VibeGen

Proteins are not static objects. Catalysis, allostery, and signal transduction all depend on how a chain moves, yet the dominant generative design tools optimize a fixed backbone geometry and leave motion to emerge as a byproduct. VibeGen, from the Laboratory for Atomistic and Molecular Mechanics at MIT with collaboration from Carnegie Mellon University, inverts that framing: it generates amino acid sequences conditioned directly on a target vibrational signature.

The design objective is the normal mode shape — the per-residue distribution of vibrational amplitude along the backbone in a protein's lowest non-trivial normal mode. This vector is invariant to the coordinate system and captures where a protein is stiff and where it is floppy, a property tied to ligand binding, conformational change, and enzyme mechanism. VibeGen learns the mapping between sequence space and this dynamics descriptor in both directions.

Rather than one model, VibeGen is an agentic pair. A protein designer proposes sequences that should exhibit a requested mode shape; a protein predictor independently estimates the mode shape a given sequence would produce, and is used to rank the designer's candidates on the fly. Designs are then validated end to end by folding, relaxation, and full-atom molecular dynamics normal mode analysis, closing the loop between a generative proposal and a physics-based measurement.

#Key Features

• Dynamics as the design target: Conditioning is a normalized per-residue vibrational amplitude vector, so a user specifies how the protein should move rather than what shape it should hold.
• Two-agent designer-predictor loop: The predictor screens the designer's candidate pool in silico, letting users trade off accuracy against diversity without invoking expensive molecular simulation for every proposal.
• Frozen language model, trainable diffusion: Both components are protein language diffusion models that keep the pretrained sequence encoder fixed and train only a conditional denoiser, which keeps training cost modest.
• Genuinely novel sequences: BLAST analysis of generated designs shows a large population with no significant similarity to known proteins, placing outputs outside the space reachable by evolutionary retrieval.
• Physics-based validation: Generated sequences are folded, energy-minimized with a full-atom force field, and re-analyzed by normal mode analysis, so reported accuracy is measured against simulation rather than a learned proxy.

#Technical Details

The designer and predictor are both protein language diffusion models: a pretrained protein language model — a 150M-parameter checkpoint from the ESM-2 series — maps between token space and a latent space, and a trainable one-dimensional U-Net with attention performs conditional denoising in that latent space. Only the diffusion component is trained. The training corpus was built by running full-atom molecular dynamics relaxation with the CHARMM force field followed by normal mode analysis on Protein Data Bank chains of at most 126 residues, yielding 12,924 monomer chains split 90/10 into training and test sets. Designed sequences are folded with OmegaFold and re-analyzed with the same normal mode protocol. Across the test set the vector-wise correlation between requested and realized mode shapes has a median of 0.53, rising to 0.72 when smoothed mode profiles are compared; selecting the predictor's best-ranked candidate out of 40 proposals raises median accuracy from 0.31 for the worst-ranked group to 0.53, quantifying the benefit of the two-agent arrangement. Code and weights are released under Apache-2.0.

#Applications

VibeGen targets engineering problems where flexibility, not just fold, is the specification: enzymes whose catalytic cycle depends on hinge motion, dynamic scaffolds and sensors, and protein-based biomaterials whose mechanical response is the product requirement. Because the conditioning is a plain per-residue vector, it slots into larger automated design pipelines as a dynamics-aware generator alongside structure-conditioned tools. The predictor is also usable on its own as a fast forward model for screening natural or designed sequences by expected vibrational profile.

#Impact

VibeGen establishes a direct, bidirectional link between protein sequence and vibrational behavior in a generative setting, extending conditional design beyond static geometry. Its accuracy is honest about the difficulty of the problem — residue-level agreement is moderate, and the inverse mapping is genuinely degenerate, since many distinct sequences and folds share a similar mode shape. That degeneracy is also the source of the method's diversity. The work was published in Matter, and the released checkpoint and companion code make dynamics-conditioned generation reproducible for other groups building multi-objective protein design workflows.