Key Takeaways

  • Conventional lab-in-the-loop drug discovery is bottlenecked by physical synthesis and assay cycles, which creates tiny datasets that lack counterfactual density.
  • Leash Bio screens massive combinatorial chemical libraries once, creating an amortized latency setup where software agents explore biology with zero physical delay.
  • Ian Quigley has screened roughly 1,000 proteins against chemical libraries since launching Leash Bio, building fixed ground-truth datasets for agent benchmarking.
  • Leash Bio is ramping production to generate 20 new biological environments every day to test if reinforcement learning policies generalize to unseen biology.
  • Agents in this framework are restricted to pre-measured actions, trading open-ended synthesis freedom for thousands of rapid exploratory rollouts.

The Wet Lab Latency Trap

Reinforcement learning thrives where feedback loops are instant. Video games, chess engines, and software sandboxes let algorithms run millions of trials an hour, learning from mistakes through cheap rollouts. Physical biology does not work like that.

In standard biotech setups, testing a hypothesis requires ordering compounds, running assays, and waiting days or weeks for data. Every physical step burns cash and calendar time. Worse, budgets force scientists to test only their best guesses. That kills counterfactual density, which is the data on what would have happened if the model picked an alternative path.

Quigley diagnosed this core failure mode directly: “Because if you're doing conventional lab in the loop, it does take a long time and because those things themselves have limited budgets, it's also not counterfactually dense. So, you got to keep running it and do different choices with every rollout to make the agents learn.”

Without dense negative examples and immediate feedback, standard reinforcement learning falls flat in wet labs.

Amortizing Physical Reality into Software

Leash Bio solves this by separating the wet lab work from the agent rollout loop. Instead of testing one compound at a time in real time, they run massive combinatorial screens upfront across thousands of molecules and targets.

They record every interaction into a fixed environment called Minos. Once the physical ground truth is captured, machine learning models explore that chemical space without waiting on a pipette. Quigley calls this amortized latency: “Broadly, the way that we think about this is it is a lab in the loop with amortized latency. So, it's like go out, you record the thing once, and then you can have a huge number of rollouts through that thing, effectively infinite rollouts.”

By measuring everything at once, the cost and time of the physical experiments are paid upfront. The downstream feedback loop becomes purely digital. Quigley noted: “Where a lot of these agents are making the best and the fastest progress is where there's instant feedback that's very scalable. It's really hard to do that in the physical world, but we think that there's probably a handful of biological questions that you could rebuild in this way.”

The Constraint of Pre-Measured Boundaries

This architecture comes with a distinct trade-off. An agent trained in an amortized environment cannot propose an entirely custom molecule outside the pre-screened library and get instant feedback. The agent is locked inside the boundary of what Leash Bio physically measured.

To make this useful for drug discovery, the library must be vast, and the environments must multiply. Leash Bio has already screened approximately 1,000 proteins against its chemical libraries. Now the team is pushing throughput to 20 new biological environments per day.

The objective is to see whether an agent that learns structure-activity relationships across hundreds of pre-measured environments can generalize its policy to completely new proteins. If the policy transfers, you get the speed of software agents paired with real physical ground truth.

What to Do With This

Audit your company's slowest external data dependencies this week. If your models wait on human annotators, third-party APIs, or physical testing cycles, stop running synchronous queries in your training loop. Batch-generate a dense, fixed evaluation sandbox of 1,000 edge cases upfront so your team can test model iterations instantly without waiting on real-world latency.