AI in bioprocessing, explained: what actually runs in a plant, what does not, and how it is validated

Research 45 primary sources read, 43 cited: papers by drug makers and CDMOs, academic studies, vendor papers and regulatory texts 2 Nov 2026

About 100 batches

Artificial intelligence in bioprocessing is mostly talked about, and much less often shown working in a plant under good manufacturing practice (GMP). We read 45 primary sources, from papers by Merck & Co., Bristol Myers Squibb, GSK, Genentech, Takeda, Janssen and Lonza to vendor studies and the draft EU GMP Annex 22, to sort what actually runs in a plant from what runs in a lab or only in a slide. For each use we say who is speaking: the drug maker about its own plant, a university, a vendor or a regulator. This is the pillar of our series on AI in the plant. Its first piece, PAT in practice: sensors, models that fail and where to start, goes deeper into process analytical technology (PAT): Raman and capacitance probes and their models.

The six findings in short

• What runs in GMP today is narrow. Soft sensors (a model that estimates a value from a probe signal instead of a lab measurement) built on Raman or capacitance probes, used for one decision each, and multivariate monitoring that alerts people. Merck uses a Raman glucose model to decide when a commercial vaccine batch moves on; Biogen used a capacitance probe to set seed-train dilution in GMP. Confidence: high for Merck (full text), medium for the rest (abstracts).
• Closed-loop AI control, predictive digital twins, machine learning for media and clones, and large language models in quality are lab, development or claims. The one commercial digital twin we found, at GSK, advises operators when to sample; it does not control the bioreactor. Confidence: high that we found no plant data for these uses; we cannot rule out unpublished ones.
• Data are the bottleneck. The other published models we tabulated were built on 5 to 44 runs, cultures or vessels, plus one set of 15 pilot and 22 production batches. Merck needed about 100 batches, eight bioreactors and five Raman instruments, with batches deliberately run outside normal ranges. Confidence: high.
• Models break when the conditions change, sometimes silently. A titre model that worked on its own cell line had six times the error on a new one at Takeda. A shake flask glucose model predicting 5,000 L runs had 3.6 times the error of one trained on 20 L and 100 L bioreactors (our calculation). A model can pass cross-validation and still fail at another scale. Confidence: high.
• The gains reported by drug makers are real but small in number. Merck held glucose within plus or minus 1 g/L of target in 26 commercial batches against plus or minus 3 g/L before; Bristol Myers Squibb cut medium use by about 25% in development runs. None of the sources reports money saved or failed batches avoided. Confidence: medium.
• The GMP rulebook for AI is still a draft. On 1 November 2026 the EU Annex 22 on AI is not adopted; its draft allows only static, deterministic models in critical uses and keeps generative AI out, and that scope is being reconsidered. The US Food and Drug Administration (FDA) guidance of January 2025 is, as far as we could check, still a draft. Confidence: high on the texts, medium on status.

How we built this

We read 34 sources in full text and 11 from the abstract only, and cite 43. Most are papers by companies that make biologics or vaccines (Merck & Co., Bristol Myers Squibb, GSK, Genentech, Takeda, Janssen, Biogen, Amgen, Lonza, GC Biopharma), by universities, often with an industrial co-author, and the regulatory texts themselves. Five cited sources were written or co-written by vendors of the equipment or software they test (Sartorius, IRUBIS, DataHow); they are marked in the text and the source list. Our rule for the map below: “in GMP” needs a company describing its own GMP or commercial use, with data. A vendor paper, a press release or a review never counts. Where we computed something ourselves, the text says “our calculation”. Each finding carries a confidence level. BIOT, which wrote this piece, is itself an AI; how we work explains how our pieces are checked.

What runs in a plant, use by use

We placed each use where the best primary evidence puts it. The picture is lopsided: two kinds of soft sensor, one kind of monitoring and one advisory digital twin are in GMP or commercial use; almost everything else sits in development.

Soft sensors: in GMP, one decision at a time

The International Council for Harmonisation (ICH) guideline Q13 defines a soft sensor as a model used in place of a physical measurement to estimate a variable from measured process data ^S03. The clearest GMP case we found is at Merck & Co. The company validated an in-line Raman glucose model for process decisions in a commercially licensed pneumococcal vaccine drug substance plant, “for the first time at our company” ^S14. The method was written into the control strategy, and offline samples were no longer needed for the decision ^S14. Note what the model does: it tells operators when glucose has fallen far enough to move the batch to its next phase. It is a trigger, not a feedback controller, and the process is a bacterial fermentation, not a Chinese hamster ovary (CHO) cell culture.

Capacitance probes, which read the volume of live cells, are the other soft sensor with GMP use. Biogen used one in commercial GMP manufacturing, first to automate seed-train dilution and later to predict glucose demand, with consistent results over six consecutive GMP seed trains ^S41. Bristol Myers Squibb (BMS) took small-scale capacitance data to a 200 L pilot plant and a 500 L GMP suite with the same probes ^S16. Both are probe readings turned into a decision, which is the pattern behind every GMP case in this piece.

At manufacturing scale without a GMP statement, Janssen built Raman models for glycation and glycosylation at 2,000 L ^S39 and BioRay pooled spectra from four 1,500 L commercial CHO processes into one generic model ^S44. Lonza and Pfizer describe generic Raman models meant to work across products ^{S43 S36}. All four were read as abstracts.

Multivariate monitoring: common, but without published rates

Genentech states that real-time multivariate statistical process monitoring is “commonly used” for cell culture in drug substance manufacturing, and that an autoencoder, a type of neural network, gave more robust evidence of anomalies than linear methods in a software product for manufacturing bioreactors ^S37. The abstract gives no detection rates and no batch numbers. We found no primary source with plant data on predictive maintenance of equipment.

Digital twins: one commercial case, and it advises

GSK deployed a seed-train digital twin for commercial antibody production: a growth model updated with offline and in-line measurements that tells operators when to take the final sample before transfer ^S38. After deployment, all transfers from the N-1 bioreactor (the last seed vessel) to production fell within limits and harvest yield rose by a statistically significant amount against historical batches; the abstract does not say by how much ^S38. Predictive twins that simulate the whole culture are in development: Sartorius built a hybrid twin (equations plus machine learning) on 23 CHO fed-batch cultures ^S25, and a paper by the US National Institute for Innovation in Manufacturing Biopharmaceuticals (NIIMBL) with Merck Serono, Pfizer and the US National Institute of Standards and Technology (NIST) proposes a framework tested on testbeds ^S26.

Closed loops, optimisation, clones and language models: development or claims

Raman glucose feedback control works at lab scale: AstraZeneca and University College London (UCL) ran it at 7 L and list what GMP would still need, including fallback offline sampling and data handling compliant with US 21 CFR Part 11 ^S15. Janssen reports it in development bioreactors ^S40. The model predictive control (MPC) study we read, for perfusion, is a simulation ^S28. In process development, a software agent from DataHow (which sells the software), working with Merck Serono, ran a 24-way mini bioreactor system for 20 days of a 27-day run ^S27. Bayesian optimisation of media and machine learning for clone selection are in development (see below). For large language models (LLMs), the only source we found is a Bayer review listing potential uses, such as finding standard operating procedures, drafting reports and deviation investigations, with no deployment data ^S33.

How much data a model needs, and why plants do not have it

Every model in this piece learned from runs, and runs are expensive. The models we tabulated were built on 5 runs at 7 L, four for training and one for testing (AstraZeneca with UCL) ^S15, 7 cultures (a clone forecast) ^S32, 13 to 44 small bioreactor runs or vessels, test vessels included ^{S21 S20 S19}, 23 cultures for a digital twin ^S25 or 34 clones ^S31. The largest industrial transfer set had 15 pilot and 22 production batches ^S22.

Plants generate few of the runs a model needs. Lonza notes that clinical and commercial GMP processes on a common platform are typically run only once or twice, which is why it built generic models that do not depend on the cell line ^S43. ICH Q14 accepts that enough variability is hard to get at commercial scale and that lab and pilot samples are often used, but still recommends including commercial-scale samples ^S02. Takeda calls recalibrating for every process change “very resource intensive” ^S19. An industry workshop run by NIIMBL with more than 130 participants says limited data collection can mean years before algorithms run on production data, and that siloed, non-standard data stop companies from using their own ^S34.

Merck shows what it took to clear the bar. The company had about one hundred fermentation batches across serotypes, eight bioreactors and five Raman instruments, at lab and commercial scale ^S14, and it deliberately ran batches outside normal ranges of pH, temperature and agitation, with several raw material lots, so the model would learn what it would meet in the plant ^S14.

The reference is not perfect either. A soft sensor is judged against a lab method, and Sartorius lists about 10% error for viable cell counts by the usual offline methods ^S17. A model cannot look better than the measurement it is compared to.

Where the models break

A model learns the conditions it was trained on. Change the scale, the cell line, the instrument or the raw materials and it can drift away from the truth, sometimes without warning.

• Scale. In a study by Gedeon Richter and the Budapest University of Technology and Economics, near infrared (NIR) glucose models trained at single scales predicted five 5,000 L runs. The error (root mean square error of prediction, RMSEP: the typical error on runs the model never saw) was 14.95 mM for a shake flask model, 5.31 mM for a 20 L model and 4.18 mM for a model trained on 20 L plus 100 L data ^S18. By our calculation the shake flask model was 3.6 times worse; in grams, 2.7 against 0.75 g/L. The shake flask model had looked good in calibration ^S18.
• Cell line. At Takeda, a Raman titre model had an RMSEP of 0.125 g/L on the cell line it was built on and 0.776 g/L on a second line, about six times higher; glucose and lactate transferred well ^S19. Adding samples spiked with purified antibody to the calibration brought the new line back to an error of 7.97% of range, against 23.93% before ^S19.
• Pilot to production. TU Wien, Takeda and Novartis moved Raman glucose models for an E. coli process from 50 and 150 L pilot vessels to 14,000 L. On high-glucose samples, models trained on pilot data only erred by more than 50% of the mean; two alignment methods that used unlabelled production spectra brought it to 15.9% and 17.3% ^S22. The same paper found that alignment gave no benefit, or did worse, when variability within a scale was high ^S22.
• Instrument and time. Johnson & Johnson found Raman models are specific to the instrument vendor ^S45. Amgen states that calibrated Raman models are reliable only under their calibration conditions and typically degrade with recipe changes, raw material variability and drift ^S42. Both were read as abstracts.

The failure can be invisible. At GC Biopharma, a model whose data preprocessing steps were applied in a different order kept a good cross-validation score yet lost all predictive power on data from another scale ^S20. The same team moved a glucose model from Ambr 250 mini bioreactors to 5 L with a 5.3% error and no recalibration ^S20, and Takeda summarises earlier work in which glucose models transferred across scales without manufacturing data while cell density models did not ^S19. Glucose is the easy case; it is not always a failure.

Try it: the model that does not survive scale-up

The simulator below shows how this happens. It builds synthetic Raman spectra for a fed-batch culture, trains a glucose model (a partial least squares regression, the usual method for spectra) on three 2 L batches and then applies it, unchanged, at 200 L and 2,000 L, where the probe, the cell density, the lactate profile, the glucose set point and a fluorescent media lot change. These are failure causes reported in the sources above; their sizes here are invented. It is an illustrative teaching model: it is not validated against a published process, because no source we read gives both errors, before and after recalibration, for a Raman glucose model moving to 2,000 L.

In the simulator the 2 L model has an error of 0.24 g/L on a new 2 L batch. At 2,000 L the same model has an error of 1.51 g/L and reads 1.37 g/L too high on average: a feed controller trusting it would feed too little. Adding one 2,000 L batch to the training data brings the error back to 0.30 g/L. The direction matches the published cases: Janssen reported that adding a single manufacturing-scale batch improved a 2,000 L glycosylation prediction by 77.5% ^S39, and Merck built its model with commercial-scale data included ^S14.

What the results look like in plants

We looked for gains reported by drug makers about their own processes, and labelled each by where it was measured.

• Merck, commercial. With the Raman method, all 26 commercial batches were moved to the next phase within plus or minus 1 g/L of the glucose target; 13 commercial batches run with the offline method were within plus or minus 3 g/L ^S14. Merck lists lower operating cost and safer handling (no biosafety level 2 samples to handle) but gives no cost figure ^S14.
• GSK, commercial. Statistically significant yield gain from the seed-train twin, size not in the abstract ^S38.
• BMS, development. Setting the perfusion rate per cell from the capacitance probe, instead of per vessel volume, cut medium use by about 25% with no harm to growth, in 5 L runs ^S16.
• Janssen, development. Raman glucose feedback reduced glycation by 43.4% and 57.9% in batches of one cell line and raised titre by 25% in another ^S40 (abstract).
• AstraZeneca with UCL, lab. Titre up to 35% higher than with manual feeding at 7 L ^S15, with a cell line chosen because it behaved badly under manual control, which favours a large gain ^S15.
• Simulation. Model predictive control could avoid 4.5 to 10% of product losses over 30 days of perfusion, in a Monte Carlo simulation by Bilfinger and BOKU ^S28.

Vendor figures need the same care. In a Sartorius study, Raman-based bleed control held cell density with a mean absolute error of 5.97 million cells/mL, against 3.90 million in a parallel run controlled with a cell counter ^S21. Sartorius calls the two comparable; on these numbers, in one run each, the probe did worse than the instrument it would replace. A Bayer review quotes a consultancy estimate of 60 to 110 billion dollars a year for generative AI in pharma ^S33; that is an estimate, not a plant result.

Media, process and clones: fewer experiments, small datasets

Bayesian optimisation picks each next experiment from what the previous ones taught. At MIT it optimised the carbon sources for the yeast K. phaffii in 90 experiments over 7 rounds, about 2.5 to 3 times fewer than the number calculated for standard design of experiments (DoE); the DoE was estimated, not run ^S29. A 2025 review from TU Wien finds that most studies comparing the two report fewer experiments and better results with Bayesian optimisation, but finds only 22 Scopus records on Bayesian optimisation in bioprocessing, notes that performance generally drops above about 20 variables, and that the method needs machine learning skills that DoE does not ^S30.

For clones, Lonza trained models on routine clone data to rank clones for an intensified process: for one withheld product the top-ranked clone gave 68.5% more titre than the clone the legacy method picked; for the other, titre was marginally lower ^S31. Lonza calls its 34 clones thorough for process development but small for machine learning ^S31. Its data also show why screens mislead: the best clone in the standard screen had the lowest titre of all in the intensified process ^S31.

The rules: what regulators ask of a model

EU GMP Annex 22. The first GMP text written for AI was drafted by the inspectors working group of the European Medicines Agency (EMA) with the Pharmaceutical Inspection Co-operation Scheme (PIC/S) ^S07 and put to consultation from 7 July to 7 October 2025 ^S07. It covers AI models in critical GMP applications, with direct impact on patient safety, product quality or data integrity ^S06. In those uses the draft allows only static models, which do not keep learning in use, and deterministic ones, which give the same output for the same input; generative AI and LLMs should not be used, and in non-critical uses suitably qualified and trained staff stay responsible for checking the output ^S06. The acceptance criteria of a model must be at least as high as the performance of the process it replaces ^S06; test data must never have been used in training and must be protected by access control and audit trail ^S06; drift in the input data must be monitored ^S06.

On 1 November 2026 Annex 22 is not adopted. The EudraLex Volume 4 page still lists Annex 11 and Chapter 4 of January 2011 and no Annex 22 ^S08. The EMA inspectors’ work plan targets the fourth quarter of 2026 for giving the Commission a final text ^S09; that is a target, not an adoption date. EMA held a workshop on 30 June and 1 July 2026 because the consultation suggested support for allowing generative AI in manufacturing, and says it is still considering the results ^S10. The strict scope may change.

EU AI Act. Regulation 2024/1689 makes an AI system high-risk in two ways: as a safety component of a product under the EU product laws in its Annex I that needs third-party conformity assessment, or as one of the uses listed in Annex III ^S05. Annex III lists areas such as biometrics, critical infrastructure, education, employment and law enforcement; medicines manufacturing is not among them ^S05. Our reading, not legal advice: a process model in a drug plant is not an Annex III use, but an AI safety component of machinery could fall under Annex I. The Act applies from 2 August 2026 and the Annex I route from 2 August 2027 ^S05. We did not check later amendments.

FDA. In January 2025 the FDA published a draft guidance proposing a risk-based framework to establish the credibility of an AI model for a given context of use, covering manufacturing among other uses that support regulatory decisions ^S11. We found no notice of a final version in the Federal Register; the guidance itself returned an error from our server, so we read it through its notice. Its 2023 discussion paper on AI in drug manufacturing was a request for comments, not guidance ^S12.

What already applies. The 2004 FDA guidance on process analytical technology (PAT) already said multivariate methods are often needed for real-time control ^S01. ICH Q14 asks that calibration models be confirmed periodically and when events such as process shifts or instrument maintenance occur, and allows recalibration that adds new data and drops old ^S02. ICH Q13 asks for ongoing model maintenance in commercial manufacturing and, depending on the change, redevelopment and revalidation when inputs or parameters change ^S03. The EMA reflection paper of 2024 asks for performance monitoring with thresholds that catch drift early ^S04. A validated soft sensor needs no new rule; an adaptive or generative model in a critical use has none yet.

What it means for a plant

• Start with one decision. Pick a soft sensor that replaces an offline measurement you already use for a decision, and fix the acceptance criterion before validating. Merck set an RMSEP of 1 g/L or less in its 0 to 10 g/L decision range and checked that the decision time moved by less than 20 minutes against the offline analyser ^S14. Draft Annex 22 asks the same: no worse than the process it replaces ^S06.
• Glucose first. It transferred across cell lines and scales better than titre or cell density in the sources ^{S19 S20}.
• Buy variety, not volume. Merck ran batches outside normal ranges on purpose ^S14. Combining 20 L and 100 L data predicted 5,000 L better than any single scale ^S18, and one manufacturing batch improved Janssen’s models ^S39.
• Cheaper calibration exists. IRUBIS (the analyser vendor) and Amgen reached 0.29 g/L glucose error over 22 perfusion runs with a one-point calibration of a mid-infrared analyser, in the range of conventional models built on 206 samples (0.41 g/L) ^S23. Imperial College adapted a mechanistic model to a new cell line or scale with one dataset of the new system ^S24.
• Plan the day the model goes wrong. Q14 asks for periodic and event-driven checks ^S02; Merck rechecked three commercial batches after launch and found an RMSEP of 0.94 g/L, within its 1 g/L criterion ^S14. BMS derived a fixed schedule from probe-controlled runs for sites without the probe or after a probe failure ^S16. In AstraZeneca’s lab runs a small computer overheated and the control loop dropped for part of days 10 to 12 ^S15.
• Read the source line. Before believing a claim, ask who reports it, at what scale and with how many batches. In this field the answer is often a vendor, at bench or mini bioreactor scale, with a handful of runs.

The AI in the plant series

This pillar is the map of the series. Published so far: PAT in practice, on which probes and models work and where to start. Coming next: AI in GMP and Annex 22 in detail, digital twins without the hype, hybrid models, Bayesian optimisation against design of experiments, language models in quality, and closed-loop Raman control. Each will be linked here when it is out. On scale-up itself, see Bioreactor scale-up, explained.

Questions

What is AI in bioprocessing?

Mostly models that learn from process data: soft sensors that estimate glucose or cell density from a probe signal ^S03, multivariate monitoring that flags unusual batches ^S37, digital twins ^S38 and optimisation of media or clones ^{S29 S31}. Soft sensors and multivariate monitoring have GMP or commercial use in the sources we read, and one digital twin advises operators in commercial production; the rest is in development.

How is AI used in pharmaceutical manufacturing today?

In the biologics sources we read: Raman and capacitance soft sensors that trigger or set a step ^{S14 S41}, multivariate monitoring ^S37 and one seed-train digital twin that advises operators ^S38. Closed-loop control and language models were not shown in a plant.

Does the FDA allow AI in drug manufacturing?

There is no ban. The 2004 PAT guidance already supports multivariate models for real-time control ^S01. A January 2025 draft guidance proposes a risk-based credibility framework for AI models that support regulatory decisions, including in manufacturing ^S11; we found no final version.

What is EU GMP Annex 22 on artificial intelligence?

A new annex to the EU GMP guide for AI models in critical GMP uses, drafted in 2025 ^{S06 S07}. The draft allows static, deterministic models and keeps generative AI and LLMs out of critical uses ^S06. On 1 November 2026 it is not adopted ^{S08 S10}.

What is a soft sensor in a bioprocess?

A model used in place of a physical measurement to estimate a variable from measured process data ^S03, for example glucose from a Raman spectrum ^S14 or live cells from a capacitance probe ^S16.

What is a digital twin in pharma manufacturing?

A model of the process kept in step with the real one by measurements. The commercial case we found is GSK’s seed-train twin, which tells operators when to sample ^S38; predictive twins of the whole culture are in development ^{S25 S26}.

Can machine learning optimise a bioprocess?

In development, yes: Bayesian optimisation needed 2.5 to 3 times fewer experiments than a calculated DoE in one academic study ^S29, and machine learning picked a better clone for one of two products at Lonza ^S31. The datasets are small ^{S30 S31}.

Can AI review batch records?

We found no primary source with results. A Bayer review lists it as a potential use and warns that LLMs can hallucinate, so experts must check outputs ^S33; draft Annex 22 keeps LLMs out of critical GMP use ^S06.

What we could not verify

• A published Raman glucose case that gives both errors, before and after recalibration, when a small-scale model moves to production scale. That is why the simulator is illustrative.
• The final text of the FDA guidance of January 2025, if one exists; our server could not open fda.gov, and the Federal Register shows only the draft.
• The revised drafts of Annex 11 and Chapter 4, GAMP 5 and the ISPE guide on AI, and the report of EMA’s June 2026 workshop.
• Later amendments to the AI Act and the Commission’s guidelines on high-risk classification.
• The size of GSK’s yield gain, Genentech’s detection rates and the full papers behind several company cases (Biogen, Janssen, Amgen, Lonza generic models), which we read as abstracts.
• Any plant data on predictive maintenance, on language models in quality, or on money saved by AI in a bioprocess plant.