EP.003 · Research talk · ICLR 2022
Designing Molecules With Rules You Can Read
Minghao Guo · Postdoctoral Associate, MIT
Hosted by Dr. Omar Costilla Reyes
Listen or watch
Audio edition, 1 hr 2 min. Keyboard: K play/pause, J back 15 s, L forward 30 s. The hands-on code demo is left out of the audio edition; it is in the video.
Full video, 1 hr 8 min.
Chapters(17)
About this episode
Most generative models for chemistry need tens of thousands of training molecules, but real datasets for a specific class of compounds often contain only a few dozen, because every data point costs lab time. In this session, Minghao Guo presents a data-efficient alternative: a graph grammar, a set of production rules learned directly from the data, builds molecules step by step. Because the rules encode chemistry, generated molecules stay valid and synthesizable, and extra chemical knowledge can be built in by optimizing the grammar. The method was presented as an oral paper at ICLR 2022. For Interpretable, the point is the form of the model: a neural network's chemistry is buried in its weights, while a grammar is a list of rules a chemist can read, check, and edit.
Listen if: You are curious how AI can help drug and materials discovery when data is scarce, and what it means for a generative model to be readable.
The research behind the episode
ICLR 2022 (oral)2022Peer reviewed
Data-Efficient Graph Grammar Learning for Molecular Generation
Minghao Guo, Veronika Thost, Beichen Li, Payel Das, et al.
In plain language
Instead of a large neural network, the method learns a graph grammar: a compact set of production rules, constructed automatically from the training molecules, that assembles new molecules step by step. Because each rule respects chemistry, generated molecules stay valid, and extra chemical knowledge can be added by optimizing the grammar.
Why it matters for health
Drug and materials discovery often starts from a few dozen measured compounds. Methods that learn from small data, with rules a chemist can read, fit how laboratories actually work.
Key findings
- 01State-of-the-art results on three monomer datasets that contain only about 20 molecules each.
- 02On a polymer generation task with only 117 training samples, the method is competitive with approaches trained on 81,000 data points.
- 03The production rules are built from the data automatically, without hand-written chemistry.
What you'll learn
- Why data scarcity is the norm, not the exception, in molecular design
- How a learned graph grammar generates molecules that respect chemistry
- How the approach compares with deep generative models trained on far larger datasets
Working on something like this?
Suggest a guest or a paper for a future episode, or talk to equ about AI for health in your own work.
Keep listening
Which Mutations Make a Virus Spread? Reading 6.4 Million Genomes
Dr. Jacob E. Lemieux
Science · 2022Paper discussed
An interpretable Bayesian model that turns millions of viral genomes into a fitness estimate for each mutation.
Your Brain on Code: What Brains and AI Models Both Learn
Shashank Srikant
NeurIPS · 2022Paper discussed
Using brain recordings to read what AI models of code have learned, and what people represent when they read programs.