A

aTftdmmw©

K

1

pb[piibpwat

dma

0d

Analytica Chimica Acta 590 (2007) 110–117

Multiplexing fibre optic near infrared (NIR) spectroscopy asan emerging technology to monitor industrial bioprocesses

Payal Roychoudhury a, Ronan O’Kennedy b, Brian McNeil a,∗, Linda M. Harvey a

a Strathclyde Fermentation Centre, Institute of Pharmacy & Biomedical Science, University of Strathclyde, Glasgow G1 1XW, UKb BioPharmCEDD, GlaxoSmithKline, Beckenham BR3 3BS, Kent, UK

Received 15 January 2007; received in revised form 1 March 2007; accepted 4 March 2007Available online 12 March 2007

bstract

The application of near infrared spectroscopy in bioprocessing has been limited by its dependence on calibrations derived from single bioreactort a given time. Here, we propose a multiplexed calibration technique which allows calibrations to be built from multiple bioreactors run in parallel.his gives the flexibility to monitor multiple vessels and facilitates calibration model transfer between bioreactors. Models have been developed

or the two key analytes: glucose and lactate using Chinese hamster ovary (CHO) cell lines and using analyte specific information obtained fromhe feasibility studies. We observe slight model degradation for the multiplexed models in comparison to the conventional (single probe) models,

2

ecrease in r values from 89.4% to 88% for glucose whereas for lactate from 92% to 91.8% and a simultaneous increase in the number of factors as theodel incorporates the inter-probe variability, nevertheless the models were fit for purpose. The results of this particular application of implementingultiplexed-NIRS to monitor multiple bioreactor vessels are very encouraging, as successful models have been built on-line and validated externally,hich proffers the prospect of reducing timelines in monitoring the vessels considerably, and in turn, providing improved control.2007 Elsevier B.V. All rights reserved.

ter ov

nrtaatfrpmu

ucn

eywords: Near infrared spectroscopy; Multiplexed calibration; Chinese hams

. Introduction

Monoclonal antibodies (MAbs) are key biopharmaceuticalroducts which represent a significant proportion of the currentusiness focus and generate revenues of £3 billion per annum1]. As with many other biologically derived products, the mainroduction step relies on an engineered cell line being grownn bioreactors with the aim of expressing the required productn large volume. One key characteristic of recombinant derivediopharmaceutical like MAbs is that most product yield androduct quality attributes are set during the “manufacturing”ithin the recombinant cell host [2] and these characteristics arefunction of the host cell genetics and physiological response

o the large scale culture environment.Existing invasive sensor technologies have a number of

isadvantages with respect to implementation in large-scaleammalian cell culture, even though many of the technologies

re based on “primary” or direct sensing methods. Existing tech-

∗ Corresponding author. Tel.: +44 141 548 4379; fax: +44 141 553 4124.E-mail address: b.mcneil@strath.ac.uk (B. McNeil).

ce(bcmd

003-2670/$ – see front matter © 2007 Elsevier B.V. All rights reserved.oi:10.1016/j.aca.2007.03.011

ary; On-line

ology, such as automated glucose/metabolite analysers, rely onemoval of a clarified sample by filtration. Although these filtra-ion devices are capable of aseptic removal, they are considereds an increased contamination risk, therefore, have restrictedpplication in large scale and commercial mammalian cell cul-ure processes. Moreover, although only small volumes are usedor analysis, the process requires considerable volumes to beemoved to purge the sample lines; this may not be an issue atilot scale processes, but in small scale vessels restriction on theaximum volume which can be sampled has limited the sensors

tility.Ideally, process monitoring sensors should be suitable for

se at all culture volumes: small scale, pilot and commer-ial scale fermentations. The technology should be robust andon-invasive, be dependable and not a potential risk to the pro-ess asepsis. Additionally, the sensor should be multiplexible,nabling multiple bioreactors to be monitored simultaneouslyparticularly in development). The sensors should also be capa-

le of measuring a number of cell culture parameters, such asell density, culture viability, glucose, lactate, glutamine, gluta-ate and ammonia, which are critical in making process specific

ecisions.

a Ch

(itmtioitrap

pttmocbtilbopip

oAlacwrm[ou

tcatdo

ibttphns

atp

mboenos(dkcabOf

iwiwdo

iuuskbur

epdhtct[stfoma

2

P. Roychoudhury et al. / Analytic

In recent years, the utilisation of near infrared spectroscopyNIRS) as a monitoring technique within the bioprocessingndustry has grown rapidly, initially from raw materials testing,hen to final product quality control, and most recently to process

onitoring. Advances in personal computers and chemometricechniques have contributed to this growth, together with anncreased realisation that NIRS can be used with a wide rangef processes operated under diverse conditions. The growingnterest in NIRS is probably a direct result of its major advan-ages over other analytical techniques: simplicity of operation,apidity of measurement, multiplicity of analysis, flexibility inpplication, and possibility of predicting chemical and physicalarameters from a single spectrum [3].

The Food and Drug Administration (FDA) outlined therocess analytical technology (PAT) initiative, which is a sys-em for designing, analysing and controlling manufacturinghrough timely measurements of critical quality and perfor-

ance attributes of raw and in-process materials, with the goalf ensuring final product quality [4]. This has led many pharma-eutical companies to re-evaluate their future requirements forioprocess analysis. The PAT initiative specifically encourageshe development of improved process monitoring technologyn the manufacturing environment. This has effectively stimu-ated the investigation of a number of novel technologies in aiomanufacturing context; and much attention has been focusedn near infrared spectroscopy [5]. Potentially, NIR offers therospect of real-time control of the physiology of cultures cellsn fermenters, leading to marked improvements in authenticity,urity and production efficiency.

However, a number of potential difficulties in the applicationf in situ NIRS to process monitoring have been identified byrnold et al. [6]. The key issues involved: the challenging rheo-

ogical characteristics of some bioprocesses, high cell densities,nd the vigorous gassing and stirring rates of aerobic microbialultures, as also observed in previous reports of NIR [7,8]. Butith mammalian cell culture processes, these problems are much

educed, since these cultivations involve low agitation rates, aodest gas phase, and clean homogeneous non-viscous broths

9,10]. Therefore, some of the potential problems of in situ NIRSutlined by Arnold et al. [6] are less likely to impact seriouslypon the quality of the spectra in these culture systems.

The recent review by Scarff et al. [11] details on the poten-ial use of off-line, at-line and in situ NIRS for monitoring andontrolling a number of industrially important organisms. It ispparent that all previous on-line studies have been restrictedo the condition of single probe monitoring the bioprocess. Toate, no study has been published involving multiplexing fibreptic probes to monitor industrial cell culture bioprocesses.

Conventional calibration methods for NIR mainly focused onnstalling a single probe in a reactor vessel and collecting cali-ration data over a series of bioreactor runs. This implies a fairlyime consuming calibration exercise, with modelling based onhe process characteristics in one vessel only. By contrast, in the

resent study, we propose a multiplex calibration method whichas a number of distinct advantages over the conventional tech-ique including: model transferability i.e. models built on smallcale reactors could be employed to monitor large scale reactors,

2

a

imica Acta 590 (2007) 110–117 111

nd since multiple vessels could be monitored simultaneously,his in turn, would expedite the calibration process reducingrocess timelines.

Logically, before implementing the fibre-optic technology toonitor industrial bioprocesses, a number of issues require to

e addressed, such as, evaluation of the optical characteristicsf the different probes, mirrors and spectrometer channels to bemployed. Despite its critically important nature, there is simplyo information available on what effect probe/mirror/channelptical differences would have on the calibration models. Howimilar do the optical properties of the measurement systemprobe/mirror/spec channel) have to be to make a model, built onata arising from one system usable on another? Therefore, ourey objectives here are to answer such questions via a systematicharacterisation of the three key factors namely: probes, mirrorsnd channels, along with the realisation of their practicality,efore implementing them for the on-line process monitoring.ur approach broadly has been to evaluate the impact of these

actors upon signal intensity.As mentioned previously, all NIR studies have been limited

n their application to monitoring a single reactor at a given time,here analyte specific information was usually obtained by spik-

ng experiments [6–8,10,12]. By contrast in the present study,e report on use of multiplexing 5–6 probes, and have appliedesign of experiment (DoE) as a diagnostic tool to develop a setf uncorrelated mixtures of glucose and lactate.

The uncorrelated mixtures were utilised in the NIR feasibil-ty studies, which involved scanning of the two key analytessing multiple probes. Subsequently, models were developedsing the OPUS software, with the purpose of achieving analytepecific information, and to remove any colinearity between theey analytes in the model building process. This exercise wasasically aimed at assessing the differences between the probessed in the study, and simultaneously identifying the wavelengthegions peculiar to the analyte of interest.

Rationally, the next step was to utilise the wavelength regionsstablished from the DoE studies to build models for the actualrocess parameters themselves in production processes. We haveeveloped models here for glucose and lactate using Chineseamster ovary (CHO) cell lines. As mammalian cell culturesend to be low biomass processes with insignificant rheologi-al changes grown in the completely soluble growth medium,he model building exercise should be relatively straightforward9,10]. Therefore, we have built models based on: data from aingle probe (i.e. conventional technique); and data from mul-iple probes in multiple reactors, and compared the outcomesrom both exercises. This provided novel insights into the effectf probe optical variability on the predictive ability of derivedodels, and finally upon the practical utility of multiplexed NIR

s a process monitoring technique.

. Material and methods

.1. Bioprocess and sampling

CHO cells were cultivated in Applikon BioConsole XLnd Bio Console ADI 1035 (Applikon Biotechnology Ltd.,

1 a Ch

W16fl3CbacttblaB

2

2

au

2

tMosi

ns

prrobTe

2

iidwsc

2

atwyaaos

3

Fm

12 P. Roychoudhury et al. / Analytic

orcestershire, U.K.) fermenters of working volume 5L and2L respectively. Inocula for fermentations were cultivated in00 mL of generic propriety shake flask medium in a 1 L shakeask, at an initial seeding density of 0.6 × 106 cells mL−1 on a–4 day cycle at 37 ◦C on orbital shakers at 150 rpm in a 5%O2 atmosphere. The inoculum used to initiate growth in theioreactor was of initial seeding density 0.7 × 106 cells mL−1

nd the medium used for the bioreactors was generic proprietyomplex animal component free seed media. Process set points:emperature 37 ◦C, pH 6.95 ± 0.1, agitation rate 263 rpm, aera-ion rate O2 of 0.1 vvm and CO2 of 0.05 vvm, pH control methody CO2 sparge, and set point DOT of 30%. Samples were col-ected twice daily over the entire time course of the seed vessels,nd whole process fluid was scanned in real-time using in situruker FT-NIR as detailed in Section 2.4.

.2. Primary/reference analysis

All reference assays were carried out in triplicate.

.2.1. Glucose and lactateSupernatant samples were used for the glucose and lactate

ssays. The glucose and lactate concentrations were determinedsing a YSI 7100 MBS analyser (Yellow Springs, Ohio, USA).

.3. NIR spectral acquisition

Spectra were acquired using Bruker Matrix F FT-NIR sys-ems (Bruker Optics Ltd., Coventry, England, U.K.). The

atrix-F NIR offers an internal multiplexer for up to six fibreptics channels, enabling the use of both conventional, and thepecially designed probes used in the present study as depicted

n Fig. 1.

The FT-NIR system was controlled by an IBM PC run-ing OPUS software version 5.5, along with the ADIO processoftware to automate the measurement and analysis of multi-

3

n

ig. 1. (A) Pictorial representation of the fibre optic technology, (B) schematic reonitoring.

imica Acta 590 (2007) 110–117

le measurement channels. A spectrum of air was used as aeference and thirty-two co-added scans of the sample wereeferenced with thirty two co-added scans of air at 8 cm−1

ptical resolution, over the first and second overtone wavenum-er region of 11,000–4500 cm−1 with a pathlength of 2 mm.he OPUS software (vs. 5.5) was used for the model buildingxercises.

.4. NIR accessories

Seven 12 mm stainless steel transflectance probes were usedn the present study, which were custom designed for applicationn autoclavable bioreactors, with sapphire window material andetachable mirrors as depicted in Fig. 1. The fibre optics usedere silica cables of 12 m length, having polymer coated steel

heathing, a core diameter of 0.6 mm and terminating in 2SMAonnectors.

.5. Data analysis

Data analyses were carried out using Statistica (vs. 6.1)nd SIMCA (Umetrics, vs. 11) softwares. Statistica was usedo carry out linear regression (General Linear Models, GLM),hereas SIMCA was used to create principal component anal-sis (PCA) scores plot to get visual interpretation of the datanalysed. The DoE software (Design Expert vs. 6.0) was useds an analytical tool for developing different synthetic mixturesf glucose and lactate to assist in the NIR (section 2.3) feasibilitytudies.

. Results and discussion

.1. Signal Intensity (SI) studies

The optical system used for multiplexing NIR consists ofumber of components, including, (a) spectrophotometer chan-

presentation of the 12 mm transflectance probes used in the on-line process

P. Roychoudhury et al. / Analytica Chimica Acta 590 (2007) 110–117 113

Table 1Regression analysis results from signal intensity studies to identify sources of variance in the NIR components: spectrophotometer channel, probe design and mirroroptical properties

Degrees of freedom (d.f.) Signal intensity (F-value) Signal intensity (p-value)

Intercept 1 1444.341 0Probe 2 68.593 0Media 2 39.989 0Channel 1 15.867 1.43 × 10−4

Mirror 5 4.882 5.65 × 10−4

Error 85

T

O

npte

otia

i[paaw

itmwc

toowcato

dfht

tic

Fa

otal 95

ptical components are ordered in descending effect of variance.

el, (b) probe design, (c) mirror optical properties, which musterform in a consistent manner over the entire time course ofhe bioreactor runs, otherwise the possibility of introduction ofrrors into the NIR signal exists.

Here we conducted a SI study to determine the effect of eachf the above factors on the overall variability in signal, and onhe spectra acquired. Accordingly, each factor has been variedn turn, and SI’s have been recorded in a range of “media” i.e.ir, methanol and acetone.

A wide range of reference standards are available for validat-ng the performance of NIRS/mid infrared spectroscopy (MIRS)13,14], by which a probe manufacturer can check the opticalerformance and similarity of the key components mentionedbove. Here we used methanol and acetone as standards, as bothre easy to handle, safe, and have distinct spectral characteristicshich aid interpretation of the spectral data.The contribution of each optical component to the variability

n SI was assessed using regression analysis. Table 1 summarises

he results from regression analysis incorporating all the above

entioned factors: channel, probe and mirrors. This exerciseas aimed at ranking each component in terms of its overall

ontribution to variance in signal intensity.

hita

ig. 2. Scores PCA plot developed for the probe test using probes (1–7) on a constans depicted in the legend.

In Table 1 each optical component is sorted on the basis ofhe F value and is listed in descending order to indicate the levelf variance contributed by individual components. The variancef each optical component was compared to variance associatedith media differences. Media was expected to show a signifi-

ant effect on SI as all the medium components; air, methanolnd acetone have distinct SI’s, and was used as a standard (posi-ive control) to determine relative difference in variance for eachf the optical components.

The Intercept is a test of average of all the components andetermines whether the overall average is significantly differentrom zero. For this, the Intercept would be expected to have theighest F value and be significantly different from zero; hencehe effect can be ignored.

The regression analysis presented in Table 1 establisheshat probe differences contribute significant optical variationn signal, followed by media, whereas mirror differencesontributed least. Logically, the next step would be to establish

ow this variability in probes affects spectra. For example, ist possible to readily distinguish between probes by examininghe spectral data for methanol and acetone? To answer thebove questions, a probe test involving seven different probes

t single channel and mirror where P0: probes, MT: methanol and AC: acetone

1 a Ch

oav

usipvnsrcoab

aeavsvtt

iaal

Fd

dcndtdoipibd

caoii

3

awcua

14 P. Roychoudhury et al. / Analytic

n a constant single spec channel and mirror was conducted,nd spectra were recorded, the aim being to capture only theariance contributed by the probes.

The spectral data for methanol and acetone were analysedsing principal component analysis (PCA) using the SIMCAoftware. The spectra were mean centred, and without scal-ng. The scores plot (Fig. 2) was developed using the first tworincipal components. The first component captures maximumariance in the dataset of 99.6% whereas the second compo-ent extracts the maximum amount of remaining variance. Thecores plot is potentially useful as it provides a summary of theelationship between each observation condition and provides alear visual representation of all parameters and distinguishedn the basis of spectral data. It is clear from the plot that therere two distinct groupings, for methanol and acetone, as showny the two circles black and red respectively.

Additionally, the loading on PC1 (Fig. 3A) is rather similar tocombination of the methanol and acetone spectra (Fig. 3B) as

xpected since PC1 captures the difference between methanolnd acetone spectra. The loading on PC2 captures the probeariance corresponding mainly to the acetone spectra, but showslight similarity to the methanol spectra. We also observe thatariance in signal is greater at lower wavenumbers both for ace-one and methanol. This was perhaps unsurprising as energyhroughput is directly proportional to the wavenumber.

Therefore, to summarise, from both these studies, we havedentified, (i) probe optical differences contribute most vari-

nce to signal, (ii) these optical differences between the probesre apparent in the scores plot, (iii) more variance occurs atower wavenumbers as shown in the loadings plot. The work

ig. 3. (A) PCA loadings plot of principal component 1 and 2, and (B) spectralata of pure methanol and acetone.

r(o

ios[w5a

TEl

P

S

1111

R

imica Acta 590 (2007) 110–117

escribed here is entirely novel, as so far there has been no studyonducted on implementing multiplexed NIR fibre optic tech-ology in bioprocesses. Most importantly, despite the care takenuring probe manufacture to ensure identical physical charac-eristics, the study clearly shows the potential impact of opticalifferences. However, the inherent advantages of this technol-gy enable the user to obtain a much higher frequency of datan comparison to the conventional in situ technique. From theerspective of the biopharmaceutical industry, this approachs beneficial as a number of reactors at different scales cane monitored simultaneously which will accelerate the processevelopment stage.

Rationally, the next step was to conduct a modelling exer-ise (a feasibility study) using synthetic samples of glucosend lactate, as this would provide information as to whetherr not the probe optical variability identified in the preced-ng section impacted significantly upon the modelling processtself.

.2. Rationale behind modelling

DoE is a powerful technique used for optimising processesnd identifying critical quality variables. In this study, DoEas used to provide an experimental design matrix where the

oncentrations of each analyte in experimental mixtures werencorrelated. There were two key objectives, firstly, to obtainnalyte specific information for glucose and lactate using uncor-elated mixtures with optically different probes as variablesfactors), and, secondly, to assess the effect of optical variabilityf the probes on the modelling exercise.

Spectral data were acquired from all the probes, by immers-ng each probe individually in the set of uncorrelated mixturesf glucose and lactate as presented in Table 2. The OPUSoftware was used to develop partial least square (PLS) models

6–8,10,12]. Lactate absorbed in the region 8921–7146 cm−1

hereas glucose absorbed in the region 7501–6097 and450–4497 cm−1. Multiplicative scatter correction was useds a pre-processing technique for both the analytes [12,15].

able 2xperimental design layout for uncorrelated synthetic mixtures of glucose and

actate used for the NIR feasibility studies

robe number 1 → 4

erial number Glucose (g L−1) Lactate (g L−1)

1 0.2 0.022 10.1 0.023 20 0.024 0.2 2.515 10.1 2.516 20 2.517 0.2 58 10.1 59 20 50 10.1 2.511 10.1 2.512 10.1 2.513 10.1 2.51

eplicate block was carried out for each probe.

P. Roychoudhury et al. / Analytica Chimica Acta 590 (2007) 110–117 115

Table 3Models built from uncorrelated mixtures of glucose and lactate for the NIR feasibility studies

Probes Analyte Number of samples Wavenumber region (cm−1) Factors Calibration Cross Validation

r2 (%) SEC (g L−1) r2 (%) SEP (g L−1)

Single Glucose I 9 5450–4497 2 99.6 0.8 93.1 0.97501–6097

Lactate I 9 8921–7146 2 99.7 0.1 99.3 0.2

Multiple Glucose II 25 5450–4497 5 99.7 0.6 98.3 0.97501–6097

Lactate II 32 8921–7146 5 99.1 0.2 98.5 0.2

C multi 2

a artial

Cuco

I9vlgott

iilqoisnlm

dpilt

3

fmmmantb

ssmTapipoa

tdwbcp

t[uli

uttipnab1pas

c

alibration and validation data for glucose and lactate developed using a single/nd SEP: standard error of prediction and multiplicative scatter correction and p

alibration models were developed for glucose and lactatesing both conventional single probe and multiple probes, andross validation of these models was performed using the leavene out approach.

The calibration models for single probe, lactate I and glucoseperformed well, having high correlation coefficient values of9.6 and 99.7 respectively (Table 3) with 2 factors, and on crossalidation had high correlation coefficients of 93.1 and 99.3 andow SEPs. Cross validation of the models from multiple probes,lucose II and lactate II, also gave high correlation coefficientsf 98.3 and 98.5, and showed an increase in the number of factorso 5. This may well have been due to these models incorporatinghe inter-probe variability.

It can be concluded from the feasibility study that NIR candentify discrete wavelength regions for glucose and lactate. Thiss important because previous NIR studies have reported over-apping wavelength regions for the key analytes [6] which raisesuestions about the model accuracy and about the chemical basisf the models. One risk of using multivariate techniques like PLSs that since in fermentation systems a stoichiometric relation-hip exists between analytes, a PLS model of one analyte mayot be based solely on the spectral characteristics of that ana-yte alone [15]: the above approach helps minimise this risk inodelling.In addition, the results of the feasibility study show that

espite the optical dissimilarities in probes successful multi-robe PLS models could be constructed for these two analytesn synthetic mixtures. The next step was to seek to use the wave-ength regions identified in the feasibility study for modellinghe bioprocess data.

.3. Modelling seed data

In mammalian cell cultures, the host cell line is generatedrom a single well characterised cell bank, and the growthedium does not undergo substantial reformulation, this aspectinimises development timelines during early phase develop-ent. Development usually entails a series of small scale seed

nd production fermentations, which are followed by a smallerumber of pilot and large scale seed and production bioreac-ors. Due to these process development characteristics, it shoulde possible to develop a NIR calibration from multiple small

sgsf

ple probes where, r : correlation coefficient, SEC: standard error of calibrationleast square has been used to create the models.

cale bioreactors and transfer the calibration to pilot and largecale bioreactors with a minimum (or ideally no) recalibration. Itay also be possible to transfer calibrations between cell lines.herefore, to systematically investigate the practicality of theseims, we have built models on: spectral data arising from a singlerobe and multiple probes. This study was carried out to identifyf there are any differences in the models (single probe/multiplerobes), and whether is practicable to implement them for then-line monitoring of the two key process parameters: glucosend lactate.

The wavenumber regions that provide information specifico the analytes were included in the present study, while regionsominated by noise or other artefacts that might distort the modelere excluded. PLS was used to build the models since it haseen widely demonstrated to be useful for extracting analyteoncentration from the complex spectra typical of fermentationrocess fluids [16].

Although previous reports have mostly involved the use ofhe second derivative when using NIRS in bioprocess fluids6,8,10], in the present study this spectral manipulation provennecessary. Mammalian cell culture systems tend have veryow biomass levels relative to microbial processes, so usefulnformation can be easily extracted from the raw spectra [16].

So ideally, NIR models should be derived from the spectra ofnmodified samples, but it is often essential to compensate forhe undesirable characteristics of such fluids including absorp-ion bands, matrix effects and light scattering differences. Thiss important both for prediction and for interpretability. In theresent study, MSC has been applied as a pre-processing tech-ique to help obtain useful information from the raw spectra,nd to reduce the effect of the scattering caused by changingiomass concentration [12]. MSC fits a straight line to a plot of00 spectral points for a sample against the 100 correspondingoints of the mean spectrum by least squares, and uses the slopend intercept of the line to “correct” the scattering effects on thepectrum [12,15].

Development of a PLS model requires two datasets: one foralibration, or training and one for prediction. In the present

tudy, two calibration models have been developed for glucose:lucose I (Fig. 4A) comprising 50 samples captured using aingle probe, and glucose II (Fig. 4B) incorporating 119 samplesrom multiple probes. On calibration, both the models performed

116 P. Roychoudhury et al. / Analytica Chimica Acta 590 (2007) 110–117

F evelop ely.

wplr

v2theod

t

lpb

B(m9wI

TM

P

S

M

SM

Cp

ig. 4. Correlation coefficient of calibration and validation datasets for glucose drobes, (C and D) test set validation using single and multiple probes, respectiv

ell (Table 4) using 4 and 6 factors (as determined by the PRESSlot), with high correlation coefficient of 98.3% and 92.8%, andow standard errors of calibration (SEC) of 0.1 and 0.2 g L−1

espectively.However good the calibration equations appear the real test is

alidation. Here external validation was undertaken using 16 and5 samples (for glucose I and II respectively) to critically assesshe practical utility of the developed models using datasets whichave not been included in the formulation of the calibrationquation. The standard error of prediction (SEP) is a measure

f the standard deviation of the residuals calculated from theifference between primary reference data and NIRS data.

External validation of glucose I and II (Fig. 4C and D) usinghe additional datasets generated r2 of 89.4% and 88% and a

a

ff

able 4odels built for glucose and lactate using spectral datasets from single/multiple prob

robes Analyte Wavenumberregion (cm−1)

Number of samplesin calibration

Factors

ingle Glucose I 5450–4497 50 47501–6097

ultiple Glucose II 5450–4497 119 67501–6097

ingle Lactate I 8921–7146 76 4ultiple Lactate II 8921–7146 152 7

alibration and validation data of the respective analytes where, glucose and lactate Irobe models and multiplicative scatter correction and partial least square has been u

ped in situ from CHO cell lines, (A and B) calibration using single and multiple

ow SEP of 0.2 and 0.2 (Table 4) respectively. Therefore, theerformance of both models (single probe and multi-probe) wasroadly similar on validation.

Further calibration equations for lactate I and II (Fig. 5A and) were built using 76 and 152 samples, using 4 and 7 factors

determined by the PRESS plot) respectively. On calibration, theodels performed well with correlation coefficients of 97% and

4.5% and low SEC (Table 4) respectively. External validationas carried out using 25 and 30 unseen samples for lactate I and

I (Fig. 5C and D) generating r2’s of 92% and 91.8% respectively,

nd low errors of prediction.

Both single and multiple probe based models could beormulated, and external validation showed broadly similar per-ormances of these models, indicating the practical utility of this

es in the seed vessels of Chinese hamster ovary cell (CHO) lines

Calibration Number of samplesin validation

Validation

r2 (%) SEC (g L−1) r2 (%) SEP (g L−1)

98.3 0.1 16 89.4 0.2

92.8 0.2 25 88 0.2

97 0.1 25 92 0.194.5 0.1 30 91.8 0.1

– developed using single probe and glucose and lactate II – built using multiplesed to create the models.

P. Roychoudhury et al. / Analytica Chimica Acta 590 (2007) 110–117 117

F evelopp ely.

av

4

Naiipim

R

[

[

[

[13] http://www.brukeroptics.com/accessories/catalog/NIR/nir ref stand.html.

ig. 5. Correlation coefficient of calibration and validation datasets for lactate drobes, (C and D) test set validation using single and multiple probes, respectiv

pproach for monitoring key analytes in multiple fermentationessels.

. Conclusion

This study seeks to investigate the utility of multiplexedIRS in monitoring key analytes in CHO cell cultures used for

ntibody manufacture. The study is both timely and innovativen view of the pressing need for effective bioprocess monitor-ng. Our findings indicate that despite optical differences in therobes used, and other contributory sources of variance, thiss a technology which has the potential to deliver real time

ulti-analyte data from multiple bioprocesses.

eferences

[1] http://www.dti.gov.uk/sectors/bioscience/igt/report/page25377.html.[2] A.J. Chirino, M.S. Anthony, Nat. Biotechnol. 22 (2004) 1383.[3] G. Reich, Adv. Drug Del. Rev. 57 (2005) 1109.

[[[

ed in situ from CHO cell lines, (A and B) calibration using single and multiple

[4] http://www.fda.gov/cder/guidance/6419fnl.pdf.[5] T.M. Larson, H. Lam, PAT J. 1 (2004) 20.[6] S.A. Arnold, R. Gaensakoo, L.M. Harvey, B. McNeil, Biotech. Bioeng. 20

(2003) 405.[7] S.A. Arnold, J. Crowley, S. Vaidyanathan, L. Matheson, P. Mohan, J.W.

Hall, L.M. Harvey, B. McNeil, Enzyme Microb. Tech. 27 (2000) 691.[8] S. Vaidyanathan, S.A. Arnold, L. Matheson, P. Mohan, G. Macaloney, B.

McNeil, L.M. Harvey, Biotechnol. Prog. 16 (2000) 1098.[9] J. Crowley, M.A. Arnold, L.M. Harvey, B. McNeil, Eur. Pharm. Rev. 5

(2000) 39.10] S.A. Arnold, J. Crowley, N. Woods, L.M. Harvey, B. McNeil, Biotechnol.

Bioeng. 84 (2003) 13.11] M. Scarff, S.A. Arnold, L.M. Harvey, B. McNeil, Crit. Rev. Biotechnol. 26

(2006) 17.12] J. Crowley, A.S. Arnold, N. Woods, L.M. Harvey, B. McNeil, Enzyme

Microb. Tech. 36 (2005) 621.

14] http://www.pro-lite.uk.com/Light/Labsphere/lab mats.html.15] M.R. Warnes, P.P. Glassey, Process Biochem. 31 (1996) 147.16] M.R. Riley, M.R. Rhiel, X. Zhou, M.A. Arnold, D.W. Murhammer,

Biotechnol. Bioeng. 55 (1997) 11.