NMR-Like: A New Molecular Descriptor for Virtual Screening Analysis

Abstract

In the most recent years, the discussion surrounding the application of deep learning to virtual screening has been focused on the use of neural embeddings in comparison to conventional descriptors to represent the structural and physical characteristics of ligands and/or targets to catch molecular features otherwise hidden by the canonical approaches. With the growing application of Graph Neural Networks, which are short-range embeddings for atomic neighborhoods, the focus on embeddings has increased. In this paper, we introduce a new descriptor called NMR-Like, which tries to address the limits of current Fingeprints while preserving computationally efficient numerical embedding. NMR-Like is the first molecular descriptor used in the Virtual Screening domain and is based on H-NMR spectra of small compounds. Unlike Molecular Fingeprints, it maintains visible molecular features and structure despite being a numerical embedding of fixed size, allowing the operator to interpret the distinct characteristics of each active molecule. It is an appropriate embedding because of this feature, as well as the low computing cost required for training neural networks that use it as input data.