Efficient Parameter Estimation of Generalizable Coarse-Grained Protein Force Fields Using Contrastive Divergence: A...

Area of Science:

• Computational Biology
• Protein Folding
• Statistical Mechanics

Background:

• Maximum Likelihood (ML) optimization is crucial for parameter inference in computational models.
• Physics-based coarse-grained protein models require accurate parameterization for reliable simulations.
• Existing models often lack generalizability to unseen proteins or peptides.

Purpose of the Study:

• To infer a generalizable, physics-based coarse-grained protein model using ML.
• To apply contrastive divergence for efficient gradient approximation with large datasets.
• To evaluate different van der Waals (vdW) potential forms for protein simulations.

Main Methods:

• Utilized a Maximum Likelihood (ML) inference scheme with contrastive divergence.
• Trained a coarse-grained protein model on native protein conformations.
• Compared hard cutoff and Lennard-Jones (LJ) vdW potentials with inferred and adopted parameters.

Main Results:

• The inferred ML protein model demonstrated generalizability, folding peptides and protein G not in the training set.
• The LJ potential with inferred vdW parameters significantly outperformed the hard cutoff model.
• Inferred parameters for the LJ potential showed superior performance compared to adopted parameters from other force fields.

Conclusions:

• A generalizable coarse-grained protein model can be achieved through ML inference.
• Accurate parameterization, especially for vdW interactions, is critical and specific to the chosen energy function.
• The developed model and inference strategy offer a robust approach for protein structure prediction.