Single-Sequence Deep Learning Delivers Crystal-Quality Models of Covalent K-Ras G12 Hotspot Complexes.

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• Structure-based drug design traditionally requires time-consuming experimental methods like X-ray crystallography.
• Predicting covalent drug-protein complexes is crucial for developing targeted therapies.

Purpose of the Study:

• To evaluate Chai-1, a structure prediction tool, for its ability to accurately model covalent protein-ligand complexes.
• To assess Chai-1's efficiency and accuracy compared to existing methods, particularly for K-Ras inhibitors.

Main Methods:

• Utilized Chai-1, a structure prediction tool accepting user-defined ligands, to predict covalent K-Ras(G12C) complexes.
• Employed a covalent-bond restraint within Chai-1 for predicting complexes with K-Ras(G12D) and K-Ras(G12S).
• Compared Chai-1's throughput and pose accuracy against AlphaFold3.

Main Results:

• Chai-1 achieved pocket-aligned RMSDs < 2 Å for diverse K-Ras(G12C) inhibitors without using multiple sequence alignment (MSA).
• The tool successfully reproduced binding poses for K-Ras(G12D) and K-Ras(G12S) inhibitors using covalent-bond restraints.
• Chai-1 demonstrated ~40-fold higher throughput than AlphaFold3 with comparable pose accuracy.

Conclusions:

• Chai-1 is an accessible and computationally efficient tool for predicting covalent protein-ligand co-complex structures.
• The covalent-restraint mode of Chai-1 accelerates covalent drug discovery for challenging targets.
• Further development is needed to account for detailed chemical properties like leaving groups and stereochemistry.