CYCLICCAE: A CONFORMATIONAL AUTOENCODER FOR EFFICIENT HETEROCHIRAL MACROCYCLIC BACKBONE SAMPLING.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Macrocycles represent a promising therapeutic class with significant potential in drug development.
• Rational design of macrocycles is hindered by challenges in incorporating heterochiral and non-natural building blocks.
• Existing computational methods lack specific tailoring for heterochiral macrocycle design.

Purpose of the Study:

• To develop a novel machine learning model for the rapid generation of energetically favorable macrocycle backbones.
• To address the limitations of current methods in heterochiral macrocycle design and structure prediction.
• To accelerate the drug design pipeline for macrocyclic therapeutics.

Main Methods:

• Development of a novel convolutional autoencoder model named CyclicCAE.
• Creation of a custom dataset of macrocycles through in-house and in silico methods due to data scarcity.
• Benchmarking CyclicCAE against the Generalized Kinematic loop closure (GenKIC) method in Rosetta.

Main Results:

• CyclicCAE demonstrates superior performance in generating energetically stable macrocycle backbones compared to GenKIC.
• The model rapidly produces designable structures, outperforming the state-of-the-art method.
• CyclicCAE offers functionalities for energy minimization, structure generation (diverse or similar), and inpainting with fixed elements.

Conclusions:

• CyclicCAE accelerates the design of stable macrocycles, significantly speeding up drug discovery pipelines.
• The novel machine learning approach facilitates the incorporation of complex chemical building blocks.
• This method is poised to advance the field of macrocyclic drug development.