Contract manufacturing (CDMO)
Making the small scale predict the large one
A microbial contract manufacturer wants its run histories at two scales to say, before a campaign, how a transferred process will behave at full scale.
The challenge
Every transfer relearns the same lesson: what the development scale showed does not fully carry to the production scale, and the gap is found by running the campaign. Oxygen transfer, mixing gradients and narrow feeding and induction windows are exactly where scales stop behaving alike.
Recombinant proteins, plasmid DNA, antibody fragments and vaccines share the trains, with few repeats per process, so experience accumulates slowly and sits with individuals. A scale-up surprise is the client’s delay, and it lands on the manufacturer.
The client
A European cGMP contract manufacturer of microbial biologics (E. coli and yeast), running a development-scale and a production-scale fermentation train.
Our approach
We build a golden-batch model per train from historical runs, then compare the two. The twin gives:
- a live warning when a campaign leaves the pattern of successful runs
- the deviating variable named, not hunted
- a quantified statement of what systematically shifts when a process moves from the smaller to the larger train
Estimated effects
54% vs 12%
deviation score of a bad batch versus a good one
Published golden-batch benchmark on 100 simulated fermentation batches, with pH and dissolved oxygen as the main contributors. Simulated data, not a plant result.
20+ years
of multivariate batch monitoring in industrial fermentation
Documented in production plants since 2001 and described as a PAT tool for biopharmaceutical cultivation.
1 model, 2 scales
a reusable scale relationship
It describes the manufacturer’s own equipment, not one client’s molecule, and improves with every new process.
We do not promise a percentage. How much of the difference between the two scales the recorded signals explain is exactly what Phase 0 measures, on the manufacturer’s own run histories.
How the PoC runs
-
01
Phase 0 · 4 weeks
Data assessment; scope and success criteria frozen.
-
02
Data buffer · 2–4 weeks
Cleaning, digitising and time-aligning the run data.
-
03
Phase 1 · 8–12 weeks
Models per train, the scale comparison, tests on held-out runs.
-
04
Next step
Using the scale model in the next client transfer.
What we asked for
-
Train description
P&IDs or drawings of both fermenter trains, with measurement points and ranges.
-
Run trends
Temperature, pH, dissolved oxygen, agitation, airflow and oxygen enrichment, feed rate, off-gas where recorded.
-
Run outcomes
Final cell density and titre or yield per run, plus in-process samples if taken.
-
Run history
Run number, scale, host organism, induction point, media lot and recorded deviations.
What you get
-
Fewer surprises
In a client’s first campaign at full scale.
-
Faster investigations
A starting hypothesis instead of a blank page.
-
A selling point
Evidence of consistency, offered to clients as part of the service.
-
Cumulative knowledge
Each new process improves the scale model.
Client data stays out: processes can be anonymised to identifiers, with molecules and sequences withheld — the model learns the equipment, not the clients’ products. The work runs on a copy of historical data, outside GxP, with no new sensors, IT integration or remote access needed to start.
Have a similar process?
Tell us about it. We’ll tell you where a model pays off first and what data you’d need.
Book a technical call