DeepMind AlphaFold 3 - This Will Change Everything!

📝 Check out AlphaFold 3 here:
dpmd.ai/yt-tmp...
📝 Or try it out through AlphaFold server for free:
alphafoldserve...
This video was made in partnership with Google DeepMind.
My paper on simulations that look almost like reality is available for free here:
rdcu.be/cWPfD
Or this is the orig. Nature Physics link with clickable citations:
www.nature.com...
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Alex Balfanz, Alex Haro, B Shang, Benji Rabhan, Gaston Ingaramo, Gordon Child, John Le, Kyle Davis, Lukas Biewald, Martin, Michael Albrecht, Michael Tedder, Owen Skarpness, Richard Sundvall, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, 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
Károly Zsolnai-Fehér's research works: cg.tuwien.ac.a...
Twitter: / twominutepapers
#AlphaFold3 #AlphaFold

[@Hitjuich]

Alphafold changed bìotechnology forever. This is huge! Looking forward to working with the new version

[@Z0MBUSTER]

Could it be also used as "chemtrails" to "eat" CO2 from the atmosphere to transform it into a harmless gas?

[@ScottzPlaylists]

​@@Z0MBUSTER God designed molecular nanomachines to do that, it's called photosynthesis❗ No human involvement needed. If we have twice the Co2 in the air one day, plants will grow twice as fast. Co2 is a harmless gas --- and we would die without it. Don't believe the propaganda. Historically, Co2 levels have been much higher in the far past. ❗❗❗❗❗❗

[@dertythegrower]

We have done this foldingathome using gpu since 2002... hardforums team, especially. We used to teach people how to do it at home, long before bitcoin mining using gpu, which we did back when bitcoins were useless except to other bitcoiners.

[@dertythegrower]

​@@Z0MBUSTER co2 is not the issue, its methane... many commercial growers ON KZbin dump co2 into grow areas, and a lot of it escapes. Methane from factories and smog byproducts is far worse....and a major concern, which breaks down into more co2, ironically.. you can capture it in a balloon shape object, and power generators. Methane is 28x worse greenhouse gas than co2.

[@juhor.7594]

​@@Z0MBUSTERone big problem with this kind of climate engineering is that it's hard to predict what long term side-effects it can have. I imagine that the plastic-digesting example can't be used outside of highly controlled environments.

[@itsjusttmanakatech1162]

I wasn’t impressed until I heard AlphaFold had implemented Dark Mode. What a time to be alive!

[@austinvw1988]

I was literally about to make a joke lol

[@Ren33469]

I've been waiting for BOINC to add dark mode for a while😂

[@hrihori_art]

"Dark Reader" extension for browser (free)

[@snooks5607]

fairly big reason for why I've been using mostly linux since the 90s was inability to control blinding background colors in other OSes. now 20+ years later windows11 and macos are finally able to display most window contents with dark background, it was a pretty momentous achievement (their highly paid UI/usability "experts" are idiots)

[@crubs83]

Just imagine where we'll be just two papers down the line!

[@Favmir]

4 minutes earlier than the official Google deepmind channel? That was indeed fast, you're right

[@TwoMinutePapers]

🙂

[@JohnDontFollowMe]

TwoMinutePapers be like: *sniffs white substance*, "lets read these papers!"

[@asdads3948]

@@JohnDontFollowMe I think our man needs the white substance to get *down* after reading a really good paper.

[@thunderinvader9031]

What a time to hold on to my papers

[@ElOroDelTigre]

I'm around 50 now and seeing that the field I love is doing real change (not in a "generating revenue" way, but in actual "improving humankind" way) brings a tear to my eye. Maybe a 20-something years old kid does not realize how big this is, but for me it is, truly and marvelously, an amazing time to be alive. Gives me a bit of hope that before I kick the bucket I'll get to see some of that utopian advancement and eradication of diseases we used to dream about.

[@_sky_3123]

I am 30 ad I too am amaized. I remember reading SCI-Fi books where we were able to simulate proteins and their interaction. And you could give a cancer genome to the computer and it would give you a blueprint for protein that is poisonous to the cancer, but not to the host. (So like a perfect cancer treatment, you just had to drink/inject the "poison" :D) And now I might live to see something like that come to life.

[@tortysoft]

@@_sky_3123 I'm 66 :-) Need I say more? I hope I can extend my time being alive. Could it fix Long Covid please? I'm less than a bit alive at the moment, I'm after quality rather than length, but I'll be happy with both !

[@ksprdk]

What could be a scenario that you could imagine happening now after this? And when? :)

[@ronankearns381]

20 smth college grads are developing AI lol

[@squamish4244]

I'm 45 and hope that this means that our generation will not have to endure the chronic illnesses of old age that plagued our grandparents and are hitting our parents now. Maybe this means Alzheimer's and dementia will not end up costing TRIPLE what it currently does by 2050, which it will according to current predictions. And maybe even reverse aging to some extent. AlphaFold - 2018. AlphaFold 2 - 2020. AlphaFold 3 - 2024. Nobody saw anyo of this coming a whole _six years ago._Who even knows what's coming now.

Bet AplhaFold 6 will be folding reality itself 😎

[@Argh0pirata]

not if it keeps being closed-source with carefully curated examples

[@globurim]

It will fold my mind into smithereens.

[@frederickmiller6431]

You are AlphaCreate 6 duhh this is what the Egyptians knew how to craft reality

[@thegreenxeno9430]

Alphafold 6? More like Betafold 2

[@NuclearTopSpot]

@@thegreenxeno9430 Sigmafold Ohio Rizz edition

[@Rajivrocks-Ltd.]

Everyone is talking about LLMs, But this is really where impact to the future is being made!

[@tingtonggamer4901]

This is made by an LLMs you know that right "Alphofold AI"

[@tingtonggamer4901]

LLM*

[@Rajivrocks-Ltd.]

@@tingtonggamer4901no, I was correct

[@Rajivrocks-Ltd.]

@@tingtonggamer4901 Oh really? I haven't read the paper. Thanks for pointing it out, I'll check it out

[@mark9294]

LLMs are at the root of all these developments. Their emergence is an epochal event

[@PranavPunuruFilms]

i'm pretty sure they're participating in CASP16 because of this. My lab is participating too. We just agreed that they've already won

[@azrael5648]

ChatGPT was fancy. Now this. This is revolutionary.

[@DreckbobBratpfanne]

It's gonna be massive once we can combine these, either a next gen chatGPT helping to create a next AlphaFold or direct chat to Protein 😂

[@w花b]

Yup. A guy that's all talk but this is all bites

[@magpielive]

Well its takes 12 to 15 years become undisputable like the journey of iPhone 1 to iPhone 15 . They are just started there main objective is to make AGI if agi come it revolutionise the humanity .

[@dexio85]

It's not. Just a few % over last technique. The guy is presenting it as if it's a second coming of JC. Probably because he partnered with Google on this. Google has sunk Billions into Deep Mind and the shareholders start to ask questions. They are trying to make their investment look good.

[@eprd313]

​It's coming faster than you think. They call it the technological singularity

[@kaspernordlund6828]

I get shivers every time i work with alphafold! to think that it is free. A true testement to the Greatness of human civilisation!

[@lucasreis6251]

Hey bud! While watching the video I had a question pop up, and seeing that you work in the field maybe you know the answer: From what I could understand, Alphafold is able to take "letters" as input (aminoacids or something like that I believe, but correct me if I'm wrong), and outputs the 3d shape of them right? If so, how does that helps with development of things like, the plastic eating enzyme, and other things? Wouldn't you already need to know which letters to use beforehand? How does knowing the shape of everything affects the development? Pardon me if I'm saying nonsense. I'm an automation engineer and know almost nothing about biology.

[@jadynhasstupid2275]

@@lucasreis6251you look slightly autistic

[@bolatm22]

​@@lucasreis6251 answer of ChatGPT: Yes, you understood correctly. AlphaFold takes a sequence of amino acids (these are the "letters" in proteins) as input and predicts their three-dimensional structure. This can help in the development of various biological processes, such as creating a plastic-eating enzyme or other enzymes. Knowing the three-dimensional structure of a protein can help scientists understand how it interacts with other molecules and what changes in its structure could improve its functionality. Thus, even if we know the amino acid sequence beforehand, predicting the three-dimensional structure aids in understanding its function and designing improved versions of the protein.

[@kaspernordlund6828]

@@lucasreis6251 So yes you are abselutly correct, you would need to know the amino acids beforehand. That is luckily a pretty standard thing to get to know. You can do a buch of differnet test (like mass spec) to figure out what amino acids and in which order they are placed. The almost impossible part to figure out is how the protein folds. It has been a core theory that the the chemistry of the aminoacid chain is what makes proteins fold in specific ways and hold their shape and activity. But breaching the gap from sequence to folding is a huge task that until alphafold many times took years of research to figure out. What most people did to get as accurate readings as alpha fold was to do crystalografi. That is a very complicated task that needs pure crystals of your specific target protein, which was almost more art than science to make work honestly. Alpha fold skips this years of wait and hard labor, letting us play with how mutations affect folding in a timeframe of hours. It is hard to overstate the impact of this ability in research. Hope this answered your question ^-^ edit: realized i didn't answer your question about how protein shape affects development. All proteins have basically the same amino acids in them. The amino acids are actually not what makes proteins able to interact in a biologically active manner. It is the shape of the protein that does. So to know how a protein interacts chemically you need to know how it folds, and not what amino-acids it has.

[@nikolavideomaker]

​​@@lucasreis6251 As a biochemist I can answer this. These "letters" are just a human code for amino acids, amino acids are molecules that form polymers (polymers are long-chain molecules that consist of a certain number of repeating connected units). In the case of amino acids, almost all organisms have 20 different of them, using these 20 every protein in your body is created. Proteins can do things such as move your body (muscles), but basically everything else in your cells is done by proteins. Signal transduction in neurons, breaking down nutrients and synthesizing new structures, transport of molecules around your cells, the immune system is also dependent on a lot of proteins (antibodies for example are proteins). Now, a type of proteins are enzymes, these are nature's catalysts, they can speed up reactions drastically (up to *10^9 !!!). For example we have enzymes that break down sugars into food. It gets interesting when you study bacteria, fungi, plants... Often times they have very specialized enzymes, for example some that can break down polymers. Plastics are polymers as well with often very similar chemical bonds to molecules that bacterial enzymes would break down. If we can mutate/adjust these enzymes to be more efficient at breaking down plastic, they could be a very useful catalyst in such a reaction. This is just one potential use, the possibilities are endless. We can have enzymes that produce lab-grown meat, help with electricity production... It's very easy to get the sequence of a protein (the amino acids it's comprised of), but how these proteins look structurally (3D) is very hard to find out by traditional methods. Using AI trained on existing 3D models, it can find patterns, for example a code of a certain 10 aminoacids would always fold into a certain structure. This is how predictions are made. So, now imagine you discovered a new enzyme from some bacteria that can do X, you get its sequence and instead of spending years and thousands of dollars in finding its structure in the lab, you can just plug the sequence into alphafold and get very likely the same result. Based on this then you can find out more about the protein's function, the mechanism of function and the sequence to structure relationship. I hope it was understandable what I wrote.

[@Sashik]

Oh I'm holding onto my papers now

[@larrychanhangpei9292]

You made me love papers and start reading it. It means a lot to me and thank you for the excellent work as always!

[@TwoMinutePapers]

This really made my day. Thank you so much! 🙏

[@ajay_krishna]

Same for me.

[@projectarduino2295]

If a protein can be manufactured to repair human dna, or to recover damaged telomeres, that would probably one of the most impactful biomedical results of all time. To help the elderly regain their strength to live healthier and longer lives by reducing cellular damage and cancer risks would be profound. I recently lost my grandfather because of age related weakening and loss of autonomy. If something like an enzyme injection could have kept him more autonomous by keeping his fatiguing body healthier, he might even have been alive today. I lost all my other grandparents to cancer, so if dna repair could help inhibit cancer development from mutation, it would be amazing to see no one ever have to go through what I have gone through again.

[@js_es209]

Death is inevitable and so is the timing of it. Will take u time to understand..

[@antongazizov8473]

any sources?@@js_es209

​@js_es209 Death may just become a probability that can be mostly mitigate. One day you'll understand.

[@Gelatinocyte2]

Our cells already contain proteins that repair damaged DNA. The reason we're even dealing with cancer nowadays is because we live long enough for one to eventually develop in our body, and our own repair systems can't keep up with the rising amount of damage we are now sustaining: carcinogens in the air and in our food, irregular lifestyles, and overexposure to substances and radiations that humanity had otherwise not have access to prior to technological advancement. Without those three conditions, our body's own proteins is probably more than good enough to protect us from cancer, although growing older still means increasing likelihood of developing cancer (because that's just the inevitably of aging).

[@valberm]

​​​@@js_es209 not really. There are ways, at least in theory, to become what they call biologically immortal. So, while your assertion may still be true on this day and age, maybe in the future it won't. Ok, probably real eternal life won't be achieved or even desired, but, periods of time which we consider to be long enough might be in the future.

[@EVILBUNNY28]

I can imagine a system that can sequence your personal genome and through tons of training be able to discern how exactly you’ll react to certain drugs and what dosage you need. It’s crazy to look back at how far Medicine has come in the last 20 years, it’s impossible to even begin to image the capabilities in another 20 years time. All advancements start somewhere, and releasing a paper is usually always the first step

[@notsojharedtroll23]

That's is a very cool concept 🎉🎉🎉

[@scoutbane1651]

You just re-invented the field of pharmacogenomics, congrats. I can literally go do said test right now. It's just currently in its baby stages. I got offered to do said test for 500€ by my psychiatrist, with said disclaimer. The tech will only keep getting better though and once your genome is sequenced, they can just update the treatment with any new way to analyse it.

[@alexandregaultier8313]

Do you have any need in mind ?

[@cureadvocate1]

Seeing AlphaFold 3 design ligands to help the brain and spinal cord regenerate would be cool, as myelin (and a few other tricks) inhibit neural regeneration Another interesting use case: exploring SIRT6 variations, its impact on aging, and finding ways to enhance its functionality.

[@alphaomega154]

it WILL HAPPEN.

[@shezcmayo]

Finding effective remyelination therapies to help people with MS like my daughter.

[@krox477]

Are you talking about Lizard man

[@dawiedekabouter5733]

If that as a medicine can be programmed into a fruit or vegetable or maybe a group of fruit and vegetables eaten together that can be grown it would really be amazing. Also changing the snakes DNA so it can have hands and legs and fruit from a tree that can help women reduce the pain at child birth and robots that can do the hard labour in the sun whose hands cannot be pricked by thorns.

[@lacklvster4512]

im so glad that i was born at JUST the right time to enter college for biotech in the infancy of all these new programs. cant wait to actually work with them later in life

[@vectoralphaSec]

Yeah AI will revolutionize Biotechnology forever.

[@CloudCoderChap]

You’re our future and I wish you luck.

[@bluesmanshoes]

I recommend working with them now already!;)

[@haianabou-karam4430]

Hary! AI will work alone when you finish the university! SUPERINTILIGANCE

[@spambird68]

Amazing work again from DeepMind! As a protein researcher, I'd love to see custom enzymes designed to bind PFAS, or the 'forever' chemicals that interrupt biological processes. Designing an active site to break them down chemically would be next level, but difficult due to the chemistry involved. However, there are other pesky environmental toxics listed as the Stockholm Convention Persistent Organic Pollutants (POPs), that may be more attainable. AI to design enzymes to breakdown environmental toxic chemicals? Sign me up! I'm in.

[@JohnVance]

I wanna dump truckloads of this on old landfills, come back in a few years and maybe it'll all be free compost...

[@alexandregaultier8313]

@@JohnVance I guess that it won't work like this before some time.

[@Rkcuddles]

Absolutely love these longer videos. Can’t get enough of your enthusiasm and the paper news

[@vitalyl1327]

The very fact that AlphaFold even works is huge - it means that behind the insanely complex Van-der-Waals physics of millions atoms interacting, there is some really good analytical approximation hidden with a tiny computational cost. If machine learning could uncover such a system, there is a good chance that we'll eventually discover it analytically, and this will be the real game changer.

[@cleanslate668]

Good day, I'm an AI undergraduate student and am interested in your statement that an analytical representation would be a real game changer. Isn't a machine learning solution already an analytical solution, just more of a black box? I have an intuition for why you would say it's a game changer, however, could you provide me some of your thought or point me in the direction of some interesting sources? thank you!

[@vitalyl1327]

@@cleanslate668 ML solution is a very large polynomial function with tons of coefficients with unknown meaning. It is far from what a proper analytical solution is, where you know where all the numbers are coming from. But, the existence of ML solution *hints* at an analytical possibility.

[@KnightsWithoutATable]

This is huge. There are so many illnesses that we know what the mechanism is right down to the proteins, but we just don't have any drugs for that are effective other than for treating the symptoms. This could also allow us to precisely target treatments to the illness much easier.

[@cyborgar15]

Curing people doesn't make money. Creating new illnesses makes money..Guess which one is coming your way..

[@JoshTheWhale]

I fold my papers... Too good... ... and actually Imagine where we'll be just two more folds down the line!

[@bvoros361]

41 more folds to reach the moon!

[@postscriptum9856]

I thought it’s impossible to fold paper more than 5 or 6 times?

[@anmol1713]

Very excited, I am working on a project where i used AF2 to predict better binding antibodies mixed with some reinforcement learning. I emailed everyone I know in university about the AlphaFold3, so excited.

[@lucasreis6251]

Hey bud! While watching the video I had a question pop up, and seeing that you work in the field maybe you know the answer: From what I could understand, Alphafold is able to take "letters" as input (aminoacids or something like that I believe, but correct me if I'm wrong), and outputs the 3d shape of them right? If so, how does that helps with development of things like, the plastic eating enzyme, and other things? Wouldn't you already need to know which letters to use beforehand? How does knowing the shape of everything affects the development? Pardon me if I'm saying nonsense. I'm an automation engineer and know almost nothing about biology.

[@福の子]

@@lucasreis6251you are right，but same ‘letters’ can consist different shape and this shape detect different fraction.

[@gumanelson2007]

​@@lucasreis6251i have a rough idea but lets wait for him His must be better

[@mehakmittal]

This is incredible! I really hope all undergrad colleges could incorporate Alphafold-3 in their research so that the students of this generation can take advantage of this revolutionary technique!

[@andynonomous8558]

Sigh, it's amazing how everything is about to 'change everything' and yet nothing ever seems to fundamentally change.

[@samuelgibson780]

If you had a president take a space race approach to this technology you'd see a lot change. Something like "let's fundamentally eliminate biological failure as a cause of death within 10 years", and throw space race money at it, and you'd probably be able to pull it off. Would require a lot of planning though. Problem is that much more is *possible* than anyone has the authority or administrative capacity to implement, when it comes to all that sci fi stuff. Would require a society making a decision, which is a bigger barrier than the technology itself at this point.

[@PhilipFranklin-l4f]

Our politicians do not work for the people. They work for corporations. That's why they're trying to be careful not to eliminate just worker jobs but also not make their industry or business obsolete in the future. They are going slow not only to be safe but to line their pockets with profit

[@adissentingopinion848]

I mean, we said the same thing about the internet and the smartphone. Sure we have hype about eventual flops like NFTs, but you can't say something like email hasn't fundamentally changed the world in 30 years. And even stuff like enzymes are all in your detergents and your cosmetics. It doesn't change until it does, all at once.

[@andynonomous8558]

@@adissentingopinion848 Sure, but one just tires of the constant rhetoric that everything is always being revolutionised while day to day life is more of a struggle than ever. I think we just need to tone down the rhetoric a little.

[@tachywubdub2469]

It is also a game of patience. Lots of these are like designer custom made parts. You'd need to market it to scale it up enough for people to afford it, and that requires large initial funding... Which requires a Monetary incentive to do it. Basically, unless the enzymes break plastic into gold or companies/people(worst case) are taxed for plastic waste accumulated, there is no incentive to break down plastic until it is a dire threat

[@mattbrandon9157]

had to mute the voice and turn on cc but in the end it didn't matter cause i really didn't see the big deal . guess i'm not well informed about this whole protein folding science.

[@jamesmorrison4976]

I think it is merely bs. Protein folding by computer is not precise enough - a lot of interactions (h-h bridges) are just crudely approximated and you still would need a super computer as a neural network even bigger as every element of the universe to make it fold like in real time. The most precise models are when you combine protein x-ray crystallography with protein folding programs.

[@scottmiller2591]

Ligands for breaking cellulose out of wood for paper manufacturing that reduce smell and waste discharge.

[@TehNetherlands]

Folding on to my papers!

[@PlaylistsOnly-ox5hp]

I've already told you, but I'll tell you again: you are doing a lot to make tomorrow a better world, an enzyme to recycle plastic. The world lacks people like you.

[@SteveRowe]

I'd like to see an enzyme that could break down interstitial amyloid plaques to prevent or slow the progress of Alzheimer's.

[@4grammaton]

Amyloid plaques appear to be a symptom of AD, not a cause. They seem to be a kind of defensive or homeostatic mechanism in response to insufficient nutrition in brain cells, with the function of downsizing brain volume to keep more important areas of the brain well-supplied. Treatments which reduce or prevent amyloid plaques have shown to be ineffective, and in fact counterproductive, in halting or reversing the progression of the disease. Tau protein tangles and buildup of protein in cerebrospinal fluid during the prodrome of the disease seem to be more material to the pathogenesis.

[@SteveRowe]

@@4grammaton Interesting. I read (5+ years ago now) that the plaques naturally build up and eventually encroach on synapses. I will refresh my knowledge.

[@4grammaton]

@@SteveRowe Please do, and feel free to inform me if what I said is inaccurate. I also remember reading that plaques are found even in the brains of young children.

[@wlockuz4467]

I can't wait for the applications of these models to never make it to general use because of politics, or have them completely monopolized by big companies.

[@bloodbound696]

I LITERALLY Made a 5-page paper.. on mealworms specifically superworms that broke down plastics and got calories from them and were able to reproduce from the plastic diet thanks to their enzymes. Finally, this is getting attention, making those enzymes into like an artificial stomach to at least turn plastics into Glycol is better than nothing!

[@sedthh]

all of life's molecules mapped

[@SevAbante]

Alpafold: we made something that will change the world Government: 🔫

[@sobbski2672]

Best part of this was them supplying all the pseudo code and training datasets. Hopefully Rosetta Commons will drop an open source version soon

[@Sonny_McMacsson]

"Letters go in and a 3-D structure comes out. You can't explain that."

[@test-uy4vc]

What a protein time to be folded alive! 🎉

[@perplexedon9834]

Dude! I'm actually flipping out at this bar chart!

[@goodtothinkwith]

For the non-biologists here (myself included), how close does this get us to designing more specifically targeted drugs without undesirable side-effects?

[@andreavitale2845]

A friend of mine is curious to know, too.

[@jamesmorrison4976]

In a nutshell: To predict correct protein folding is absolutely insane. In reality: You need actually to concentrate the real protein and use x-ray crystallography and a simulation like this to get a quite precise glimpse how it fold. I think this advertising for this program just ignores that. A solely simulation can’t predict folding of a protein (not even close). Peptides (very short proteins) are eventually possible so take this advertising with a grain of salt. Yes it will fold your proteins and give you some insights but no real scientist would put their hand into fire that he/she trusts a program like this.

[@EaziGX]

Meanwhile Apple has made a squeezable pencil

[@gumanelson2007]

😂😂😂😂

[@vincentpelletier1246]

I love how in a couple of years, all of this channel's content will definitely be done by AI. Even his voice sounds already auto generated already. What a time to be "alive" !

[@leejerrett8268]

How dare you! Only a human is capable of sounding this stilted and unnatural! XD

[@brexitgreens]

I want Dr Károly's voice in GPT-4o.

[@brexitgreens]

I want Dr Károly's voice in GPT-4o.

[@andyc8707]

Having choose to Fold@home in the early 00s rather than mine bitcoin, I am glad to see this field has grown so much.

[@redstrat1234]

I have been keeping up with developments in AI for the last couple of years, and have lost count of the new developments that will 'change everything'. I look out the window and nothing's changed, I go to my place of work, nothing has changed, I watch TV and sports, same thing, no change. Nothing has changed in 2 years apart from some online AI programs that'll give you pictures, videos, music, essays from a prompt, it's fine and all, but it hasn't 'changed everything' It's like crying wolf every 5 minutes, the 'changes everything' hype just becomes white noise.

[@Johnithinuioian]

The only thing that's really changing is the increasing level of fakeness.

[@Broken_Mesh]

It feels like where in the bubble phase of this technology where people see potential everywhere but there is no actual use, so there’s a race to find it. It reminds a lot of the dotnet bubble. People claimed the Internet would change everything for years and only after the bubble popped and a small number of products survived did change start to happen. Well and now its hard to find a place in everyday life where the internet didn’t change something.

[@danypell2517]

this is crazy exciting for the next update damn

[@viniciusmoura9105]

In 2004 I was ingressing University to course Biological Sciences. If you asked anyone there how much time they'd think it would take for mankind to see an applicable tool with the capabilities and accuracy of Alphafold you wouldn't hear "20 years" even from the most optimistic.

[@David-ct]

Honestly I would like to see some annotations about the function of the differents parts of each protein structure. That would be so cool!

[@debayondharchowdhury3001]

Holy Mother of all Papers...

[@jmkyarrow]

Although I'm tremendously excited about this advancement, as a biotechnologist myself, it ought to have been mentioned that right now the list of available ligands that are compatible with AF3 is actually very small and so if you're trying to model interactions between proteins and any molecules that aren't incredibly mainstream you're out of luck right now. Certainly no scope for users outside of DeepMind and Isomorphic to use it for ligand design at the moment.

[@Filaxsan]

The "tree bark" truly killed me

[@TetrahedreX]

How many times has something changed everything without anything actually changing?

[@SOTPOD]

what a time to be alive! thank you so much!

[@mikesbasement6954]

I worked in the plastics industry. In order to maintain the desired properties in the final product you can only include around 10% recycled material. I suspect this would not change that.

[@DreckbobBratpfanne]

These papers give so much optimistic hope for the not so distant future 😁

[@ingoos]

looks like the next Nobel prize awardee!!

[@DJVARAO]

Drug discovery faces a significant hurdle with target validation. If AlphaFold 3 simplifies this, it could save billions and years of research typically wasted due to inadequate in silico validation. At our lab, we are testing a new quantum computing tool specifically for that purpose. It would be interesting to see how AlphaFold 3 performs on the same tasks.

[@weishenmejames]

I "held on to my papers" and am impressed

[@RoulDukeGonzo]

I'd like to repeat the seminal work of Chothia and Lesk, laying the groundwork for protein structure biophysics with a couple of dozen protein structures in the 1970s, re-applied to the vast universe of new protein structures. This would be a tribute to the late Cyrus Chothia, who saved my PhD (and probably my life).

[@MrSchweppes]

Demis Hassabis and the AlphaFold team deserve a Noble Prize in medicine for this system!

[@adityarao2134]

Amazing! Can someone shed light on why these ml models work better than the ones more rooted in physics? To me it seems like this should be a first principles based problem considering the causal non-entropic physicality of it. For example, we similarly have calculators to do math since the rigid rules are part of the ground truth (E.g. transformer based LLM models predicting the next token are worse at adding numbers than a calculator).

[@jdirksen]

I ain't a guy that does much personal work with AI, but from what I gather it's mainly because this AI model can "intuit" to approximations and skip over many steps, then go back and verify it, and adjust its answer until it's as close as it can get. rather than having to work through the calculations done for a physics based interaction with many points of failure that can cause a cascade of inaccuracies to develop due to small errors and the like. Again, take it with a grain of salt, that's just my best guess

[@ThisMoth]

I'm flattered on behalf of humanity that you decided to interact with humans before consulting with an AI, My human reply would be that I guess the physics based algorithms are based on physics as we understand them and are more limited as they're only as good as we understand physics. A full general AI can take inspiration from any element in the world that would boost the chance of success. So less precise because it's not rooted in physics yet more precise because of the insane amount of datapoints as inspiration. Chatgpt 3.5 Machine learning models like AlphaFold 3 outperform physics-based approaches in protein folding due to their ability to learn complex patterns from vast amounts of data. While physics-based methods rely on rigid rules and principles, they often struggle with the inherent complexity of biological systems. ML models, on the other hand, can capture subtle relationships and nuances in the data, leading to more accurate predictions. However, it's essential to note that both approaches have their strengths and weaknesses, and a hybrid approach combining the best of both worlds may yield even better results in the future. PS, this is just me trying to come up with an answer to your question because it was a good question and I'm curious.

[@Sycamore_flaw]

I'm definitely no expert but from my engineering experience when it comes to simulating complex processes with algorithms errors tend to stack quickly and so your result loses reliability the more iterations are ran. Artificial intelligence systems have the benefit of "knowing" roughly what the output result should be expected as, whilst an algorithm has to be tuned either by the operator or a separate algorithmic process (With much the same issues as before). The intelligence model is probably never going to be exactly accurate as it doesn't solve the problem as a series of fundamental physical rules of nature but instead through user defined rules. The physics model will probably never be exactly accurate because noise is inherent in the system. A weird analogy I think of is if the two systems are both set out to solve a jigsaw puzzle the intelligence algorithm takes premade data (That may not be neccesarily accurate) in the form of puzzle pieces and tries to make the picture as you and I would, but will always lose data in the gaps and overall accuracy. On the other hand the physics algorithm attempts to cut out its own pieces which can sometimes be effective but compounds errors as it goes. As I say tho, I am no expert, just having some fun with the ideas

[@josmulder2667]

We simply do no have the compute, because these calculations scale unfavorable. These new models seem to somehow compress the data. Its like the game of life, simple rules can create complexity

[@leejerrett8268]

From what I understand it isn’t a problem of accuracy but of computing power. If an AI gets decent at learning the higher order ‘rules’ which govern protein folding it can look at a sequence of amino acids and give an educated guess on the shape it would likely take under certain conditions without having to calculate millions interactions between hundreds of thousands of physics objects. It makes it viable for a researcher to identify promising novel drug candidates within a search field of millions of possible amino acid sequences without needing to invest in thousands of years of supercomputing time or the obscene amount of lab time and resources that would be required in order to to test each novel substance in vitro.

[@AndreasLindful]

Aging gracefully without pain and be fit and young at the same time. Constant improvement in quality of life. (And wars must be banned.)

[@valentinmitterbauer4196]

Finally, i can efficiently engineer more deadly prions to sell to the highest bidder

[@htopherollem649]

you can have mine, but you must take all of them (CTE sufferer)lol

[@StandardName562]

I have been using AlphaFold 2 for my thesis and it helped me so much understanding the protein I was researching. I am so hyped for this ❤

[@tortysoft]

- bringing the singularity yet closer !

[@koosswart112]

Perhaps I'm too old school. But I would get really excited if we could learn principles or form methods or theories from these improved results that can improve our understanding. I inherently dont like the idea of trusting complex purely to ai without any understanding of the results

[@_ayohee]

The thing is, we already understand a lot - if not most - of the mechanics at play here. We just haven’t had computationally efficient algorithms to do it.

[@koosswart112]

Okay, obviously I dont know anything about protein folding algorithms, so to simplify and generalize my comment - i wish ai would make us smarter, enhance us, make us able to produce whatever wonderful results

[@jan7356]

Protein folding in the real world isn’t as simple as presented. For example, immediately when seeing this, I was asking myself: how do they deal with post translational reshaping. Sometimes other proteins come after the ribosome and modify the protein’s shape. Also: proteins sit in (ionic) water, but most protein shapes we know of have been estimates using the crystalline shape. So being trained to predict those structures will never give you the true structure. I am sure there are hundreds of things like this, and I am sure they know about this. What I want to say is that this probably isn’t as close to a solution to the problem as you would think, and the solution to getting rid of the inconsistencies, that we don’t even know how big they are as we don’t measure the shape of the protein in its natural state anyway, so we have no benchmark, will require a lot more than just a bigger network and more training data.

[@jamesmorrison4976]

Exactly! The complexity is just insane and just brute force AI won’t give any precision to it. They might get better when it comes to longer peptides but protein chains and yes post translated ones will be a real pain to computerize and simulate.

[@Pattythebiologist]

If multimodality is taken seriously at OpenAI, then in 10 to 15 years tops we will have a general artificial intelligence that can be integrated smoothly with some of the brightest minds in the field to combat diseases that we can only dream of fixing now. Ready to see what the future has to offer in the medical field!

[@simonpenelle2574]

WOW this is amazing. Wonder if we can build a text to protein transformer with this technology to find new proteins based on textual descriptions of use cases and associated receptors...

[@marcosfuentes980]

I would love to see it use to create a nootropic, some thing like the NZT-48 of limitless

[@m.b.9061]

The first question I have about enzymes, that decompose plastic is. How do you make sure that the enzymes only decompose the plastic, Which one do you want to have decomposed? Especially when they are used in the environment and not just in the laboratory. But the AI is very Impressive.

[@ajaykumar-ve5oq]

great progress lets see if it can bring back hairs

[@derasor]

Definitely looking forward for other molecules, along with their interactions (the dynamism you referred to) is imo the most exciting possibility. The fact that this may be doable in the next few years, given these truly amazing developments, is just totally mindblowing.

[@nickrobinson7096]

I was in my final year or biochemistry in 2018 and we were discussing the protein folding problem where the lecturer said estimates for solving it are around 25 years aways. Next year alphafold went mainstream.

[@marinomusico5768]

AMAZING ❤

[@killbubbatm5983]

Can't wait to not be able to afford whatever breakthroughs this produces!

[@robertoaguirrematurana6419]

Just a few more papers down the line and it'll figure out abiogenesis, what a time to be alive!

[@abowden556]

The user interface, on it's own is awesome, but the model, oh, the model... I didn't think anything could be more revolutionary than the original, but I was wrong. the breadth of applicability has increased VASTLY.

[@egehosgungor2897]

Holding my papers !

[@Sancarn]

Not sure about applications of alpha fold, but the design of the AI is a similar problem to that of drawing subcatchments in waste water modelling, which is currently an extremely labour intensive exercise.

[@Naxt366]

Hoping for a quick industrial scale solution of plastic recycling. Please make it open source!

[@n0o0b090lv]

This is gone be wild

[@E2_1B]

Thankyou for sharing!

[@ShanilPanara]

Absolutely huge!

[@hardwareful]

Károly had to add quite a few more buttons to his speech synthesis sound board for this video.

[@KilgoreTroutAsf]

Original Two Minute Papers: lets explain this incredible advancement and discuss the algorithms in detail Two Minute Papers now: let's soyface for ten minutes at an extended press release from a corporation's marketing department while looking at pretty animations

[@gdehms]

Can it tell how I'm going to fold my papers while I hold them?

[@flencko9290]

2 minute papers* * - but explaned more like in 20 minutes, but it's so awesome you don't mind to stick around till the end

[@tschaboo19]

The future is now, old man

[@KCJbomberFTW]

As a Beer transport researcher I got excited at the low score of 7:33 no IPA😂😂

[@theh1ve]

My goodness this is so exciting I let go of my papers!

[@Dom-zy1qy]

Crazy how we beat biochemistry with the derivative.

[@pooritech]

The fact that generalist AI has more performance doesnt seem surprising. Machine learning seems to be a way to describe intuition. Abstracts weights that do not directly correlate to physical parameters but that together allows us to draw a bigger picture. The AI has no concept of what it is doing, it is just doing it : input, update wights, chrun out whatever. This is analogous to observing, thinking about it for a bit, and saying whatever it lead you to. The more informed your intuition is the best.

[@dennyw2383]

In 5 years, this paper will have as much impact to human live as Attention is All paper.

[@EternalDensity]

When you mentioned the generalist game-playing AI, I thought "I wonder whether it's any good at Satisfactory" cos I've played that a lot this week. Cut to footage of Satisfactory, lol

[@animanaut]

forget about mangled fingers on a grid of pixels. this is what it's about!

[@JamesAwokeKnowing]

would like to see alphafold assisting in creating metamaterials where genetic sequences form desired shapes and the shapes are preserved/hardened via enamel (tooth enamel)

[@cam_8528]

yeah dude alphafold is a mad good investment Google made

[@goodtothinkwith]

So a transformer based diffusion model for 3D molecular structures? (Is that right?) The question is: how much closer does this get us to individualized medicine?

[@aperson2730]

Thanks for the info 👍🙂

[@Old299dfk]

This is really nice for the rich people who'll be able to benefit from this 😊 Good for them.

[@dref_one]

Amazing! What a time to be alive! 🧬

[@eric.is.online]

Can't wait until this hits localColabfold