The next frontier in protein design will not be defined by structure alone, but by the capacity to…
@ProfBuehlerMIT28K followers
89,984 views
32 Comments
Comments from the original post on X (Twitter)

@ProfBuehlerMIT
Code: Model weights: Paper:
39 likes

@ProfBuehlerMIT
MIT News Article:
23 likes

@LifeNetwork_AI
A shift from structure to dynamics reframes protein design from predicting static folds to engineering motion as function itself. If proteins are dynamic ensembles on rugged energy landscapes, it raises a deeper question: how much of the possible dynamical design space has biology actually explored?
13 likes

@ProfBuehlerMIT
Biology has only explored a tiny fraction of what is possible, so lots of opportunities for engineers!
2 likes

@samsenchal
Almost like the state space has been reduced by evolution to the most efficient per unit cost to the organism - because ai gives us essentially free compute what's efficient broadens and we get novelty
6 likes

@HangZheng855161
Huge implications for drug discovery: allosteric binding and induced-fit are inherently dynamic, yet structure-based tools treat receptors as rigid. Does the functional degeneracy result suggest 'undruggable' allosteric sites could be accessed by targeting specific motional modes?
6 likes

@ProfBuehlerMIT
Exactly the right question! We found that functionally similar proteins can have very different folds but share vibrational signatures. That 'functional degeneracy' suggests motional modes are the deeper design variable. If we can target the dynamics, we may be able to reach sites that look inaccessible in a crystal structure.
4 likes

@ChristosA89
My view this is the entire point of a ‘metabolism first’ view of abiogenesis. Thermodynamics required both protein topology and their dynamics/motion integrated as one, cofactors within said reactions requiring no code. Such that chemomechanical transduction was powered by enthalpic contribution, configurational entropy, equilibrium constant K, autocatalysis dX/dt=kAX^2; and as such integrated reactions WERE those which manipulated protein conformation. There was no biochemical division of labour, the later canalisation ATP hydrolysis acting the raw energy needs powering nucleotide base manipulation. In this way, early selection was a brutal regime, and far more sensitive/immediate than later gene redundancies. It would be an entirely alien concept to propose in early life the view proteins and their adaptation were differentiated by topology versus dynamical function, these were always a singularity before RNA.
5 likes

@MusicOfOldGods
Brilliant work... and the functional degeneracy finding is the most interesting part. Nature sampled a tiny corner of what's physically possible. VibeGen starts mapping the rest. But here's the question I keep going back to: VibeGen learns vibrational modes conditioned on existing PDB structures. If I am not wrong, the training distribution still anchors it to what evolution happened to discover. What about the regions of design space that have no natural analogue to learn from? That's where a physics-first approach becomes interesting. If you can derive vibrational modes directly from geometry... no training data, no ML, deterministic... you're not bounded by the evolutionary sample at all. Novel chemistry works on day one because the engine runs on the physics underneath, not on patterns learned from what already exists. The two approaches could be genuinely complementary... VibeGen to generate candidate sequences, a physics kernel to evaluate their physical properties instantly at scale. If there are lingering questions from the paper that a deterministic physics engine might help answer... I'd be happy to let you run some experiments on and see what comes back. No pitch! Just curious what happens when the two approaches talk to each other.
3 likes

@JediScale
@burny_tech
3 likes

@theodorus5
@niroshajmurugan maybe some useful context here re: Cosic resonance?
2 likes

@LAMM_MIT
Congrats to our postdoc @_Bo_Ni!
2 likes

@thereal_akashab
This looks super cool – where does the predictor get the ground truth target motion profile from? I was under the impression that we didn’t have dynamics data for a lot of proteins already. Is it mostly from MD sims?
2 likes

@ProfBuehlerMIT
Physics comes, as you say, from molecular dynamics. Of course we can also fine-tune the model with experimental data on top of that, so there is flexibility.
3 likes

@thereal_akashab
Makes sense, and yeah, training on MD and fine tuning on experimental data is sound given wet lab costs. Super cool that you’re also doing language diffusion, I’ve been working on getting flow matching to work with protein sequences myself! Will read the paper
3 likes

@blixem22
Very insightful, thanks!
2 likes

@ProfBuehlerMIT
Thank you!
1 like

@ThinkticaAI
Amazing job congrats!
1 like

@ProfBuehlerMIT
Thank you @ThinkticaAI !

@sandeepdixit22
Interesting.
1 like









