Prediction of molecular field points using SE(3)-transformer model

Abstract

Due to their computational efficiency, 2D fingerprints are typically used in similarity-based high-content screening. The interaction of a ligand with its target protein, however, relies on its physicochemical interactions in 3D space. Thus, ligands with different 2D scaffolds can bind to the same protein if these ligands share similar interaction patterns. Molecular fields can represent those interaction profiles. For efficiency, the extrema of those molecular fields, named field points, are used to quantify the ligand similarity in 3D. The calculation of field points involves the evaluation of the interaction energy between the ligand and a small probe shifted on a fine grid representing the molecular surface. These calculations are computationally prohibitive for large datasets of ligands, making field point representations of molecules intractable for high-content screening. Here, we overcome this roadblock by one-shot prediction of field points using generative neural networks based on the molecular structure alone. Field points are predicted by training an SE(3)-Transformer, an equivariant, attention-based graph neural network architecture, on a large set of ligands with field point data. Resulting data demonstrates the feasibility of this approach to precisely generate negative, positive and hydrophobic field points within 0.5 Å of the ground truth for a diverse set of drug-like molecules.

Water position prediction with SE(3)-Graph Neural Network

“…An encoding module was used to encode node features of 10 atom types into 16 scalar features. Node and edge features are processed with an interaction module, which performs interaction between nodes with 6 SE(3)-Transformer blocks 50 followed by a ConvSE3 50 operation. SE(3)- Transformer blocks consist of a sequence of SE(3)-Transformer, layer normalization(LayerNorm) 51 , and exponential linear unit(ELU) activation 52 , where SE(3)-…”

“…After SE(3)-Transformer, node features are processed into 64 scalars, 64 vectors, and 64 traceless matrices, and ConvSE3 transforms node features into 64 scalars and 32 vectors. Finally, the structure module converts node features into two scalars and two vectors via six linear blocks, where linear blocks consist of a sequence of LayerNorm, ELU, and LinearSE3 50 , where LinearSE3 performs a 1x1 convolution-like operation on each node. The resulting node features on probe nodes are trained to predict the probability of neighboring water molecules near the probe and the displacement vector from the probe to the predicted position.…”