Short answer: Dov Gertz says cheap genome sequencing has put 400 times more sequenced DNA online than text, but high-quality labeled data is scarce, leaving molecular AI five to ten years behind text, image and audio models.
Watch this answer from Episode 32 (starts at 5:03) · Episode 32 notes and transcript
From Episode 32: From DNA to Drugs: How AI Is Rewriting Human Biology, starting at 5:03.
Dov Gertz [5:03]: So you can imagine if we look at text, image, and audio, then each one of those kind of types of information can be represented easily by computers. So text, we know that computers can comprehend. Images are just pixels. Audio are wavelengths that we can comprehend as a in computer science. So how do we comprehend molecular data? So in essence, if we look at, for instance, DNA, then our genome, every single person that's listening to this and in this audience currently, me, Jordan, and Sam, you can represent as a 3,200,000,000 nucleotides, in essence, the letters that represent our genetic code. And the code is built out of four bases, a, c, g, and t. So if computer code is built out of zeros and ones, then we are built out of these four different basic building blocks. So that's 3,200,000,000 DNA nucleotides in our full genome that represents us as humans. So using these language models that similar architecture is trained on text image and audio, we can now start comprehending the textual kind of representation of the genetic code.
Dov Gertz [6:16]: So same goes for DNA and RNA and also protein data. All can be represented as text in essence, and then we harness the foundation models that are trained on similar architectures that are trained on text image and audio for the molecular data, and then we can use them to, predict molecular, activity to generate novel molecules, generate alterations in molecules, predict given patient samples if they will respond or not respond to particular drugs. So it's just opening an explosion of value of what we can do with this data. Another interesting thing to think of is because of the drastic reduction in cost of sequencing genomes, there's 400 times more sequenced DNA on the Internet than text on the Internet. So there's a massive amounts of data for model training. But with that said, there's very, very low amount of high quality labeled data, so data that we actually know what it means. So there's a huge challenge, and we're about five to ten years behind what's happening in text image and audio. But, yeah, really pushing the frontier in terms of what's possible with these models. So very exciting. We're, like, in the open AI, Anthropic, DeepMinds kind of race just, like, seven or five to seven years behind, and exciting times.
Dov Gertz [7:38]: Yeah.
