IR SPECTROSCOPE CELL CULTURE ANALYSIS

Abstract

A method of detection of a status of cells within a cell culture is described. The method is based on comparing an IR spectrum of a test sample obtained from the cell culture with an IR spectrum of a control sample or samples or with an IR spectrum of a sample or samples obtained at an earlier time point in the cell culture and correlating differences between the spectra with the status of the cells. The status may be classified into healthy or unhealthy. The test sample may or may not contain cells. The test sample may contain cell fragments, cell components or cell media, or may be cell supernatant. The comparison may be performed using a predictive model based on pattern recognition algorithms, such as support vector machines SVM, linear discriminant analysis LDA, principal component discriminant function analysis PC-DFA, neural networks, or random forests). The analysis results may be used to monitor and/or control the cell culturing process.

Claims

1-19. (canceled)

20. A method for detecting a status of cells within a cell culture, the method comprising: i) providing an IR spectrum from a test sample obtained from a cell culture; and ii) comparing the IR spectrum with an IR spectrum of a control sample or samples, or a sample or samples obtained at an earlier time point of the cell culture to the test sample, in order to detect any difference between the test and control or earlier samples, which can be correlated with the status of cells within the test sample obtained from the cell culture.

21. The method of claim 20, wherein the sample or samples is obtained from the cell culture manually, by a semi-automated, or fully automated system associated with the cell culture.

22. The method of claim 20, the method being implemented and/or controlled by a computer system with integrated software in order to conduct the comparison step and detect any differences between the test sample and control or earlier samples obtained from the cell culture.

23. The method of claim 20, wherein the status of cells within the cell culture relates to the health of the cells.

24. The method of claim 23, wherein the status of the cells is detected as being healthy, or unhealthy.

25. The method of claim 24, wherein the unhealthy status of the cells, is due to the cells being infected by a virus or bacteria, or the cells have developed an abnormality.

26. The method of claim 25, wherein the abnormality has occurred due to cell culture conditions, such as a change in pH, oxygen levels, toxic component build up and/or nutrient levels.

27. The method of claim 20, wherein the cell culture comprises cells which produce a product, such as a recombinant protein or the like; or, the cells are the product

28. The method of claim 27, wherein the cells are the product and are intended for in vivo and/or in vitro application.

29. The method of claim 20, wherein the cells are bacterial or eukaryotic, such as mammalian or other eukaryotic cell types.

30. The method of claim 20, wherein the IR spectrum is an FTIR spectrum, or a portion or portions thereof, such as including transmission, transflection, and/or attenuated total reflection (ATR).

31. The method of claim 20, wherein the test sample or control sample is a wet or dry, which has been obtained directly or indirectly from the cell culture.

32. The method of claim 20, wherein the step of comparing the analytical results is performed using a predictive model, optionally which has been developed by a training database of pre-correlated analyses.

33. A computer installed with computer software configured to operate the computer to perform a predictive cell status analysis method based on a spectroscopic IR spectra or processed IR spectra/data of a cell culture sample.

34. The computer of claim 33, which incorporates a predictive model derived from one or more pattern recognition algorithms applied to a plurality of pre-correlated IR spectra or processed IR spectra/data, the computer optionally further installed with a database, for correlating results with known cell status samples and/or to store training set data.

35. A computer-implemented method of correlating the results of an IR spectroscopic analysis method with cell status within a cell culture, the method comprising: collecting data from said spectroscopic analysis; and employing a predictive model, suitably based on pattern recognition algorithms conducted upon pre-correlated spectroscopic analyses (optionally in conjunction with a database, as defined herein) to correlate said data with cell status of cells within the cell culture.

36. An integrated cell culture and IR analysis system comprising: a) cell culture apparatus for growing cells in culture; b) a sample handling system for obtaining one or more samples from a cell culture within the cell culture apparatus, the sample handling system comprising a sampler for obtaining a sample from the cell culture and transporting the sample to an IR spectrometer in order that an IR spectrum of the sample or samples may be obtained; and c) conducting a method to detect the status of cells within the cell culture.

37. The method according to claim 36, wherein the integrated system further comprises the ability to, alert a user, alter and/or halt the cell culturing process in response to any result(s) obtained from the method.

Description

[0052] The present invention will now be further described by example, with reference to the following figures which show:

[0053] FIG. 1 shows the spectral analysis of dried CHO cells which are healthy, grown under nutrient deficient conditions or are virus infected;

[0054] FIG. 2 shows PCA processed analysis of dried CHO cells obtained from the spectral data displayed in FIG. 1;

[0055] FIG. 3a shows the PC1 loadings referring to the PCA scatterplot displayed in FIG. 2, and FIG. 3b shows the PC2 loadings, showing where spectral differences arise between healthy and virally infected cells;

[0056] FIG. 4 shows projected PCA scatterplot analysis of healthy, virally infected and nutrient depleted (′old) samples;

[0057] FIG. 5 shows results from a random forest classification for first a multi-class classifier of all cell classes (A), and binary classification based upon healthy and non-healthy (B);

[0058] FIG. 6 shows the spectral analysis of dried CHO cells that have been fixed during sample preparation;

[0059] FIG. 7 shows PCA processed analysis of dried, fixed, CHO cells obtained from the spectral data displayed in FIG. 6;

[0060] FIGS. 8a and 8b shows the spectral sources of variance in PC1 (A) and PC2 (B) across the spectrum between healthy and unhealthy cells, following fixation as an optional sample preparation step;

[0061] FIG. 9 shows projected PCA scatterplot analysis of healthy, virally infected and nutrient depleted cells following fixation;

[0062] FIG. 10 shows the random forest binary classification results of the cells in FIG. 6, following fixation as a sample preparation step;

[0063] FIG. 11 shows processed spectra from supernatant material from healthy, virally infected and nutrient depleted cells;

[0064] FIG. 12 shows PCA scatterplot analysis from supernatant material from healthy, virally infected and nutrient depleted cells;

[0065] FIGS. 13a and 13b shows variance arising from PC1 (a) and PC2 (b) following PCA showing clear differences in the amide regions;

[0066] FIG. 14 shows processed spectra in the fingerprint region of supernatant samples derived from healthy and two types of virally infected CHO cells;

[0067] FIG. 15 shows PCA scatterplot analysis of processed spectra in the fingerprint region of supernatant samples derived from healthy and two types of virally infected CHO cells;

MATERIALS, METHODS AND RESULTS

Sample Preparation

[0068] It is possible to obtain a number of samples from cell culture systems, in the form of cell suspensions, or pellets, supernatant, which may or may not contain cell fragments, or as media directly from the cell culture. Sampling points could be either directly from the active culture system, or in between passage stages in the seed train process. In this instance, concentrated cell suspensions were obtained during the passage process of producing cells, where cell monolayers are removed from the culture flask using trypsin, and concentrated using centrifugation. In this state with additional cell culture media, cells can be immediately seeded into flesh culture flasks for continuous growth, or stored at −80° C. for later analysis. For this analysis, cells were analysed following freezing.

[0069] Initially cells were thawed prior to preparation for spectroscopic analysis. Cells were analysed immediately following thawing, and following cell fixation. For immediate analysis, cells were thawed in a water bath at 37° C., gently inverted, and immediately centrifuged for 3 minutes at 1000 rpm to concentrate the cells into a pellet. The supernatant was then aspirated and contained for separate analysis. Three microlitres of the cell sample was then deposited on each of the three wells of the ClinSpec Dx Optical Sample Slides. The cell sample maintained sufficient viscosity for pipetting, as some residual solution remained. For cell fixation studies, cells were also thawed in a water bath at 37° C. and gently inverted, then the thawed solution was first mixed 1:1 vol with ethanol, gently mixed with a pipette, and subsequently centrifuged at the same parameter settings; the supernatant was then discarded and only the cell pellet analysed. For all cell samples, the slides were analysed immediately after deposition to investigate the samples in a wet state, and in full following 30 minutes drying. For supernatant deposition a larger sample volume of 6 microlitres was used so that the SIRE surface was sufficiently covered.

Sample Analysis

[0070] Prior to analysis, spectra are usually pre-processed in order to remove unwanted variance from the dataset, such as sample thickness, that can mask the true biological variance within the sample. There are numerous approaches to this process, and these have been extensively checked within this study. Generally, the presented spectra are processed by baseline correction and normalisation, across the full spectrum or cut to the fingerprint region.

[0071] Generally, spectral analysis can be split into three parts; (i) spectral observation, (ii) variance exploration and (iii) classification modelling. In overview, each spectrum is first observed by eye to identify any differences in the IR spectra of the samples and try to discern any differences between the treatments, in this case, cell health. As sometimes differences cannot be observed, multivariate techniques, such as principal component analysis (PCA) permits looking at the variance alone within the dataset, which can unearth subtle differences between cells. PCA works by reducing the dimensions of the spectral dataset into principal components (PCs) that account for the variance in the data. These can be visualised as scatterplots, where spectra are single points and clustering can infer similarity, and separation between points can suggest differences. PC loadings plots can then suggest where in the spectrum this variance is arising from. Finally, classification models can be used to statistically separate the data, and to see how well the data can be accurately predicted. This can provide a level of sensitivity/specificity to analyses, similar to that of a disease prediction.

[0072] Initial results from unprocessed spectra, displayed evidence of unwanted variance in the form of baseline differences that should be minimised, in order that spectra can be directly compared. The inventors observed that baseline correction and normalisation reduced these effects so spectral differences can be more clearly seen. In this instance, a rubberband baseline correction and a subsequent vector normalisation step was applied using the mathematical software R.

[0073] FIG. 1 shows the spectral differences between healthy cells and those which have been infected by a virus and/or subjected to nutrient limitation. These samples have been analysed immediately following thawing from the frozen sample and dried onto the surface of the IRE. This nutrient depletion was induced by simply allowing the cells to grow in media for longer than recommended, in this case up to 10 days, so that nutrient levels were low. Some clear differences in the water (OH) content of the cells (3500-3000 cm.sup.−1), and the carbohydrate region (1300-1000 cm.sup.−1) in the fingerprint region, can be seen between healthy and ‘unhealthy’ cells respectively. Thus, following processing of the spectral data in order to reduce unwanted variance, spectral differences between healthy and unhealthy cell cultures can be seen. These difference however can be subtle to the user, and not always visible by eye.

[0074] Principal component analysis (PCA) can further show variance, which serves to distinguish how the different cell culture conditions effect the spectra obtained, as the different data classes begin to cluster together. This process reduces a dataset into key sources of variance, known as principal components (PCs), which encompass spectral differences. Variance can be visualised by comparing these PCs as a scatterplot, where different PCs can be plotted comparatively. Separation on a scatterplot between classes infers differences, and clustering infers similarities. Often scatterplots are presented as PC1 versus PC2, as these two components account for the most amount of variance in the dataset (PCs have descending variance values as the PC number increases, with PC1 accounting for the greatest source of variance in the dataset). The variance can be further explored as a PC loadings plot, which correlates these differences to the original IR spectrum. For example, if two classes separate visually on the PC2 axis on a scatterplot, a PC2 loadings plot will subsequently show where that variance is arising from in the IR spectrum.

[0075] As shown in FIG. 2, PCA begins to differentiate the different cell statuses based on the spectral variance between them. In this example, PC1, accounting for 60% of variance in the dataset, begins to split virus infected cells, nutrient deficient from healthy and the other virally infected cells. PC2 on the other hand, begins to split healthy from virally infected cells. As the cell classes appear as distinct clusters in this PCA scatterplot, it is clear that there are biological difference that are being elucidated by spectral analysis.

[0076] The loadings for these PCs can show where this variance arises from in the spectrum. Negative in PC1 that differentiate the nutrient deficient and virally infected cells, appears to be associated with water (O—H) content and positive in PC1 shows more significant differences in the fingerprint region, associated with proteins and carbohydrates specifically (FIG. 3a). Positive in PC2 separates healthy cells from virally infected cells, which the loadings suggest is again a difference in the O—H stretching peak, associated with water content (see FIG. 3b). The larger amount of noise in this plot is representative of the high noise and low variance value of PC2. Whilst spectral differences were observable in FIG. 1, this analysis shows further detail and sources of these differences between healthy and abnormal cells.

[0077] Projecting these samples as a test into a PCA (see FIG. 4), shows that it is possible to correctly separate the cell types. Projected PCA, and PCA-DFA are commonly implemented analysis methods, that can be accessed in programs such as R or Matlab. Here we used the Biocluster Toolbox for Matlab. Highlighted labels are the test spectra that were projected onto the training data. Closeness represents similarity between the spectra. This process is similar to an envisaged commercial application of this technology, where unknown test samples will be compared against an existing database, with known signals of cell status, and predicted against it.

[0078] Next is the process of classification. There are a large variety of classification algorithms available, such as SVM, LDA, Random Forest, and Neural Networks, each of which have different benefits and constraints. Generally, the main requirement of these algorithms are a sufficient number of samples to generate meaningful results.

[0079] In the examples presented here, the current dataset is limited as the data has been obtained from one sample vial, split into 9 technical replicates, of four treatment types. However, preliminary attempts have been made here to look into sample classification, It is possible to conduct a multi-class classifier on data with multiple comparators, however due to limited data, this is not immediately possible for this study. This can however be simplified to just healthy versus unhealthy cells/cell conditions, which is a simpler approach. The initial results show 100% sensitivity and specificity which is extremely promised. However, the results have been obtained from a small dataset that needs to be expanded in order to generate statistically significant results (see FIG. 5).

[0080] Similar patterns are observed with our other analyses too, with fixed cells showing subtle differences across the spectrum (FIG. 6). Fixation does appear to alter the biochemical composition of cells, which is largely seen in the lipid regions of the cells. However, this change is consistent between samples, and fixes them in their pre-frozen state so that the treatment response can be compared. PCA shows begins to highlight spectral differences more clearly. FIG. 7 displays variance between healthy and unhealthy cells, particularly with virally infected cells appearing more negative in PC1. Looking at the loading of PC1 (FIG. 8a), this is associated with widespread differences across the fingerprint region of the spectrum compared to healthy cells. Nutrient deficient cells appear predominantly more positive in PC2 (FIG. 7), which the relative loadings plot suggests is associated with lipid changes compared to healthy and virally infected cell lines. Without being bound by theory, this may be due to the fixation process.

[0081] A full PCA projection is shown in FIG. 9 and interestingly good separation between the classes, particularly in PC1 and PC3 is observed. Again, the test projections highlighted as a box shadow, indicate that test sets work well. Classification was also relatively simple with regards to healthy versus unhealthy (see FIG. 10). Again, the caveat is that this is derived from a small dataset.

[0082] It is clear that both fixed and un-fixed cell samples can be used for the purpose of determining cell health, with some further studies in order to support these earlier investigations. As well as using samples of frozen cell samples, the inventors also studied cell supernatant. In FIG. 11, aspirated supernatant of a cell pellet containing cell media, fragments and related products, from the previous nutrient deficient and virally infected cell lines was analysed. Spectral differences can be seen immediately. Virally infected cells in particular display large spectral variances through the fingerprint region. Without wishing to be bound by theory, it is possible that some cellular material may be present in these samples to a varying degree. As shown in FIG. 12, these spectral differences contribute to good split in PC1 between virus infected cells, and also in PC2, discriminating nutrient deficient cells. PC1 is largely displaying amide I alterations because of infection (FIG. 13a), indicating protein differences, whereas PC2 describes more subtle variances across the spectrum (FIG. 13b).

[0083] In another instance, supernatant was analysed in a separate virus focused study, investigating the impact of Reo virus, and also the EMC virus. Sample acquisition, preparation and analysis are the same as previously described. In FIG. 14, the raw spectra from healthy and virally infected cells are shown. The fingerprint region was focused upon as differences are more pronounced when focusing in this region, with a clear difference arising in the virus infected cells between the amide I and amide II peaks (see FIG. 14). Clear differences are evident between supernatants obtained from various cell culture conditions following PCA, particularly with EMC infection appearing more negatively in PC2 (FIG. 15).