Can you talk to us a little bit about what the Paraphrase project is about?

Well, the paraphrase project was an exploration of the power of AI design tools for protein design to create or reformulate, we use the analogy paraphrase, to paraphrase the amino acid sequence that comprises proteins.

The question that I had several years ago that I engaged a larger team about was, could AI tools that are used for incredible purposes to design new proteins, new structures aimed at disease, at new kinds of pharmacologics, new kinds of solutions to health problems, for example.

Could they be used to take a toxic protein like ricin, which is a well-known toxic protein, and be asked to change out the sequence a bit.

In other words, redesign the protein, but try its very best to hold the structure in place and to potentially keep the activity unchanged, to retain the activity while changing the amino acid sequence.

And the question we had was, Would today's, at the time, the existing sequence screening tools that are used to screen orders that scientists make to make DNA so they can make the proteins in their labs, could these synthetic homologs, as we call them, or redesigns, slip through the screening tools? And we created several thousand variants, all done in silico.

These are computer sequences.

Are there like proactive guardrails that you think should be put on AI when it comes to developing new proteins like this?

It's assuming that there's going to be both, you know, that AI isn't going to replace doctors, which is also something you hear about, that there's going to be both and how do you get them to collaborate for the best outcome? So, you know, I'd love to hear from you guys kind of what prompted you to start this study and that sort of thing.

Yeah, I'm happy to just jump in and give you a little bit of background on what started the study.

So we were all involved in a prior study that actually was featured in the New York Times with big headlines saying things like AI chatbots defeat doctors at diagnosing illness.

And being so involved with that study, we saw what was going on, some of the results where We had six complicated diagnostic and therapeutic vignettes.

We had a scoring rubric and we basically had a set of practitioners, active physicians use the GenAI tool or use conventional resources.

And even those people using the tool got around 75% right on the zero to 100% metric.

GenAI alone got 92% correct, and that's all the headlines.

Like, is this really, can you imagine this just being like consumer tech or is this going to fall by the wayside?