Designing Molecules With Rules You Can Read
Minghao Guo
ICLR · 2022Paper discussed
A generative model for chemistry built from learned rules: valid by design and readable by chemists.
Topic
AI methods for designing and evaluating molecules, especially when experimental data is scarce and every data point costs lab time.
1 episode
Drug discovery and materials design often begin with only a few dozen measured compounds, because every data point costs laboratory time. This topic covers how the podcast treats AI methods for that setting. Its episode is Episode 003, in which Minghao Guo presents Data-Efficient Graph Grammar Learning for Molecular Generation, an oral paper at ICLR 2022.
The paper begins from a mismatch. Deep generative methods for molecules usually require tens of thousands of training samples, while datasets for a specific class of chemicals are often limited to a few dozen. Its approach is a learnable graph grammar that builds molecules from a sequence of production rules, constructed automatically from the training data. Further chemical knowledge can be added by optimizing the grammar. The authors report state-of-the-art results on three monomer datasets of about 20 samples each and, for polymer generation with 117 training samples, performance competitive with methods that use 81 thousand data points. The talk walks through the grammar's components, how it is learned bottom-up, and a demonstration of the code. The episode and the paper behind it are listed below.
Minghao Guo
ICLR · 2022Paper discussed
A generative model for chemistry built from learned rules: valid by design and readable by chemists.