# Bioprocess analytics and PAT

The logic of the FDA's process analytical technology framework, the spectroscopy that makes in-line measurement possible in a living broth, and why a chemometric model trained at bench scale is the weakest joint in the chain.

Why a process can be measured while it runs — and why the calibration model, not the probe, is the part that fails.

Source: https://en.bioecon.ru/docs/bioproduction-equipment/analytics-pat/bioprocess-analytics-pat/
Updated: 2026-09-04



Process analytical technology is not a class of instrument. It is an argument about where quality comes from, set out by the FDA in its 2004 guidance and carried into ICH Q8, Q9 and Q10: quality cannot be tested into a batch at the end, because a release assay on a finished lot only sorts good material from bad. It has to be built in, which means the process must be understood well enough that the variables driving quality are known, measured while the process runs, and controlled. End-of-batch testing then confirms a conclusion already reached rather than discovering it.

## In-line, on-line, at-line

The distinction is about sample handling, and it decides what a measurement can be used for. An **in-line** probe sits in the broth itself and returns a value continuously with no sample removed — a Raman immersion probe, a pH or dissolved-oxygen electrode. An **on-line** measurement diverts a stream through an analyser and may return it; the sample leaves the vessel but not the loop. An **at-line** measurement is a sample pulled and carried to an instrument beside the suite, minutes away. Only in-line and fast on-line data close a control loop, because a feedback controller cannot act on a value that describes the process as it was twenty minutes ago. At-line data still has real value, but it informs decisions, not automatic control.

## Why spectroscopy, and what it costs

A mammalian culture is water with cells, sugars, amino acids, salts and protein in it, and it must not be disturbed. That rules out most chemistry and leaves optical methods. Near-infrared reads overtone and combination bands of C–H, N–H and O–H stretches: strong signal, but the bands are broad and heavily overlapped, and water dominates them. Raman reads inelastic scattering, where a photon exchanges energy with a molecular vibration. Its bands are sharp and water is a weak Raman scatterer, which is exactly why it suits an aqueous broth — but only around one incident photon in ten million scatters inelastically at all, so the signal is faint and easily buried under fluorescence from media components and from the cells themselves.

Neither technique gives a glucose concentration directly. The raw spectrum is a sum of everything present, so the concentration is inferred from a multivariate calibration — usually partial least squares — trained on spectra paired with reference assays.

## The weak joint

That model is correlational. It learns the covariance structure of its calibration set, and it is valid only where that structure holds. A model built on scale-down runs in 3-litre glass has been trained on a particular relationship between the analyte of interest and everything that moves with it: cell state, lactate accumulation, feed timing, bubble population. At 2,000 litres the mixing time is longer, gas hold-up and shear differ, the culture takes a different metabolic trajectory, and the covariance the model relied on quietly changes. The probe still reports; the prediction drifts, and nothing in the spectrum announces it.

This is why PAT programmes stand or fall on model lifecycle management rather than hardware: calibration transfer between instruments, deliberate design-of-experiment runs that break correlations instead of reproducing them, residual and Hotelling statistics to flag when a new spectrum lies outside the model's training space, and a defined revalidation trigger. A model treated as a fixed asset rather than a maintained one is the standard failure in this field.