Stanford CS224W: ML with Graphs | 2021 | Lecture 15.4 - Applications of Deep Graph Generation

For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: stanford.io/3EwmakW
Lecture 15.4: Application of Deep Graph Generative Models to Molecule Generation
Jure Leskovec
Computer Science, PhD
Finally, we discuss an important use case for deep generative model for graphs, that is to conduct molecule generation. This application belongs to the goal-directed graph generation task which we mentioned in Lec 15.1. Here, we want to generate valid and realistic molecules with optimized property scores. We introduce Graph Convolutional Neural Networks (GCPN) as a solution to the task. The idea of GCPN is to generative desirable graphs via reinforcement learning, where the reward is defined by the goal of graph generation, and the policy network is parametrized as a Graph Neural Network (GNN). We compare the differences between GCPN and GraphRNN. GCPN is able to generate molecules with high drug-likeness score, which reveals a new direction for in-silico drug discovery. More information can be found in the paper: “Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation” arxiv.org/abs/1806.02473
To follow along with the course schedule and syllabus, visit:
web.stanford.edu/class/cs224w/
#machinelearning #machinelearningcourse

[@doyleBellamy03]

Excellent lecture accompanied by excellent visual aids. Thank you, Jure.