Research

Peptides can be thought of as ~50-letter words made with a ~20-character alphabet. Some peptides have cool properties like killing bacteria. There are labs that will synthesize arbitrary peptides for you, but 20^50 possible peptides is obvs way too big to search with brute force. My brother figured we could train a model to predict bacteria-killing ability. We scraped a bunch of experimental data from all the online databases we could find and tried a few different architectures before settling on a CNN. It SOTAed on the binary yes/no activity measure and on predicting exact amount of activity. Paper here: Deep learning regression model for antimicrobial peptide design

Talks

Extracting Structured Data from Legal Documents

External Blogs

Teaching an AI to Draft Magic: the Gathering

Atrium Records

Misc

One Third Blue