MIT Introduces VibeGen: The First End‑to‑End Dynamic Protein Generator Linking Sequence and Vibration

Motivation: Protein Dynamics as Functional Determinants

Proteins function through dynamic conformational changes across femtosecond‑to‑millisecond timescales; aberrant dynamics underlie diseases such as p53‑related cancers and CFTR‑mediated cystic fibrosis. Consequently, designing proteins with prescribed dynamics is emerging as a frontier in structural biology and bioengineering.

Limitations of Existing Experimental and Computational Methods

Traditional techniques (NMR, HDX‑MS, cryo‑EM) and computational approaches (molecular dynamics, normal‑mode analysis) are either low‑throughput, costly, or limited in temporal scale, preventing large‑scale dynamic‑aware design.

Deep‑Learning Gap

Recent generative AI models (AlphaFold2, RFdiffusion, AlphaFold3) excel at static structure prediction but largely ignore intrinsic dynamics, treating proteins as rigid bodies. A unified sequence‑structure‑dynamics‑function mapping remains unsolved.

VibeGen Overview

MIT and Carnegie Mellon propose VibeGen, an agentic framework that couples a protein‑design diffusion module (PD) with a prediction diffusion module (PP) to achieve bidirectional mapping between amino‑acid sequences and low‑frequency normal‑mode shapes.

Low‑Frequency Normal‑Mode Database Construction

Researchers filtered the January 2024 PDB for single‑chain proteins ≤126 residues, cleaned structures with VMD, MMTSB, and SCWRL4, performed CHARMM energy minimization, and computed block normal modes. After discarding the six rigid‑body modes, the lowest‑frequency non‑trivial modes were retained, yielding a dataset of 12,924 chains. Vibration amplitudes were extracted for Cα atoms, normalized, and used as dynamics descriptors; the set was split 9:1 into training and test subsets.

Model Architecture

Both PD and PP share a 150‑million‑parameter ESM‑2 pretrained protein language model as the backbone. Each diffusion model employs a U‑Net with multiple channels to inject dynamics conditions during denoising. The PD generates sequences conditioned on target vibrational features, while the PP inversely predicts normal‑mode shapes from candidate sequences using diverse sequence embeddings.

Generation‑Evaluation‑Selection Loop

During inference, the PD first proposes candidate sequences. The PP then evaluates their predicted dynamics in real time, allowing users to filter results by accuracy or diversity and iterate until satisfactory designs are obtained.

Performance Evaluation

On 1,293 test cases, VibeGen achieved a median Pearson correlation of 0.53 and median relative L2 error of 0.57 between designed and target normal‑mode shapes. After low‑pass filtering to retain overall shape, correlation improved to 0.72 and L2 error dropped to 0.37, indicating strong capture of global vibrational contours.

Diversity analysis showed that for a single dynamics target (e.g., U‑type or L‑type modes), VibeGen produced multiple structurally distinct yet functionally equivalent designs, typically featuring a dense core (α‑helices or mixed helix‑sheet) and flexible termini matching high‑amplitude regions.

BLAST similarity distributions exhibited a bimodal pattern, with the primary peak corresponding to truly de novo sequences, confirming the model’s propensity for generating novel proteins.

Structure‑Dynamics Correlation

Across experiments, low‑amplitude regions consistently aligned with dense secondary structures (α‑helices, β‑sheets), while high‑amplitude regions corresponded to loops or flexible termini, demonstrating that VibeGen internalizes the physical relationship between structure and dynamics.

Broader Impact and Future Directions

The study situates VibeGen within a growing ecosystem of dynamic‑aware protein design tools, noting complementary efforts that integrate normal‑mode analysis with advanced diffusion models to mitigate design degeneracy. Extensions toward enzyme active‑site engineering, binder optimization, and industrial protein applications are discussed, highlighting the method’s potential to accelerate biomedical and biomanufacturing innovations.

Key References

VibeGen: Agentic end‑to‑end de novo protein design for tailored dynamics using a language diffusion model, Matter (2026). DOI: https://www.cell.com/matter/abstract/S2590-2385(26)00069-X