1. Background
Predicting how small molecules (ligands) bind to their protein receptor is an essential part of drug discovery. Previous ML-based scoring methods for protein-ligand binding have an overt featurization problem and require expensive data augmentation strategies. Using an SE(3)-transformer can avoid these problems. The SE(3)-transformer model utilizes a novel self-attention mechanism which enables the model to be guaranteeably equivariant to global rotations and translations of its input. In this project, we trained a SE(3)-transformer model for ligand-receptor pose scoring and affinity prediction.

2. Experimental Design and Methodologies
2.1 Training and Test Set
The poses used for training and testing the model are from the CrossDocked2020 Down-sampled Data Set. The CrossDocked2020 Data Set was generated by using the structures from PDB. It includes poses of ligands docked not only to its cognate receptor but also to other similar receptors

2.2 Converting examples to K nearest neighbors graph
2.2.1 Choosing Nodes for the Input Graph
Each node ∈ V in a graph G = (V,E) represents one of the atoms in the ligand-receptor complex, and it has a 3D coordinate x_i∈ R. The nodes in the input graph consist of all atoms of the ligand and the atoms of the receptor that are within cutoff distance d of the ligand. The cutoff distance d is the shortest Euclidean distance between two 3D coordinates.

2.2.2 Building K-Nearest Neighbors Graph
Based on the atoms chosen from a complex in the previous step, we built a K nearest neighbors of the ligand-receptor complex. The neighborhood of a node is the set of K nearest nodes where the distance is the Euclidean distance between 3D coordinates. We created undirected edges between a node and all its neighbors. In the KNN graph, node features are a one-hot encoding of the type of the atom and an extra column indicating if the atom is from the ligand or not. In this way, We constructed a knn graph for each sample in Cross-Docked downsampled dataset.

2.3 Optimizing SE(3)-transformer
Pose scoring and affinity prediction are trained with a Sigmoid Binary Cross Entropy loss and L2-like pseudo-Huber loss respectively. The pose predictions are evaluated via the test set's accuracy and the AUC-ROC curve and the affinity predictions are evaluated via RMSE and Pearson's R. The hyperparameters that will be optimized are the number of layers, the number of channels, the number of degrees, dropout probability and so on.

David Ryan Koes

University of Pittsburgh

Appearances