Enhancing Ligand-Based Virtual Screening with 3D Shape Similarity via a Distance-Aware Transformer Model

Following the assumption that chemically similar molecules exhibit similar biologcial properties, ligand-based virtual screening can be a valuable starting point in drug discovery projects. While 2D-based similarity metrics generally focus on similar scaffolds or substructures, 3D-based methods can capture the shape of a molecule, allowing for the identification of compounds with different scaffolds. We recently published a proof-of-concept study which demonstrated how a Transformer model can be adapted to preserve 2D similarities in latent space in the form of Euclidean distances. In this work, we extend this research and prove that the approach can be adapted to 3D similarities. We use pharmacophore-based shape similarity as 3D similarity measure. We show that the model is able to enrich the predicted most similar hits with compounds with different scaffolds that are indeed similar in 3D space. Whereas classical pharmacophore- or shape-based 3D similarity methods rely on expensive alignment processes, in our approach, we identify similar compounds directly by the Euclidean distances in latent space. This enables for the first time the 3D screening of ultra-large databases with high efficiency.