Decoupled coordinates for machine learning-based molecular fragment linking

• Published in: Machine learning: science and technology Vol. 3; no. 1; pp. 15029 - 15040
• Main Authors: Fleck, Markus, Müller, Michael, Weber, Noah, Trummer, Christopher
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.03.2022
• Subjects: Ablation; Coordinates; deep generative models; drug design; fragment linking; Machine learning
• ISSN: 2632-2153; 2632-2153
• DOI: 10.1088/2632-2153/ac50fc

Abstract Recent developments in machine learning-based molecular fragment linking have demonstrated the importance of informing the generation process with structural information specifying the relative orientation of the fragments to be linked. However, such structural information has so far not been provided in the form of a complete relative coordinate system. We present a decoupled coordinate system consisting of bond lengths, bond angles and torsion angles, and show that it is complete. By incorporating this set of coordinates in a linker generation framework, we show that it has a significant impact on the quality of the generated linkers. To elucidate the advantages of such a coordinate system, we investigate the amount of reliable information within the different types of degrees of freedom using both detailed ablation studies and an information-theoretical analysis. The presented benefits suggest the application of a complete and decoupled relative coordinate system as a standard good practice in linker design.