Has Protein Folding Been Solved?
Вставка
- Опубліковано 14 тра 2024
- NordVPN special deal: Just go to nordvpn.com/sabine and use our coupon code sabine at checkout.
Recently Deepmind made big headlines with its AlphaFold success. Did it really "solve" protein folding? What did actually happen? In this video I explain what the protein folding problem is, why it's important, and what the current situation is.
The protein animation shown at 1 mins 13 seconds goes back to this publication:
Structure of the Cdc48 segregase in the act of unfolding an authentic substrate.
Cooney I, Han H, Stewart MG, Carson RH, Hansen DT, Iwasa JH, Price JC, Hill CP, Shen PS.
Science, 365(6452): 502-505 (2019).
doi.org/10.1126/science.aax0486
Reused with permission.
If you like this video and want more like this, you are helping us a lot if you support us on Patreon:
/ sabine
#science #education #biochemistry
0:00 Intro
0:35 What's the problem?
1:49 Why does it matter?
5:07 How to try and solve it
6:27 The CASP Competition
8:10 Alphafold 2
9:07 What does this mean?
10:42 Sponsor Message - Наука та технологія
NordVPN special deal: Just go to nordvpn.com/sabine and use our coupon code sabine at checkout.
Hello Sabine, I am 12 years old and my dream is to be an astrophysicist and solve the theory of quantum gravity.... But can you please answer me these questions ???
If the graviton boson really exist can it develop quantum gravity and it will be found in the standard model of particles ????
What do you believe more , is it String theory or loop quantum gravity or a new theory????
Thank you very much 🌹🌸🌷🌻🌼🌺
thats SAD
in one of the interview with You , you sad that the reason for yt channel was something else then money from commercials...
Well, Now I finally understand how to pronounce your last name. I shorted you a syllable prior, but no more. Nice to see you again Sabine. :) Thank you for keeping us educated on what's going on in science without the nonsense. Have a great day.
@@anko6999 would you turn down money if I paid you to do what you were gonna do anyways and all you gotta say is a nice few words about a genuinely good product?
Why do you have to waste my time and attention with commercial advertising, basically spam.
Why do I waste money on you via patreon when you're now forcing me to watch advertising here?
Any more advertising and I'm unsubscribing.
Refreshing to hear someone use “exponentially” when it actually applies.
I wish I had more than one upvote to give to this comment :-)
What like a theoretical physicist? Would be sort of surprising if she didn't.
@@martinda7446 Sure, but for every Sabine that understands it and uses it correctly, there are hundreds, thousands, who use it incorrectly. It’s a pet peeve of mine and, judging from the other comment and the many “Likes” I’m not alone.
@@TIO540S1 except it shouldn't be a surprise that Sabine is using it correctly. She's a theoretical physicist, after all.
Your comment would make sense if we were on a video not related to math or physics and the presenter had a totally different background where you would normally except the word to be used incorrectly, but actually wasn't.
In other words, it's not surprising at all that Sabine knows how to use these terms properly. I would be much more surprised if somebody from a cooking channel, for example, got it right.
@@TIO540S1 Yes, you are right of course. It just at that moment seemed strange to praise a mathematician for using a common mathematical term.
I though have made much the same comment myself so apologise... I have too much time on my hands..
The silver rule of broadcasting: Whenever there is a question in a headline, the answer is always NO.
Is this called the Betteridge's law of headlines?
Always seems like the least likely answer, not "no or "yes" specifically...
Tell that to the guys who wrote the arxiv papers I'm reading...
Your conjecture seems to work.
Daily Twerp - " What's half of mid-day? "
Protein folding is not that much a holy grail, nor is 10^50 or 10^300 or 10^500 a big number. The AI methods work. Your comment is not intelligent.
A few clarifications:
CryoEM doesn't generally give better resolution (it was actually worse than crystallography until recently), but it also works on many proteins that are not amenable to crystallization (e.g. membrane proteins). That's why it's so useful.
Also, oligomers are not several amino acids interacting, but several distinct amino acid chains (i.e. proteins).
U work in Biophysical Chemistry?
@@aniksamiurrahman6365 no, but a related field.
Thanks for the clarification! Sorry about the sentence with the oligomers. I now realize that it can be misunderstood. The phrase "in which several amino acids are interacting" refers to the "interesting cases", not to the oligomers.
@zeawoas I believe the point being made here is that the amino acid groups that make up the oligomer chains can have complex intermolecular interactions between each other
Q: How are proteins folded? A: DNA
Sabine - really appreciate you taking the time to explain these complex topics to us. Thank you!
Happy you find it useful!
@@SabineHossenfelder Unfortunately, there is no official website explaining how the Big Bang responsible for creating time and space occurred 14 billion years ago . So
We ask physicists to develop a nine spatial dimensions membranes theory This is in order to describe the four theories, which are supersymmetry, the Higgs boson, (Gluon-quark plasma) and (electrons and a positron)
We request the development of the mathematics of the theory of membranes with nine spatial dimensions in order to describe the four theories
Submit the application as well as this necessary link on membrane theory to physicists and mathematicians This is the link
theory.cern/events/recent-developments-m-theory
.
@@SabineHossenfelder very very usefull madam, much appreciate your work (edit:typos)
@@user-pu8ch1lh3f wtf?
My headphone cord thinks it's a protein...very annoying.
Plastics come from crude oil which come from organic matter. The plastic coating on the headphone cords may have proteins in them :) Wishing you a Blessed New Year, Lord-Jesus-Christ com
So, you need an A.I. to unfold it!
@@v.gedace1519 That would be very helpful!
The Protein folding problem was almost solved by Alphafold; the headphones folding problem was 100% solved by Apple. :-)
i know exactly what you mean...
As a biochemist who worked both with x-ray crystallography and cryo-EM, a few comments: a) this is an excellent video, but i also never expected anything other than awesome from Sabine b) even if alpha fold performs less well for complexes or unusual folds, still it would be great if it could releive the burden of having to determine the structures of hundreds upon hundreds of similar proteins. This would free up valuable people and equipment to concentrate on the very unusual and complicated structures c) even if it doesn't revolutionize basic biochemistry, it can be of tremendous help to pharmaceutical research as they often work on relatively common folds (e.g. spike protein of Covid when you already know the spike structures of several other coronaviruses)
Was thinking the same. The better in silico (computer) modeling we can do, the more we can rely on it for hypothesis generation and to help determine where to dedicate resources of time, effort, and lab time. It doesn't replace the bench scientist. Real world experimental data will always be needed, but this helps us get to meaningful real world data faster.
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
@@hyperduality2838 There is always one...
would you mind explaining why the protein takes the same shape in crystal form and in solution? this is not at all intuitive.
@@nraynaud This is actually not necessarily true. It depends a lot on the environment, that is mostly represented by interacting water molecules, which can play decisive structural bridging between aminoacids. This makes the crystal formation a key point in the process. Sometime you cannot even obtain crystals from a protein. For instance membrane receptors have always been a nightmare, since their functional 3D shape depends on the lipid layers in which they are embedded.
Sabine's social work in disseminating knowledge on such a wide range of scientific topics is invaluable.
She only did forget that all science allowed is always the science for the riches interests only. Dont be a silly puppet.
@@miguelferreiramoutajunior2475 I'm absolutely confident Sabine does not share that sentiment. Science is for everyone. It makes all our lives better. Anyone and everyone should be free to contribute what they can when they're able UwU
@Manmeet Colon well, her work is invaluable in the sense that it's priceless but she's definitely very valuable.
English is crazy
@Manmeet Colon invaluable things don't lack value - their value simply can't be quantified ;)
@Manmeet Colon what does the government have to do with the value of Sabine making science accessible to the public?
"That's almost as many vacua as there are in String Theory" -- priceless 🤣
Sabine going savage mode
I'm glad I wasn't take a sip of anything at that moment. It definitely would have been a spit take.
shots fired
Michio Kaku: someone talking about me?
Can someone explain what she means?
Hi Sabine, I'm echoing what a lot of other folks have already stated. I really enjoy the diversity of topics that you present, and your well thought out presentation too. Thank you.
Excellent. I learned so much in 12 minutes about a subject I was totally ignorant of.
Thank you for clarifying this difficult information for lay people like me. Yours is a vital service and you are
so good at it. Getting the general public to 'somewhat' understand the amazing discoveries coming so fast
now is urgent.
I just started my PhD in peptide folding and chemistry and then comes this news!! 😅
I think it's great as it should open up new opportunities for protein engineering. There is still plenty of stuff to do at the bottom!
@@thatyougoon1785 Ya, it will eventually save a lot of time and money, plus structure elucidation is only a part of my research not the whole. So I hope I won't be jobless after my long arduous PhD.
Congrats on your new journey. Best wishes! 😊
@Heloise O'Byrne I will have to read a lot before I can answer that.
@@subhrodeepsaha9245 What else do you you do for your PhD? I'm currently doing my master in nanobiology (which is best described as biophysics + system biology, the name is a bit weird)
This was just an absolutely fascinating topic. I had heard of this problem years ago, but never bothered looking into it. Sabine's ability to unravel the complexity of such issues is remarkable.
''some of you may have folded proteins yourself''
YES I HAVE
WeLL
IM ALIVE, that's why!~😂🤣
I'm like constantly folding proteins. :) Literally day and night!
@@ListenToMcMuck deep philosophy. 🤣
Yeah this deep mind crap isn't impressive. I can fold my proteins at 100% accuracy in microseconds, and I can do proteins I've never even been trained on!
@@ListenToMcMuck Yes.
I’m still doing it, doing work on the distributed computing project, Folding@Home. I have had their software on computers for over a decade.
Congratulations on 200k, much deserved
You deserve a lot of respect for exposing us to different domains and questions in science. In a very understandable way too.
CASP updated the scoring algorithm in 2014 and the top scores have increased linearly since then.
The C-alpha “IDDT” distance scoring is based only on C-alpha positions - that is one atom out of all other amino acid atomic positions. There are other scoring calculations too.
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
I am glad the scores did not increase exponentially!
Email
*Squints eyes and reads "IDIOT" scoring system*
"That's as many as there are vacua in string theory" do we live in a world where string theory is more widely understood than protein folding?
Possibly. But almost certainly Sabine lives in one.
The difference is that proteins are very real.
I presume that she made a joke that I can't understand. She could probably tell us we are complete idiots and wouldn't know. I just watch her because she is pretty.
There are more protein fold possibilities than there are atoms in the universe.
How does one count units of vacuum?
With String theory!
More vacuum counted = more evidence gathered?
I'll see myself out.
This is a very, very well done video. The protein folding problem was put into perspective without over or under dramatization. Excellent delivery.
I've seen a few of your videos in passing before, they're always very good. Subscribed today :)
Thanks. I really do enjoy my weekly fix of interesting stuff. I was actually wondering how they could test that claim of solving the protein folding problem - now I know.
Great explanation and nice seeing more Biology content on here, as a Biologist myself!
Thanks for the feedback :) I had some help for this one though.
@@SabineHossenfelder more biology content please!
I suspect Sabine would say that biology is merely a specialized branch of physics.
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
@@SabineHossenfelderany chance of a video on your take on 'Quantum Biology' sometime Sabine?
Love the subtle String Theory dig. Good job!
The fact so much of these problems were being tackled in the 90s and before is mind blowing. The possibilities of a collective mind
The problem with AI is that even when it understands something, it can't teach us anything about what it's learned.
That is a good point as far as learning nothing about how proteins actually function as 'molecular machines'.
Its a problem with a multilayered neural network, aka deep learning.
Not with the buzzword AI itself. Deep learning and AI are not really the same. One could claim that AI is an umbrella term that includes neural networks and deep learning.
Many other learning algorithms can easily give us viewable solutions we can learn from tho.
Machine learning is analogous to our brain’s intuition or unconscious behaviors. Those are not interpretable either.
e.g. a professional basketball player can aim at a net from a distance even if in motion. This takes a highly complex coordination of visual and motor functions to achieve, yet it’s seemingly effortless. The brain forms a model of its environment (an internal representation / embedding) and uses it to navigate/interact with it. This is how ML works in essence. It converts a complex ‘environment’ into a more manageable representation. A task that’s intractable in one space, becomes almost trivial in another.
Is interoperability important? I think we are giving it more credence than it deserves.
Interoperability with humans is a good idea, regardless of whether it is a concern. I mean, eventually these things will be doing science experiments, too, and what is the point of having an AI figure out quantum gravity if this knowledge dies with it? Sooner or later, they will be writing or maintaining other AIs and managing a missile defense network; do we really want to make it impossible to audit their decisions?
@@davidhand9721 I agree with the overall sentiment. I think, the problem with AI is the use of the term AI. It gives people the wrong impression.
Also, It’s not that these things are black boxes to us. We architect them to purpose with desirable inductive biases. Take AlphaFold for example. It has two parts, the first one being ML the second is a minimization function (basically a kind of force directed graph physics model). In a nutshell, the ML part predicts a pair-wise proximity matrix. This is a square matrix with rows=columns=number of amino acids in your protein, so the diagonal is constant (comparing an amino acid to itself). Given we are dealing with chains, the distances should grow as we are moving away from the diagonal. However, due to 3D geometry and folding, we find amino acids that are distant, along the chain, to appear close in 3D. We train the network to predict these distances. It’s basically translating the input features (base pairs and equivalence info etc) into these distances. Once you have those, you run a simple physics model that tries to conform to those predicted distances, allowing the chain to fold in 3D (the second part of AlphaFold). This is highly interpretable. Sure, you can’t interpret the individual weights of the neural network, but it’s the emergent behavior we are mostly interested in. More often than not, NNs are over-parametrized, thus giving multiple equally performing solutions. So individual weights are of no particular importance. Or in other words, how and why the network converges is important and heavily studied, what the network converges to (the weights), less so.
By the way, I mistyped earlier, I meant interpretability not interoperability. The latter is of course vitally important. I’m arguing that the former is not as much of a problem as some make it to be.
I just love the way you explain all these things and I easy that sounds thanks
This kind of channel is the best use of the internet. Thank you Dr Sabine!
You are one of the best and honest youtuber. No hype, to the point explanations.
I really appreciate how you out the science back into popular science, you make real science popular, not that phony science that is found in self help books or on certain pod-casts that shall not be names 😶
Joe rogan lol
Superb as always. My lay-person understand is significantly enhanced! I am thinking that the folding of engineered proteins might still be much harder to predict as they will not have run the gauntlet of natural selection.
There are proteins whose task is to specifically help fold proteins as they are being created.
Most proteins that are made in the body are actually folded improperly. They will be disposed of in lysosomes.
Cells prefer quality over quantity.
Love your videos! I’m a biomedical engineering student and I find your channel really informative.
You have become one of my favourite content creators
Thanks for the clear cut explanation, Sabine 👍
Great video! I would love to see more Biology and Chemistry content here.
Protein structures can also be studied using NMR. This has the advantage of not having to crystallize proteins and having them in their "native" form in solution.
Great video as usual!
There is still not an easy way to get NMR structures of proteins bigger than 50-90 kDa. Very few larger than 35 kDa have been solved.
Thanks again Dr. H.! You are helping build critical context for public grasp of important and relevant science, industry and social action.
I love your sense of humor. I laughed loudly when out of the blue you said ten to the three-hundred possibilities, that’s almost as many as vacua there are in string theory.
How do you decide your dress and hair style? You are fantastic in this video!
Interesting and clearly explained as always. Many thanks.
Wonderful video, Sabine! 😎
Thanks for the feedback!
I clicked on your song Outer Space right after your lucid explanation about protein folding. Imagining proteins fold while watching you dance created weird mental images ;-)
Anyways thank you for teaching us this important topic. May you and your family have a great weekend!
Another great video - I appreciate the wide range of topics that you explain in your videos. Thanks
Proteins need chaperones to fold, the correct enzymes.
For example, for human mitochondria, after synthesis on cytosolic ribosomes, nuclearly encoded mitochondrial precursors are imported into the organelle in their unfolded state ( due to the limited pore size of the import machinery. ) After import, mitochondrial precursors must efficiently fold into their functional structure, to avoid aggregation due to exposure of hydrophobic surfaces.
In the mitochondrial matrix, [ human mitochondria contain their own genome encoding 13 core subunits of the oxidative phosphorylation (OXPHOS) machinery ] two highly conserved chaperone systems, HSP60 and mitochondrial HSP70 (mtHSP70), are critical in facilitating the folding reaction of mitochondrial precursors.
Proteins don't fold in isolation of their environment or intended functions. Sorry to invoke intentionality.
"Proteins need chaperones to fold, the correct enzymes"- wrong at different levels. Only a few massive proteins need help in folding. Chaperones are proteins themselves and the question will be asked as who will help the chaperones to fold correctly? Also not all proteins are enzymes but most globular proteins have a well defined folding pattern and can bind some substrate. Also, many well folded proteins have random coil regions that are not seen in the x-ray diffraction (because they are random).
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
I had a project at NIST where my job was to write a program that could model how proteins plump up in an aqueous solution! Maybe it finally can be put real use haha!
You never cease to amaze me. So far I thought that you were a rebellious theoretical physicist occasionally singing pop music and uttering catchy quips. Now you come out to explain an intricate problem of biochemistry. 👌👏🏻👏🏻 Awaiting more such pleasant surprises.
Botton down, botton up and not anywhere closer to solving this problem. I have always been worried about the folding problem myself, since I first heard of the problem in elementary school. Anyway hope we are almost there and can solve this mystery before I departure. Fantastic lecture and I appreciated every minute.
A few months after this video was made, Alphafold made a little more progress, becoming usefully accurate for the easy cases. Although far from solving the overall problem, this still saved several man-centuries over older cataloging efforts, which could lead to much faster drug discoveries, and shorter clinical trials.
She knows how to stimulate our minds. Keep it up please.
I am 13 years old and i am learning organic chemistry and biochemistry, i think you're a kid and interested in biology :) right!??
Amazing !Clarity and comprehensiveness in this brilliant video.Thanks!
Such a great video. The purity of the educational content is so high.
.. I like and appreciate your infomative, matter-of-fact, neutral views.
Good string theory burn!
This is a helpful intro to protein folding. Thanks Sabine!
Been getting this in bits and pieces. Thanks for bringing it together for me.
How wonderful it would be if Sabine were my next door neighbor....with a LOT of patience for questions :) She is just brilliant and so enjoyable to listen to.
Everybody's your next door neighbor on the internet UwU
The Ai team just exposed how inefficient the other computational techniques have been. Molecular dynamics in particular, is such an inflated but inaccurate tool, that maybe is time to start thinking about the seriousness of "bottom up" approaches. Starting by the fact that you cannot simulate non-covalent interactions at the required accuracy even with quantum mechanics for big systems just yet. Much less with Newtonian mechanics, as 99% people in this field foolishly do. But they get grants, so why they will gonna change that? Out of embarrassment? Certainly not.
I sometimes donated hundreds of hours of my devices' CPU time for folding simulations. Was that all a waste?
"The Ai team just exposed how inefficient the other computational techniques have been. "
Indeed.
I do MD. Its use is very limited.
I do not like the idea of AI as well: "no one really knows how A.I arrived at the decision". Physics is lost. We have magic only at our disposal. And Sabine have been against magic, religion... ;(
This is the first time I gave "hand down" to Sabine.
Dr B: if you are involved in these thought issues (I guess you are), perhaps we could write together some very short article against that mess around against AI? I am at nanophysics.pl
Well, simulating the whole protein with quantum mechanics is Impossible. So they choose to simulate only the active site with QM and rest with Newtonian mechanics. Deep mind, didn't even try to simulate those at all. It just figured out patters from existing database and made some educated guess. Their educated guess is better, but it didn't reveal anything about the underlying physics.
As for how Deep mind's algorithm is better? It has more to do with the fact that they can employ AI researchers worth of a million dollar each, people only available to companies like Apple or Google. After all u can't expect top-notch result with old, low-power hardware and heavily underpaid postdocs, can you? Factor that in and Alphafold actually falls behind a lot in "bang for buck" than most research groups.
If AI produces results that are much superior to the ones obtained with the “physical” approach then the funding for that approach should dry up very quickly. Perhaps future AI developments (in a decade or two?) will even be able to show us how to solve such problems using a physical approach.
Thank you Sabine for another wonderful video, also for mentioning the need for a protein that breaks down plastics. Any help with that is needed right now.
You really open my horizon, I am deeply appreciated your teaching.
The “protein folding” problem(s) are one thing - what appears to be shown here is a “protein folded” problem - various protein constructs folded in a plethora of expression systems and crystallized or otherwise isolated in states conducive to structure determination - notably not including NMR methods - and deposited in the RCSB. That’s a magnificent endeavor but the programs are not showing *how* or *what* the protein went through to fold - it is showing the likely final result in the RCSB except for NMR structures. I think fold space is filling up, so this might indicate the methods to determine structures are not discovering highly unique structures at the moment.
Is the work peer-reviewed yet?
That annoys and concerns me a lot. I now fear that research groups that are investigating the underlying physics - how protein structure emerge from myriad of non-covalent structure will get less and less fund as industry and government will invest in this superficial pattern matching BS more and more.
You do realize we're not all biochem postgrads, right? Despite my ignorance, my intuition tells me you are making a valid and interesting point. I just wish I could understand it better.
@@alexh1524 I'll try to write down a longer-winded interpretation of me reading @Bryan Lepore's comment, from the perspective of a master's student in Bioinformatics:
There is a very clear-cut difference between measuring or predicting the *process* of what happens as a protein folds (i.e., as it somehow goes from a chain of amino acids to a structurally functional machine) on the one hand, and measuring or predicting the typical final state of such a folding process, i.e., measuring a single "folded protein structure". There is much to learn from both of these, but of course there is much less overall information contained in just the final structure as opposed to the entire folding process.
Another issue is that proteins, inside a living organism, are not structurally static even after they are fully folded. The degree to which this can be important varies from protein to protein, so some proteins are pretty stable overall and just have atoms wiggling around mostly locally, while other proteins critically depend on their folded structure changing over time to achieve their respective function.
AlphaFold (and AlphaFold 2) just predict that final structure. They don't tell you much, if anything, about the folding process. So, if they don't know about the process (if they did, they could just tell you it), how can they perform well in predicting the final outcome?
They're trained on the RCSB PDB (the biggest and most important protein structure database) and the data in the PDB is, from a certain standpoint, overall pretty homogenous: a huge portion of entries consist of exactly one structure from X-ray experiments. This makes for a pretty huge bias in the data, towards proteins that crystallize well (and maybe also proteins that are structurally similar to these well-crystallizing proteins, but don't crystallize well themselves).
So, one explanation for AlphaFold 2 doing this well might be that "whatever AlphaFold 2 knows is mostly all there is to know about protein folding"; but another explanation, the one that Bryan seems to be implying to me in their second-to-last sentence, is this: "the data in the PDB today maps out quite well the kinds of structural motives and structural interdependencies we see in measured proteins, so a well-designed model trained on it will work well, *on the kinds of protein structures we can measure* ". The most important point being that there is probably still an _ocean_ of protein structures and protein dynamics that we, for the most part, are unable to measure and map out today. So even with AlphaFold's progress, understanding proteins and protein folding is still far from being solved.
@@gooblepls3985 I don't think the guy understood it any better. But wow! I'm a Biochemist and you write really well.
@@aniksamiurrahman6365 Well thank you for answering for me, but let me give it a try. At the risk of oversimplifying, it seems to me that the problem boils down to "measurement." You guys call it measurement but in laymen's terms it's more like shape observation. The database is full of protein shape samples that are readily observable because these are proteins that crystallize. The missing inference being that somehow crystallization lends itself to proper and accurate shape observation. So the database is severely lacking in other protein categories whose shapes can't be readily measured/observed, and since there is 10 Exp. 140 possible shapes this is a real bummer.
With regards to the AI program that has shown remarkable improvement in predicting shapes, I was under the initial impression that this improvement would somehow lead us to new insights into the process by which an amino-acid chain transforms and folds into a final protein structure. Again, I am oversimplifying, but a successful algorithm, heuristic, and/or neural net should be able to tell you how a given set of initial conditions leads to an ultimate result. You guys throw shade at this assumption, and I still can't wrap my head around how a successful method for predicting ultimate outcome does not need to incorporate knowledge regarding the process that led to the outcome.
In any case, it seems to me that if we had better tools for measuring/observing all types of proteins, we would make significant inroads into understanding the folding process. I imagine that with the proper tools, the folding process could be observed at any point in time and valuable knowledge could be gained regarding the transformation process.
"That's as many as there are vacua in string theory"
And this is why I love Sabine
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
@@hyperduality2838 "All energy is dual"? What is that even supposed to mean? And where are you going to find duality in the three body problem?
@@SpectatorAlius Energy is dual to mass -- Einstein.
Dark energy is dual to dark matter.
Gravitation is equivalent or dual to acceleration -- Einstein's happiest thought, the principle of equivalence (duality).
Potential energy is dual to kinetic energy -- gravitational energy is dual.
Apples fall to the ground because they are conserving duality.
Electro is dual to magnetic -- Maxwell's equation.
Positive charge is dual to negative charge -- Electric charge.
North poles are dual to south poles -- magnetic fields.
Electro-magnetic energy is dual -- waves are dual to particles -- quantum duality.
Energy is duality, duality is energy.
The conservation of duality (energy) will be known as the 5th law of thermodynamics!
"Always two there are" -- Yoda.
The three body or mass problem is a side issue to deflect you from finding a great truth.
@@SpectatorAlius DNA is made out of atoms hence energy, it is storing information in the form of duality hence it should be called the dual helix as it contains two strands which are equivalent or dual to each other. The code of life is fundamentally dual.
Mind (the internal soul, syntropy) is dual to matter (the external soul, entropy) -- Descartes.
Concepts are dual to percepts -- the mind duality of Immanuel Kant.
@@hyperduality2838 As someone with a background in physics...just no. Don't call something energy when it clearly isn't; energy has a very specific definition, and you are misusing it.
This is the first time I am hearing of this. Thank you, Sabine 🙂
5:50 I had no idea. Thank you for sharing!
Very interesting! A couple of question: what makes a fold stable? How can software determine if a fold is stable or not?
The stability problem is a question I hoped Sabine would talk about. If I'm correct, a fold (like any other chemical structure) follows the "local minimal energy" requirement: for example H2 and O2 are more stable than atomic H or O, less stable than water. Why? atoms in molecules are directly bound by electrons "orbiting" mre than one atom (i.e. covalent bonds or as I call them "shepherd electrons"), or by electromagnetic fields (ionic bonds, van del Walls forces, etc.) A molecule/bond is then in a stable state when the energy required to stay in that distribution is a local minimum (I imagine that's a good reason for physicists to try to solve this problem).
A problem here, which Sabina hasn't talked about but I have read in some other comment, is about "protein denaturalization" i.e. protein changing its structure followinf a change in the environment (something we do day to day when we cook and eat). And that opens an even more important question: does crystallography/deep-freezing microscopy alter the protein fold before the imaging process?
As for your second question. Again I could be wrong, but many AI algorithms don't know they're providing a good solution. In the training phase, humans are inputting known data and "train" (directly or indirectly modify the AI model) the software until it produces the desired output (a stable fold in this case). Then in production phase the software is feeded with unknown input, it follows its training/model to produce an output, and it's up to humans to decide if the output is good (in this case, by running the CASP contest). In short, an "AI" is indeed artificial, it has 0 creativity (outside some "variations on a theme"), and it lacks the concept of correctness (that's the domain of humans - borrowing from an old book, it's up to us to tell good from bad).
Mostly a protein is stable when it's entropy is minimized. In practice, this is (for the most part) a function of burying amino acids that don't like water on the inside where water can't get, and to a lesser extent, reducing the amount of similar charges (positive or negative) that are right next to each other
One way to answer is that a fold is stable if changing the fold (conformation) to another fold (conformation) requires a lot of energy. (heat).
To computationally test the stability of a fold you can run molecular dynamics simulations in the temperature range of interest. If the general shape of the protein stays basically same with only minor wobbling and oscillating near "the" stable fold, the fold is indeed stable.
Note that a protein can have more than one stable conformation. These can be thought as local minima of the total structrular potential energy of the molecule. Typically enzymes turn other proteins "on" or "off" by lowering the energy barrier between two possible stable folds. On the other hand, allergens like nickel typically attach themselves to a healthy protein, forming a complex with a new stable way to fold. If the immune system classifies the new stable conformation as an alien protein, it means being allergenic to nickel. Note that the immune system ignores nickel alone, but reacts to the nickel protein complex.
@@rikulappi9664 You can also directly assay the denaturation (unfolding) of a protein directly with experiments called differential scanning fluorimetry. The general principle is to monitor either the fluoresce emission of tryptophan changing as a function of its local environment (ie. it unfolded), or the occurrence of fluorescence from a dye that only fluoresces in the presence of hydrophobic residues (which are normally buried and inaccessible) as temperature increases
Then unnerving thing about AIs is that they mostly work, but they always have edge cases where they fail, and don't know they've failed. In fact they're often quite confident about their erroneous prediction, and fixing them is difficult. I imagine optical illusions are this phenomenon in humans, but we at least understand we're wrong.
@Heloise O'Byrne At the present moment I would say that AIs are worse than us. Most of the cases where they are better than us have narrow applicability, and as you point out they require training by programmers. Computer Scientists talk about an artificial general intelligence that can learn on its own, but we have no roadmap from current state to it.
AI is in its infancy. It will soon surpass us in every category.
@@bozo5632 They said the same thing when I took undergraduate AI in 1986.
@@mheermance has any progress been made since 1986?
@@mheermance I'm not one of those "exponential growth forever, singularity tomorrow" zealots, but I do see it coming. "Soon."
Such a good presenter, few people combine both knowledge and no-nonsense charisma like this. Its refreshing to get an outright take.
I first came for the music, then stay for the science. Doctor you are amazing at explaining things without the gobbledygook. And I thank you for that.
Protein folding? Sound interesting, thanks, but that remind me I need to do the laundry...
9:28 That is actually called machine learning, not artificial intelligence. And it's not entirely correct that it can only learn on existing data, AlphaGo learned without any data, governed by a set of rules which helped to determine whether one option was more preferable than the other. From evaluating those rules on sets of data from initial runs with a random values, the machine learned which one values are better for the task. That's is probably just how it learned folding too, since it's pretty easy to determine which resulting fold is better than the other...
You mean reinforcement learning?
The Reinforcement learning
It seems you don't need to train on data from every event.
Example: GPT-3 was trained on a lot of text, but training did not include all possible 3 digit addition.
It still learned to predict addition without actually implementing addition. That's the fun - it can find hidden relationships.
@@JohnBoen Yeah but the dataset nevertheless has to be diverse. For example an AI trained to distinguish a male and a female should look at all possible variations of men like african man, asian man, man with a blonde, red headed man and every other possible kind of skin tone and other facial features and the same goes for women. We need not to show it each and every man to have ever existed on earth but still the dataset should have all possible features.
@@KishoreKumar-uz8ir thats why I mentioned GPT-3. Huge data set...
finally I understood this problem, thanks for the nice explanation
I used to do Folding@Home, back in the day. It felt good to help science, even in such a small way.
It’s gotten a new lease on life, what with COVID-related proteins being simulated...
As a former combichem drug researcher, I must admit, this makes me want to be more rational.
more rational?
@@happylittlemonk If you want a serious evaluation from a reputable biologist, may I propose asking a professor at a university instead of the general youtube comment section public? Also, you should probably be the one, who writes the paper for your own theory. If you struggle with writing, there are courses to improve that. If the reason you struggle with writing down your theory is really that your theory is more of an unclear convoluted hunch, (just a guess of mine, maybe unfair, but triggered by your unorthodox attempt to search publicity for a world-moving theory) it is your job to put in the work to make sense out of it.
In a 4 dimensional system DNA (A, C, G, T) there are 20 amino acids + 1 stop amino acid.
In a 4 dimensional space/time (T, X, Y, Z) there are 20 components of Riemann curvature + 1.
Protein folding is based upon Riemann geometry or curvature!
Positive curvature is dual to negative curvature -- Gauss, Riemann geometry. Curvature is dual.
Amino acids, proteins and the DNA code are based upon duality as all energy is dual.
DNA or the double helix should be called the DUAL helix, it has two dual strands.
Duality creates reality!
Matthijs Slijkhuis
Good point! Well done...
@@mslijkhuis Two. The one strand is dual to the other.
Gravitation is equivalent or dual to acceleration -- Einstein's happiest thought, the principle of equivalence (duality).
Energy is dual to mass -- Einstein.
Dark energy is dual to dark matter.
Space is dual to time -- Einstein.
Time dilation is dual to length contraction -- Einstein, special relativity.
Potential energy is dual to kinetic energy -- gravitational energy is dual.
Apples fall to the ground because they are conserving duality.
Action is dual to reaction -- Sir Isaac Newton.
Certainty is dual to uncertainty -- the Heisenberg certainty/uncertainty principle.
Electro is dual to magnetic, electro-magnetic energy, light, photons are dual.
Positive charge is dual to negative charge -- electric fields.
North poles are dual to south poles -- magnetic fields.
Energy is duality and everything in physics is made out of energy.
Syntropy (prediction) is dual to increasing entropy -- the 4th law of thermodynamics.
Teleological physics (syntropy) is dual to non-teleological physics (entropy).
"Always two there are" -- Yoda.
A pattern of duality exists within physics which means that DNA or the code of life is storing information in the form of duality -- the DUAL helix.
These videos are extremely disturbing. Intelligent, rational explanations in UA-cam videos? Talk about crushing my world view.
I hope your humor is appreciated.
Now we need an acoustician to study this echo chamber.
@@joedart1465 "appreciated" Nope, not by my ex-wife it wasn't.
Thank you for the videos, I will never get enough of them.
Excellent, succinct, communicates a complex understanding of progress on a very stubborn problem. Reminds me of the Imagination and sweat that led to the description of the structure of the heme in the 1950s.
Sabine is really fantastic. She is the best.
It has not been solved. This is just another tool for numerical optimization.
That's a good way to summarize it, yes.
@@SabineHossenfelder They did solve some protein folding problems for which traditional technique had failed to generate solutions.
That's underselling it a bit, no? It makes it sounds like "eh, you can use this numerical tool or another one, doesn't really matter", when the point is that at the moment there is no non-AI tool that comes even close in predictive performance. Of course, it doesn't "solve" protein folding in a classical sense of having a complete theory of protein folding. But honestly I think it is quite reasonable to assume that some problems may just not be solvable in the classical sense.
@@MrGeometres We folded protiens before this tool existed. The search space for protein folding is large. Machine learning was applied are there doesn't seem to good heuristics to navigate the search space. This made some folding problems intractable.
Awesome Sabine! Answered a lot of my questions. Thanks!
Thanks again! The topics you choose are very interesting and you cover it with quality erudition.
But please, please, increase the recorded sound volume 2 more steps. Thanks.
if by "solved" you mean "not solved", then yes, its' been "solved"
😆
aside; I find the high number of shot changes (almost every sentence, and hopping left, right, forwards, backwards) to be quite distracting. Fewer would be easier to follow, I think.
The brief string theory blip at 3:30 was hilarious! I laughed out loud. Nice one Sabine, loved it.
I love your channel you explain complex concepts and making them understandable 👍🏾🙌🏾👍🏾🙌🏾
I dont think folding should be the action term used for proteins because when I fold my clothes this way, i always get into trouble with the wife.
This was actually quite fascinating. Thank you Sabine
An excellent video. Just one small point:- multidimensional NMR spectroscopy is also used to determine protein folding, in addition to X-ray crystallography. The NMR spectroscopy approach makes use of internuclear distances.
I'm more confused about the protein folds than before this video. You bring the supposed problem of protein folding, not exactly, and then you failed to explain it by confusing it further. I don't think you full grasp what the issue really is. Or, you failed to explain it
Thank you . By far the best discussion of this subject that I have seen
Thank you for explaining such complex problem so clearly.
Thank you for helping me understand a scientific problem that I never knew existed. Intriguing.
A biologist here,
6:21 Cryo-EM doesn't give better resolution. Actually, relatively low resolution is one of the drawbacks of Cryo-EM.
In the past, resolutions were very bad, but in the past few years, great development in algorithms used to 3D reconstitute the structure (the electron microscope gives you 2D projections of the structure, and you need to reconstitute the 3D structure out of these projections) and developments in the electron microscopes themselves have improved resolutions a great deal and made cryo-EM a very attractive approach (personally I believe along with predictive AI software, Cryo-EM is the future of structural biology).
On the other hand, the advantage of CryoEM is that it is much better at imaging large protein complex compared to X-Ray chlystarograhpy. You don't need to crystalize the protein (which is a pain in the ass), it is better at capturing biologically relevant structures because the proteins are in solution (before being flash-frozen) compared to the crystallized proteins, and Cryo-EM is much better at capturing protein dynamics (which is very relevant to the function)
Thank you for giving clarification to the claims of protein folding solutions.
I was playing Fold it in highschool Legit randomly remembered and watched this to catch my memory up!
Methods for protein crystallography, and molecular simulations, was my field for many years. I would like to add to your nice explanation why even with the good answers ,deep mind is not quite there; and it comes down to the precision of the answer, and the accuracy to the true answer. If we take catalysts, much of the protein is a scaffold, a frame work to hold a few atoms in the right orientation, but with enough flexibility to enable a reaction to occur. To understand these reactions and how the proteins work it is necessary that the atom positions are well determined, or we cannot tell if a reaction is SN2 or Elimination - and without that knowledge we cannot determine how that protein fits into the metabolism. The the fundamental process of solving the structure of a protein is therefore to discover the function of the protein and we can only do that if the atom positions are well enough defined, and the flexibility can be observed that allows understanding.
Therefore, protein folding must be good enough to discover that, and a score of 90% for the whole protein (or group of proteins) says nothing about the small number of critical atoms. That problem will be solved at some stage, but I would say, not quite yet, as we still cannot replace the complex process of structure determination for folding programs even for quite simple structures.
I feel like I just watched a very well written publication. I'm curious to know what an acceptable percentage would be and whether A.I. will be able to successfully guess the appropriate folds needed for all (or most of) the different forms of proteins. I'm excited to see what the future holds.
Sabine - love your videos. You do a great job of explaining the complex. Besides protein folding, a video on contoured bedsheet folding would be nice. ;-)
Excellent explanation of the problem and potential solutions. Thank you.
Wonderfully fascinating, molecules, evolution, physics and some measure imperfections over Spacetime! Thank you for what you do and share in your videos! My faint understanding is always very much more illuminated during and ever after such encounters.
Thanks so much - new next to nothing about this problem! What a great introduction - and so interesting! I can't get over how fascinating biology really is - unfortunately, they didn't make it so at my school! ;)
Thank you for producing this video Dr. Hossenfelder. Duly liked and shared.
I've been looking into the field a decade ago, before the attempts with AI.
We start with a protein codified by the linear, 2-D code of the DNA. As the ribosomes in the cell finish to assemble the protein, as soon this sequence of linked aminoacids leave the cell, within few milliseconds it folds into the required shape. This happen due to the complex interactions of electric fields the molecule possesses.
My (informal) research consisted in locating, within the molecule, a number of electric dipoles, and elaborating a matrix which completely define the molecule.
Each dipole is a 3-D vector, and the matrix gets quite big quite fast; I just assumed this isn't a problem if a computer with the right power is used for the calculations.
Interesting how harmonic patterns emerges from the molecule electric field.
You rightly pointed out that AI require some initial pattern definitions in order to conduct a recognition.
This is why AI wouldn't initially work in the field of proteomics. But that kind of AI is what Google excels at: picture it, find pattern in the picture. I believe is not going to work well. They don't need X-ray crystallography or electron microscopy, but a nanometer-sized array of Hall sensors, so they can map the dipoles by moving the molecule within the Hall's sensors array.
Regards,
You're the best, Sabine! Keep up the great work! Thank you.
Wow. I just love this explanation. Thank you 🙏🏻