Converge Bio founder Dov Gertz on training AI on DNA, RNA and proteins, why labeled data is the bottleneck, and how AI could make drugs cheaper and faster.
AI can design new drugs by reading DNA, RNA and proteins the way language models read text, and then predicting how molecules will behave before anyone runs a lab experiment. That is the work of Converge Bio, and its founder Dov Gertz explained it on Episode 32 of Built This Week. His view: the field is roughly five to ten years behind text and image AI, held back by labeled data rather than chips, but the payoff could be faster, cheaper and more personalized medicine.
Dov Gertz describes Converge Bio as a frontier AI lab that trains foundation models directly on DNA, RNA, proteins and small molecules. On the episode he said the company was founded a year and ten months earlier, had raised $33 million and had a team of 40. Its customers are drug developers, and the models are delivered through systems that fit into the design and analysis work those teams already do.
He traced the idea back to ChatGPT. Once transformers changed what computers could do with text, images and audio, the scientific community started applying the same architectures to molecular data.
Watch Dov Gertz explain AI on DNA in Episode 32 (starts at 5:03) · Episode 32 notes and transcript
Computers already understand text, images as pixels, and audio as wavelengths. The question Dov set out to answer is how a computer understands a molecule. His answer starts with the genome: every person can be represented as about 3.2 billion nucleotides, written with four letters, 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.
— Dov Gertz, Episode 32
Because DNA, RNA and protein data can all be represented as text, models built on the same architectures as language models can learn from them. Dov says that lets Converge predict molecular activity, generate new molecules or alterations to existing ones, and predict from patient samples whether someone will respond to a particular drug.
There is no shortage of raw data. Dov says sequencing has become so cheap that there is far more DNA online than text:
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.
— Dov Gertz, Episode 32
The catch is quality. Very little of that data is labeled, meaning we know what it actually does. That is why he places molecular AI about five to ten years behind the text, image and audio models, in its own version of the race between the big labs.
Jordan asked who the customers are. Dov says they are biotech and pharma companies, with academic labs now being added. They use three systems, each built to plug into an existing workflow:
The point, Dov says, is that customers keep the workflow they know from the lab and run it computationally instead. When Jordan asked whether the gain is speed or cost, Dov said it is both, plus quality. He frames value as faster, cheaper and better, and says most products deliver one or two of those, while AI lets Converge deliver all three.
Before predicting the future, Dov listed the bottlenecks: data, model architecture and compute. In his view the industry is not yet at the scale where compute limits it. The real limits are data and architecture, so Converge pays for large data collection to build what it hopes will be the best dataset in the world and keeps working on new model designs.
If those bottlenecks are solved, he expects three things: drugs reaching patients much sooner, lower prices as more competing drugs reach the market, and true personalization. Today, he says, drugs are sometimes called personalized but are not, because FDA approvals require strict inclusion and exclusion criteria. As time and cost to market fall, he expects that to change. He was clear that these are predictions from an optimist, and closed with this:
But can you actually improve your quality of life and length of life? I think those are the real questions that we should be asking ourselves.
— Dov Gertz, Episode 32
Sam asked which diseases might benefit first. Dov's hypothesis, based on where the data is, is cancer, because it has the most data. He also pointed to diabetes, inflammatory bowel disease and autoimmune disease as large markets, and expects neurodegenerative disease to come later because of its molecular complexity. For more on AI in medicine, see our AI in healthcare page.
Dov splits drug development into two parts. Discovery, which covers target discovery, drug discovery and preclinical studies, takes roughly the first five years and carries little regulation because drugs are not yet tested on people. Clinical trials take roughly another five years and are where regulation starts. The easiest place to save time is discovery.
He does not expect clinical trials to get shorter for the foreseeable future, possibly not within a decade. Jordan summed it up as a ten year path becoming six to eight years, and Dov agreed. What he does expect is higher trial success rates: better molecules going into trials, and better prediction of how patients will respond, so trials can choose the right population. Fewer people would get drugs that do not help them, and more would get ones that do. For a related build, see our clinical trial site selection dashboard.
Jordan asked whether these models run on the same Nvidia hardware as language models. "Exactly the same," Dov said. The difference is that compute is not the bottleneck for molecular AI yet. He believes the market is one breakthrough away from hitting the same scaling laws seen in text, image and audio, and then compute would matter a lot more.
Demand is already growing. Dov says pharma companies are sharply increasing their token budgets for text and image models, used for regulatory filings, data analysis and clinical trial design. Molecular-level demand, he says, is the part that will explode next. Jordan noted that GPUs are already hard to get, and Dov added that he does not think chips will be the massive bottleneck long term, even if they are one today.
In the news segment we talked about OpenClaw. Jordan installed it early, when it was still called Clawdbot, and finds it fun but still low in value, mostly useful for scheduled, task-style work. What interests him is people becoming managers of agent teams. He has been using multi-agent teams in Claude, and connected it back to Dov's work: how many agent scientists could run drug discovery experiments that nobody tries today because they look too risky? We covered multi-agent work in more depth in Episode 30.
Dov asked whether Jordan bought a separate computer to run it, as many people have. He did not. He hosted it on DigitalOcean in about five minutes for around $6 a month, and argued that work is moving to cloud desktops, not back to local machines. His own laptop is maxed out on storage, RAM and energy, and he would rather have a light machine that connects to cloud desktops and agents.
Dov agreed that most of Converge's computing runs in the cloud, with some small experiments on local machines. Jordan added that he finds it shocking Apple and Meta have not launched their own clouds, and that he tells his developers to be ready to change how they work almost every week as new models arrive.
Dov walks through the three Converge systems on screen, and Sam opens the show with a retro cell defense game he built that Converge could use at conferences. Watch Episode 32 of Built This Week, where the full transcript is also on the page, or read more about Dov Gertz.
