BRENNEN

Industries / Health / Pharmaceuticals / Biotech

Growing cells at scale

This work grows microbes or animal cells in large steel or plastic tanks, called bioreactors, so that they make a medicine or material that is then purified.

7

Milestones achieved, from 1944 to 2023.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1944

    Deep tanks for penicillin

    Penicillin mould had been grown on the surface of shallow trays, which gave low yields. Pfizer grew it throughout deep, stirred tanks instead, in a converted Brooklyn ice plant with fourteen tanks of 7,500 US gallons (about 28,400 litres) each. The American Chemical Society (ACS) says the plant opened on 1 March 1944; The Chemical Engineer magazine says 1 March 1943. By the ACS's account, most of the penicillin that went ashore on D-Day came from this plant.

    Lessona process that works on a lab bench has to be redesigned before it works in a factory.

    eurekalert.orgthechemicalengineer.com

    Credit

    Pfizer, with chemist Jasper Kane and engineer John McKeen

  2. 1986

    Hamster cells as drug factories

    Genentech's clot-dissolving drug tPA (tissue plasminogen activator, sold as Activase) was made in Chinese hamster ovary (CHO) cells, a lab-grown animal cell line that can fold and finish complex human proteins. A 2013 review calls tPA "the first recombinant therapeutic protein produced from mammalian cells", and a second 2013 review dates its approval to 1986. That second review says the protein yield per litre of culture, called the titre, has risen about 100-fold since the 1980s to grams per litre; a third review gives 2 to 6 g/L for antibodies made in CHO cells. Florian Wurm of EPFL credits the gain mainly to better nutrient mixes and process control.

    pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govebi.ac.uk

    Credit

    Genentech for tPA; reviews by Wurm, by Lai, Yang and Ng, and by Bandaranayake and Almo

  3. 2006

    Copies of cell-made drugs

    A biosimilar is a close copy of a biologic medicine, made by another company in its own cells. It is not identical to the original, so regulators judge it on comparable quality, safety and effect. Times Higher Education reported Sandoz's growth hormone Omnitrope as the first biosimilar authorised in the EU, after a positive EMA opinion in January 2006. The US followed on 6 March 2015 with Sandoz's Zarxio, a copy of Amgen's Neupogen, reported as the first US biosimilar.

    timeshighereducation.compmc.ncbi.nlm.nih.govhealio.com

    Credit

    Sandoz, EMA, FDA

  4. 2007

    Corn sugar replaces oil

    DuPont Tate & Lyle Bio Products opened a USD 100 million plant in Loudon, Tennessee, that uses fermentation to turn corn sugar into 1,3-propanediol (Bio-PDO), a building block for plastics and other products, at 100 million pounds a year. The companies said it uses 40% less energy than making the chemical from petroleum, and a trade report called it the first facility in the world to make the product this way.

    reliableplant.comrenewable-carbon.eu

    Credit

    DuPont and Tate & Lyle; the ACS gave its 2007 Heroes of Chemistry award to the team, including Charlie Nakamura, Cathy Babowitz, Dennis Adkesson and Gregg Whited

  5. 2009

    A virus stops a plant

    In June 2009 Genzyme halted production of Cerezyme and Fabrazyme, enzymes for rare inherited diseases, at its Allston Landing plant in Massachusetts after a contamination with Vesivirus 2117, a virus not known to cause disease in people. The EMA issued temporary advice to give some patients lower or less frequent doses while supply was short. A 2020 study by MIT's Consortium on Adventitious Agent Contamination in Biomanufacturing collected 18 viral contamination events since 1985 from 20 companies; 12 involved CHO cells, and the most common test takes at least two weeks.

    Lessonone infected production line can cut off the world's supply of a drug made in one place.

    ema.europa.eunews.mit.edu

    Credit

    EMA (shortage advice); Paul Barone and colleagues at MIT (the study)

  6. 2021

    A bio-made film fails

    Zymergen had launched Hyaline, an optical film for electronics made with a biomolecule, in 2020. On 3 August 2021 it said several prospective buyers had trouble fitting the film into their own manufacturing, and that it expected no product revenue in 2021. CNBC reported that the shares fell 68% after hours, and co-founder and chief executive Josh Hoffman stepped down.

    Lessona material that works in the lab still has to fit the customer's production line.

    cnbc.com

    Credit

    reported by CNBC

  7. 2023

    Shared rules for continuous production

    Most cell-grown medicines are made in batches: fill a tank, grow, harvest, clean, repeat. In perfusion culture, fresh nutrients flow in and product flows out for weeks. Sanofi opened a plant in Framingham, Massachusetts, in October 2019 to run production continuously. In March 2023 the FDA issued ICH Q13, the international guideline on continuous manufacturing, which covers therapeutic proteins and works through an example linking a perfusion bioreactor to continuous purification.

    fda.govbiopharmadive.com

    Credit

    the International Council for Harmonisation (ICH) and the FDA; Sanofi for the plant

Watch for

  1. Continuous antibody production at scale

    Enzene says its fourth monoclonal antibody made with its "fully-connected continuous manufacturing" process has reached commercial supply: pertuzumab, a breast cancer drug, made in Pune, India, and approved by India's Drug Controller General. Enzene calls the process the industry's first of its kind.

    Achieved whenindependent figures for yield and cost per gram are published, or a second regulator such as the FDA or EMA approves a product made this way.

    cphi-online.com

    Claimed by

    Enzene, September 2026

  2. Software steering bioreactors

    Two Pfizer scientists describe an in-house control layer on top of existing bioreactor controllers. It runs Python models that combine machine learning with physical models to set nutrient feeding, and it is connected to more than 150 bench-scale bioreactors. They say it improved titre and consistency but give no numbers, and that it is not yet used in regulated production.

    Achieved whenthe system runs in regulated commercial production and comparison data are published or reviewed by a regulator.

    bioprocessonline.com

    Claimed by

    Shu Yang and Edwin Shen of Pfizer, August 2025

Open

  1. 01

    A shared test for predicting titre

    Gather public data from real cell-culture bioreactor runs into one format, fix the training and test splits, and score models that predict the end-of-run titre from the first days of a run. Publish simple baseline models and a public leaderboard.

    The brief
    Credit · unclaimed
  2. 02

    A public record of contamination

    A searchable, tagged table of US regulatory actions that blame microbial or viral contamination for a problem in a cell-grown medicine, drawn from recalls, warning letters and inspection findings. Each row is tagged with the product, cell type, contaminant, date and source link.

    The brief
    Credit · unclaimed

Open 01 · the brief

A shared test for predicting titre

Credit · unclaimed

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

What to build

Gather public data from real cell-culture bioreactor runs into one format, fix the training and test splits, and score models that predict the end-of-run titre from the first days of a run. Publish simple baseline models and a public leaderboard.

Why it is still open

A 2023 review says comparing machine-learning methods on bioprocess data is hard "due to the lack of publicly available benchmark datasets", because companies treat the data as trade secrets. A December 2025 comparison used three datasets but does not say they are public. The nearest open benchmark, from the Alberta Machine Intelligence Institute (Amii) in 2026, scores methods for choosing cell-culture nutrient recipes; it does not forecast titre during a run.

arxiv.orgarxiv.orggithub.com

Built with what exists

Python, pandas and scikit-learn; Zenodo to archive data; GitHub for code and the leaderboard; Amii's benchmark code (MIT licence) as a pattern for the format.

  1. First step

    List every openly licensed cell-culture dataset that has a titre column, with its licence and size.

  2. Done looks like

    At least three public real-run datasets in one schema, a fixed split, two baseline models, and a reported mean absolute error in g/L for predicting final titre from day-5 data, all rerunnable by a stranger with one command.

Open 02 · the brief

A public record of contamination

Credit · unclaimed

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

What to build

A searchable, tagged table of US regulatory actions that blame microbial or viral contamination for a problem in a cell-grown medicine, drawn from recalls, warning letters and inspection findings. Each row is tagged with the product, cell type, contaminant, date and source link.

Why it is still open

The largest study of viral contamination, MIT's 18 events, got its data only "on the condition that the data would be released anonymously", and most events had not been reported publicly. The Microbiology Forum compiles FDA actions on microbiological issues, but as periodic PDF lists across all product types, and says it is "not guaranteed to be complete".

news.mit.edumicrobiologyforum.org

Built with what exists

The openFDA drug enforcement API, which holds recall records from 2004 onward and is updated weekly; FDA warning-letter pages; Python; a language model such as Claude to sort the free-text reason fields, with a person checking a sample. The note "Seven more flagged" (Claude at Enveda) covers Claude reviewing drug-making records.

open.fda.gov

  1. First step

    Pull all drug enforcement records from openFDA and count how many name contamination, sterility, mycoplasma or a virus.

  2. Done looks like

    A public table covering 2004 to 2026 with a source link on every row, published tagging rules, and a hand check of 100 random rows finding at least 95% correct contamination tags.