BRENNEN

Industries / Health / Pharmaceuticals / Biotech

Lab automation

This work covers the machines and software that run biology experiments: plates with many small wells, robots that move liquids, and AI programs that plan experiments and read the results.

7

Milestones achieved, from 1951 to 2023.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1951

    The many-well plate

    During Hungary's 1951 flu epidemic, supplies were short. Gyula Takátsy machined rows of small wells into a plastic block so that many tests could run at once on tiny volumes, moving the liquid with wire spirals. He settled on 8 rows of 12 wells (96), the layout labs still use.

    Lessona shortage drove the shrinking, and the shrinking later made robots practical.

    hackaday.comen.wikipedia.org

    Credit

    Gyula Takátsy

  2. 2004

    One shape for every plate

    Standards fix a microplate's footprint, height, bottom edge and well positions, so that any maker's plate fits any maker's robot. Four were published in 2004 and a fifth in 2012, drawn up by plate makers, instrument makers and users and published through the American National Standards Institute (ANSI).

    slas.org

    Credit

    Society for Biomolecular Screening (now SLAS), with ANSI

  3. 2009

    A robot forms hypotheses

    The robot scientist Adam made its own hypotheses about the genes of baker's yeast and tested them with lab automation, and people then confirmed its conclusions by hand. The project logged more than 6.6 million growth measurements in a formal, machine-readable language.

    ebi.ac.uk

    Credit

    Ross King, Stephen Oliver, Amanda Clare and colleagues, in Science

  4. 2012

    Landmark results did not repeat

    Between 2002 and 2012, scientists at Amgen tried to repeat 53 landmark cancer papers and, by Glenn Begley's account, could not reproduce 47 of them. Begley and Lee Ellis reported this in Nature.

    Lessonan experiment that cannot be repeated cannot be automated; methods need full, exact written steps.

    obamawhitehouse.archives.gov

    Credit

    C. Glenn Begley (Amgen) and Lee M. Ellis

  5. 2014

    An open-source pipetting robot

    OpenTrons launched the OT.One on Kickstarter on 31 October 2014, starting at USD 2,000, when liquid-handling robots cost tens of thousands of dollars. The makers published the code, design files and parts list.

    synbiobeta.com

    Credit

    Will Canine, Chiu Chau and Nick Wagner (OpenTrons)

  6. 2021

    An outside check of cancer biology

    The Reproducibility Project: Cancer Biology repeated 50 experiments from 23 papers. For positive effects, the median repeated effect was 85% smaller than the original, and 46% of effects (51 of 112) passed most of five tests of success.

    Lessonan independent, public study bore out the Amgen finding.

    ebi.ac.uk

    Credit

    Timothy Errington, Brian Nosek and colleagues (Center for Open Science and Science Exchange), in eLife

  7. 2023

    A language model runs experiments

    Coscientist, a system built on the large language model GPT-4, searched the internet and equipment manuals, wrote and ran code, and drove lab automation to plan and carry out chemistry experiments, including tuning palladium-catalysed reactions. The work was peer-reviewed and published in Nature.

    ebi.ac.uk

    Credit

    Daniil Boiko, Robert MacKnight, Ben Kline and Gabe Gomes (Carnegie Mellon University)

Watch for

  1. An AI model runs a cloud lab

    A cloud lab is a robot lab that customers run remotely through software. OpenAI and Ginkgo Bioworks say GPT-5 designed six rounds of experiments in Ginkgo's Boston cloud lab over six months, testing 36,000 recipes for making a protein without living cells on more than 580 plates of 384 wells, and that the cost fell to USD 422 per gram by their own numbers. People still prepared reagents and loaded plates, and the result is a preprint covering one protein.

    Achieved whenthe work passes peer review or another lab repeats the cost gain, ideally on more proteins.

    rdworldonline.com

    Claimed by

    OpenAI and Ginkgo Bioworks, February 2026

  2. AI agents pick a drug candidate

    FutureHouse's Robin, a set of AI agents that read papers and analyse data, proposed ripasudil, an existing glaucoma drug, as a candidate for dry age-related macular degeneration, an eye disease, and proposed and analysed a follow-up test. The authors say Robin produced every hypothesis, plan and analysis in the main text.

    Achieved whenthe paper passes peer review and an outside lab repeats the cell result.

    arxiv.org

    Claimed by

    Ali Essam Ghareeb, Andrew White, Samuel Rodriques and colleagues at FutureHouse, May 2025

  3. "AI Science Factories"

    Lila Sciences says its first factory, which combines AI with custom lab hardware, has run "hundreds of thousands of AI-driven experiments" and produced "thousands of discoveries" in life science, chemistry and materials, by its own count.

    Achieved whenspecific discoveries appear in peer-reviewed papers or are confirmed by outside groups.

    lila.ai

    Claimed by

    Lila Sciences, September 2025

Open

  1. 01

    Protocols to robot scripts, scored

    A public test set. Each item pairs a published protocol, limited to liquid-handling steps such as dilutions and plate-to-plate transfers with water or buffer, with a hand-checked Python script for the Opentrons robot. A scorer runs any AI-written script in the Opentrons simulator and compares its steps with the reference. Nothing runs on a real robot.

    The brief
    Credit · unclaimed
  2. 02

    Lab records into a standard format

    A converter from human-made lab records, such as plate-layout spreadsheets, electronic lab notebook exports and supplementary tables from papers, into the Allotrope Simple Model (ASM), a shared JSON format for lab data, with a hand-labelled test set so others can score their own converters.

    The brief
    Credit · unclaimed
  3. 03

    Link protocols to repeat attempts

    An open table that links each original experiment to its written protocol (the paper's methods or a protocols.io DOI) and to every known repeat attempt and its outcome, starting with the 50 experiments from 23 papers in the Reproducibility Project: Cancer Biology.

    The brief
    Credit · unclaimed

Open 01 · the brief

Protocols to robot scripts, scored

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set. Each item pairs a published protocol, limited to liquid-handling steps such as dilutions and plate-to-plate transfers with water or buffer, with a hand-checked Python script for the Opentrons robot. A scorer runs any AI-written script in the Opentrons simulator and compares its steps with the reference. Nothing runs on a real robot.

Why it is still open

BioProBench (22,413 protocols) tests reasoning about protocols, not robot code. BioPlanner turns protocols into plain step lists, not runnable code. AEGIS checks robot code against 24 protocols, and ABC-Bench includes writing liquid-handling robot code as a biosecurity test, but neither page says its protocol set is released. No public protocols.io-to-Opentrons test set was found.

arxiv.orgarxiv.orgarxiv.orgarxiv.org

Built with what exists

The Opentrons Python Protocol API and its simulator; Opentrons' open-source protocol-evaluation service; public protocols from protocols.io; any code-writing language model. Models that write code and repair it from simulator error messages are what make this practical.

docs.opentrons.comgithub.com

  1. First step

    Pick 20 dilution or plate-transfer protocols, write and simulate reference scripts, and publish them with the scorer.

  2. Done looks like

    At least 100 protocol-script pairs released under an open licence, every reference script passing the simulator, and a scorer that reports, per model, the share of scripts that pass and the share of steps that match the reference.

Open 02 · the brief

Lab records into a standard format

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A converter from human-made lab records, such as plate-layout spreadsheets, electronic lab notebook exports and supplementary tables from papers, into the Allotrope Simple Model (ASM), a shared JSON format for lab data, with a hand-labelled test set so others can score their own converters.

Why it is still open

The main open-source ASM converter, Benchling's allotropy (MIT licence), reads output files from instrument software; its page does not mention notebooks or free-form records.

github.com

Built with what exists

allotropy and the published ASM schemas; Python's jsonschema for checking; a language model to pull fields out of messy tables. The SiLA 2 device-control standard can supply instrument names and terms.

sila-standard.com

  1. First step

    Collect 25 plate-reader spreadsheets from open-access paper supplements and hand-write their ASM versions.

  2. Done looks like

    100 public records hand-converted to ASM files that pass schema checks, and a scorer that reports field-level precision and recall (the share of extracted fields that are right, and the share of true fields found) for any converter.

Open 03 · the brief

Link protocols to repeat attempts

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open table that links each original experiment to its written protocol (the paper's methods or a protocols.io DOI) and to every known repeat attempt and its outcome, starting with the 50 experiments from 23 papers in the Reproducibility Project: Cancer Biology.

Why it is still open

The FORRT FLoRA library links repeat attempts to original papers and records outcomes, but its page does not mention links to protocols or methods.

forrt.org

Built with what exists

The FLoRA dataset on OSF (CC BY 4.0 licence); the Europe PMC, Crossref and protocols.io public APIs.

  1. First step

    Pull the 23 papers' records from Europe PMC and match each to its eLife replication paper.

  2. Done looks like

    All 23 papers and 50 experiments entered with original, protocol and replication identifiers, and a second person's check of 10 random rows agreeing on all 10.