Automating Scientific Operations with AI

By Julie Penzotti, PhD

Experimentation is the engine of scientific discovery, and all scientists know that a negative experiment, even though it's not the result you hoped for, is often more informative than a positive one. Whether it's drug discovery, cell engineering, or antibody optimization, the work is fundamentally iterative: generate hypotheses, run assays, interpret messy data, adjust parameters, and try again. Failure is a natural part of this method, sharpening our understanding of where not to explore, and often revealing more promising directions by ruling out dead ends. Yet this test-and-learn cycle remains slow, manual, and constrained by both physical throughput of the lab and the cognitive load of managing complex, interdependent protocols. Each experiment requires juggling dozens of conditions and contextual quirks, making even routine tasks like optimizing media or tuning culture conditions stretch into weeks of trial and error.

Where Biology Struggles to Scale

Biology is inherently noisy and context-dependent, with nonlinear, deeply interconnected systems driving its behaviors. Cells respond to subtle changes in ways we still don't fully understand; tiny perturbations can cascade across pathways much like emergent behavior in social networks. And unlike software systems, biology doesn't scale cleanly. Moving a cell-based assay from a 96-well format to 384 wells, for example, is rarely a simple task. Factors such as surface area-to-volume ratios change, gas exchange rates are different, edge effects may become more pronounced, and mixing, shear stress, or dispensing tolerances can break protocols entirely. Liquid-handling robots help, but their protocols are often brittle: slight variations in viscosity, cell density, or reagent quality can throw off a workflow. Metadata is frequently incomplete, environmental conditions drift, and instrument interoperability is still far from seamless. Taken together, these challenges have kept experimentation cycles slow, expensive, and difficult to reliably automate.

As someone who started in drug discovery and later worked in tech, I've seen firsthand how these constraints cap the speed of iteration for biology. And even though we have powerful technologies that push us closer toward higher throughput, like combinatorial chemistry , high-throughput screening , microfluidics, next-generation sequencing , and automated clone selection , we are still far from the continuous, rapid experimentation cycles that tech takes for granted. Wet-lab experimentation remains a critical bottleneck.

By contrast, experimentation in the tech world has evolved dramatically. The test-and-learn cycles I helped build in marketing and e-commerce operate at a velocity that biology can only dream of. Digital platforms, from e-commerce to consumer apps, run thousands of experiments simultaneously, enabled by an enormous volume of data and high-frequency interactions: millions of users, billions of clicks, and instantaneous feedback loops. This scale enables sequential tests, multi-armed bandits , CUPED-based variance reduction , and Bayesian optimization to operate continuously behind the scenes to optimize for experiment efficiency. Human behavior is still noisy, sometimes irrational, and influenced by network effects, but with millions of interactions per day, the sheer quantity of data overwhelms the noise and uncertainty. Biology, in contrast, lacks both the high-frequency loop and the experimental volume needed to "average out" its inherent variability.

How AI Can Transform Scientific Operations

AI gives us a path to make biological experimentation faster, more efficient, and more iterative by reshaping the operations of science. Unlike traditional experiment development, AI systems can learn from patterns across hundreds of assay development protocols — including those buried in lab notebooks, supplementary materials, internal documents, and decades of scientific literature. An AI-driven workflow can synthesize this knowledge, recognize contextual cues, and propose protocol variations at a scale that's extremely difficult for any human scientist to manage alone.

When paired with lab automation, AI becomes a scaffolding layer: ingesting literature, extracting key parameters, generating protocols, designing experiments, and translating plans into executable robotic scripts. For the first time, we can realistically imagine an end-to-end scientific operations pipeline that is both data-driven and adaptive.

AI in this context is a collaborator, not a replacement—handling the cognitive load of planning and orchestration so scientists can focus on interpretation and research strategy. This is where Design of Experiments (DoE) becomes essential. DoE provides a statistical framework for systematically exploring complex spaces like media formulation, cell-culture conditions, or multi-step assay workflows. And because full factorial designs quickly explode in biology, modern practice increasingly relies on computational DoE, Bayesian optimization, and adaptive sampling to identify the most informative experiments.

Take media optimization as an example: maximizing cell growth requires tuning dozens of variables: amino acids, metals, glucose, vitamins, and even mixing parameters that affect gas exchange. The combinatorics of these interacting factors, combined with the need to learn from each round to choose the next, make this problem ideal for AI-driven exploration. DoE becomes the bridge between human reasoning and AI, giving scientists structure and giving AI direction. When AI leverages broad protocol knowledge and scientific literature to propose DoE strategies, it transforms experiment planning from manual guesswork into a guided search grounded in accumulated scientific insight. The result is a more efficient, information-rich process, one that helps scientists extract maximum knowledge from every experiment.

Closing the Loop with Automation

Of course, establishing an effective human-AI collaboration for DoE is only the first step. The real challenge is connecting that design intelligence to the physical lab. Right now, that connection is fragile. Designs often live in Jupyter notebooks while the robots—intended to reduce human error and speed up experimentation—operate in isolation. Metadata goes missing, conditional logic is hard-coded, and real-time feedback is rare. To make experimentation truly AI-driven, we need closed-loop orchestration: systems where AI designs, robots execute, and data automatically flows back for model updates.

By embedding DoE directly into the AI workflow, we begin to unlock a true closed loop: literature → protocol → experimental design → robotic execution → data ingestion → model update. The more robust this loop becomes, the more biological experimentation starts to resemble the continuous optimization cycles seen in digital technology. In this framework, lab automation becomes the physical analog of high-frequency user interactions, enabling more rapid iteration, controlled comparisons, and reproducible execution at greater scale. As each executed experiment feeds data back into the model, uncertainty estimation, reproducibility checks, and calibration become essential safeguards, ensuring the system knows when a result is reliable, when a model is confident, and when additional experiments are necessary to reduce ambiguity. As the robots run more experiments, the AI model learns not just which conditions work, but how to design better experiments and refine hypotheses.

The Future of Experimentation

This vision is the first real step toward meaningful AI—scientist collaboration. By automating more of the experimental workflow and giving AI systems the ability to propose, run, and learn from experiments, we begin to generate the kind of data volume needed to unlock more sophisticated strategies—similar to the experimentation frameworks that transformed the tech world. With stronger DoE-driven exploration, clearer measures of scientific value, and better-structured experimental data, AI collaborators can start to reason about biology with increasing depth and reliability. Lab automation provides the speed; AI provides the judgment. And as Edison noted, success is measured by "the number of experiments that can be crowded into 24 hours."

As we move forward toward this new way of doing science, we need to remember that scientific progress has never been about speed alone— it's about curiosity guided by rigor. The goal isn't to replace scientists but to build systems that think experimentally and learn scientifically, amplifying our capacity to discover. In the long run, autonomous research ecosystems may emerge, networks of AI scientists and robotic collaborators continuously running, analyzing, and refining experiments. Human researchers will frame the questions and interpret meaning; the AI will handle the iteration and exploration. If success is measured by the number of experiments in 24 hours, the next era will measure success by how well our AI collaborators learn from each one.

Ready to experience the future of science?