Target-Focused Library Design by Pocket-Applied Computer Vision and Fragment Deep Generative Linking.

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• Publicly available structural data from protein-ligand complexes is vast but underexploited for library generation.
• Generating focused compound libraries for specific protein targets remains a challenge in drug discovery.

Purpose of the Study:

• To introduce a novel computational approach, POEM (Pocket Oriented Elaboration of Molecules), for the de novo design of target-focused molecular libraries.
• To leverage existing structural databases for efficient generation of potential drug candidates.

Main Methods:

• POEM utilizes a computer vision method to align 31,384 PDB-derived images of fragment-bound microenvironments to a target cavity.
• Fragments from similar PDB subpockets are positioned in the target cavity using image transformation matrices.
• A deep generative model links oriented fragment pairs to construct fully connected molecules.

Main Results:

• POEM successfully generated a library of 1.5 million potential inhibitors for cyclin-dependent kinase 8.
• Synthesis and testing of only 43 compounds yielded nanomolar inhibitors.
• The process required limited resources and was completed within two iterative cycles.

Conclusions:

• POEM is an effective computational strategy for accelerating the identification of potent inhibitors.
• This method significantly reduces the experimental effort required in early-stage drug discovery.
• The approach demonstrates the power of integrating structural databases with generative models for targeted library design.