Build on every culture run.

Upstream development is iterative: each culture run is an experiment, and the next should start from what it showed. The result depends on the cell source and the inoculum’s history, the media and feeds, the vessel and its control strategy, the additions actually made, the process signals, the samples and assays, and the state of the culture at harvest.

A page-based ELN is built to record what was done. It can hold the recipe name, a pasted feed table and an overlay of two runs, but the material lots, actual additions, control history and sample results behind the overlay are not records it can follow. When the process moves to a larger scale or another site, someone reconstructs the recipe and its rationale from those files, and the reason a range was chosen is the part most often lost. Development guidance expects that reasoning to survive: each study cited in support of the commercial process should be summarised with its purpose, the data collected, how it was analysed, the conclusions reached and its impact on the process.¹

Seal is where the culture protocol runs, or it reads the run from the systems that produced it. A difference between runs can then be examined against what actually happened rather than what the recipe intended, and a successful run becomes reusable process knowledge rather than a graph in a presentation and a recipe reconstructed for transfer.

A protocol in Seal carries instructions, fields, calculations, linked samples and review steps. During execution the team records actual values, additions, observations, variations and instrument references as the study runs, not only in the final comparison. The fictional USP-⁠029 ran under recipe UP-⁠08 v03 with a split feed; its harvest is recorded and its review is still pending.

A recipe expresses intent and permitted variation: phases, setpoints and ranges, transitions, control modes, additions, sampling, interventions, alarms, calculations, harvest criteria and optional branches. Development flexibility is then explicit and attributable; a change made during a run is recorded against it rather than hidden in free text.

Neil can help prepare the configuration and analyse the results. The team tests the workflow and defines the review and approval requirements before its intended use. A recipe that works can become the controlled process for scale-up and transfer with its development evidence attached, and a later revision of the protocol does not rewrite the runs made under an earlier one.

The development programme defines the biological objective: product and modality, host or cell system, target quality profile, development stage, intended scale and process mode, facility assumptions, timeline, study strategy, risks, decisions and owners. Each experiment then answers a planned question rather than becoming an isolated run number.

Bring the selected cell line or strain into the study with its bank identity, characterisation, passage limits and known sensitivities. Host, parental line, construct, clone, passage, generation, bank, vial, storage, thaw, expansion, restrictions and use history stay connected, so a production culture can be traced to the exact source and manipulations that created it. Cell banks are the starting point for manufacturing a biotechnological drug substance, with guidance of their own in ICH Q5A, Q5B and Q5D.¹ Candidate screening, clonality and selection are covered in cell line development; the evidence stays connected without confusing clone selection with process optimisation.

Media and feed formulations are versioned: components, manufacturers, grades and lots, concentrations, preparation sequence, pH, osmolality, filtration, storage, stability and intended use. A feed strategy adds the rate, the bolus logic and the timing. A formulation change creates a new effective version and identifies every run that used the prior state, and media and feed preparation can run in Seal as well, so each run links to the prepared lots it consumed.

Development guidance asks the programme to identify which material attributes and process parameters should be controlled.¹ That question is answered from runs that kept their materials. Here USP-⁠028 and USP-⁠029 differ only in recipe version and feed schedule, so that is the difference the comparison can examine.

Vial thaw, vessel sequence, inoculation density, passage, media, volume, growth, viability, culture age, holds, transfer time, pooling, deviations and acceptance determine the state of the inoculum. Recorded as one lineage, they let production-reactor performance be compared against the complete inoculum history rather than only the final N-⁠1 value.

The cells’ age in culture starts at the thaw. In manufacturing, estimates of in vitro cell age start from the thaw of the master cell bank containers,² and the limit of in vitro cell age for commercial production has to be assessed.¹ A seed train that keeps its passages, timings and holds supplies the development record behind that limit, and a run with an unusual inoculum is visible as such before its results are compared with the others.

Vessel, sensors and calibration, control loops, recipe version, actual setpoints and process values, gas and liquid additions, agitation, pressure, temperature, pH, dissolved oxygen, weight, alarms, manual actions and the audit trail remain aligned by phase and time. The record distinguishes requested, delivered, measured, corrected and invalid signals, so an overlay is not mistaken for proof of comparable conditions.

Additions and interventions are causal events. Feed, base, acid, antifoam, supplement, induction, bleed, media exchange, sensor replacement, sampling, manual correction, pause and recovery each record the material lot, quantity, rate, reason, operator, time, verification and the process interval affected. In the fictional BR-⁠MAB14-⁠2L-⁠084, a two-litre confirmation run, the dissolved-oxygen sensor replacement is bounded to the interval it affected rather than discarding the run or passing unnoticed.

Sampling preserves process context. Sample identity, vessel, port, process age, phase, volume, handling, dilution and preparation, requested tests, chain of custody, method, source data, result and review remain connected, so titre, metabolites, cell density, viability, product quality, impurities and microbiology are read against the conditions at collection. Bioreactor and historian files keep their source versions beside the run, as scientific data management describes, and samples keep their methods, units and review status in the LIMS.

The fictional study compares two 2 L fed-batch runs from cell source CL-⁠04 / WCB-⁠02, in basal medium M-⁠06 v01 with feed formulation F-⁠03 v02. USP-⁠028 used a daily feed under recipe UP-⁠08 v02; USP-⁠029 used a split feed under v03. Each point is one reported observation, with no replicates. Viable cell density and viability use CELL-⁠02 v03, glucose MET-⁠04 v02 and harvest titre TTR-⁠02 v03.

USP-⁠029 keeps more viable cells after day 10: 16 against 12 × 10⁶ cells/mL at day 14, and 89% viability against 78%, with a harvest titre of 4.1 g/L against 3.6. USP-⁠028’s sampled glucose falls to 0.4 g/L at day 6, while USP-⁠029 is at 1.8 g/L at the same sample time. That is a pattern worth testing, not evidence that split feeding is better. Glucose was sampled every two days, so the samples do not show the minimum between measurements, and no complete product-quality panel is supplied, so titre and quality still need review before a harvest strategy is chosen.

Run comparisons use common context: recipe, scale, vessel, cell source, media, feed, material lots, seed state, actual phase timing, parameter summaries, interventions, samples, assays, harvest criteria and outcomes form the comparison set. Differences in that context stay visible before any statistical or model output is interpreted. Compare actual additions and control histories before treating two runs as a matched experiment; an overlay alone does not establish that scales or conditions are equivalent.

Bring Ambr or bench-scale exports, historian windows, sample results and the run records together, then ask Neil for a run comparison, a reconciliation of the recipe against actual execution, the open analytical questions or a draft follow-up study. Its output retains source versions, calculations, missing evidence and open decisions, so a reviewer starts from what was measured rather than from a summary of it.

For USP-⁠028 and USP-⁠029 the useful proposal is a study, not an answer. Reconcile the delivered feed, seed history, pH, dissolved oxygen and temperature records for both runs. Design a replicated feed comparison with matched cell source, media, sampling and harvest criteria. Assess product quality and harvest suitability alongside titre and viable cell density. The upstream scientist decides the design, and the process lead decides which process to advance.

The draft does not approve a feed strategy, establish a causal effect or prove scale equivalence. Everything in the middle column is preparation; the study design and the process decision need scientific review.

In line with the EU’s draft GMP Annex 22 on AI, Neil is not used in GMP execution. It helps set up configuration, which your team verifies and releases under change control. Customer data is not used to train AI models, and the requester’s permissions govern what Neil can read. See how Neil fits Annex 22 and more about Neil.

Cell-retention device, exchange and bleed rates, viable cell density, cell-specific perfusion rate, steady-state criteria, filter or membrane state, harvest stream, product residence, interruptions and recovery remain on the timeline. One long perfusion run can contain distinct operating intervals without becoming one undifferentiated batch record. The fictional PERF-⁠INT-⁠084-⁠D08-⁠D14 is the qualified operating interval of BR-⁠MAB14-⁠2L-⁠084 after its recovery from a filter-pressure disturbance.

Calculation bases follow the mode. A perfusion harvest interval is not interchangeable with a final fed-batch harvest, and a cell-specific rate needs the cell density and exchange rate of the same interval. ICH’s guideline on continuous manufacturing describes transient events, planned such as start-up, shutdown and pause or unplanned such as disturbances, through time, process parameter and quality attribute values. It notes that a limit of in vitro cell age established for a bioreactor run in batch mode may not apply to one operating continuously under different culture conditions, and that the cell bank vials used should be traceable to the drug substance batches they produce.³

Geometry, working volume, mixing, power input, tip speed, oxygen transfer, carbon-dioxide removal, heat transfer, gas strategy, pressure, probe dynamics, control capability, raw-material effects, the scale criterion, model version, uncertainty and verification define the translation. The model states what is held constant, what changes and where prior runs support the prediction. The fictional SCALE-⁠MAB14-v07 is an approved 2 L to 2,000 L model whose carbon-dioxide-removal assumptions were revised in that version.

Matching setpoints do not by themselves establish equivalent performance, and development guidance sets the same expectation. Small-scale models should account for scale effects and be representative of the proposed commercial process.¹ A design space developed at small or pilot scale needs its relevance to production scale justified, and one meant to span scales is better described in scale-independent terms, such as shear rate rather than agitation rate.⁴ Data from commercial-scale batches should confirm the results of the small-scale studies used to support process validation.¹

Ask Neil to compare the receiving vessel’s geometry, gassing, mixing and control with the model and to list the evidence needed to resolve each difference. The model’s assumptions, predictions and uncertainty are versioned, so a change to the model is reviewed like a change to the recipe.

Process age, viability, productivity, quality trend, impurity burden, contamination state, volume, downstream readiness, hold capacity, sample results, forecast, deviations and the reviewer determine whether to harvest, continue, terminate or divert. The resulting harvest pool retains the culture and interval evidence that justified it. USP-⁠029 reached 89% viability and 4.1 g/L at day 14, but its quality results are not yet reviewed, so its harvest decision stays open.

Many upstream processes split or pool harvests to make a drug substance batch, and development guidance expects the manufacturer to explain how batches are defined from them.¹ A harvest that keeps its source runs and intervals makes that definition traceable.

Downstream sees which culture, conditions and samples produced its starting material, and purification findings return to the development decision. The fictional DSP-⁠019 purification study starts from USP-⁠029’s harvest; downstream process development follows it from load to pool. Process-quality relationships become reusable knowledge the same way: a parameter, material or event’s relationship to glycosylation, charge, aggregation, potency, impurities, titre, viability or downstream behaviour keeps its direction, magnitude, confidence, scale, cell line, media and operating-region boundaries.

The selected recipe, formulations, equipment requirements, seed train, sampling, parameter ranges, control responses, expected profiles, harvest criteria, known sensitivities, exceptions, models and supporting runs make up the transfer package. The fictional RECIPE-⁠US-⁠MAB14-v18 is a pilot-ready recipe with a revised day-five feed and temperature-shift window, and its transfer to Site 02, TT-⁠US-⁠MAB14-⁠SITE02, keeps a restricted antifoam alternative with its restriction.

Process characterisation receives the true development history: the variables selected, prior ranges, failures, interactions, scale evidence, material effects, analytical methods, model versions, unknowns and residual risks feed its strategy and designed studies, as process characterisation describes.

Technology transfer moves product and process knowledge between development and manufacturing, and that knowledge forms the basis for the manufacturing process, control strategy, process validation approach and continual improvement.⁵ A transfer report written at the end summarises the knowledge; a transfer built from the development records carries it, with the runs behind each range. Later evidence returns the same way: engineering, PPQ and commercial batches, deviations, trends, changes, raw-material shifts, site differences, complaints and improvements attach to the transferred knowledge as lifecycle evidence. The tech transfer blueprint covers the receiving side.

Upstream knowledge spans biology, execution, equipment, materials, samples, analytics, modelling, characterisation and transfer. Seal keeps the lineage across them, from source vial or strain to harvest decision and downstream handoff. Connect permitted exports or scope an agreed connection, and keep vessel geometry, working volume, cell source, media and feed versions, process age, actual control profiles and sampling methods in view. Source interfaces, configuration and permissions determine the available connection, and setpoints stay separate from measured values. See all integrations.

The first implementation should follow one clone or strain from bank vial through the seed train, two media or feed versions, small-scale experiments, a perfusion or fed-batch run, connected samples and assays, one intervention, a scale model, pilot confirmation, a harvest decision, a characterisation question and transfer to manufacturing.

Include a mis-timed feed, a sensor replacement, a material-lot effect, an invalid sample, a failed scale assumption and a restricted platform relationship. The system should show what changed, what was learned and what the evidence supports transferring. Then extend the same records downstream into purification development.