AlphaFold 3 deep dive

[@LookingGlassUniverse]

This video is very long and detailed, so please don't feel you have to watch this one. I've been spending a lot of time understanding the AlphaFold 3 algorithm and this video is the result of that. But if you just want to get the gist of the algorithm then I'd recommend these much shorter videos I've made: AlphaFold 2: kzbin.info/www/bejne/aZi2qpKvg56MeNE AlphaFold 3: kzbin.info/www/bejne/eYrRlIF4grV_gs0 Here are the notes I was writing in the video: lookingglassuniverse.substack.com/p/deep-dive-into-alphafold-3

[@MsSonali1980]

This is fun, I started studying maths, switched to a biology heavy BSc and MSc but ended as software dev and this is the perfect combination of all three fields.

[@zephyrlibs]

Epigenetic attractor network based medicines and cryoEM rocks

[@theprettyrao]

This is me!

[@omnijack]

Pretty inspiring that you studied it enough to be able to explain it in an approachable way (speaking for myself anyway). How long would you say you spent on this learning project overall?

[@LookingGlassUniverse]

Thank you!! About 2 months, but some of that time was learning other relevant algorithms like diffusion

[@LookingGlassUniverse]

Yup. But I mean I first learnt about more general diffusion models etc before looking into the specifics in AlphaFold

[@markglover6971]

Very nice description of AF3 - you did a great job describing this important finding for us structural biologists who are not well versed in AI. If you have time, a follow-up on the ways they assess the reliability of their models would be great. Thanks!

[@zephyrlibs]

cryoEM rocks

[@abhishekdurgude6296]

A really one video i was looking to get a broad picture overview of the alphafold algorithm , i watch several other video but no can has included the details in that depth, thanks once again i really appreciated your work and time, its bit challenging to understad at very first time.

[@richards7502]

Excellent video, as mentioned in other comments, I was looking for actually how AF3 works, and you have explained it extermely well. Now onto a deep dive of diffusion transformers.

[@michaelm358]

A day with a new video by Mithuna is always a GOOD day. These detailed deep dives are just the best! I have a genius autistic son, maths prodigy etc, and we always look forward to new vids. Sending thanks for all your hard work.

[@LookingGlassUniverse]

Thank you for the lovely lovely message!!

[@larsyt80]

Thanks for the great content!

[@LnlyCloud]

Wow this was a great video, despite not understanding all of what was going on and the terms and processes I saw on screen, I felt that you explained it really well and made it accessible to someone who has not previously delved deeply into any sort of "AI" process. Thanks so much!

[@sebek12345]

Do you think this technology is on track to model entire cells and eventually organisms? If not, why not? What will it take to get us to that point?

[@LookingGlassUniverse]

Thank you for the super thanks! That's very kind of you! I think that is the goal for sure, and there's at least one company I've heard of claim that they've done it (www.nature.com/articles/d43747-022-00108-3). I can't say how successful that's been, but I think it will eventually be possible to do. Cells are much much much more complicated than a single protein though.

[@zephyrlibs]

cryoEM, dft and ab initio + ai/ml is the way to go. The human genius of maths is that an mit mathematician has modelled fractal structure of chromatin bundles in nucleus as a fractal. Then proven correct and the ideal state can be a good symmetry check. Chromatin bundles act like attractor networks. Sometimes we needlessly worry about broken and empty places while it could also imply adaptation is change of surface folds and orientation. Helmholtz attractors have flexibility issues that's why we got a brain and neurons to teach our cells by counterfactual evidence. Epigenetics is back. Waddington drew a network below the landscape and that's big deal. You can also auto heal. But also create intelligent medicine

[@sebek12345]

@@zephyrlibs what do you think about Penrose’s theory that microtubules in the cells are components of a vast, entangled quantum processor representing the true intelligence of cells, especially cell morphologies?

[@igNights77]

That was amazing. Please do more deep dives like this.

[@anakimluke]

So, if it is predicting the position of the molecules well does that imply that it must have picked up from the training data a heuristic for physics? Surely whatever it learned wasn't a conventional molecule simulation, otherwise the results would be the same. Could we poke its insides and find out if it has learned at least things like 'opposite charges attract'? Is this something that is talked about in those papers at all? Or is this a case where it really isn't known exactly what it is learning but it IS working(!)? I wonder if it'd be possible to dissect the parts of physics alphafold thought to be the most relevant and then apply it to a "barebones" molecule simulation. Thanks for the video

[@LookingGlassUniverse]

Such a great idea! I would love to know what exactly it’s learning. You’d assume it’s at least some physics/ chemistry! But it also learns how to incorporate the templates and the evolutionary data. It’d be so interesting to find out what it learns, since it does better than pure physics simulations

[@zephyrlibs]

Barebones is possible with ab initio and dft related methods. cryoEM has pushed microscopy to next level so enrichment of data is very much possible but one no longer depends on alpha fold a lot but writes their own custom models over publicly funded protein databases

[@artemisgaming7625]

Excellent questions. I think if we applied mechanistic interpretability methodology similar to the techniques applied on LLMs we could learn fascinating new insights on what exactly it's learned. It's quite likely imo we'll even find previously unknown physics which might potentially lead to a complete understanding of say why proteins fold how they do and how to calculate their folded state given a protein with more traditional algorithms.

[@zephyrlibs]

@@artemisgaming7625 I don't know, so alphafold was beaten by humans in 2017 and cryoEM drove the stuff forward. Likewise chaos dynamics, topology may give all the machine learning stuff a run for their hype. Humans are too good, a mathematician predicts fractal structure of nucleus way before, 10 years before verification and breakthrough. And that stuff is like some fractal from Alexandria back in the day. So there's no sense of progress here as you may classically think. You're already full blown chaos made chaos observer lol. Enjoy the experience but also understand what sorta maths we struggle with and is tyrannical and why imagination fuelled topology fixes cancer better with morgify algorithm etc.

[@cianjones2728]

Please make more deep dive videos. Amazing video, thanks!

[@EviLPlayeR04]

Very informative for a new learner like me! Thanks for the explanation.

[@jinghu5718]

perfect presentation！It would be great if u can introduce more details about the training part

[@jurjenbos228]

Interesting to see that the diffusion module basically works the same as a picture generator like dall-E. I didn't see that coming.

[@LookingGlassUniverse]

Me neither!

[@zephyrlibs]

cryoEM also has some very simplistic image improvement techniques, besides offering whatever we were waiting for at almost atomic resolution. Now dft and ab initio + hpc/gpu can do some massive simulations hopefully soon

[@Petch85]

What, another Looking Glass video, you are on a roll. 🙂

[@Petch85]

This was a tuff video to get through. I maybe got like 20-30% of it. AlphaFold need not worry that I will create a competing algorithm right away 😂

[@philiplinderberg]

Does anyone know the cutoff date for AlphaFold 3's training set? It doesn't seem to be mentioned in the article or elsewhere that I've looked.

[@KilianMandon]

Just found this incredible channel! I found your video on understanding the codebase of OpenFold so relatable. I recently tackled the same, but didn’t think to use Claude. Did you get through the whole code by now? I just started a KZbin series on implementing AlphaFold 2 from scratch (code’s ready, videos hopefully by October). I feel like our interests are super aligned right now!

[@LookingGlassUniverse]

That’s awesome!! I’d love to see that when you’ve done it!

[@josephpareti9156]

I don' t understand the loss function as weighted sum of Ldiffusion, Ldistogram and Lconfidence. When the processing is in the trunk there is no diffusion yet which occurs later in time. When the processing is in diffusion, the trunk does not seem to execute again. Is there any backpropagation? where is it documented ?

[@LookingGlassUniverse]

I haven’t looked at the paper again to check, but I imagine all these loss terms are calculated at the end of the full run after both parts of the algorithm. Then you do indeed do back prop on the entire thing

[@gelly127]

I understand that the triangle inequality is important for alignment in 3D but how is it actually carried out in the algorithmic level? In regular attention I guess you have an nxn matrix but here you have to represent nxnxn? Is that the way it works?

[@SCKD3035]

This is so informative thank you so much for making this!!!

[@davidaugustyn9234]

What math do you need to make an alphafold

[@seanvw7905]

thanks for a making this excellent explanation

[@kevincardenas6629]

Question a bit out of context: what app are you using?

[@LookingGlassUniverse]

It’s Goodnotes for the iPad!

[@kevincardenas6629]

@@LookingGlassUniverse Thanks!

[@nicksamek12]

Thanks for the new video!

[@Dr.KrishnaKantGupta]

So much helpful..Thank you so much

[@davidaugustyn9234]

I feel bad i didnt watch this sooner

[@Paul-Wonder-Universe]

[@TheThomasAaron]

This is awesome, and I do DNA/RNA/MRNA/Ribosomal/Dendrite reacted research of chain reactions//// often.//// in that pattern of cause and effect as the ball rolls down the steps of order and trackitude on its path and course./... anyways... you get it lol//. Enough said for that//////. What do you think about the MRNA Vaccine..... would the Ribosomal tampering or mingling be rendered a mutation on the dendrite? Was the dendrite its self objectively mutated therefore? Or is this different hap stance a coat that can be shed? Later on?? As the cells replicate thru there natural cycles.... of renewal.? Plz reply If you know. Thx and stay learning things..... I see you're interested in things like sciences and the way things are and such.... that's awesome. I admire that/

[@zephyrlibs]

cryoEM and biophysics. Reading through 1800-2024 is what I like. Including mediaeval paintings, religion and what not to form a grand veiw

[@zephyrlibs]

Adhd is super power. Google books is awesome. Autism makes me penny poor but 3 years some fun experiments and mountain climbing on my epigenetic landscape networks to auto alter them

[@zephyrlibs]

Checking Peschek's paper - life is work - meta stability, cyanobacteria origins of life

[@zephyrlibs]

Bioenergetic Processes of Cyanobacteria: From Evolutionary Singularity to Ecological Diversity

[@zephyrlibs]

Life Implies Work: A Holistic Account of Our Microbial Biosphere Focussing on the Bioenergetic Processes of Cyanobacteria, the Ecologically Most Successful Organisms on Our Earth

[@jean-philippeemond7638]

Wooooow! You are smart!!!!

[@irenerayne7332]

Hey I'm 16F Saw your video bout being mad at math..then going on to learn math in uni and wanting to be good at it . I'm the same and found the video very relatable to my present condition. You have really inspired and motivated me and as a fellow women in stem thank youuu❤

[@Lolleka]

beeter hi jay

[@GodOfTetris]

Why are u into these kinda stuffs? Thought you're Quantum Computing kinda gal?

[@guild6343]

people's interests can change/evolve

[@tcaDNAp]

People used to say that protein folding was one of the biggest uses for quantum computing, so now we can all think about other cool computing stuff!

[@zephyrlibs]

Specially quantum chemistry

[@lumpyspaceprincess6335]

Protein folding problem finally solved ?

[@zephyrlibs]

cryoEM

[@zephyrlibs]

Dft and ab initio remain in the simulation pipeline

[@JxH]

Isn't it interesting that a protein can fold itself up in a nanosecond ? Point being, explain the gap between 100,000 PS3s running 'Fold At Home' all week and a nanosecond. There's a deep point there... A fundamental truth. Some won't get it. I can't explain it, but I see it there screaming at us.

[@zephyrlibs]

I guess light harvesting complex proteins don't do too much of this