The paper: Unlocking de novo antibody design with generative artificial intelligence.
What they did: Used sequence generation models to generate from scratch candidate antibodies for a specific target and then filtered them through automated testing to get a few ones that look good (for preclinical testing values of "looking good") but are very different from known antibodies for the same target.
Why this is important: Antibody-based therapeutics are potentially very powerful but antibody design is hard and implementation is expensive. If we can make the process faster and more efficient, and the output better, it has large downstream impacts across multiple medical fields.
Ideas to steal: If you're in biotech, the obvious ones. If you aren't: you can use generative models to create weird things - profitably weird things (training a model to create things just like the ones in the data isn't a good way to solve hard problems or outcompete anybody). But large-scale candidate generation via AI requires large-scale candidate testing via high-throughput laboratories or their equivalent. In most interesting fields you're going to get better results from AI technology with a state-of-the-art experimental setup and mediocre data than with state-of-the-art data and a mediocre experimental setup.