In-Process Sampling Strategy for Bioreactor Operations: Off-Line, At-Line, On-Line, and In-Line Methods Compared

What Is In-Process Sampling and Why Does It Matter?

In-process sampling is the planned collection and measurement of material from a running bioreactor so that operators can confirm the culture is on track and make decisions before the batch ends. A good bioreactor sampling strategy decides which analytes you measure, how often, by which method, and how much culture you are willing to spend and how much contamination risk you are willing to accept to get each data point.

Most teams inherit their sampling plan rather than design it. A process development group samples twice a day because that is what the previous project did, a manufacturing site adds tests after each deviation, and ten years later the plan holds 40 tests nobody can justify. The cost is real: every manual event is a contamination opportunity, every milliliter removed is product you will not harvest, and every result that arrives 90 minutes late is a decision made on stale data.

This article is about the strategy layer. It does not re-explain how a PAT technology such as Raman or capacitance works, and it does not review sensor hardware, which is covered in the bioreactor instrumentation guide. Here we ask a different question: given a process, a scale and a regulatory context, what should the in-process control (IPC) plan actually say?

The regulatory backbone is short. ICH Q6B expects in-process tests at critical decision-making steps. EU GMP Annex 1 (2022 revision) requires that sampling methods must not pose a contamination risk to the product or the process. The FDA process validation guidance expects documented IPC sampling plans in the Stage 2 process performance qualification (PPQ), and ICH Q8 design space work depends on sampling data to link critical process parameters (CPPs) to critical quality attributes (CQAs). None of them prescribes a frequency. The justification is yours to write, which is why strategy matters.

Four Sampling Modalities Compared: Off-Line to In-Line

Bioreactor analytics fall into four modalities: off-line, at-line, on-line and in-line. They differ in where the measurement happens, how long you wait, how much culture you consume, and how likely each event is to introduce contamination.

Off-line means a sample is withdrawn and carried to a remote laboratory. Turnaround is typically 30–60 minutes, and sample volume is 1–5 mL. It has the highest contamination risk because a human handles an open sampling port, but it supports the most comprehensive analytics: HPLC, LC-MS, capillary electrophoresis (CE-SDS), glycan mapping and bioburden testing. At-line means a sample is withdrawn manually and analyzed beside the process on a benchtop analyzer such as a BioProfile FLEX2, Cedex Bio HT or a Nova BioProfile. Turnaround drops to 5–15 minutes and volume to 0.5–2 mL, with moderate contamination risk because the sampling step is still manual.

On-line means an automated bypass loop withdraws sample, delivers it to an analyzer, and in some designs returns the unused portion to the vessel. Lag is 2–5 minutes, volume is 0.1–1 mL, and contamination risk is low because no operator touches the port. Systems include Seg-Flow, MAST and Numera. In-line (in-situ) means a probe sits inside the vessel or in a closed recirculation path and no sample is removed. Raman, NIR, capacitance, pH and dissolved oxygen (DO) probes all belong here. Volume is zero, contamination risk is added only at installation, and data are real time.

No modality wins on every axis, which is why mature processes use all four. In-line sensors carry the continuous load (pH, DO, temperature, and often capacitance or Raman). At-line or on-line analyzers carry the metabolite and cell-count trends that drive feeding decisions. Off-line laboratories carry the measurements that cannot be moved closer to the process: product quality, adventitious agent testing, and anything that must be release-grade. The review by Fung Shek and Betenbaugh (2021) reaches the same conclusion for mammalian cultures: at-line and in-line tools complement rather than replace each other.

Which Analytes Need Which Sampling Modality?

Match each analyte to the cheapest, lowest-risk modality that still delivers the result before the decision it supports. If a number only matters at harvest, off-line is fine. If it drives a feed pulse in the next two hours, it belongs at-line, on-line or in-line.

Three questions decide most placements. First, what is the decision latency, meaning how quickly must the result arrive for it to change an action? Second, is the result release-relevant, meaning it must come from a validated, compendial or otherwise regulated method? Third, can a calibrated in-line model predict it well enough? Glucose and lactate by Raman often can, which is why they migrate in-line first. Glycan distribution cannot, so it stays off-line.

Two patterns stand out in the table. The first is that the closer an analyte sits to a control decision, the closer its modality sits to the vessel. The second is that sample volume concentrates in the off-line column. Titer, amino acid and product quality methods use the most culture per result, so they are the first place to economize when volume is tight. For VCD measurement details see cell counting methods for bioprocess, and for the amino acid panel see amino acid analysis in bioprocess.

How Often Should You Sample a Bioreactor?

For a CHO fed-batch, sample metabolites, cell counts and titer once or twice per day, rising to the upper end during exponential growth and the production shift, and measure pH, DO and temperature continuously. Frequency should follow how fast the analyte changes and how costly a missed excursion would be, not habit.

Typical CHO fed-batch practice looks like this. pH, DO and temperature are in-line and logged every 1–5 seconds. VCD and viability are measured once or twice daily by manual sampling, or every 2–6 hours when automated. Glucose and lactate are measured 1–2 times per day off-line, or every 15–30 minutes with in-line Raman. Ammonia is measured 1–2 times per day, osmolality once per day, and IgG titer once per day at-line or off-line. Amino acids are measured 2–3 times per run, product quality attributes such as glycan profile and charge variants at harvest and key timepoints, and sterility and bioburden once per batch.

Frequency is not constant across the run. During inoculation and lag, little changes and samples are mostly confirmatory. During exponential growth, glucose and lactate move fast and VCD is the main input to feed-rate decisions, so sampling peaks. In production phase, titer and product quality become the focus while growth analytes can relax. At harvest, a dense burst of samples characterizes the final material.

A few rules make frequency defensible in an audit. Sample at least often enough that the fastest relevant dynamic is covered by three or more points; a lactate shift that develops over 12 hours cannot be characterized by a 24-hour interval. Increase frequency around known events such as temperature shifts, feed start and induction. Reduce it when development data show an analyte is stable. And always record why: the justification belongs in the control strategy, not in someone's memory. Fed-batch feeding logic interacts directly with sampling cadence, and the Fed-Batch Calculator helps check how much glucose can be consumed between two sampling points.

Sample Volume Budget: When Sampling Steals Your Working Volume

Cumulative sample volume should stay below about 5% of working volume, and at large scale that limit is irrelevant while at bench scale it is often the binding constraint. Every sample also removes cells and product, so volume loss changes mass balance and titer calculations if it is not tracked.

The arithmetic is simple but easy to forget. A typical manual sample is 3–5 mL once you include dead-volume purge. For a 14-day CHO fed-batch at 2,000 L, a minimal strategy (1 sample per day) removes about 14 × 5 mL = 70 mL, which is 0.004% of working volume. A standard strategy (2 per day plus extras for amino acids and QC) removes about 200 mL, or 0.01%. An intensive strategy (4–6 per day) can reach 500–800 mL, or up to 0.04%. None of these approaches the 5% guideline, and the volume penalty is negligible.

At 2 L the picture reverses. Twice-daily sampling at 5 mL removes 14 × 2 × 5 mL = 140 mL, which is 7% of working volume. Even minimal once-daily sampling takes 70 mL (3.5%). This is why bench-scale teams adopt micro-volume at-line analyzers, return-to-vessel on-line loops, and in-line sensors: they are the only way to keep data density high without consuming the experiment.

Three practical tactics protect volume. First, drop the purge: closed sampling devices with minimal dead volume avoid discarding 2–3 mL per event. Second, share a sample: take one 2 mL aliquot and split it between the VCD counter, the metabolite analyzer and a retained sample, instead of three separate draws. Third, shift analytes in-line. Metze et al. showed that capacitance biomass sensors track viable biomass across scales in single-use bioreactors, which allows manual VCD checks to drop from daily to every second day without losing the trend. The Bioreactor Data Dashboard is useful for comparing capacitance trends against the discrete counts that remain.

Automated Sampling Systems: Numera, Seg-Flow, and MAST

Automated sampling platforms withdraw culture from several bioreactors on a schedule and route it to one or more analyzers without operator contact. They raise data density from a few points per day to dozens, and they remove most manual aseptic handling.

In a published evaluation, Hofer et al. (2020) showed a reliable automated sampling system could monitor CHO cultures on-line and in real time, with sample integrity good enough to feed biomass and metabolite analyzers. Automated sampling can replace roughly 80% of the manual off-line events in a typical metabolite and cell-count plan. The remaining 20% are release-relevant or specialized tests that still need a human and a laboratory. For cell and gene therapy, Liu et al. (2024) describe a device for automated aseptic sampling aimed at future manufacturing, where small culture volumes make manual draws especially costly. On-line analytics can extend beyond standard metabolites: Cortada-Garcia et al. (2024) demonstrated on-line targeted metabolomics for real-time monitoring in fermentation.

Choosing among them is mostly about context. For parallel scale-down models in process development, high channel count and fast intervals (Numera, Seg-Flow) matter most. For a GMP manufacturing suite, qualification package, data integrity and the sampling path (MAST) matter most. For a lab already standardized on one vessel platform, native integration (Eppendorf) saves integration cost. In every case, budget for the analyzer, not only the sampler: a sampler that cannot be matched to a validated analyzer has no IPC value. The Sigma-Aldrich technical article listed in the resources section gives a vendor-side view of how automated aseptic sampling accelerates access to data.

Aseptic Risk Assessment for Bioreactor Sampling Events

Every sampling event is a small, quantifiable contamination risk, and the sum of those risks over a run is a design input, not an afterthought. EU GMP Annex 1 (2022) requires that sampling must not pose a contamination risk, and treats Grade A zones as zero tolerance.

For a manual event in a Grade A environment, a reasonable planning estimate of sterility breach probability is 10^-4 to 10^-3 per event. These values are planning figures from industry experience rather than regulatory limits, and your own media fill and contamination history should replace them. The point is the multiplication. A run with 40 manual events at 5 × 10^-4 per event has about a 2% cumulative probability of at least one breach (1 − (1 − 0.0005)⁴⁰ ≈ 0.0198). At 5 × 10^-5 per event, using a closed automated path, the same run drops to about 0.2%. With 200 batches per year, that is the difference between roughly four contaminated batches and fewer than one.

A practical assessment follows the structure of an FMEA. List each sampling event type, score the severity (loss of batch, loss of product, none), occurrence (open versus closed path, operator count), and detectability (will contamination be caught before harvest?). Then redesign the highest-risk steps: replace open syringe draws with closed single-use devices, replace repeated draws with a shared aliquot, and replace discrete checks with in-line sensors where a model exists. The detailed hardware options for closed devices are covered in the aseptic bioreactor sampling guide. For media fill validation of the whole aseptic process, see the article on aseptic process simulation.

A cumulative risk budget is a useful governance device. Agree with quality assurance on a maximum number of manual open-port events per batch (for example 30) and treat any plan above the budget as needing either design change or documented justification. Then every proposed new test has a visible cost: it either fits in the budget, replaces another test, or triggers a conversation about automation.

Building an IPC Sampling Plan: Step-by-Step Framework

Build the plan in six steps: list decisions, list analytes, assign modalities, set frequency, check the volume and risk budgets, then document and challenge the plan. Doing it in this order stops teams from starting with instruments they already own and working backwards.

1. List the decisions the data will support. Typical decisions: adjust feed rate, trigger temperature shift, set harvest time, release the batch to downstream. Tie each to a CPP or CQA from your PAT and control strategy.
2. List the analytes that inform each decision and mark which are release-relevant under ICH Q6B (in-process tests at critical decision-making steps).
3. Assign a modality using decision latency, regulatory status, and in-line model availability (Table 1).
4. Set frequency per phase using the dynamics of each analyte and the events in the run (Figure 2).
5. Check the budgets. Cumulative volume under 5% of working volume (Figure 3) and manual open-port events within the agreed risk budget.
6. Document, justify and review. Capture each choice in the control strategy and PPQ protocol, then review it after every campaign. Remove tests that never changed a decision.

The same logic scales down. For a 2 L bioreactor, the plan would use the automated or micro-volume at-line path and in-line sensors for most trends, reduce discrete samples to 1 mL, and accept only once-daily off-line titer. The Growth Curve Fitter can then fit the discrete VCD data from that plan to estimate growth rate, and the OTR/kLa Estimator helps verify that oxygen transfer is not the unseen limit behind a growth deviation your samples revealed. Where gas composition matters for a sampling-driven control action, the Gas Mixing Calculator handles the overlay arithmetic.

Common Mistakes in Bioreactor Sampling Plans

Several errors recur across sites. Teams over-sample stable analytes out of habit, burning volume and risk for no decision value. They under-sample at event boundaries, missing the response to a temperature shift or feed change that is exactly the data a design-space argument needs. They change modality without bridging, so that a trend plotted across a switch from off-line to at-line contains a hidden bias. They forget dead volume, so the real draw is double the nominal. And they write plans around instruments rather than decisions, which leaves expensive analyzers producing numbers nobody uses.

A last point concerns timing. A sample is a snapshot of the culture at the moment of withdrawal, but results from off-line methods arrive later. Time-stamp both the draw and the result in your data system, and make sure control decisions use the draw time. A 45-minute off-line delay on glucose in a fast-growing culture means the number describes a vessel that no longer exists.

Frequently Asked Questions

Related Tools

• Growth Curve Fitter — Fit growth data from your sampling results
• Bioreactor Data Dashboard — Visualize process data across runs
• Gas Mixing Calculator — Gas overlay calculations

References

1. Hofer A., Kroll P., Barmettler M., Herwig C. (2020). A Reliable Automated Sampling System for On-Line and Real-Time Monitoring of CHO Cultures. Processes, 8(6), 637. doi:10.3390/pr8060637
2. Fung Shek C., Betenbaugh M. (2021). Taking the pulse of bioprocesses: at-line and in-line monitoring of mammalian cell cultures. Current Opinion in Biotechnology, 71, 191–197. doi:10.1016/j.copbio.2021.08.007
3. Cortada-Garcia J. et al. (2024). On-line targeted metabolomics for real-time monitoring of relevant compounds in fermentation processes. Biotechnology and Bioengineering, 121, 1170–1181. doi:10.1002/bit.28599
4. Liu D. et al. (2024). Device for automated aseptic sampling: Automated sampling solution for future cell and gene manufacturing. Frontiers in Bioengineering and Biotechnology, 12, 1452674. doi:10.3389/fbioe.2024.1452674
5. Metze S. et al. (2020). Monitoring online biomass with a capacitance sensor during scale-up of industrially relevant CHO cell culture fed-batch processes in single-use bioreactors. Bioprocess and Biosystems Engineering, 43, 193–205. doi:10.1007/s00449-019-02216-4