international comparability in circular dichroism

International comparability in spectroscopic measurements of protein structure by circular dichroism: CCQM P59 Jascindra Ravi, David Schiffmann, Ratna Tantra, Simon Cox, Jonathan Eady, Christopher Jones1, John S Vrettos2, Richard P Affleck2, Julie DeSa Lorenz3, Yasushi Shigeri4, Sheng Linghui5, Liu Jun5, Robert Willows6, Philippe Charlet7, Yves Dupont8, Curtis W Meuse9, Marc J A Bailey and Alex E Knight10

Analytical Science Division, National Physical Laboratory, Teddington, Middlesex, TW11 0LW 1Laboratory for Molecular Structure, National Institute for Biological Standards and Control, South Mimms EN6 3QG. 2Human Genome Sciences, Inc., Rockville, MD. 3Olis, Inc. Bogart, GA. 4National Metrological Institute Japan (NMIJ)/National Institute of Advanced Industrial Science and Technology (AIST), Osaka, Japan. 5NIM, Beijing, China. 6Macquarie University, Department of Chemistry and Biomolecular Sciences, North Ryde, NSW, Australia. 7Laboratoire National d'Essais, Paris, France. 8Département de Réponse et Dynamique Cellulaires, Laboratoire de Biophysique Moléculaire et Cellulaire, Grenoble, France. 9Biochemical Science Division, National Institute for Standards and Technology, Gaithersburg, MD, USA 20899-8312

1. Abstract

Circular dichroism is a spectroscopic technique that is widely used to obtain information about protein structure, and hence is an important tool with many applications, including the characterisation of biopharmaceuticals. However, there is a lack of confidence in the technique, arising from an observed lack of comparability in the data obtained by different laboratories, or even different operators. In this study, we set out to determine the extent of comparability in the technique, by comparing the results obtained from identical protein samples by a panel of worldwide laboratories. The laboratories chosen were either national measurement institutes (NMIs) or expert laboratories nominated by an NMI. We also aimed to identify the main factors contributing to any lack of measurement comparability. Data were analysed using PCA11 and SIMCA methods, and we show these statistical techniques are ideal for analysing large amounts of this type of spectroscopic data. We found a startling lack of comparability among laboratories, but we also found that most of the variability arose from relatively simple problems, which can be avoided by following simple guidelines. We believe that the lack of an absolute reference or measurement traceability in circular dichroism contributes to a lack of confidence in the technique.

10 Author for Correspondence. E-mail address: [email protected]. Web page: http://www.npl.co.uk/biophysics11 Abbreviations: CD: Circular Dichroism; PCA: Principal Component Analysis; SIMCA: Soft Independent Modelling of Class Analogies; NMI: National Measurement Institute; NPL: National Physical Laboratory; ACS: Ammonium d-10-camphorsulfonate; SI: Système International d’Unités; CCQM: Comité consultatif pour la quantité de matière – métrologie en chimie; BIPM: Bureau International des Poids et Mesures ; FTIR: Fourier Transform Infra-Red (spectroscopy); ROA: Raman Optical Activity; BAWG: Bio-Analysis Working Group; FDA: Food and Drug Administration; EMA: European Medicines Agency; UV: Ultra-Violet; SDS-PAGE: Sodium Dodecyl Sulfate Polyacrylamide Gel Electrophoresis.

of 27

mailto:[email protected]

http://www.npl.co.uk/biophysics

Spectroscopic measurements of protein structure: CCQM P59

2. Introduction

2.1. Circular Dichroism Spectroscopy

Circular dichroism is defined as the difference in the absorption of left and right-handed circularly polarised light by a sample [1, 2]. This technique is widely applied for the study of chiral molecules. There is a particularly wide range of applications in the spectroscopy of biological molecules. The most common use in biology is to obtain a priori information about the secondary structure of protein molecules [3]. CD spectra in the far UV region (typically 180-260 nm) can give an overall indication of the secondary structure content, because the principal chromophore in this region is the peptide backbone of the molecule. A wide variety of algorithms exist for extracting structural information from these spectra [3, 4].

Useful information about protein structure can also be obtained from other spectral regions. The visible region is useful for those proteins that have prosthetic groups or ligands that absorb in this region, such as haems. However, much more widely used is the near UV region, typically 240-320 nm. The dominant chromophores in this region are the side chains of the amino acids, particularly aromatic residues such as tryptophan, and also cystines. Although the structural interpretation is difficult, the spectral signature in this region provides a useful structural “fingerprint”, which may be very sensitive to changes in protein structure. This region is therefore very useful for stability studies or in the characterisation, comparability assessment and process development of biopharmaceuticals.

Circular dichroism in the infrared region is known as vibrational circular dichroism, and is a distinct technique, measured with different instrumentation, and is therefore beyond the scope of this article.

Unfortunately, CD is a challenging phenomenon to measure. The greatest difficulty arises from the relatively small size of the effect. Typically the difference in absorption between left and right-handed circularly polarised light is of the order of 1 part in 103 or less [5]. Commercially available instrumentation takes a number of different approaches to measuring the effect, some of them rather indirect [5]. Therefore CD measurements are usually not absolute, but are calibrated using a reference material.

Regrettably, the reference materials that are used are not truly fit for purpose. The reference values used are based on the literature [6] and have unknown uncertainties. Ideally, such reference values should be traceable to the SI. Furthermore, the standards typically only provide calibration at one or two wavelengths, which are not necessarily the critical wavelengths for the measurements being made. Evidence from the literature shows that there may be considerable wavelength-dependent variation of calibration between instruments [7, 8]. Also there are issues relating to the correct formulation and storage of these materials [9].

Another issue that appears to bedevil the field is a lack of good practice in CD measurement. The technique requires a degree of knowledge to be carried out and interpreted correctly. Unfortunately, it is often the case that users are inadequately trained and poor measurements are common, even in the published literature [10]. There have been several previous attempts to formulate good practice in the measurement of CD [10, 11].

With all measurement techniques, the ideal is that they should be traceable to the SI. For example, UV/Visible absorbance spectroscopy is traceable, allowing absolute measurements to be made, and there are a number of standards to support this. Whilst circular dichroism spectroscopy of proteins has a relatively long history (dating back to the 1960s [12]), it does not have such well-established measurement traceability. Indeed, the measurement problems of the technique have been recognised for many years, but few solutions have been found.

What are the consequences of such a lack of traceability? If measurements are not traceable to the same absolute reference, then measurements made in different laboratories, or with different instrumentation, or at different times, or under different conditions, cease to be comparable. If the data are not comparable, then this limits the applicability of the data. For example, scale (magnitude) differences to CD spectra can cause different results in secondary structure predictions [4]. Such scale differences cannot be unambiguously assigned to calibration, path length, concentration or structural differences, without additional data. Furthermore wavelength differences, or distortions in spectral shapes, will also influence results [4, 7, 13].

of 27


Similar problems will also occur in other applications of circular dichroism. For example, where CD is used as a tool to compare different production batches of biopharmaceuticals, artefactual differences between spectra that are collected at different times, or in different laboratories, will cause structurally identical samples to appear to be different. This could lead to the rejection of a batch of a drug that is actually sound – at huge expense to the manufacturer. Conversely, if spectral comparability is poor, tolerances on spectral differences could be set too high, resulting in problems being missed. The net result is to undermine industry confidence in a technique that should be extremely useful.

Given the importance of this technique in such a commercially and medically significant sector, and our concerns about the reliability of the measurement, we set out to investigate the seriousness of the problem on an international scale. The international system of units – the SI – is administered by the BIPM. Development and support of the SI in various scientific areas is the responsibility of the consultative committees; the committee responsible for chemical measurement is the CCQM, which is responsible for the mole, the SI unit for amount of substance. The Bioanalysis Working Group (BAWG) is responsible for international metrology activities in bioanalysis. The Bioanalysis Working Group and the CCQM approved this study as a pilot study, designated CCQM-P59.

2.2. Aims of Study

The aim of a pilot study is typically to compare the ability of NMIs to make a particular measurement. This, indeed, was largely the goal of this study, with the addition of some expert laboratories. Furthermore, we also attempted to identify the source of any errors in participant measurements. We were also keen to investigate the utility of pattern recognition techniques in analysing large spectral data sets of this type. In this study, we chose to achieve these aims by distributing samples to the participating laboratories for them to measure.

Because we were primarily interested in the biological applications of the technique, we chose to distribute biological samples to the study participants. Measuring biologicals (in this case, proteins) introduces additional

concerns, such as sample stability and light scattering. These issues are exacerbated by the necessity of shipping the samples internationally. A key element of the study design, therefore, was to control for changes in the samples.

3. Methods and Materials12

3.1. Study Design

A significant concern informing the design of the study was that changes in the samples during shipping or storage over the course of the study would be responsible for differences between the data from participants. To eliminate this concern, a number of steps were taken. Firstly, all the samples were subjected to a battery of stability tests (see below). Secondly, a master set of samples was retained at NPL and measured at regular intervals throughout the study. Thirdly, participants were asked to return unused samples to NPL for re-measurement. This last step allowed us to place an upper limit on the changes in the sample at the time of measurement in the participants’ laboratory.

3.2. Collection of data from study participants

Prior to the study, all participants were asked to complete an electronic questionnaire detailing their CD equipment and associated technical matters. This questionnaire assisted in designing the detailed protocol for the study. Similarly, when participants made the measurements, they were also asked to complete and return a questionnaire with their data, describing the actual measurement parameters used. All questionnaires and CD data were returned by electronic mail.

3.3. Protocol for study participants

Participants were provided with a detailed protocol for the experimental work in the study.

12 Certain commercial materials, instruments, and equipment are identified in this manuscript in order to specify the experimental procedure as completely as possible. In no case does such identification imply recommendation or endorsement by the National Physical Laboratory or the National Institute of Standards and Technology, nor does it imply that the material, instrument or equipment identified is necessarily the best available for the purpose.

of 27


The detailed parameters used are given in Table 4. Participants were encouraged, but not required, to follow these parameters. However, variations from these parameters were captured in the electronic questionnaire.

3.4. Stability trials

Due to concerns for the stability of the protein samples, particularly when shipping internationally, tests were carried out to evaluate the stability of similar samples in respect of various conditions the samples were thought likely to encounter. These tests included multiple freeze-thaw cycles; extended exposure to high temperatures (37°C, 4 days); and extended storage at 4°C. Stability was evaluated by comparison of spectra of all the samples before and after the trials, both visually and using PCA.

3.5. Preparation of study samples

The samples used in this study are listed in Table 3, and included three protein samples; one sample of a commonly-used CD calibrant (ACS11); and corresponding solutions for use as blanks. All proteins were obtained from Sigma (UK). ACS was obtained from JASCO (Great Dunmow, UK). All weights were measured using a Sartorius (Göttingen, Germany) “Genius” or CP2P balance, as appropriate, and solution volumes were measured using volumetric flasks and calibrated pipettes. All solutions were filter-sterilised to prevent bacterial or fungal growth during the study. Proteins were made up as concentrated stock solutions at 5 mg·ml-1 in the phosphate buffer, before dilution into the final solutions13. Samples were stored at 4°C and protected from light.

3.6. Distribution of study samples

Samples were packed using ice packs (Sorba-Freeze, Aberdeen) in polystyrene boxes, together with all the appropriate documentation. The package also included a temperature monitor (Tinytag Plus, Gemini Data Loggers,

13 Since all participants received aliquots of the same solutions, uncertainties in protein concentration were not a factor in the comparability of participant results, and were therefore not evaluated. The purity of the proteins was however checked qualitatively by SDS-PAGE.

Chichester, UK) so that the environment of the samples could be monitored during the time they were being shipped and stored at the participant’s laboratory. Temperatures were recorded at hourly intervals.

3.7. Processing of study data

All CD data were returned to NPL by electronic mail as text files. Data processing, including collation, baseline subtraction, interpolation, curve fitting and amplitude corrections, were performed in MATLAB 7 (The MathWorks, Natick, MA, USA), using specially written scripts. Data was stored in Excel spreadsheet format (Microsoft, Redmond, WA, USA). Typically an average was taken of the two baselines provided by a participant, and this was subtracted from each spectrum. Interpolation was performed where data was collected on a pitch other than 0.1 nm, so that data could be compared directly.

Curve fitting was used to derive calibration values for the participants from the ACS spectra. Spectra were truncated to 280-300 nm and fitted with a Gaussian model using the MATLAB Curve Fitting Toolbox.

The NPL data set was corrected for the measured path length of the cells used. No calibration correction was performed as the ACS calibration values were within the normal range for the duration of the study, as confirmed by repeated measurements. ACS spectra were corrected for cell path lengths where participants provided a measured path length value.

All data were collected and processed in the technical units of millidegrees (mdeg) for convenience. The ellipticity, θ, in mdeg is simply related to the CD (expressed as ΔA) by the following equation:

AΔ⋅= 32980θ

3.8. Multivariate analysis

Essentially, five sets of protein spectra were produced, for use in the multivariate analysis:

a) The reference data set, which was used to build the models

b) The re-measurement data set, which was used to control for sample changes during shipping

of 27


c) The participants’ data as reported (after baseline subtraction, interpolation, etc.)

d) The participants’ data with the intensities corrected for the stated path length of the cells used

e) The participants’ data with the intensities further corrected for the calibration values derived from the ACS data

All multivariate analysis was performed in SIMCA-P 10 (Umetrics, Umeå, Sweden). Far and near UV data were analysed separately. For each data set, the reference data was imported into SIMCA and transposed such that each spectrum was treated as an observation, and each wavelength point was treated as a variable. No mean centring or scaling was applied to the data set. Observations (spectra) were grouped into three classes, corresponding to the sample which was measured. For each spectral data set, two types of analysis were performed:

1. A Principal Components Analysis (PCA) of all classes of spectra, to identify clusters of similar spectra.

2. A SIMCA analysis, where a separate PCA model is built for each class. This allowed new spectra to be classified according to the model.

Therefore four principal component models (one global and three class-specific) were built for each spectral dataset (see Table 5). Once the models had been built, scores plots were used to visualise the clustering of the data. “Scree plots” were used to evaluate how many of the principal components were useful for further analysis. Loadings plots were also used to assess how physically meaningful each component of the model was. For the class models, 1 component was used for the Far UV spectra and two for the near UV spectra (see Table 5). For further details, please consult the Umetrics documentation and references therein [14].

The re-measurement and participant data were then compared against the models. Predicted scores plots against the global model were used to visualise the clustering of the new data in relation to the reference data. Scores plots for the individual class models were used to assess whether the NPL data formed an outlier when compared to the participant data. The relationship between the new data and the

models is summarised by the combined distance to model statistic (DModXPS+). These distances were normalised in units of standard deviation, to permit comparison between different models, and were unweighted. Threshold or “critical” distances (Dcrit) were calculated for each model using a significance level of 0.05 (or 95% confidence). Contribution plots were used to investigate why some spectra were further from the model than others [14].

4. Results

4.1. Study practicalities

Data sets were successfully returned from all participants, and were, in most cases, complete. These included the CD data, experimental questionnaires, and temperature logs. In most, but not all, cases the samples were also returned successfully. Temperature logs (not shown) showed that in most cases, extremes of temperatures were not encountered during shipping or storage, although in most cases the samples reached ambient temperatures within 1-2 days. One participant stored the samples in a freezer, which reached temperatures as low as -21°C, both before and after measurement.

4.2. Calibration status of participants

The ACS data returned by the participants are shown in Figure 1. It is immediately apparent from panel a) that there is significant variation among the participants’ calibration state, both in terms of peak intensity and wavelength. There were no significant changes in the samples returned to NPL, or in the reference samples returned to NPL (data not shown). This is consistent with previous results, which show that the ACS calibrant is sufficiently stable for the purposes of this study, provided it is stored and measured under the correct conditions [9]. The results of curve fitting to the peaks are summarised in panel b) of the figure. This shows that many of the participants fall outside the nominal range of the calibration values, i.e. 190.4 ± 1 mdeg at 291 ± 0.8 nm [6]. The peak heights from this data were used for the calibration correction of the participants’ data; no attempt was made to correct for wavelength errors. The NPL data showed that there was a slow change in the ACS sample, leading to an increase in the measured CD after 167 days to 193.7 ± 0.18 mdeg (a 3.3 mdeg increase). The instrument was not recalibrated during this time,

of 27


and showed a slow calibration drift to 188.2 ± 0.18 mdeg (a 2.2 mdeg decrease, as measured with a fresh ACS sample) after 244 days, which would influence the previous measurement. Both sample and instrument remained within the specified limits for peak wavelength. The changes observed are relatively small, and cannot explain the relatively large calibration discrepancies observed in some cases.

4.3. Stability trials

No significant changes, or very small changes, were detected in the spectra of the test samples after freeze-thaw cycles, room temperature storage, or elevated temperature (data not shown). This indicated that if the samples were shipped as planned, no significant changes in the spectra were likely to be observed. Any changes that did occur were controlled for by the re-measurement of the returned samples.

4.4. Reference data set

Reference spectra for the three samples were collected on 7 different occasions throughout the duration of the study (a period of five months). Each spectrum was collected in triplicate, using the standard conditions for the study (Table 4). This data set served two purposes: firstly, it acted as a reference data set which was used to build statistical models; secondly, it acted as a stability control for the samples. The spectra from this data set are shown in Figure 2. Note that in both spectral regions, samples 1 and 3 give very similar spectra, whereas sample 100 is clearly distinct.

These spectra were analysed by PCA (Model 1) in both near and far UV; see Table 5. Figure 3 illustrates how PCA can clarify the structure within the data set; similar spectra are represented by clusters of points. It can be seen that samples 1 and 3 form overlapping clusters, whereas sample 100 is clearly distinct. Throughout the duration of the study, there is no change in the spectra apparent from either visual inspection or from the PCA analysis. The data set was further analysed using SIMCA (Models 2-4; Table 5) where separate models were built for each class. The distance to model statistics (DModXPS+) were, with only one exception, below the critical distance for both near and far UV data. This indicates that there were no outliers from the model, and no trend was

apparent with time (not shown). This shows that the samples underwent no significant structural change during the study, when stored under the recommended conditions.

4.5. Re-measurement of participant samples

Where samples were successfully returned by the participants, they were re-measured and the data compared to the reference dataset by SIMCA. Unfortunately, some materials were lost in shipping, which meant that a complete picture of the effects of shipping and storage could not be obtained. An example of the spectra obtained from the returned samples is shown in Figure 4(a), where it is apparent that one sample differs from the others (Participant 2, Sample 3). The re-measurement data was compared to the baseline data set using SIMCA. In the far UV data (Figure 4(b)), virtually all observations were consistent with the baseline models (that is, their distance from the model was less than the critical distance). The exception is, unsurprisingly, Participant 2, Sample 3, which shows a statistically significant difference from the baseline data.

In the near UV data (Figure 4(c)), more changes are apparent. Sample 3 is again a strong outlier for participant 2, and (to a lesser extent) for participants 1 and 3. Sample 1 is also above the critical distance for participant 2; whereas sample 100 is an outlier for all the participants. In summary, statistically significant changes have occurred in all of participant 2’s samples, and in all of sample 100.

4.6. Participant data: Far UV

The participants’ far UV data, as reported, are shown in Figure 5(a) for sample 1. Similar results were obtained with the other samples. The immediate impression is that the spectra clearly show a great deal of variability. Most strikingly, three spectra are of much greater intensity than the others.

Inspection of the experimental questionnaires returned by the participants provides a partial explanation. These three participants used cells with longer path lengths than those specified in the protocol, even though the returns from the initial questionnaire indicated that they had the correct cells. CD and absorbance both follow the Beer-Lambert law:

of 27


lcAlcA

⋅⋅=⋅⋅Δ=Δ

εε

(ΔA: CD; Δε: CD coefficient; c: concentration; l: path length; A: absorbance; ε: extinction coefficient). Therefore, an incorrect path length leads to a proportionate error in both CD and absorbance. This explains why these three spectra are so different from the others – the stated path lengths are a factor of 10 or 5 larger than the nominal path length (see legend to Figure 5). This error in path length is readily corrected for by scaling the data, as shown in Figure 5(b). However, this reveals a more serious problem. The spectra for participants 1 and 2 are so noisy below 200 nm that they become effectively meaningless. This is a consequence of the way that these CD instruments work; meaningful CD data can only be obtained over a certain range of absorbance values [5, 10, 11]. The absorbance of the sample (protein plus buffer) in cells of 1 mm is so high that insufficient light reaches the detector.

The data from participant 9, in contrast, does not suffer from this problem, and the quality of the data obtained is unaffected. However, it is apparent that there is still a discrepancy in the intensity of the CD spectrum, although this has been reduced to approximately twofold. What could be the explanation of this discrepancy? Examination of Figure 1 reveals that this instrument is apparently calibrated to read too low by ~15% (relative to the reference value). The most likely explanation, therefore, is that the path length of the cell differs from the value stated by a factor of approximately twofold.

The other spectra are much more similar to each other, although a high degree of variability is still apparent. Such plots where many spectra are superimposed, however, rapidly become difficult to interpret. PCA and SIMCA can be useful in identifying and characterising outlying spectra. SIMCA can also be used to objectively score spectra on how well they fit a reference model. Such an analysis is shown in Figure 6, based on the path-length corrected data.

As would be expected from the previous discussion, the analysis shows that the spectra from participants 1, 2 and 9 are clearly distinct from the reference data set, even after path length correction. However, it is also obvious that most of the other participants’ data also shows significant differences from the model.

Inspection of the spectra and of the contribution plots (not shown) indicated a number of ways in which the spectra differed.

• noise or “spikes” in the spectra, particularly at short wavelengths or near zero intensity;

• overall spectral intensity;

• wavelength shifts, which are particularly apparent in the contribution plots (not shown) where zero crossings occur in the spectra

Only participant 6 consistently produced spectra that were below the critical distance; participants 7 and 8 managed this in some cases; the other spectra were above the threshold. Examination of scores plots showed that the data from the different participants did not form a cluster.

Although the data used in this analysis was corrected for the reported path length, in most cases the participants reported a nominal path length, rather than one they had measured themselves. It is likely that many of the cells differed appreciably from the nominal path length [10, 11, 13]. This factor alone could explain many of the discrepancies noted above. We have also noted previously that calibration varied widely between the participants. We therefore investigated whether correcting for these calibration differences improved the fit of the participants’ data to the reference model. This correction gave improved results for some participants, but worse results for many (not shown).

4.7. Participant data: Near UV

The participants’ data for sample 100 are shown in Figure 7; similar results were obtained with the other samples. Although the gross errors seen in the far UV data are not apparent, there are a number of problems apparent in the data. These include:

• Participants 5 and 9 show a lack of detail in the spectra; this may be due to excessive smoothing.

• Significant intensity differences are apparent, particularly at the broad peak around 260 nm and the two minima at approximately 282 and 289 nm. These differences are not consistent across different wavelengths.

of 27


• There is a major shift in the spectrum from participant 9, apparently increasing towards longer wavelengths.

• Wavelength shifts are also apparent in a number of the spectra.

The differences between the spectra can be further illuminated through a SIMCA analysis, and this is shown in Figure 8, where the distance to model statistics are plotted for the participants’ spectra. It is again immediately apparent that the majority of the spectra exceed the critical distance, and are therefore statistically different from the reference data set. This is consistent with the conclusions from the visual inspection of the data. A factor which is picked out particularly in the SIMCA analysis is the variability in position of the zero-crossings in the spectra, which is principally due to wavelength calibration differences between the instruments.

Since only one participant reported a non-nominal path length, the effect of path length correction is minimal. The absence of gross differences in the intensities of the spectra is likely to be due to the typically smaller deviations of 10 mm cells used from their nominal values [10, 11, 13]. Calibration correction gave somewhat better results than for the far UV data (not shown) but there were still significant differences, and in some cases the results were worse.

5. Discussion

5.1. Validity of Reference Data set

The spectra returned by the study participants show a high degree of variability and exhibit a number of measurement artefacts (Figure 5 and Figure 7; discussed in more detail below). They are almost all significantly different from the reference data set (Figure 6 and Figure 8), and this difference cannot be explained by changes in the samples (Figure 4). Some of the changes may be due to differences in instrument calibration (Figure 1). However, since there is no absolute traceability in CD measurement, it could be that the differences between the participant and the reference measurements are due to errors in the reference dataset itself. Since the aim of the study was to investigate measurement comparability, it is important to eliminate the possibility that the participant measurements form a cluster that excludes the

reference data. This would imply that there was comparability between laboratories other than the reference laboratory. Examination of the PCA scores plots (not shown) of the class models (M2-M4; Table 5) for the far and near UV data shows that the participant data does not form a cluster distinct from the reference data; we therefore conclude that the observed differences arise from a lack of comparability in the data between laboratories.

5.2. Far UV Data

The most significant errors in the far UV data were due to the selection of cells of the wrong nominal path length (1 mm rather than 0.1 mm). In principle, this effect could be corrected for simply by scaling the data, but in practice the overall absorbance of the sample becomes too high for some instruments, and the data recorded are not meaningful (participants 1 and 2).

The next most significant effect relates to the discrepancy between the nominal path length (as stated by the manufacturer) and the actual path length of the cell. This discrepancy can become quite significant for cells with short path lengths. For this reason, it is good practice to measure the actual path length of the cells being used [10, 11]. This effect can also be minimised by choosing the longest suitable path length for the measurement.

Another factor, which is likely to be contributing to the variability in spectral intensities in many cases, is the variation in instrument calibration. The spectra of ACS reported by the participants were used in an attempt to correct the spectra so as to remove this effect. This was not successful, and there are two possible reasons for this. Firstly, the calibrant and far UV measurements were made using different cells. Therefore the errors in the path lengths of both cells propagate into the calibration corrected data. Secondly, intensity calibration of an instrument at one wavelength may not be sufficient to remove all scale differences, as instrument sensitivity may vary with wavelength [7].

The spectra also differ in terms of their noise levels. Noise is a function of many aspects of instrument performance, but the most significant factor is likely to be the light throughput of the instrument. The mirrors in CD instruments are particularly vulnerable to attack by ozone, and

of 27


their reflectivity tends to diminish with time; various other factors can also contribute to a decline in performance.

A final, and very significant, difference between the spectra is that there are shifts of the wavelength scale between the different instruments. These shifts can have a significant effect on the interpretation of the resulting spectra, such as secondary structure predictions [4]. Indeed, we found dramatic differences between secondary structure prediction results for the participant spectra (not shown).

Therefore, in at least five respects, there are significant differences between the spectra returned by the participants and the reference data set. However, it is worth noting that most of these issues are reasonably straightforward to rectify, and can be eliminated by the following steps:

1. Choice of the correct path length cell for the measurement

2. Measurement of the actual path length of cells used (particularly for cells of 1 mm or less)

3. Regular, careful calibration of the intensity scale of the instrument

4. Regular, careful calibration of the wavelength scale of the instrument

5. Regular instrument performance checks, corrective and preventative maintenance

These recommendations are largely in line with previous guidance on improving CD measurement [10, 11].

5.3. Near UV Data

The variation observed in the near UV region is less dramatic than in the far UV, but no less significant. The intensity variations that are observed are difficult to attribute directly, but are likely to be a combination of small differences in cell path lengths and calibration. Applying a calibration correction gave better results than for the far UV region. This is likely to be because the calibrants were measured in the same 10 mm cell, and so the path length errors cancel and are removed. However, this calibration corrected data was not used for analysis, as there was not an improvement for all participants. One reason for this is that the ACS calibration is a single point measurement

and there may be wavelength dependent trends in instrument sensitivity [7].

In addition, a number of other types of variability were observed in the spectra, which were also seen in the far UV spectra, including wavelength shifts and differences in noise levels. The same steps laid out in the previous section would help to eliminate or reduce these effects.

In addition, other discrepancies between spectra are observed. Excessive smoothing is apparent in some spectra. The spectral features observed in near UV spectra have narrower peak widths than in the far UV, so it is important that this should be taken into account if smoothing is to be used. One of the sets of spectra (participant 9, sample 100) shows an apparent baseline shift increasing towards longer wavelengths. This was not seen in spectra from other samples with this participant, and the cause is not clear.

5.4. Applicability of Multivariate analysis methods

Large spectroscopic data sets can quickly become difficult to visualise – see, for example, Figure 5. Furthermore, it is difficult to objectively compare spectra and draw robust conclusions about whether or not they represent the same underlying protein structure. Both of these problems can be addressed by using multivariate statistical approaches. Such approaches are far from being new to the analysis of spectral data; for example, similar methods may be used to extract information about protein secondary structure from CD spectra [3], FTIR spectra [15-17], or ROA spectra [18]. Similar ‘chemometric’ approaches are used in analysis of complex mixtures in FTIR spectroscopy [19, 20]. Our approach is simply to use PCA and SIMCA for the objective comparison and classification of CD spectra, without any attempt to interpret them. This approach could be particularly useful for the analysis of near UV spectra, which are difficult to interpret but may contain detailed structural information.

Any data analysis method is only as good as the data that is fed into it. Here we have used multivariate methods to investigate the measurement comparability of data. However, where the questions to be asked are of a more biological nature, the comparability of the data is usually assumed. In many cases, this

of 27


assumption may be justified, because often all the data will be collected on one instrument, by one operator, within a short space of time. However, in some contexts, such as when spectroscopic techniques are used for industrial purposes, comparability of data becomes a critical issue – if the data are not comparable, then any analysis method will give potentially misleading results.

The obvious solution is to attempt to standardise all instruments to record identical spectra. However, given the complexity of CD instrumentation, this is unlikely to be achieved. A more practical approach is to characterise the behaviour of instruments using reference materials or standards. This data could then be used to “pre-process” spectra before comparison by approaches similar to those described here. A similar approach was recently described by Miles et al. [7].

5.5. Conclusions

These results showed that international comparability in circular dichroism was poor, with a number of significant sources of variation in measurements. However, it is also clear that many of the causes of the variation are readily preventable by the application of good practice. This mirrors the situation we observed in a similar study within the UK [21]. Good practice guidance can be found in the following references: [10, 11] for example. A number of more subtle differences between CD instruments are known to occur, such as wavelength dependent changes in intensity [7], but are masked in this study by other sources of variation. There is, therefore, a clear role for a follow-up study, to see if measurement comparability can be improved when the lessons learnt from this study have been applied. The results of just such a follow-up study are presented in the accompanying paper [8].

A significant difference between many of the spectra, which is not always readily apparent to the eye, but is brought out by the multivariate analysis methods used, is the existence of wavelength shifts between the spectra. Quite small wavelength differences can have a significant effect on the interpretation of spectra [13]. Since CD instruments typically record both CD and absorbance, wavelength standards such as holmium filters developed for the calibration of absorbance instruments are

equally useful for CD and could readily address this problem.

One of the difficulties encountered in this study has been the lack of any absolute reference for CD. This has made the interpretation of some of the results more difficult. One possible solution to this is the development of some form of measurement traceability for CD. This is still, we feel, a pressing requirement.

There is also a clear need for improved reference materials which provide calibration across the spectral region of interest for protein measurements. Such a material was recently described [22] and should be commercially available in the near future.

The uncertainties involved in a CD measurement are rather complex, involving components relating to calibration, path length, sample concentration and instrumentation. In this study we treated the uncertainties in a statistical manner. However, we have recently developed a more detailed model for the uncertainties of the measurement [23], and applied it to the results of our follow-up study [8].

Such improvements to the metrological support of CD measurements will greatly increase the reliability of the technique, and increase confidence among academic and industrial users, and in particular regulatory authorities such as the FDA and the EMA.

In this report we have also shown how multivariate statistical methods can be used to investigate the measurement comparability of spectroscopic data. In a similar way, these techniques could also be used in a industrial scenario; for example, where production batches of a biopharmaceutical are compared to a reference batch (for another approach to this problem see [24]). The robust and objective character of these methods is ideal for such applications. However, as we have shown, this application would require greatly improved comparability of CD data. To achieve this will require better measurement practice, new reference materials, and pre-processing of spectra to correct for measurement artefacts.

Acknowledgements

We would like to acknowledge the contribution of the study participants, without whose efforts this study would not have been possible.

of 27


Thanks are also due to Kerry Emslie of NMIA for supporting the study by nominating an expert laboratory, and to an anonymous expert laboratory for their contributions to the study.

We would also like to thank the members of the CCQM Bioanalysis Working Group for supporting the study and for their input and suggestions.

This work was funded by the Measurements for Biotechnology Programme of the UK government’s Department of Trade and Industry (now the Department for Business, Innovation and Skills).

6. References 1. Fasman, G.D., ed. Circular Dichroism and

the Conformational Analysis of Biomolecules. 1996, Plenum Press: New York. 748.

2. Rodger, A., and Norden, B., Circular dichroism and linear dichroism. 1997: Oxford University Press.

3. Venyaminov, S.Y. and J.T. Yang, Determination of Protein Secondary Structure, in Circular Dichroism and the Conformational Analysis of Biomolecules, G.D. Fasman, Editor. 1996, Plenum Press: New York. p. 69-107.

4. Miles, A.J., L. Whitmore, and B.A. Wallace, Spectral magnitude effects on the analyses of secondary structure from circular dichroism spectroscopic data. Protein Science, 2005. 14(2): p. 368-374.

5. Johnson, W.C., Jr., Circular Dichroism Instrumentation, in Circular Dichroism and the Conformational Analysis of Biomolecules, G.D. Fasman, Editor. 1996, Plenum Press: New York. p. 635-652.

6. Takakuwa, T., Konno, T., Meguro, H., A New Standard Substance for Calibration of Circular Dichroism: Ammonium d-10-Camphorsulfonate. Analytical Sciences, 1985. 1: p. 215-218.

7. Miles, A.J., et al., Calibration and Standardisation of Synchrotron Radiation Circular Dichroism and Conventional Circular Dichroism Spectrophotometers. Spectroscopy, 2003. 17: p. 653–661.

8. Ravi, J., et al., International comparability in spectroscopic measurements of protein structure by circular dichroism: CCQM P59.1. Metrologia, In Preparation.

9. Vives, O.C., et al., Val-CiD Appendix B: The use of chemical calibrants in circular dichroism spectrometers. 2004, National Physical Laboratory.

10. Kelly, S.M., T.J. Jess, and N.C. Price, How to study proteins by circular dichroism. Biochimica et Biophysica Acta, 2005. 1751: p. 119 – 139.

11. Jones, C., et al., Val-CiD Best Practice Guide: CD spectroscopy for the quality control of biopharmaceuticals. 2004, National Physical Laboratory.

12. Yang, J.T., Remembrance of Things Past: A Career in Chiroptical Research, in Circular Dichroism and the Conformational Analysis of Biomolecules, G.D. Fasman, Editor. 1996, Plenum Press: New York. p. 1-24.

13. Miles, A.J., Wien, F., Lees, J. G., Wallace, B. A., Calibration and standardisation of synchrotron radiation and conventional circular dichroism spectrometers. Part 2: Factors affecting magnitude and wavelength. Spectroscopy, 2005. 19: p. 43-51.

14. User's Guide to SIMCA-P, SIMCA-P+. Version 11 ed. 2005, Umeå: Umetrics AB. 158.

15. Surewicz, W.K., and Mantsch, H. H. , Infrared absorption methods for examining protein structure, in Spectroscopic methods for determining protein structure in solution, H.A. Havel, Editor. 1996, John Wiley & sons. p. 135-162.

16. Dousseau, F. and M. Pézolet, Determination of the secondary structure content of proteins in aqueous solutions from their amide I and amide II infrared bands. Comparison between classical and partial least-squares methods. Biochemistry, 1990. 29(37): p. 8771-8779.

17. Lee, D.C., et al., Determination of protein secondary structure using factor analysis of infrared spectra. Biochemistry, 1990. 29(39): p. 9185-9193.

18. Zhu, F., et al., Raman optical activity of proteins, carbohydrates and glycoproteins. Chirality, 2006. 18(2): p. 103-115.

19. Martens, H. and T. Næs, Multivariate Calibration. 1992: John Wiley & Sons.

20. Malinowski, E., Factor Analysis in Chemistry. 2002: Wiley.

21. Schiffmann, D.A., et al., Val-CiD Appendix A: CD spectroscopy: an inter-laboratory study. 2004, National Physical Laboratory.

of 27


22. Angeliki Damianoglou, E.J.C., Matthew R. Hicks, Suzanne E. Howson, Alex E. Knight, Jascindra Ravi, Peter Scott, Alison Rodger, A new reference material for UV-visible circular dichroism spectroscopy. Chirality, 2008. 20(9): p. 1029-1038.

23. Knight, A.E., et al., Uncertainty in measurements of protein structure by circular dichroism. In preparation.

24. Bierau, H., et al., Higher-Order Structure Comparison of Proteins Derived from Different Clones or Processes. BioProcess International, 2008: p. 52-59.

7. Figures

See following pages.

of 27


a)200 250 300 350 400

-300

-200

-100

0

100

200

300

Wavelength (nm)

CD

(mde

g)

123456789

b)288 289 290 291 292 293

150

160

170

180

190

200

210

220

230

Peak λ (nm)

Pea

k H

eigh

t (m

deg)

123-13-24-14-256789Nominal ValueNominal Limits

Figure 1 Calibration Status of Participants’ instruments

a) ACS spectra as measured by the study participants. Numbers are participant numbers as given in Table 1. Samples were returned to NPL and re-measured; no significant changes were detected (data not shown).

b) Summary of peak wavelengths and intensities derived from the above spectra by curve fitting. Error bars derived from the fits are too small to display at this scale. Participants 3 and 4 used more than one instrument; instrument numbers are appended to the participant numbers in the key. Note that few spectra fall within the “nominal range” for both wavelength and intensity.

of 27


a)

180 190 200 210 220 230 240 250 260-6

-4

-2

0

2

4

6

8

Wavelength (nm)

CD

(mde

g)

001003100

b)

240 260 280 300 320 340 360-60

-50

-40

-30

-20

-10

0

10

20

Wavelength (nm)

CD

(mde

g)

001003100

Figure 2 Reference data set: spectra

a) Far UV spectra; b) Near UV spectra. These spectra were acquired at NPL throughout the duration of the study, and were used to build a model for each sample, and to test for sample stability. Note that samples 001 and 003 give similar spectra, as anticipated. Far UV spectra were collected in 0.1 mm cells and near UV in 10 mm cells (other parameters are given in Table 4).

of 27


a)

- 4 0

- 3 0

- 2 0

- 1 0

0

1 0

2 0

3 0

4 0

7 0 8 0 9 0 1 0 0 1 1 0

t[2]

t [ 1 ]

131 0 0

b)

- 3 0 0

- 2 0 0

- 1 0 0

0

1 0 0

- 5 0 0 - 4 0 0 - 3 0 0 - 2 0 0 - 1 0 0 0 1 0 0 2 0 0

t[2]

t [ 1 ]

131 0 0

Figure 3 Reference data set: PCA analysis

Scores plots from PCA analysis of data shown in Figure 2: a) Far UV; b) Near UV. Each point represents a spectrum (observation). The x- and y-axes represent the scores for the first and second principal components, respectively; similar spectra will be plotted close together. The ellipse indicates the Hotelling T2 statistic, which indicates the boundary outside which points are considered outliers. Note that samples 1 and 3 form overlapping clusters in both near and far UV regions, but are clearly distinguished from sample 100.

of 27


a)

180 190 200 210 220 230 240 250 260-6

-4

-2

0

2

4

6

8

Wavelength (nm)

CD

(mde

g) 123467

180 190 200 210 220 230 240 250 260-6

-4

-2

0

2

4

6

8

Wavelength (nm)

CD

(mde

g) 123467

b)

1 2 3 4 5 6 7 8 90

1

2

3

4

5

6

D c r it = 1 . 2 6 2

P a r t ic ip a n t

DM

odXP

S+[

1](N

orm

)

0 0 10 0 31 0 0D c r i t

1 2 3 4 5 6 7 8 90

1

2

3

4

5

6

D c r it = 1 . 2 6 2

P a r t ic ip a n t

DM

odXP

S+[

1](N

orm

)

0 0 10 0 31 0 0D c r i t

of 27


c)

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

3 0

3 5

4 0

4 5

D c r it = 1 . 2 6 5

P a r t ic ip a n t

DM

odXP

S+[

2](N

orm

)

0 0 10 0 31 0 0D c r i t

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

3 0

3 5

4 0

4 5

D c r it = 1 . 2 6 5

P a r t ic ip a n t

DM

odXP

S+[

2](N

orm

)

0 0 10 0 31 0 0D c r i t

Figure 4 Typical sample re-measurement data

a) Plot of far UV data for Sample 3. Note the missing spectra from participants 5,8 and 9, due to sample losses during shipping. The spectra appear to overlay, except for that from participant 2, which appears to have undergone a change during shipping, storage or measurement.

b) Plot of distance to model (DModXPS+) for far UV spectra, when compared to the relevant baseline models using SIMCA. Note that participant 2, sample 3 is a strong outlier (mean DModXPS+ is 5.6). Participant 4, Sample 100 has only one observation that is a weak outlier. All other observations are within the critical distance and are therefore not significantly different from the model data set.

c) Corresponding plot for near UV data. Note that the distances to model are higher, with a greater proportion of the observations falling outside the model. As before, Participant 2 sample 3 is a strong outlier, with a mean DModXPS+ of 41.2. The same sample is also a relatively weak outlier for participants 1 and 3. Sample 1 is also an outlier for Participant 2 (7.6). Sample 100 is an outlier for all the participants.

of 27


a)

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 6 0

- 4 0

- 2 0

0

2 0

4 0

6 0

8 0

W a v e l e n g t h ( n m )

CD

(mde

g)

123456789

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 6 0

- 4 0

- 2 0

0

2 0

4 0

6 0

8 0

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 6 0

- 4 0

- 2 0

0

2 0

4 0

6 0

8 0


CD

(mde

g)

123456789

b)

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 1 0

- 8

- 6

- 4

- 2

0

2

4

6

8

1 0


CD

(mde

g)

123456789

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 1 0

- 8

- 6

- 4

- 2

0

2

4

6

8

1 0

1 8 0 1 9 0 2 0 0 2 1 0 2 2 0 2 3 0 2 4 0 2 5 0 2 6 0- 1 0

- 8

- 6

- 4

- 2

0

2

4

6

8

1 0


CD

(mde

g)

123456789

Figure 5 Participant data: Far UV

a) Participants’ far UV spectra of sample 1, as reported. The majority of spectra overlay closely; however, some show clear discrepancies. Participants 1,2 and 9 show spectra that are clearly of a radically greater amplitude (roughly ten-fold) than the others. Participants 1 and 2 also exhibit very high noise below 200 nm. Participant 5 shows an apparent wavelength shift. Similar results were observed for the other samples (not shown).

b) The same data, corrected for participants’ reported path lengths. Path lengths were reported as 1.0 mm for participants 1 and 2; 0.5 mm for participant 9; and 0.105 for participant 6 (nominal path length was 0.1 mm, see Table 4). Scale agreement is now better for participants 1 and 2, although the high noise level is still apparent. Participant 9’s spectrum is still roughly a factor of two too intense.

of 27


1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

D c r it = 1 . 2 6 2

P a r t ic ip a n t

DM

odXP

S+[

1](N

orm

)

0 0 10 0 31 0 0D c r i t

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

D c r it = 1 . 2 6 2

P a r t ic ip a n t

DM

odXP

S+[

1](N

orm

)

0 0 1

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

D c r it = 1 . 2 6 2

P a r t ic ip a n t

DM

odXP

S+[

1](N

orm

)

0 0 10 0 31 0 0D c r i t

Figure 6 Distance to Model: Far UV data

Distance to model plots from SIMCA analysis of the path-length corrected participant data. Uncorrected data showed distances for participants 1,2 and 9 greater than 150. It is immediately apparent that the spectra for participants 1-5 and 9 are all above the critical distance, with distances in most cases much higher than the corresponding samples re-measured at NPL (see Figure 4). Participants 7 and 8 have some data above and below the critical distance, whereas participant 6 has all but one spectrum below the critical distance.

of 27


2 4 0 2 6 0 2 8 0 3 0 0 3 2 0 3 4 0 3 6 0- 3 5

- 3 0

- 2 5

- 2 0

- 1 5

- 1 0

- 5

0

5

1 0

1 5


CD

(mde

g)

123456789

2 4 0 2 6 0 2 8 0 3 0 0 3 2 0 3 4 0 3 6 0- 3 5

- 3 0

- 2 5

- 2 0

- 1 5

- 1 0

- 5

0

5

1 0

1 5

2 4 0 2 6 0 2 8 0 3 0 0 3 2 0 3 4 0 3 6 0- 3 5

- 3 0

- 2 5

- 2 0

- 1 5

- 1 0

- 5

0

5

1 0

1 5


CD

(mde

g)

123456789

Figure 7 Participant data: Near UV

The participants’ CD spectra for sample 100, as reported. All path lengths were reported as nominal, with the exception of Participant 6 (10.013 nm). The effect of the path length correction is therefore minimal. However, a significant degree of variability is apparent. Participants 5 and 9 show an excessive degree of smoothing and intensity differences; Participant 3 also shows a different intensity; and Participant 2 shows a different intensity and a wavelength shift. Similar results were obtained with the other samples.

of 27


1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

3 0

3 5

4 0

4 5

5 0

D c r it = 1 . 2 6 5

P a r t ic ip a n t

DM

odXP

S+[2

](Nor

m)

0 0 10 0 31 0 0D c r i t

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

3 0

3 5

4 0

4 5

5 0

D c r it = 1 . 2 6 5

P a r t ic ip a n t

DM

odXP

S+[2

](Nor

m)

0 0 10 0 31 0 0

1 2 3 4 5 6 7 8 90

5

1 0

1 5

2 0

2 5

3 0

3 5

4 0

4 5

5 0

D c r it = 1 . 2 6 5

P a r t ic ip a n t

DM

odXP

S+[2

](Nor

m)

0 0 10 0 31 0 0D c r i t

Figure 8 Distance to Model: Near UV data

Distance to model plots from SIMCA analysis of the path-length corrected data. Most of the spectra fall above the critical distance, indicating that they are significantly different from the model data set. In most, but not all, cases, the distance to model is greater than that for the re-measured samples, indicating that a change in the samples is not responsible for the differences between the participant measurements and the reference data set. The exceptions are participant 2 sample 3, and participant 1 sample 100. We have already seen that these samples have undergone significant changes, and it may be that the changes occurred after the samples were measured by the participant. The path length correction results in a very slight reduction in the distance for participant 6, who quoted a measured path length of 10.013 mm.

of 27


8. Tables Table 1 Participants in the study.

“Status” indicates whether the laboratory in question is a National Measurement Institute or a nominated expert laboratory. “Role” indicates whether the laboratory was an active participant, a sponsor of an expert laboratory, or an organiser of the study.

Country Organisation Status Role

National Physical Laboratory (NPL) NMI Organiser UK

National Institute for Biological Standards and Control Expert lab Participant

National Institute for Standards and Technology NMI Organiser and participant

Human Genome Sciences, Inc. Expert lab Participant USA

Olis, Inc. Expert lab Participant

National Metrological Institute Japan (NMIJ) NMI Participant Japan

Anonymous laboratory Expert Lab Participant

China National Institute of Metrology NMI Participant

National Measurement Institute NMI Sponsor Australia

Macquarie University Expert Lab Participant

Laboratoire National d'Essais NMI Sponsor France

Département de Réponse et Dynamique Cellulaires Expert Lab Participant

of 27


Table 2 Information about participants.

Most participants had one instrument, others two. Three instrument manufacturers were represented (shown as A, B, and C). Some participants returned additional data, and where possible, this was incorporated into the analysis. Participant numbers are assigned on a random basis.

ParticipantNumber

Number ofInstruments

Manufacturer(s)

1 1 A

2 1 A

3 2 A, B

4 2 A

5 1 C

6 1 A

7 1 A

8 1 A

9 1 C

of 27


Table 3 Samples distributed to the study participants.

The ACS sample was made up in ultra-pure water. The protein solutions were made up in the phosphate buffer. Samples of the water and buffer were also distributed to act as blanks.

Sample ID Description

Water Ultra-pure water

ACS ACS 0.6 mg·ml-1

Buffer 30 mM Sodium Phosphate Buffer, pH 7.2

001 Lysozyme, 1.00 mg·ml-1

003 Lysozyme, 0.98 mg·ml-1; Cytochrome C, 0.02 mg·ml-1

100 Cytochrome C, 1.00 mg·ml-1

of 27


Table 4 Instrument Parameters Used.

The parameters given are those used on manufacturer A instruments. Similar parameters were used for the C instruments, with the following caveats: the bandwidth was set to 1 nm; the data pitch was 1 nm on one instrument (Participant 9) and 1.02 nm on another (Participant 5); the scan speed was variable, 0.5 to 10 s per point; a 13 point digital filter was applied to the data. For the B instrument, the bandwidth was 2 nm; data pitch 0.5 nm; scan speed 60 nm·min-1; sensitivity ±30 mdeg for the protein samples and ±300 mdeg for the ACS; the response time is set automatically on this instrument.

Sample

Proteins

Parameter Far UV Near UV ACS

Cell path length (mm) 0.10 10 10

Wavelength range (nm) 180-260 240-360 200-400

Accumulations 6 4 1

Response time (s) 1 1 1

Bandwidth (nm) 1 1 1

Sensitivity Standard Standard Standard

Data pitch (nm) 0.1 0.1 0.1

Scan speed (nm·min-1) 50 50 50

of 27


Table 5 Models used in PCA and SIMCA analysis.

Four models were built for each data set. Each observation is a spectrum; each variable a wavelength point in a spectrum. The R2X statistic [14] is the fraction of the sum of squares of all the data explained by the current component; components that were used in further analysis are highlighted in bold type.

R2X (component) Dataset Model Samples Observations Variables

1st 2nd 3rd

M1 1,3,100 63 801 96.46% 3.36% 0.04%

M2 1 21 801 99.82% 0.04% 0.02%

M3 3 21 801 99.83% 0.04% 0.02% Far UV

M4 100 21 801 99.78% 0.05% 0.03%

M1 1,3,100 63 1201 78.08% 21.83% 0.01%

M2 1 21 1201 99.94% 0.02% 0.01%

M3 3 21 1201 99.94% 0.01% 0.00% Near UV

M4 100 21 1201 99.81% 0.07% 0.02%

of 27


of 27

international comparability in circular dichroism

Documents