Novos métodos para determinar a qualidade da representação quiral em modelos baseados em GNN

Abstract

How do graph neural networks, especially models such as SphereNet and ChIRo, in- corporate the perception of molecular chirality? This study aimed to analyze and evaluate the capacity of graph neural networks to incorporate the perception of molecular chirality, focusing on specific models such as SphereNet and ChIRo. We hypothesized that current graph neural networks, including SphereNet and ChIRo, exhibit limitations in fully in- corporating the perception of molecular chirality, particularly in complex and continuous cases. To test this concept, we conducted tests and developed datasets that more com- prehensively reflect chiral incorporation, including (I) RSA classification as an alternative to traditional RS classification; (II) classification of complex chiralities; and (III) creation and use of a continuous chirality dataset (CCM). We conclude that RSA classification demonstrated potential to improve model accuracy, while the CCM dataset revealed the importance of geometry as a determining factor in chiral classification. Furthermore, we identified limitations in current models for classifying complex chiralities. Scientifically, this work points to the need for developing new neural network architectures that better incorporate the properties of molecular chirality. Socially, there is potential for improved prediction and design of chiral molecules, impacting areas such as drug development and materials science.