Standardized Experiments Enable Predictive Models

This idea is the core of insitro’s strategy, because in drug discovery the hard part is often not the model, it is creating training data where every cell experiment was run the same way, measured the same way, and linked back to human disease signals. Old pharma data is usually a patchwork of assays from different teams, tools, and goals, which makes it noisy for prediction. insitro was built to spend upfront on automated labs and purpose built assays so its models learn from cleaner cause and effect data, not lab history.1 2 3

• Recursion follows a similar logic. It says public biology data is not reliable enough for training and has built proprietary standardized datasets in automated wet labs, running millions of experiments per week. That makes purpose built data a sector level requirement, not just insitro’s opinion.2
• The product implication is concrete. insitro is not selling a general AI assistant to scientists. It runs a closed loop where genetic and clinical data suggest a target, lab systems generate matching cell data, and models help decide which gene or molecule moves to the next experiment or into a partner program.1
• This is also why workflow systems like Benchling matter. Their value is turning experiments into structured, consistent records tied to inputs, outputs, and model results. Even when Benchling distributes third party models, the bottleneck is still whether the underlying experimental data was captured in a standardized way.3 4 5

Going forward, the winners in AI drug discovery are likely to look less like software companies trained on legacy records and more like data factories that design biology experiments for machine learning from day one. For insitro, the next proof point is whether that cleaner data advantage produces better targets and drug candidates that partners and regulators treat as real clinical assets.1 2