Back to Home

VibeGen: AI for protein design by vibrations

MIT engineers developed VibeGen — an AI model for generating protein sequences from specified vibrational and dynamic patterns. The system uses diffusion models with two agents for iterative design. Results confirmed by simulations, opening applications in medicine and materials science.

MIT VibeGen: proteins by vibe-coding vibrations
Advertisement 728x90

# VibeGen: Generating Proteins from Dynamic Patterns Using MIT AI

MIT engineers have created VibeGen—a generative AI model that synthesizes protein sequences based on target motion patterns: vibrations, bends, and oscillations. The model solves the inverse problem, determining amino acid chains that deliver the specified dynamics, rather than static shapes. This is akin to vibe-coding, where describing desired behavior generates a functional molecule.

Traditional approaches, including AlphaFold, focused on predicting and generating 3D structures. However, dynamics—bending, stretching, pulsing—determine functionality: ligand binding, load resistance, pathogen interactions. VibeGen integrates molecular dynamics into the design process, using diffusion models for iterative optimization.

VibeGen Architecture: Designer and Predictor

The system consists of two agents:

Google AdInline article slot
  • Designer: generates candidate amino acid sequences tailored to the target motion profile.
  • Predictor: simulates dynamics and evaluates alignment with the goal.

The agents interact in a loop: the designer proposes variants, the predictor checks them via physical simulations, and iterations continue until convergence. The core technology is a diffusion model adapted for protein space: starting from initial noise, it forms a sequence with the specified vibrational characteristics.

Key features:

  • Inverse Problem: from dynamic profile to sequence, not the other way around.
  • Iterative Stabilization: feedback loops ensure precision.
  • Entirely Novel Designs: generated proteins are absent from natural databases.

Validation and Functional Degeneracy

The generated proteins underwent molecular dynamics simulations. Results confirmed fidelity to target patterns: vibrations, bends, and oscillations were reproduced accurately. An unexpected finding—multiple sequences and structures produce identical dynamics, which the authors term "functional degeneracy".

Google AdInline article slot

This points to an underexplored design space: evolution has realized only a fraction of possible dynamic solutions. VibeGen expands this range, offering non-evolutionary options with predictable behavior.

Applications in Biotechnology and Materials Science

Dynamics control unlocks new directions:

  • Medicine: therapeutic proteins with precise binding and minimal off-target effects.
  • Materials Science: fibers and coatings with programmable mechanics—strength, elasticity.
  • Synthetic Biology: molecular actuators that respond to stimuli in real time.

Dynamics as a protein's "vibe"—its physical pattern defining functionality. The model enables designing molecules for specific tasks, blending physics into AI generation.

Google AdInline article slot

Key Takeaways

  • VibeGen uses diffusion models to generate amino acid sequences from dynamic patterns, bypassing the focus on static structures.
  • Two agents—designer and predictor—drive iterative optimization via molecular dynamics simulations.
  • Generated proteins show functional degeneracy: different sequences yield the same dynamics.
  • Potential in medicine, materials science, and synthetic biology through control of vibrations and oscillations.
  • Publication in Matter (Cell Press), March 24, 2026, leads—Marcus Bühler (MIT).

— Editorial Team

Advertisement 728x90

Read Next