PharmRL: pharmacophore elucidation with deep geometric reinforcement learning

Abstract

Background: Molecular interactions between proteins and their ligands are important for drug design. A pharmacophore consists of favorable molecular interactions in a protein binding site and can be utilized for virtual screening. Pharmacophores are easiest to identify from co-crystal structures of a bound protein-ligand complex. However, designing a pharmacophore in the absence of a ligand is a much harder task.

Results: In this work, we develop a deep learning method that can identify pharmacophores in the absence of a ligand. Specifically, we train a CNN model to identify potential favorable interactions in the binding site, and develop a deep geometric Q-learning algorithm that attempts to select an optimal subset of these interaction points to form a pharmacophore. With this algorithm, we show better prospective virtual screening performance, in terms of F1 scores, on the DUD-E dataset than random selection of ligand-identified features from co-crystal structures. We also conduct experiments on the LIT-PCBA dataset and show that it provides efficient solutions for identifying active molecules. Finally, we test our method by screening the COVID moonshot dataset and show that it would be effective in identifying prospective lead molecules even in the absence of fragment screening experiments.

Conclusions: PharmRL addresses the need for automated methods in pharmacophore design, particularly in cases where a cognate ligand is unavailable. Experimental results demonstrate that PharmRL generates functional pharmacophores. Additionally, we provide a Google Colab notebook to facilitate the use of this method.

Keywords: Machine learning; Pharmacophores; Protein-ligand interactions; Virtual screening.

Fig. 1

Pharmacophore model with several pharmacophore features that matches the caffeine molecule (caffeine molecule included for illustration). The colors of the feature points are as follows: aromatic—purple, hydrogen acceptor—blue, hydrophobic—green

Fig. 2

Pharmacophore prediction pipeline. CNN is used to predict pharmacophore features from gridded binding site (top). Protein—pharmacophore graph is built by sequentially adding feature and protein nodes to it using RL framework (bottom)

Fig. 3

CNN architecture for predicting pharmacophore feature points. The CNN takes the local grid around the query point as input and provides confidence scores on the presence of the 6 classes at that point

Fig. 4

Steps followed to obtain pharmacophore feature points from a CNN predictions on a binding site

Fig. 5

MDP process used for iterative construction of the protein-pharmacophore graph. At each time-point t, the action is to chose the next graph ($G_{t+1}$). The environment takes this and provides a F1 score for that pharmacophore, along with possible graphs to choose from ($\{G_{t+2}\}$) for the next iteration

Fig. 6

The SE(3)-equivariant neural network takes a protein-pharmacophore graph as input and predicts the Q-value

Algorithm 1 Deep Q-learning algorithm to train Q function network

Fig. 7

F1 scores divided by the max F1 score attainable from ligand features for RL models trained and tested on ligand derived features (PharmRL_Ligand) and all CNN features (PharmRL_CNN)

Fig. 8

Performance of RL models on COVID screening experiments

Fig. 9

Performance of pharmacophore models on LIT-PCBA targets