CyclicCAE: A Conformational Autoencoder for Efficient Heterochiral Macrocyclic Backbone Sampling

Abstract

Peptide macrocycles are a promising therapeutic class. The inclusion of heterochiral and non-natural amino acids allows far more folds and functions to be accessed, but creates challenges for rational design -- particularly for sampling plausible mainchain conformations. We developed a conformational autoencoder called CyclicCAE to rapidly generate energetically favourable macrocycle scaffolds for heterochiral design and structure prediction. Given the absence of large, available macrocycle datasets, we created a custom dataset in silico using physics-based simulation methods. Trained on this, CyclicCAE produces energetically stable mainchain conformations and designable scaffolds more rapidly than the current state-of-the-art method, the Rosetta software suite's Generalized Kinematic Closure (GeneralizedKIC) method. We show that, despite being trained exclusively on synthetic data, CyclicCAE accurately captures the conformations accessible to peptide macrocycles found in the Protein Data Bank, with considerable speed advantages over GeneralizedKIC. We also demonstrate that CyclicCAE enables users to perform energy minimization in isolation or in a target-bound context, to generate structurally similar or diverse outputs by Markov Chain Monte Carlo sampling, and to conduct inpainting with fixed motifs. This method, which we release under a free and open source licence, will accelerate macrocycle design pipelines, speeding the development of peptide therapeutics.