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
How artificial intelligence is changing prevention, diagnosis, treatment, and research, told by the people building and studying it.
3 episodes
On Interpretable, AI in health is approached through three working examples rather than a general survey. Episode 001 looks at mental health care, where Thomas R. Insel, a former director of the National Institute of Mental Health, argues that clearer signals and measurable recovery are needed, and where digital biomarkers from everyday devices could make care more continuous. Episode 002 looks at genomic surveillance, where a Bayesian model called PyR0, described in a 2022 Science paper, was fit to 6.4 million SARS-CoV-2 genomes to estimate which mutations help a lineage spread. Episode 003 looks at chemistry, where a learned graph grammar generates valid molecules from datasets that hold only dozens of examples, as reported in an ICLR 2022 oral paper.
What links the three is the question the show is named for: can the people who rely on a result see why it says what it says? In these episodes that means a clinician reading a signal, a public health team reading a fitness estimate for each mutation, or a chemist reading the production rules of a grammar. The episodes in this topic show how each field approaches that question, and the papers behind them are listed below so each claim can be traced to its source.
Minghao Guo
ICLR · 2022Paper discussed
A generative model for chemistry built from learned rules: valid by design and readable by chemists.
Dr. Jacob E. Lemieux
Science · 2022Paper discussed
An interpretable Bayesian model that turns millions of viral genomes into a fitness estimate for each mutation.
Dr. Thomas R. Insel
Penguin Press · 2022
Why mental health care needs signals people can understand, and where data and digital biomarkers fit.