Modeling protein-small molecule conformational ensembles with PLACER.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Modeling conformational heterogeneity in protein-small molecule systems is complex.
• Atomic-level descriptions offer advantages in speed and generality for folded-state interactions.

Purpose of the Study:

• To develop a novel graph neural network (ChemNet) for atomic-level modeling of protein-small molecule systems.
• To leverage ChemNet for predicting conformational ensembles and improving enzyme design.

Main Methods:

• Developed ChemNet, a graph neural network, trained on atomic positions from the Cambridge Structural Database and Protein Data Bank.
• Utilized ChemNet to generate atomic structures of small molecules and protein side chains for docking.
• Generated conformational ensembles using ChemNet's rapid and stochastic predictions.

Main Results:

• ChemNet accurately reconstructs atomic positions from corrupted inputs.
• The model successfully generates structures for diverse organic small molecules and protein-small molecule complexes.
• Enzyme design using ChemNet for active site assessment yielded higher success rates and activities, including a retroaldolase with a k_cat/K_M of 11000 M^-1min^-1.

Conclusions:

• ChemNet provides a rapid and general method for modeling conformational heterogeneity in small molecule and protein systems.
• The approach significantly improves enzyme design by enabling the creation of pre-organized active sites.
• ChemNet is anticipated to be a valuable tool for computational chemistry and structural biology research.