A Deep Generative Model for Molecular Graphs by Niloy Ganguly

Explore a deep generative model for molecular graphs in this 23-minute conference talk from the International Centre for Theoretical Sciences. Learn about designing random graph models using variational autoencoders to discover new, plausible drug-like molecules. Examine the limitations of current models and delve into NeVAE, a variational autoencoder for graphs. Understand the probabilistic encoder and decoder, their properties, and how they guarantee structural integrity. Discover how the training process is permutation invariant and efficient. Review experimental results showcasing a smooth, meaningful space of molecules and evaluate the model's performance using metrics for validity, novelty, and uniqueness. Investigate methods for predicting and optimizing molecule properties using sparse Gaussian processes and Bayesian optimization. Conclude with key takeaways and participate in a Q&A session to deepen your understanding of this innovative approach to molecular graph generation.

Start
Designing Random Graph Models using Variational Autoencoders...
Discovering new, plausible drug-like molecules
Generative models for molecule design
Limitations of current models
NeVAE: A variational autoencoder for graphs
The probabilistic encoder
Encoder has desirable properties
The probabilistic decoder
Decoder guarantees structural properties
Training is permutation invariant
Training is efficient
Experimental setup
Smooth, meaningful space of molecules
Quantitative evaluation metrics
Competing methods
Validity of the discovered molecules
Novelty of the discovered molecules
Uniqueness of the discovered molecules
Predicting & optimizing for molecule properties
Property prediction Sparse Gaussian Process
Property maximization Bayesian Optimization
Conclusions
Thanks!
Q&A