on the way actually. i cant remember who figured it out but certain nanostructures collect co2 when given an electrical charge and they also release that co2 when the charge is turned off. so if we can mass produce these, co2 collection will be as easy as giving a premade device access to electricity that isn't releasing co2. then its not a terribly complicated process to turn the co2 into something you can shove underground so it doesn't warm the planet.
What about peptides tailored to any disease. This along with generative adversarial networks and monte carlo simulation using annotated data structures will open floodgates.
@@demonz9065 not just CO2; there are bacteria that assimilate methane as well. if we take inspiration from these proteins that allows that bacterium to essentially "eat" CH4, we can greatly reduce the greenhouse effect; because a not-insignificant amount of CH4 leaks out of oil refineries and natural gas extractors.
oh cmon nearly is an overstatement for some of these things!
@@regulate.artificer_g23.mdctlsk methane doesnt last as long in the atmosphere despite being a more potent greenhouse gas. so I think as long as we stop emitting it, taking it out of the atmosphere isn't quite as pressing an issue.
Just one small correction: Deep Blue *was* an AI. Just not one relying on ML techniques.
A correction to your correction: Deep Blue was made by IBM, who doesn’t refer to “weak AI” as an AI, but as an expert system. IBM only calls “strong AI” as an AI, which is to say, a general-purpose program that can think everything a human thinks, but better.
@@cmyk8964 And yet, IBM doesn't decide what is an AI and what is not.
@@Ceelvain IBM does decide what IBM products are called.
I heard about protein folding years ago. I am so very excited to have this in our generation.
To achieve such goal would supposed hundred of work hours for is humans, as a general practitioner I see this paper as a part of an integrated strategy for the human to evolve through genetic and AI mastery. As seen in movies such as Gattaca, after discussing all ethic concerns obviously. Great video!!
Can it solve sudoku?
Its hard to appreciate coz I don’t understand what’s going on 😅
this will indeed help humanity a lot. not only to fight certain diseases. it is basically the same impact of breakthrough as the decoding of human genome.
two very important fields
This honestly gives me some sorely needed hope
why? do? you? do? so many? stops? in between? sentences?
damn the quality of your voiceover surely has went down a big way. i could barely comprehend what you saying with all that pauses and intonations
Sir you are a gift to humanity thank you for this channel thank for you integrity. I wish you all the best
Yeah, I had that moment at the begining of the month with my algo. It was getting fractionally correct answers against the dataset. While enough to work. It was not reaching it's potential. Finally straightened out the complex architecture and, booom. Maximum potential reached.