This is the most surprising and exciting result of my career: we were running simulations of NaCl with a neural network potential that implicitly accounts for the effect of the water, ie a continuum solvent model (trained on normal MD) when Junji noticed something strange: 1/n
A crystal was nucleating and growing! despite the NNP only being trained on solution data (4 M). Amazingly it has the correct crystal structure, i.e., FCC. This is a phase transitions, an emergent phenomena, totally out of distribution, supposedly where AI is no good.
Impressively this is actually more correct than the all-atom MD as the true solubility of this model is actually very low, all atom MD is just too slow to see it. I think this pretty conclusively disproves the idea that NNPs are only interpolating on the data they are given.
I think neural network potentials are the most important scientific tool of the next decade. The ability to simulate systems at the molecular scale starting from nothing but quantum mechanics will be transformative for a vast range of problems throughout biology and chemistry 1/n
Here's another pretty incredible example of neural network potentials extrapolating outside their training data in a way I wouldn't expect. We were simulating an electrolyte with an NNP when this happened:
This is the most surprising and exciting result of my career: we were running simulations of NaCl with a neural network potential that implicitly accounts for the effect of the water, ie a continuum solvent model (trained on normal MD) when Junji noticed something strange: 1/n
I have always assumed it would be impossible to study crystal nucleation etc with a continuum solvent model. Here it just dropped out for free. They're incredibly cheap to run too. ie. tens of cpu hours. More info on continuum solvent models here:
I want to explain a statistical mechanical concept known as coarse graining which I think might be useful for thinking about things like AF3. Especially a special case known as continuum or implicit solvent models.
The lattice parameter is a bit high compared to experiment, so we took a KCl coarse grained force field trained on quantum chemical solution MD data and showed that it could simulate crystal dissolution. It maintains the correct lattice spacing and appears to dissolve in the same
So the race is really heating up to build a truly universal force field. This is one of those powerful ideas that people in the field of molecular simulation have been dreaming about for decades. What exactly is it and how far away are we? 1/n
You know a field is taking off when you regularly see multiple order of magnitude improvements in performance. Thats whats happening in deep learning for molecular simulation. This field is going to be transformed. Latest example: 1/n
Ok so what is a neural network potential concretely? It's just a very flexible function with many adjustable parameters that you fit to the 'potential energy surface.' This is just the energy as function of the position of the atoms in your system. 1/n
I think neural network potentials are the most important scientific tool of the next decade. The ability to simulate systems at the molecular scale starting from nothing but quantum mechanics will be transformative for a vast range of problems throughout biology and chemistry 1/n
We also simulated LiCl,LiBr and KCl at quantum chemical accuracy and reproduced the ion specific pairing affinities and matching experimental activity coefficient derivatives. Activity coefficients are the most important property of electrolyte solutions, they are ubiquitous
I want to explain a statistical mechanical concept known as coarse graining which I think might be useful for thinking about things like AF3. Especially a special case known as continuum or implicit solvent models.
Thrilled to announce AlphaFold 3 which can predict the structures and interactions of nearly all of lifeโs molecules with state-of-the-art accuracy including proteins, DNA and RNA. Biology is a complex dynamical system so modeling interactions is crucial
Ok so what is a neural network potential concretely? It's just a very flexible function with many adjustable parameters that you fit to the 'potential energy surface.' This is just the energy as function of the position of the atoms in your system. 1/n
So isnโt it strange that diffusion models and the human brain are the two best methods for conjuring up images out of nothing and they both happen to operate on the same fundamental algorithm ie molecular dynamics? How many ways of conjuring up images do we think the universe
It seems we're learning that deep learning is mostly about the data. If you want to know where it will really take off look to areas where you can continuously generate increasingly diverse but consistently high quality data. That leads you to quantum chemistry:
Ok so the new AlphaFold model relies in large part on a "relatively standard diffusion approach" turns out you can think of this as just a special case of a neural network potential, it just uses experimental data not quantum chemistry to train on. 1/n
Thrilled to announce AlphaFold 3 which can predict the structures and interactions of nearly all of lifeโs molecules with state-of-the-art accuracy including proteins, DNA and RNA. Biology is a complex dynamical system so modeling interactions is crucial
We also showed you can extract infinite dilution pairing free energies and diffusivities and get almost perfect radial distribution functions with the continuum solvent model and more.
So what are the most exciting potential applications of a universal forcefield? Weโre already starting to see one emerge that could result in a more general and more useful form of alphafold3. 1/n
So the race is really heating up to build a truly universal force field. This is one of those powerful ideas that people in the field of molecular simulation have been dreaming about for decades. What exactly is it and how far away are we? 1/n
Thrilled to announce AlphaFold 3 which can predict the structures and interactions of nearly all of lifeโs molecules with state-of-the-art accuracy including proteins, DNA and RNA. Biology is a complex dynamical system so modeling interactions is crucial
This is one of those beautiful ideas that took me a long time to see, but in retrospect seems obvious. If correct, it implies a massive step up in the number of problems molecular simulation can be fruitfully applied to.
Wow this is incredibly motivating and rewarding to see how many people are excited about salt crystallisation and the potential for AI to accelerate basic science! Thank you so much! Some people asked about why this is an important topic to study? 1/n
This is the most surprising and exciting result of my career: we were running simulations of NaCl with a neural network potential that implicitly accounts for the effect of the water, ie a continuum solvent model (trained on normal MD) when Junji noticed something strange: 1/n
I think this is a profound paper โฆ This is what โgrokkingโ is right? A sharp jump downward in energy/loss? Itโs just a phase transition right? Stat. mech. must have the tools to explain the success of deep neural networks.
This paper nicely demonstrates the point I was making about AF3. It can be interpreted as a forcefield as it is learning the gradient of the log probabilities ie the score, but log probabilities are just free energies. So itโs learning an approximation to the true free energy.
Is there a machine/deep learning textbook anywhere that teaches the Boltzmann/Gibbs distribution? Is it called something else? Have looked at three so far with no mention of it. It is the entropy maximising distribution! Surely it is important to know?
Couldnโt agree more with this excellent piece by
@gdefabritiis
A pioneer in this field. These tools are already transformatively useful for molecular simulation and will therefore be important for many closely related fields. There are improvements that still need to be made
Machine Learning Potentials: A Roadmap Toward Next-Generation Biomolecular Simulations
Neural network potentials are very powerful, but they appear too slow to be ever used for macromolecules.. this work provides a possible roadmap...
P:
Interesting how deep learning for generating equilibrium distributions seems to be converging back to molecular dynamics. Like this is just langevin dynamics with a learnt score. So just NNP-MD with many runs in parallel right? Or am I missing something?
Wow Microsoft patented the coarse grained diffusion model work. Anyone know of any precedents/thoughts on what this means? So they can stop anyone from using this approach?
Very exciting. Another excellent example of the power of combining quantum chemistry, statistical mechanics and neural network potentials. This is just the beginning.
Ever wondered if we could model
#water
to autoionize and correctly predict pH = 7? ๐ค Well, now we can! ๐
In our latest
@ChemRxiv
preprint, we introduce a
#deepneuralnetwork
potential trained on density-corrected
#DFT
that predicts the autoionization constant of water to be Kw
Another nice recent papers showing the power of combining MD with diffusion models. This one from
@therealpeterobi
where they show you can use it to make umbrella sampling more efficient.
You know a field is taking off when you regularly see multiple order of magnitude improvements in performance. Thats whats happening in deep learning for molecular simulation. This field is going to be transformed. Latest example: 1/n
Yeah seems backward to me how much compute is spent using exact calculations to generate Boltzmann distributions, ie langevin dynamics for diffusion models/MD simulations when nature will give you one for free if you just provide it with an energy landscape and let the atoms move
Saturday morning [not very original] thought: Given the incredible noise-resistance of deep models, it is extremely bizarre that specialized deep learning chips stick to the traditional "perfectly exact" computing paradigm, given the complexity and cost it involves.
So pleased to get this preprint out. Feel like weโve finally worked out how to do something Iโve been trying to do for 13 years since the start of my PhD: Build an accurate continuum solvent model of ion-ion interactions in solution.
I want to record a prediction: ML acceleration of molecular simulation will transform all of physical science. From quantum scale all the way up to climate. Justification: 1/n
Yes, this is the ultimate way ML will help accelerate physical sciences. By constructing custom MCMC operators (eg proposal distributions) to accelerate traditional MD/MCMC simulations in combination with existing tools. This can be done while preserving all error bars.
This is self-ionisation! Water splitting apart into hydronium and hydroxide ions. Amazingly it somehow knows to put the hydronium in the correct pyramidal structure. There are no examples of this in the training data, yet it remains stable and then something even cooler happened:
Another very interesting NNP for electrolyte solutions paper. People have spent decades trying to build accurate classical force fields of calcium carbonate one of the most fundamental substances in existence. Turns out even very sophiticated polarisable models overestimate
Exactly, for example quantum chemistry simulations and ML go together like hand and glove. One gives accuracy and reliability, the other speed and scale.
This is the Grotthuss mechanism! Chains of proton hopping events give rise to a much higher diffusivity of acid in water. The NNP has rediscovered that this is a plausible mechanism of ion transport, without even being asked to.
Granted the rates of these processes are too high
Interesting discussion. The idea that diffusion models are going to replace MD for generating probability distributions pops up again though. I do not get this: diffusion models are MD ie langevin dynamics on a free energy surface. If anything deep learning got replaced with MD.
Even highly educated / plugged-in people, myself included, know very little about the latest biotechnology
I hope this episode can help change that, because things are about to get crazy!
Full episode is here:
This is just an amazing photo of a neuron!
โA single neuron is shown with 5,600 of the nerve fibers (blue) that connect to it. The synapses that make these connections are in green.โ
Credit: Google Research & Lichtman Lab, Harvard University. Renderings by D. Berger, Harvard
Interesting comment in a fascinating piece from
@RuxandraTeslo
. I agree with the essay overall. But I often see this idea expressed about biology: itโs impossible to understand from first principles.
I think itโs become so ingrained no one questions it. But I think we should. I
Very nice. This is the way to do it. Combine diffusion to give you the stability far from equilibrium with forces to give you the accuracy near equilibrium.
Super excited to release a set of models for computational chemistry - my last 2 years of work
@OrbMaterials
.
Post ELMo, I'm at some risk of becoming a "one trick pony" career wise, but we've managed to make pre-training work nicely for 3d crystal structures.
More below!
Another very interesting NNP for electrolyte solutions paper. People have spent decades trying to build accurate classical force fields of calcium carbonate one of the most fundamental substances in existence. Turns out even very sophiticated polarisable models overestimate
Exactly this is why Iโm so excited about neural network potentials. They are the only approach where you can computationally generate your own high quality, targeted training data on the fly in an automated fashion.
*AI for BioChem is data-starved.* A theme of every session
@icmlconf
was that they were all trying to deal with this fact.
e.g. improving model efficiency, data preparations from public sources and using synthetic data where possible
I think this is a profound paper โฆ This is what โgrokkingโ is right? A sharp jump downward in energy/loss? Itโs just a phase transition right? Stat. mech. must have the tools to explain the success of deep neural networks.
Was thrilled to appear on Cognitive Revolution, one of my favourite podcasts, was a great discussion. Neural network potentials are set to significantly impact many areas of science and engineering imo. This is just the beginning.
Many of the most important scientific advances have followed a simple recipe: Adopt a tool developed in another field to your own. The most beautiful demonstration of this is actually diffusion models which have done this three times already!
Firstly tools and ideas from
@carnot_cyclist
Diffusion models use langevin dynamics which was an algorithm invented to simulate the behavior of molecules with time. The score is just a time dependent forcefield.
No great question. First task is to provide more accurate data for large scale models of chemical processes ie activities, diffusivities, reaction rates etc. This can be done with homogeneous single phase simulations. Second task is to provide direct physical insight into the
@TimothyDuignan
Amateur question: what's the end state for ML-enabled molecular dynamics simulation? Is it to replace current activity coefficient models, achieve much more accurate simulations, or both?
Quantum computing experts claim computing properties of Femoco is impossible with classical computing and if you could do it you could revolutionize fertilizer synthesis. Turns out you can do it with DFT fine but almost no one cares.
Awesome to drop by Rowan on my way back from the Gordon water conference in Holderness. Such an awesome tool theyโre building canโt wait to see what they do next!
It was fantastic to host
@TimothyDuignan
today for our third Rowan Seminar and hear about how neural network potentials are revolutionizing electrolyte simulation and ab initio MD!
Love the flow of ideas back and forth between molecular simulation and deep learning. Diffusion models originally inspired by molecular dynamics algorithms (langevin dynamics) now inspiring new approaches to accelerate MD.
Oke, the AlphaFlow paper is awesome: AlphaFold Meets Flow Matching for Generating Protein Ensembles
Just watch how AlphaFlow's ensemble reproduces details of MD.
Weights + code
We have it in the reading group on Mon 11am EST!
1/2
Holy shit, is this real?
Stanford, MIT, and Toyota found a 50% lithium-ion battery cycle life improvement simply by changing the power level of its first charge after manufacturing?
This is a compelling argument for investing in neural network potentials which donโt suffer from the same data limitations. The problem with trying to use โwet lab innovationsโ is that techniques for obtaining simultaneous femtosecond/picometer scale resolution experimental data
Wet-lab innovations will lead the AI revolution in biology
i feel like i keep repeating this argument to people so i decided to just write it out
1.9k words, 9 minutes reading time, very short!
New post on the near frontier of Neural Net Potentials and the implication for company building and industry competitive dynamics
These models are smaller than GPT1. Excited to see them scale
Thoughts on:
* role in computational stack
* materials science vs drug discovery
*
Notice how similar to MD this is conceptually. It is actually mathematically essentially the same also. The only difference is the force field is learnt from the PDB where you know the forces are 0 because they are equilibrium states. Really its an implicit solvent force field.
RFdiffusionAA generating a small molecule binding protein against an experimental FXIa inhibitor (OQO), a ligand which is significantly different than any in its training dataset.
This is a beautiful paper from Aleksander Durumeric, Yaoyi Chen,
@FrankNoeBerlin
,
@CecClementi
where they combine denoising with forces to train a coarse grained neural network potential. This is an idea we are playing with too. It nicely demonstrates the deep connection between
But thereโs a new way people are starting to get very excited about using neural network potentials. Around 2017/2018 we saw things like ANI-1 and Tensormol-0.1 which could do this for a 4 atom types for a range of structures.
Some people are not impressed by this. Maybe im just incompetent but I spent literally years trying to build continuum solvent models of this exact thing and couldnโt do much better itโs really hard to model without explicit water! ๐คฃ
Excellent piece. Completely agree we should be trying to build scientific foundation models. I suspect we might need to new organizational structures to develop them though as they need large groups of dedicated full time experts which academia is not great for.
We see exactly the same thing for simple electrolytes. If you cannot get sodium chloride pairing free energy right you are not going to get protein folding right. I often donโt point this out because I donโt want to offend senior researchers.
Take a look at this
#OpenAccess
paper ๐ from the latest issue of Journal of Chemical Theory and Computation
#JCTC
๐ The Role of Force Fields and Water Models in Protein Folding and Unfolding Dynamics ๐ฆ๐ฌ
๐
#thermodynamics
This is the future of generative AI for chemistry. It will merge with molecular dynamics as at the core they are doing the same thing. And there is only so much information you can get from minimum energy structures. The PDB is tapped out surely?
Good take as always. I donโt think this axis makes sense really though. I would argue a diffusion model is more physics based than a lennard jones forcefield. Harmonic approximation about the minima is in every physics text book but Iโve never seen a 1/r^12 repulsion.
Alex Zhavoronkov, PhD (aka Aleksandrs Zavoronkovs)
This is a beautiful clear explanation of diffusion models. The cool thing is they are actually really easy to understand if you know molecular simulation. There is a direct analog for almost every concept. 1/n
New blog post about the geometry of diffusion guidance:
This complements my previous blog post on the topic of guidance, but it has a lot of diagrams which I was too lazy to draw back then! Guest-starring Bundle, the cutest bunny in ML ๐
This rapid progress is incredibly exciting to me and indicates that we may be approaching a point where we might have a tool we could reasonably call a true universal force field. There are still many challenges though but also lots of new ideas to try:
Fascinating thread. Highlights exactly why PDB structures alone can only get you so far. So much is determined by dynamics. The central question is how to get data on the dynamics processes so you can train on it.
Hereโs a paper that will get zero press because it looks totally specialist, not to say obscure. Itโs about how an important class of transcription factors regulate genes. But I think it's worth dissecting because it raises a wider question.Bear with me...
Just imagine one day we will be able to go to a website like this and run accurate dynamics on any system of atoms we want. This will transform all of science and society. We will finally be masters of the molecular scale.
@Dr_Gingerballs
Great questions. I didn't expect it because the continuum solvent free energy is highly concentration depedendent so I would not expect it to generalise to stable crystals.
In some sense it is more accurate than the MD as it predicts the correct global minima (crystal) whereas
Awesome paper. Shows how we can train on many different levels of theory simultaneously will be very important as we make DFT databases bigger and bigger. We need to build a PDB equivalent but for quantum chemistry.
Yeah this is very consistent with what I generally see. The NNPs are now so good that the error is determined by the accuracy of the underlying DFT. This means getting improved performance may just mean better fitting to the noise in the underlying data.
interesting new work from
@johnkitchin
& co-workers studying errors in the (now-ubiquitous) OC20 dataset - they argue that NNP errors are approaching 0.2 eV because that's the intrinsic error of the dataset.
Another nice ion pairing paper on NaCl with NNPs. Look at the spread on those classical force fields in comparison! This is the fundamental medium in which all of biology occurs and we haven't been able to predict even its most basic properties until now!
Ok Iโve been trying to get people excited about neural network potentials aka machine learning force fields. But what are the key challenges for using this tool to try and build the universal force field?: 1/n
So the race is really heating up to build a truly universal force field. This is one of those powerful ideas that people in the field of molecular simulation have been dreaming about for decades. What exactly is it and how far away are we? 1/n
So byte dance have entered the universal machine learned force field race with a very impressive paper starting with the right problem imo: liquid electrolytes. I think this could be a critically important technology.
NNPs are a particular class of machine learning potential. There are many types but a particularly important recent breakthrough has been equivariance which greatly reduced the amount of training data needed. (Alphafold2 used this idea)
Very nice explicit formulation of self attention as minimizing an energy function, something I alluded to in my perspectives article. So LLMs and diffusion models are essentially just NNPs?
This can be done to well below thermal noise for many small systems today but the problem is it's slow. Here NNPs come to the rescue, they predict E from the positions of all the atoms much faster. We train them on examples of Schrodingers equation we have solved previously.
People have been trying to build this for a long time with point charges and Lennard Jones potentials. That approach has built some very useful tools but unfortunately in general you still have to do a lot of case by case fiddling and testing of parameters to get these working.
The idea is that for any set of atoms in a given position there is a potential energy. You can get this by solving Schrรถdinger equation. If you could calculate this energy and its gradient (-force) accurately and fast enough it would be transformatively useful.
This is really important as the energy is so fundamental. It determines the probability of observing particles in a given arrangement and its gradient is the negative of the force so it can tell you how the particles move too.
Very cool. Diffusion models use a molecular simulation algorithm (thermally annealed langevin dynamics) so of course you see phase transitions directly analogous to the sudden changes that occur when you cool/heat a system of molecules, i.e., crystallization.
So cool! I assume this is the same thing that goes on at phase transition boundaries in stat mech: โSchramm-Loewner curves appear as domain boundaries between phases at second-order critical points like the critical Ising modelโ
Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.
Nice. Rowan is the perfect tool for someone looking to pick up quantum chemistry I reckon, not just for experts, particularly with the excellent visualisations.
how'd i end up with hundreds of gifs in my downloads? funny you ask ...
(video is of an xtb optimization i was struggling with, looking at this paper from
@BroereDaniel
on dicobalt complex spin states: )
This is because you could run simulations to directly see what atoms are doing at the molecular scale for important systems and sample from the equilibrium probability distribution (Boltzmann distribution) to calculate important experimental quantities.
This means that Google's claim that they have "surpassed physics based tools" is kind of strange. In fact there is a ton of physics baked into it how diffusion models work!
Caveat: The reality, as always, is obviously significantly more complicated then I've presented eg excited states, non adiabatic and quantum nuclear effects etc. Which will need to be included in various cases using additional tools.
Check out this really nice collaboration with
@alisterpage
and two awesome students where we show you can resample from DFTB MD, compute forces at a higher level of theory and run stable MD with equivariant neural network potentials.