Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time.

Area of Science:

• Computational biology
• Structural biology
• Biophysics

Background:

• Protein function relies on dynamic conformational changes, which are challenging to predict computationally.
• Existing methods struggle to capture the full spectrum of protein conformational distributions.
• Predicting conformational ensembles is experimentally difficult and computationally intensive.

Purpose of the Study:

• To introduce ConforMix, an inference-time algorithm for enhanced sampling of protein conformational distributions.
• To enable efficient discovery of conformational variability using diffusion models without prior knowledge of degrees of freedom.
• To improve the prediction of dynamic biomolecular structures.

Main Methods:

• ConforMix combines classifier guidance, filtering, and free energy estimation.
• The algorithm enhances existing diffusion models, applicable to both static structure prediction and conformational generation models.
• It operates at inference time, upgrading pre-trained models.

Main Results:

• ConforMix successfully captures biologically relevant structural changes like domain motion and cryptic pocket flexibility.
• The method avoids generating unphysical protein states.
• Applied to static structure prediction models, it reveals dynamic processes such as transporter cycling.

Conclusions:

• ConforMix offers a scalable, accurate, and versatile method for exploring protein conformational landscapes.
• The algorithm is orthogonal to model pretraining and benefits any diffusion model, even hypothetical perfect ones.
• It advances the computational prediction of biomolecular dynamics and function.