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.
Speaker
Postdoctoral Associate, MIT
Minghao Guo is a postdoctoral associate at MIT and first author of "Data-Efficient Graph Grammar Learning for Molecular Generation," an oral paper at ICLR 2022 written with collaborators from MIT and the MIT-IBM Watson AI Lab.
Minghao Guo works on computational design. When the paper in his episode was presented in 2022, MIT News described him as an MIT graduate student and the lead author. His own website says he received his Ph.D. from MIT CSAIL in September 2026, advised by Wojciech Matusik, and that he is now a postdoctoral associate at MIT. The website describes his research as Scientific World Models for computational design, with applications that include synthesizable membrane polymers.
His episode covers Data-Efficient Graph Grammar Learning for Molecular Generation, written with Veronika Thost, Beichen Li, Payel Das, Jie Chen, and Wojciech Matusik and presented as an oral paper at ICLR 2022. The paper starts from a practical limit: existing molecular generation methods usually need tens of thousands of training samples, while datasets for a specific class of chemicals are often limited to a few dozen. Its method uses a learnable graph grammar that builds molecules from a sequence of production rules. The rules are constructed automatically from the training data, and 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 a polymer task with 117 training samples, performance competitive with methods that use 81 thousand data points. The code is public on GitHub.
In Episode 003, Designing Molecules With Rules You Can Read, the talk moves from the limits of deep learning and data scarcity to the benefits of formal grammars, the components of a grammar, and the experimental results. The grammar is the point for this show: its production rules can be read one by one by a chemist, rather than staying buried in network weights. The episode is filed under AI in Health, Drug Discovery, and Interpretable AI.
Minghao Guo
ICLR · 2022Paper discussed
A generative model for chemistry built from learned rules: valid by design and readable by chemists.