Bayesian-Steered Structure Prediction of Mechanical Biomolecules Using Twisted Diffusion.

Area of Science:

• Computational biology
• Structural biology
• Artificial intelligence

Background:

• Deep learning significantly advanced protein structure prediction, primarily by learning from experimental data.
• Current models excel at recapitulating known conformations but struggle to predict functionally relevant, yet experimentally rare, states.
• Generative artificial intelligence diffusion models are increasingly used in structure prediction.

Purpose of the Study:

• To develop a method for sampling underrepresented and functionally relevant molecular conformations.
• To adapt diffusion models for exploring non-equilibrium states beyond experimentally observed structures.
• To demonstrate the utility of conditioned diffusion sampling for macromolecular systems.

Main Methods:

• Reframing conformation prediction as sampling a diffusion distribution conditioned by a Bayesian likelihood.
• Implementing a twisted diffusion sampler within the Boltz-2 framework.
• Developing a diffusion analog of steered molecular dynamics simulations.

Main Results:

• Successfully reproduced experimentally consistent stretched states of DNA fragments, titin, and protocadherin-15.
• Reproduced open states of the MscL ion channel using the novel diffusion approach.
• Demonstrated the ability to sample non-equilibrium conformations without retraining neural networks.

Conclusions:

• The conditioned diffusion sampling method effectively explores underrepresented molecular conformations.
• This approach offers a powerful tool for investigating functionally relevant, non-equilibrium states in various macromolecular systems.
• Steered structure predictions are expected to broaden the scope of conformational sampling in computational structural biology.