Biology is a steering problem.

In short.

• Biology is a control problem, Lucas argues: out of all possible arrangements, find those that support what we want, and steer the system toward them.
• What holds biology back, in his account, is not theory but running the experiments. Automation took hold where one workflow repeats, much less where the workflow keeps changing.
• The instruments exist, he says, and language models can already draft protocols, though they still get some wrong. What he saw as missing: turning a protocol into code for the devices a given lab has.
• That is where we work. ProviGen builds lab-in-the-loop AI that chooses each next experiment, plans it for the supported instruments a lab already has, and learns from every result. Today with partner labs, on their instruments; their scientists decide what runs.

From predicting to steering.

Lucas opens with the frame that computer science, medicine and robotics gave him: the core question is how to move a system from one state to another, which is what all of engineering is about 1:29.

“If you look at biology as a discipline, it’s all about finding, out of all possible arrangements, those arrangements that support what we want […] and then steer the system in the direction that is desirable versus undesirable.”

We think this is the right starting point. Most AI for biology learns to predict how a living system behaves. A lab has to answer the inverse question: which interventions, in what order, move a living system from where it is to where it needs to be. Predictions narrow the options. In biology, the experiment settles which one works, and choosing that experiment well is how a system is steered. That is why we put the lab in the loop.

The experiment is the bottleneck.

Scientific theories, Lucas says, are not what holds biology back; conducting the experiments is 14:23. People trained for years to think about the science, he says, spend most of their research time carrying it out by hand 16:06.

He puts much of this down to how academic research is funded and rewarded 18:18. A lab that does buy a liquid-handling robot meets the next problem:

“Even if you would get such a device, you don’t have the allocated funds to pay a programmer or a computer scientist to program that robot. So that would fall to the scientists, who, because of the incentive structure in academia, are actually pressured to publish as soon as possible.”

The pressure comes from outside, he stresses, even though automating “will definitely pay off” in the long term 21:34.

Where automation has been used most, he says, the workflow stays the same. Some pharma companies and contract research organisations program a screening workflow once and run it across very large screens 23:39. Research is different:

“For this kind of labour, that has always-changing workflows, laboratory automation didn’t lend itself at all. It was simply not feasible, from a time-investment perspective, to really build systems that support that strongly.”

His remedy: cut the friction of understanding the devices and programming them for each protocol “to as close as possible to zero” 25:35, bring as many uses as possible onto the same instruments 26:46, and solve access and trust 27:06.

In our reading, biofoundries, CDMOs and many CRO teams work between the two: they run automation every day, and each new project, cell line or process still changes the protocol.

This is the gap we see. A development campaign changes its protocol every round, because each result changes what is worth testing next. It needs a way to choose the next experiment and turn it into a protocol the lab’s instruments can run.

Why now.

As we read the conversation, three things have come together.

• The instruments exist. Walking through one workflow step by step, Lucas says that everything inside an automation setup is “very much automatable”; the least automatable step is unboxing reagents and loading the devices 49:53. Some steps still need custom set-ups, and whether a step is worth automating depends on its capital cost 57:50. The basic operations, he says, are much the same from bacteria to mammalian cells 41:37.
• Models can draft protocols. Language models can already design an experiment step by step at the protocol level, he says 1:00:56, though they still sometimes hallucinate or miss the specific use case 1:02:00.
• Programming is getting easier to learn. In software, he notes, people have begun to program in natural language and let translation layers handle the levels below 1:06:02. AI assistants have made programming somewhat easier to learn, he says, but programming, testing and iterating on a robot still take a lot of time 20:39.

What was still missing when the episode was recorded, in his words:

“What is missing much more is not this protocol level but actually the coding level below it: how do you translate this natural-language protocol into code for your laboratory robotics? For that you need to understand which devices you have. You need to know how to access the devices, which programming language. Is there even any programming language? That’s very much still missing, I would say.”

This is where we work: turning a drafted protocol into one that runs on the supported instruments a lab actually has.

What we build.

ProviGen builds lab-in-the-loop AI for experimental biology. A lab brings a process and a target. Our system chooses each next experiment for what it will teach, turns it into a protocol for the lab’s supported instruments and workflows, and learns from every result. Each round updates its model of that process, and the next experiment goes where the model is least sure or the gain is largest. Every result is recorded with the conditions, protocol version and instrument data that produced it, failures and deviations included.

Today we do this with partner labs, on their instruments. Scientists set the target, approve each batch and can override any decision. Automated scheduling and execution are in development, and autonomy grows only where it has been shown to work. Partner to every lab, competitor to none.

Instruments and models are both improving. What we build joins them to one lab’s process: choosing the next experiment, turning it into a protocol for the supported instruments that lab has, and recording how each result was produced. Why we think it compounds:

• Better models make it better, not obsolete. A stronger model proposes better experiments; they still have to be run in a lab and recorded.
• With each experiment run on ProviGen globally, we improve our pipeline.
• Trust compounds. A system that has earned a lab’s trust is given its hardest problems.

Where it leads.

Lucas also talks about access. Five years ago, he says, few biologists could code; today almost every biologist dabbles in Python to some degree 1:03:48. He expects the same for experiments: less effort handling liquids, more on deciding what is worth testing 1:04:18. That is his view of where the field could go.

If he is right, many more people will run experiments, and choosing them well will matter more, not less. That is the company we are building: the system a lab trusts to choose its next experiment.

The episode, chapter by chapter.

1. 1:17Biology as a steering problem
2. 5:31From medicine and robotics to lab automation
3. 14:23The bottleneck is running experiments
4. 18:18Why labs stay manual: funding and incentives
5. 22:54Automation built for one repeated workflow
6. 25:22Friction, utilisation, access and trust
7. 34:39Walk-away time and what can be automated, step by step
8. 57:58From predicting proteins to programming experiments
9. 1:06:56How language models learn biology
10. 1:22:43Why this problem: choice and access

Also on Substack · Apple Podcasts · Spotify