Autonomous Discovery Starts with Automation

Two weeks ago, I wrote about an emerging category of software using LLMs to help labs translate manual experiment protocols into automation scripts. Since then, I’ve learned of some other entrants to this field, one of which - a company called Tetsuwan - broke my theory for how this category was evolving. In fact, after talking to their Co-founder and CEO, Cristian Ponce, I thought that his reasoning and vision for the future was different enough to justify its own post. So that’s what this is.

---

Thanks for reading Scaling Biotech! In between posts, I’m working on a guide to evaluating the impact, feasibility and cost of different use cases for AI in drug discovery. If you’d like to read an early copy when it’s ready, send me an email at jesse@merelogic.net or fill out the form at merelogic.net.

---

Two weeks ago, I argued that the companies that have gotten into the AI-for-automation space all started out trying to help scientists design manual protocols by analyzing protocol descriptions in the literature. But they ended up pivoting into automation, which is a fundamentally much harder problem, because that’s where the actual demand is.

Tetsuwan’s founders also didn’t start out wanting to solve automation, or at least not by the standard definition of automation. Before founding Tetsuwan, Cristian and his co-founder, Théo Schäfer, were interested in autonomous discovery - the idea of handing over the full end-to-end process of defining, planning and running experiments to an algorithm.

Many folks in this space see the first step to autonomous discovery as building a cloud lab - basically, a fully automated CRO in which scientists upload detailed protocol descriptions, wait, then get their results without any other humans in the loop. Once a person can do that, letting an algorithm do the same through an API is the easy part.

Cristian noticed that the reason many cloud labs, or attempts at cloud labs, have failed was automation engineering: translating the experiment design that scientists have in their mind into concrete automation scripts that work reliably.

Designing automation scripts is complex and difficult. There is an entire profession of automation engineering. And automation engineers are well paid and hard to find for a reason.

Because there has been no way to automate that work, or even to do it cost effectively with humans, all attempts at cloud labs so far have essentially delegated that work to the customer. Instead of solving the hardest part, they just punted it.

Tetsuwan was founded on the hypothesis that the only way to solve autonomous discovery (with a successful cloud lab as a potential step along the way) was to actually solve the hard part by building an autonomous automation engineer.

So they ended up in roughly the same place where the companies Potato and Briefly Bio from last time eventually joined them. But because of how they got here, their approach is a bit different.

First, rather than broadly addressing all possible steps in an experiment, they decided to focus on getting liquid handlers right before expanding. Many labs have liquid handlers, but they are notoriously under-used for reasons that Briefly recently explored on their blog. Tetsuwan believes that if they can crack the problem for liquid handlers, they can do it for the rest of the lab next.

Second, Tetsuwan has deliberately modeled their software around how automation engineers work. Before writing a single line of code, they talked to more than 200 scientists and automation engineers. They studied the questions the engineers were asking the scientists and how they were using the answers. Then they broke it down into individual steps that they could model and replicate.

What they ended up with is a “virtual automation engineer” that asks scientists a series of questions based on a free text description of their protocol. Each block of questions has specific logic - both LLMs and other algorithms/models - that interpret the answers and determine the next questions. The overall process currently has about 50 nodes, though many of these are abstracted from the user.

Like all the players in this space, Tetsuwan is rapidly developing their approach. So it’s probably too early to predict how the category will evolve, and who will be the first to fully crack this problem. But it’s exciting to see so many options emerging, and the range of visions that they’re pursuing.

---

Thanks for reading Scaling Biotech! In between posts, I’m working on a guide to evaluating the impact, feasibility and cost of different use cases for AI in drug discovery. If you’d like to read an early copy when it’s ready, send me an email at jesse@merelogic.net or fill out the form at merelogic.net.