Lab automation meets computational biology: closing the loop

Reasoning models crossed a threshold. Give a modern AI agent a research question and it will search the literature, propose hypotheses, design an experiment, and analyze the data, often at the level of a good graduate student. Computational biology has never had more reasoning power per dollar.

And yet none of it counts until a cell, a protein, or a mouse agrees. A reasoning model can be confidently, eloquently wrong; a binding assay cannot. Biology’s ground truth lives in the wet lab, which means the bottleneck has moved: the question is no longer whether AI can think about experiments, but whether it can reach a bench that runs them.

We have been trying to roboticize that bench for decades. The results were real but lopsided: accurate, expensive machines that most labs underused. Every biotech veteran knows the liquid handler standing in the corner of the lab, bought for one high-throughput campaign and idle ever since. What changed is not the robots; it is the arrival of software that can keep them busy.

Laboratory automation is older than most people assume. Clinical chemistry was automated first: Leonard Skeggs’ AutoAnalyzer began running blood samples through continuous-flow chemistry in 1957. The 1980s brought robotic arms for sample handling, and the 1990s wave of high-throughput screening standardized the microplate and the pipetting robot. The 2000s integrated instruments into workcells; in 2009, the robot scientist “Adam” became the first machine to generate and experimentally confirm its own hypotheses about yeast genetics. The 2010s moved the lab into the cloud, with Emerald Cloud Lab and Transcriptic (now Strateos) selling remote experiments through an API, and in 2020 a mobile robotic chemist in Liverpool ran 688 experiments in 8 days on its own. The 2020s added the missing ingredient: AI agents capable of deciding what those robots should do next.

Frohm et al. (2008) classify automation into seven levels, from totally manual work (level 1) through flexible hand tools and automated workstations up to totally automatic operation (level 7). Cost of implementation rises with each level. On top of this scale sits an eighth level that the original framework never anticipated: the self-driving lab, distinguished from level 7 not by better robotics but by the ability to predict the next best experiment from the results of the last one.

Levels 1 to 7 after Frohm et al. (2008); level 8, the self-driving lab, adds autonomous experiment selection.

The potential is much larger than the adoption. A text-mining study by Groth and Cox (2017) analyzed 1,628 biomedical papers and found that 86 to 89 percent used at least one method that a commercial robotic lab could already run. Yet adoption of automation above level 6 remains rare: the machines exist, the protocols are automatable, and most labs still pipette by hand.

The case for automation rests on reproducibility and throughput (Holland & Davies, 2020):

• Lower variability between experiments: robots do not have good days and bad days.
• Greater scope and accuracy of data capture: every timestamp, volume, and lot number logged automatically.
• Lower contamination risk: fewer open-tube manual steps.
• Improved safety: hazardous chemistry and repetitive-strain pipetting move off human hands.

The limitations deserve equal attention before you buy anything:

• The same protocol can produce different results on different robotic workcells; automation moves variability, it does not always remove it.
• Incorrect application: automating a bad protocol gives you reproducibly bad data.
• Input-material variation still propagates, and errors propagate faster: a mislabeled reagent contaminates hundreds of samples instead of three.
• Obsolescence: workcells are built around today's methods. The original PCR machine was scrapped when thermostable Taq polymerase removed the need to add fresh enzyme every cycle (Hawker et al., 2018).
• Creativity suffers when flexibility is low: exploratory science rarely fits a fixed deck layout.

The two hardware families of lab automation and how they trade off cost, throughput, and flexibility.

Two hardware families dominate. Robotic liquid handling scales from sub-$10k pipetting assistants and open-source robots like the Opentrons OT-2 up to room-scale Hamilton and Tecan workstations. Microfluidics miniaturizes instead: digital microfluidics moves droplets on electrode grids, channel-based chips run assays in nanoliter volumes, and acoustic liquid handlers like the Echo dispense without tips ever touching the liquid. Cost, throughput, and flexibility trade off along both branches; nothing is high on all three.

In-house automation

Labman and Automata represent two philosophies. Labman builds custom robotic systems across industries (biotech, agriculture, paints and coatings, food), with deep expertise in physical and chemical handling: powders, dosing, formulation. Automata’s LINQ benches focus on the molecular biology stack: liquid handling, cell cultures, genomics, assays, and screening, with a modular design that can absorb existing automation-friendly instruments. In our own scoping exercise for a formulation-science project, Labman was flexible enough to design bespoke formulation-preparation and erosion systems, while Automata declined the formulation work but offered stronger cytotoxicity-assay expertise. The pattern generalizes: Labman adapts to unusual physical processes; Automata is the stronger fit for standard molecular biology workflows.

Cloud laboratories

Emerald Cloud Lab and Strateos are the two names in cloud laboratories. In our evaluation, ECL was responsive and could support formulation preparation and erosion assays; its catalog leans chemical, with no cell-culture capabilities at the time. Quoted pricing was roughly $40k per month for two parallel workflows, with a 5-to-8-hour queue between submission and execution that drops to about 15 minutes if you ship your own instruments to their facility (idle instruments can even be rented out to other ECL users). Strateos, by contrast, was hard to reach: repeated demo requests went unanswered. On paper Strateos is the more versatile platform, covering both cell biology and chemical synthesis, offering “control our lab” (cloud), “control your lab” (software), and “build your lab” (in-house smart-lab construction), and claiming a closed design-make-test-learn cycle across 14 modules and 200+ instruments.

Orchestration

Biosero (a BICO company) approaches the problem from the opposite direction: instead of building robots, its Green Button Go software orchestrates the instruments, workcells, and people a lab already has. Device adapters connect instruments from different vendors, scheduling software sequences manual and automated steps in one workflow, and mobile robots ferry plates between workcells. Orchestration is the pragmatic middle path for labs that are partially automated, which is nearly all of them.

Hit-to-lead platforms

The hit-to-lead platforms show what full integration looks like when you control the whole stack. Recursion runs phenomic screening at industrial scale, imaging millions of cellular experiments per week to build “phenomaps” of biology, with Roche, Bayer, and Nvidia as partners. LabGenius uses machine-learning-guided robotics to evolve antibody engagers for cancer, a genuinely closed loop from design to assay. Arctoris pairs a cloud lab with drug-discovery consulting. The catch is the same for all three: the loop is closed inside their pipeline, for their modality. You cannot rent Recursion’s loop for your own research question.

The incumbents above were designed in the pre-AI era. A new generation treats the reasoning layer as native. Ten worth watching (funding verified against public announcements, August 2026):

Three patterns separate this cohort from the incumbents. First, natural language is becoming the control interface: you describe the experiment, the system programs itself. Second, teleoperation and foundation models are new routes around the old “robots only do what you script” limit. Third, several of these companies sell validated data rather than machines (Adaptyv is the clearest example), which turns lab automation from a capital purchase into an API call.

The four spaces above solve execution. What none of them solves is the thinking layer: deciding what is worth running, designing the experiment, analyzing what comes back, and deciding what to run next. Hit-to-lead platforms close that loop, but only for their own pipeline, in their own modality. For everyone else, the design-make-test-learn cycle still has a human bottleneck exactly where reasoning models are now strongest.

That is how Hydra fits this landscape. Hydra runs PhD-level bioinformatics: give it a research direction and it plans the analysis, executes it across ~100 scientific databases and 200+ codified skills, and validates every result. Through integration skills it can hand designed experiments to your chosen automation provider and pull the data back for analysis. Among the skills already available is an Adaptyv Bio skill: Hydra sends protein designs to Adaptyv’s cloud foundry for expression and binding measurements, then returns the assay data into the next design round. Your lab’s loop, closed: reasoning from Hydra, validation from the bench. See worked examples in our case studies or try Hydra on your own research question.