G Model
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Contents lists available at ScienceDirect
Artificial Intelligence in Medicine journal homepage: www.elsevier.com/locate/aiim
Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis Gopal Krishna Dhondalay a,b,1 , Katherine Lawrence c,1 , Stephen Ward d,2 , Graham Ball b,∗ , Michael Hoare c,∗ a
National Heart and Lung Institute, Imperial Centre for Translational and Experimental Medicine, Imperial College of London, London W120NN, UK School of Science and Technology, Nottingham Trent University, Clifton Campus, Nottingham, Nottinghamshire NG11 8NS, UK c Department for Biochemical Engineering, Advanced Centre of Biochemical Engineering, University College London, Torrington Place, London WC1E 7JE, UK d Cell Therapy Catapult, Biomedical Research Centre, Research and Development, 16th Floor Tower Wing, Guy’s Hospital, Great Maze Pond, London SE1 9RT, UK b
a r t i c l e
i n f o
Article history: Received 28 February 2014 Received in revised form 29 April 2014 Accepted 12 July 2014 Keywords: Artificial neural networks Cells for therapy Bioprocessing Membrane integrity Surface markers Ultra scale-down
a b s t r a c t Objective: The successful preparation of cells for therapy depends on the characterization of causal factors affecting cell quality. Ultra scale-down methods are used to characterize cells in terms of their response to process engineering causal factors of hydrodynamic shear stress and time. This response is in turn characterized in terms of causal factors relating to variations as may naturally occur during cell preparation, i.e., passage number, generation number, time of the final passage stage and hold time in formulation medium. Methods: To investigate the influence of all of these causal factors we have adopted a non-linear, multivariate predictive artificial neural network (ANN) based modeling approach to help create clearer insights into their effect on cell membrane integrity and surface marker content. A prostate cancer cell line candidate for cancer therapy (P4E6) was used and cell surface markers CD9, CD147 and HLA A-C were investigated. Results: All causal factors studied were found to be significant in establishing an ANN model for the prediction of cell quality parameters with the extent of exposure to shear stress being the most significant and then passage number (range 57–66) and generation number (range 10–19) determining most strongly the cells’ resistance to shear stress. Both the operation of the final cell passage and the hold time of the cells in a formulation buffer also determine the cells’ resistance to shear stress. The processing parameters related to cell handling after preparation, i.e., shear stress and time of exposure were found to be the most influential affecting cell quality. Conclusion: CD9 surface marker loss was the most sensitive indicator of the effects of shear stress followed by loss of membrane integrity and then HLA A-C, while CD147 remained unaffected by shear stress or even prone to increase. Also greater stability of cell surface marker presence was noted for cells generated at greater passage numbers or generation numbers or for reduction in hold time in formulation buffer. © 2014 Elsevier B.V. All rights reserved.
1. Introduction The biopharmaceutical industry has taken the first steps into production of ‘therapeutic cells’ where cells are introduced to the patient that are capable of restoring function and helping to combat disease. This type of therapy is particularly complex and the
∗ Corresponding authors. E-mail addresses: [email protected] (G. Ball), [email protected] (M. Hoare). 1 Co-lead authors. 2 Previously at: Onyvax Limited, St Georges Hospital Medical School, Cranmer Terrace, Tooting, London SW17 0RE, UK.
industry faces the challenge to manufacture cells at different scales in a reproducible and cost effective manner. All phases from initial cell sourcing to cell expansion and processing to formulation, storage and administration require a great deal of research and development to achieve the required cell quality [1]. Allogeneic cell therapies offer the opportunity to manufacture at scales of up to10s and even 100s L of culture. At such scales processing considerations become even more of a key issue than for bench scale, e.g., for cell transfer, hold stages, etc. The example studied in this paper is a cell candidate for a therapeutic whole cell vaccine [2] with cells selected to promote a potent immune stimulus [3]. For example, several prostate cell lines representing different stages of prostate tumor progression may be incorporated
http://dx.doi.org/10.1016/j.artmed.2014.07.003 0933-3657/© 2014 Elsevier B.V. All rights reserved.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9 2
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
into an allogeneic whole cell tumor vaccine [2,4] to target multiple tumor specific antigens [5]. An overriding important factor is to ensure the required cell quality is maintained throughout the culture and manufacturing processes. Methods of cell characterization range from cell membrane integrity analysis which is of importance for the responsiveness of the immune system [6], surface marker analysis [7–9] the cells’ cytokine release profile [10] and their potency in creating an appropriate immune response [11]. Similar processing and characterization challenges are faced in the production of cells for regenerative medicine therapies [12–16] as well as their ability to act as an inoculum for continued growth [17]. A record of cell history is important when producing a therapeutic whole cell product [18,19]. The most common record is the passage number assigned to the culture which increases by one each time sub-culturing occurs. Other information such as cell generation number, the extent of cell expansion especially at the final stage, time between passages, time of exposure to a dissociative enzyme (at passage) and confluency at passage all provide the user with further quantitative information about the history of the cells prior to bioprocessing and subsequent administration to the patient. Cell preparation for patient administration involves a sequence of operations during which the cells are exposed to a wide range of hydrodynamic forces. Such forces can often lead to membrane rupture and death [20–22]. Physical conditions that must be considered include the extent and duration of shear stress (or intensity of power dissipation), e.g., during mixing and vialling and the use of holding stages, e.g., during preparation for centrifugation or for formulation. Ultra scale-down devices of rotating disk [8,9] and capillary [7,23] configurations have been developed to help inform of the effects on cells of the process environments which may exist during large-scale preparation. The nature of the interaction of biological systems with physical bioprocess conditions is inherently nonlinear, making their effects difficult to interpret. This is compounded when the effects of multiple parameters are considered. Such nonlinearity and cocorrelation rule out the use of linear based statistical methods such as multivariate regression. Nonlinear predictive approaches such as artificial neural networks (ANNs) can overcome these problems; for example for the identification of “biologically relevant” molecules, in pyrolysis mass spectrometry [24]; genomic microarraying of tumor tissue [25]; discrimination between multiple classes for blind data for MALDI mass spectrometry [26,27]; or SELDI-MS-derived data containing a high level of background noise [28]. Generalized models may be produced without relying on predetermined relationships (e.g., in medical diagnostics [29], prostate cancer [30]) and can be subjected to further development through continued interpretation to improve the accuracy of outputs [31].
The research presented here uses an ANN to mine sets of experimental data parameters comprising cell culture and cell bioprocessing operating variables and resulting cell properties, e.g., membrane integrity or surface marker content, and help identify relationships between the parameters studied. 2. Materials and methods The input and output data analyzed in this paper are based on the study of bioprocessing of a prostate cancer cell line [8]. A brief description is given of the cell preparation and processing along with the nature of the data recorded during experimentation. 2.1. Cell preparation and processing A prostate cancer cell line, P4E6 (provided by Onyvax Limited, London, UK) was grown at 37 ◦ C in a humidified atmosphere containing 5% CO2 using keratinocyte serum free medium (Invitrogen, Paisley, UK) supplemented with 2% (v/v) fetal calf serum (PAA laboratories, Linz, Austria) and 5 g/L epidermal growth factor (Invitrogen, Paisley, UK) and sub-cultured when the cells reached approximately 70% confluency [8]. Cells were harvested by removal of spent medium, detachment in Accutase (Sigma–Aldrich, Poole, UK), quenching in complete growth medium, separation by centrifugation at 500 g for 3 min followed by washing twice in Hank’s balanced salt solution (HBSS; Sigma–Aldrich, Poole, UK), and final resuspension in HBSS containing 8% (v/v) dimethyl sulfoxide (DMSO; Sigma, Poole, UK) to the required concentration range of 2–5 × 107 cells/mL [8]. In all cases the cell passage number and the number of cells at the start and end of each passage stage was recorded. The cells were held in static suspension at 21 ◦ C for 1.5–3.5 h, and then gently resuspended and then exposed to a defined hydrodynamic stress environment in a 20 mL rotating disk shear device, disk diameter 40 mm operating at 4000 rpm (equivalent maximum shear rate 90 × 103 s−1 or maximum power dissipation, ε, 14 W/mL), 5000 rpm (130 × 103 s−1 or 31 W/mL) or 6000 rpm (175 × 103 s−1 or 52 W/mL) for up to 3 h [8,9]. Cells held in static suspension were used as a non-stressed control. Cells were sampled every 45 min in triplicate and each sample analyzed in triplicate for cell membrane integrity (trypan blue exclusion method using a haemocytometer) or for cell surface marker analysis, CD9, CD147 and HLA A-C using flow cytometry (FACScan; Becton Dickinson, New Jersey, USA with data analyzed using FCS Express software, De Novo Software, Ontario, Canada). 3. Data sets for analysis The process sequence studied is described in Fig. 1 along with the classification of the various input factors and resulting
Fig. 1. Process sequence studied and experimental data collected for the development of an ANN model linking output values to input factors related to cell preparation and cell testing. The flow sequence is related to factors and experimental data sets as described directly below the individual operating stages. The key to the factors used: Fixed values: passage number (PNWCB ) and generation number (GNWCB ) for working cell bank (see Eq. (1)). Input factors for cell manufacture: passage number (PNF ) and generation number (GNF ) (Eq. (1)). Proportion of time spent in stationary phase, time (tDF ) and concentration [NfF ] for final phase (Eq. (2)) and holding time after formulation (HT). Input factors and output values for cell testing: maximum power dissipation (ε) and time of exposure to shear (ST).
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
output values related to cell quality. A working cell bank is used as a source of cells of fixed history. A set number of these cells are cultured through a series of passage steps and generations with the final cells produced being defined by the resulting passage number, PNF , and generation number, GNF . The majority of the cells are produced on the final passage, the operation of which is characterized by the final cell density, [NfF ], and the time taken to complete the final passage, tDF . The cells are then formulated and subjected to a hold period, HT, until further processing is possible. The cells are then exposed to a pre-specified set of shear stress levels, ε, for set times, ST, to provide a measure of their robustness to processing conditions, e.g., centrifugation and resuspension, vialling, mixing, transport, etc. The range of stress conditions used encompasses those which may apply in production as well as some extreme conditions to help increase significance of changes in cell properties measured. The cell properties were originally characterized in relation to the testing of cell processibility, i.e., for a range of ε and ST values with the assumption that all cells produced for the study would be considered the same [8]. The considerable variations noted in the comparison of predicted and measured output data were attributed to potential differences in cell preparation and it is these differences which are explored here. 3.1. Input factors relating to cell preparation – passage number (PNF ) and generation number (GNF ) Cells are supplied for production from a working cell bank with a set passage number. The cells are inoculated, expanded and harvested for a number of additional passages until the number of cells required is achieved. In this study the working cell bank is at passage 50 (PNWCB ) resulting in: 57 ≤ PNF ≤ 66. The total number of cell doublings is given by GNF =
i=F ln(Nf /N0 )i i=1
ln 2 + GNWCB
(1)
where (Nf /N0 )i is the increase in cell number for a particular passage stage. In this study GNWCB = 10 and 10.9 ≤ GNF ≤ 18.7. The ratio PNF :GNF = (PNF − PNWCB ):(GNF − GNWCB ) reflects the history of cell preparation following recovery from the working cell bank. Greater values represent shorter times between passages or higher extents of dilution prior to the next passage stage. This ratio will be used as a record of procedures carried out, rather than as input data for modeling. 3.2. Input factors relating to cell processing decisions – time of final passage (tDF ), final cell number density ([NfF ]) and hold time after cell formulation (HT) The cells were seeded at constant cell density ([N0F ]) for the final passage stage (excess cells were discarded). Decisions when to harvest were moderated by the need to achieve a certain cell count. These two variables are analyzed together as a proportion of the total time spent in stationary phase, PF : PF =
tDF − DT ln ([Nf /N0 ])F / ln 2 tDF − TE = tDF tDF
(2)
where DT is the cell doubling time in exponential phase (=24 h) and TE is the estimated time in exponential phase. In this study PF ranges from 0.11 to 0.61. Evidently tDF and [NfF ] might be considered as separate variables, this is discussed during interpretation of results. Some manufacturing methods use an apparent doubling time to represent the final phase, i.e., tDF ln 2/ln([NfF ]/[N0F ]); this ranges from 27 to 61 h in this study (Onyvax Limited, private communication). Cells were resuspended immediately after final passage into
3
Fig. 2. Process of application of the ANN algorithm for identification of models to predict cell quality factors. The experimental data sets compared values of input data of ε, ST, GNF , DT, HT, PF and output data of cell membrane integrity and CD9 and CD147 and HLA A-C. For each partition a random selection of 60% of the data sets was used for “training”, i.e., development of a model. Various combinations of some or all of the input data were used in a multilayer perceptron approach to predict separately each of the cell quality parameters, i.e., predicted outputs. Back propagation (see Section 2) was used to minimize RMS; this procedure was stopped after 3000 epochs or sooner within 1000 epochs if, using the test data sets the mean squared error >0.1 (i.e., RMS > 0.32). The resultant model was further developed via validation and repartitioning to yield a final trained model for particular combinations of input data (e.g., ε and ST in Fig. 3 and ε, ST, GNF , DT, HT, PF in subsequent figures).
the formulation buffer (HBSS + 8% DMSO). The hold times at 21 ◦ C prior to exposure to hydrodynamic stress varied from 1.5 to 3.5 h. 3.3. Input factors relating to ultra scale-down testing of cell processibility: maximum power dissipation (ε) and time of exposure (ST) During processing cells will be exposed to a wide range of flow conditions ranging from static (hold stages) to gentle (mixing, membrane concentration) to harsh (vialling, sediment resuspension during centrifugation). An ultra scale-down shear device is used to mimic these conditions [8,9]. The applied hydrodynamic stress is noted in terms of the maximum power dissipation, ε, which exists at the tip of the rotating disk and for times of application of the stress (ST). ST is in addition to the hold time (HT) and ranges from 0 to 3 h reflecting rapid operations (e.g., vialling, sediment resuspension) to longer term operations (e.g., mixing, membrane separation). 3.4. ANN model development and interrogation The ANN model (Fig. 2) consisted an input, an output and a sandwiched hidden layer with two nodes, with feed forward flow of information and backpropagation of error algorithm for iterative learning from data [32]. The model development commenced with random data partitioning into three subsets: training, test and validation sets with partition ratio of 60:20:20, i.e., 60% of the data is used to train the network, 20% act as early stopping criteria for the network learning and the remaining 20% is used for validation of the network model. The whole process of random data partitioning is repeated for 50 predictive sub models with learning rate of 0.1 and momentum of 0.5 [32]. The sigmoid activation function was applied in the hidden nodes at hidden layer. The predictions will iterate until the termination criteria is met under each model
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
4
Fig. 3. Correlations of cell quality parameters using maximum power dissipation (ε) and time of exposure (ST) only. Parity plots (—) and lines of best fit (—) comparing predicted and observed fractional changes of membrane integrity and surface markers using an ANN model. See Table 1, line 1 for details of RMS error of parity and R2 value for line of best fit. The line of best fit offset (intercept, y-axis) and skewedness (slope) values are for: membrane integrity, 0.106, 0.873; CD9, 0.384, 0.517; CD147, 0.613, 0.435; HLA A-C, 0.431, 0.525.
which is 3000 epochs or a threshold window of 1000 epochs without improvement of root mean square error of more than 10% [33]. The resulting ANN models were used to predict, as a function of particular combinations of values of operating parameters within the ranges studied, the properties of the cells with respect to their membrane integrity and three surface markers. 4. Results Fig. 3 presents parity plots obtained comparing predicted and observed cell properties resulting from ANN models derived using just exposure to hydrodynamic stress (ε and ST) as the input data. Reasonable correlation of fit (see line 1, Table 1) is gained for membrane integrity (R2 = 0.891) but poor correlation for the surface markers (R2 = 0.501 for CD9, 0.488 for CD147 and 0.586 for HLA A-C). Also the line of best fit markedly deviates from the parity line for all surface markers. For CD147 there is evidently little effect on predicted values except at exposure at high levels of ε and ST. For the other two markers some loss is predicted for cells exposed to these conditions. To study further the impact of cell preparation conditions the ANN model was used to explore how PNF , GNF , PF and HT impact on how the resultant cells respond to ε and ST. Table 1 relates the result of increasing the number of variables, added in a stepwise fashion, used for constructing the ANN model. Each variable was added in order of its ability to improve prediction of response. The subsequent reduction in the RMS error upon addition of parameters relates with the increasing R2 values for correlations for membrane integrity and for the cell surface markers. The most important parameters and the sequence in which they demonstrated increasing quality of fit were PNF > GNF > PF > HT (Table 1) to yield finally a good correlation for
Fig. 4. Correlations of cell quality parameters using ε, ST, PNF , GNF , PF and HT as variables for determination of the ANN model and the resultant parity plots (—) and lines of best fit (—) comparing predicted and observed fractional changes. See Table 1, line 5 for details of RMS error of parity and R2 value for line of best fit. The line of best fit offset and skewedness values (see Fig. 3 legend) are for: membrane integrity, 0.028, 0.965; CD9, 0.172, 0.793; CD147, 0.449, 0.580; HLA A-C, 0.328, 0.644.
membrane integrity (R2 = 0.978) and moderate correlations for CD9 (0.853), CD147 (0.746) and HLA A-C (0.760). A comparison between the equation of a best fit line and a parity line is also shown in Table 1 and Fig. 4. The presence of offset and skewedness indicates that any extrapolation of the model outside the envelope of values of the variables studied needs to be avoided as confidence in true prediction will be low. When using the ANN model to predict values of changes in membrane integrity and the surface markers the line of best fit rather than the parity fit is used to give the predicted relationships. This approach is to recognize that unaccounted variables are affecting the degree of fit achieved with parity. Evidently, for membrane integrity and CD9, the relatively close agreement of the parity line and the line of best fit (Fig. 4) gives greater confidence in the predicted responses to hydrodynamic stress. The effect of cell history, i.e., PNF and GNF , is studied first. The relationship between the number of cells to be produced from the given working cell bank stock is reflected in the overall number of passages to be used, i.e., greater numbers of cells will require an increase in PNF with the accompanying increase in GNF. This approximately linear relationship is described in Fig. 5 but with variations both of smaller or larger GNF values for the same PNF value. The first set of conditions studied are for combinations of GNF and PNF which best follow the linear relationship. This combined increase of PNF and GNF leads to greater resistance to change in membrane integrity on exposure to hydrodynamic stress (Fig. 6a–d). This result may be a consequence of selective pressure leading to preferred growth of the more robust cells, i.e., cells which better survive the interstage processing between passages including enzyme treatment for cell detachment followed by low RCF centrifugation, cell pellet resuspension and replating for further growth.
Table 1 Development of ANN model with addition of input data variables, i.e., bioprocessing variables determining the impact of cell preparation and processing on cell quality parameters. The input data variables are included in order of decreasing significance of impact of confidence of fit obtained. Input data variables
1. 2. 3. 4. 5.
ε + ST ε + ST + PNF ε + ST + PNF + GNF ε + ST + PNF + GNF + PF ε + ST + PNF + GNF + PF + HT
Membrane integrity
CD9
CD147
HLA A-C
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
0.099 0.073 0.058 0.053 0.045
0.891 0.941 0.964 0.969 0.978
0.196 0.124 0.121 0.113 0.107
0.501 0.803 0.816 0.842 0.853
0.267 0.238 0.226 0.204 0.200
0.488 0.628 0.657 0.734 0.746
0.170 0.174 0.153 0.149 0.133
0.586 0.563 0.683 0.690 0.760
For completeness the offset and skewedness values are for: membrane integrity line 2 0.067, 0.919, line 3 0.055, 0.933, line 4 0.050, 0.944; CD9 line 2 0.219, 0.736, line 3 0.220, 0.729, line 4 0.201, 0.759; CD147 line 2 0.564, 0.469, line 3 0.519, 0.517, line 4 0.461, 0.567; HLA A-C line 2 0.395, 0.562, line 3 0.409, 0.555, line 4 0.381, 0.593. See Figs. 2 and 3 for lines 1 and 5 values, respectively.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Fig. 5. Combinations of final stage generation numbers (GNF ) and passage number (PNF ) studied in this paper. Both the experimentally studied cells and the ANN modeled cells are represented: experimental cell batches used; + combinations of GNF and PNF examined in the derived ANN models (with appropriate figure letter identified for reference to Figs. 6–9); — line of best fit for experimentally prepared cells, GNF :PNF = 0.85.
The way in which passages are carried out, e.g., seeding density or point of harvest, will affect the resultant relationship between PNF and GNF . This is explored in the sequence of Fig. 6e to a to f where for increased GNF for the same PNF value, cells are obtained which are more resistant to hydrodynamic stress. This combination of increased GNF and fixed PNF is achieved either by seeding cells at lower cell density and/or allowing cells to grow to higher proliferation density. Again it is likely that selective pressure is such that the more robust cells will be preferably recovered after extended growth. The converse effect is noted when keeping GNF the same and increasing PNF (Fig. 6g to a to h) where the cells become less resistant to hydrodynamic stress. (It is evident that GNF and PNF could be controlled to yield
5
cells of unchanged robustness to hydrodynamic stress. However this approach to achieving consistency in cell properties does have to be offset against the value of routine in carrying out passaging operations, e.g., fixed time of passage and/or fixed split numbers.) PF reflects decisions made relating to the operation of the final passage of cell preparation e.g., the point at which the operators will be ready to start the final harvest of the cells. The second is holding time (HT) in formulation media before processing. Within the scope of this study, this variable has little effect on the resistance of the cells to hydrodynamic stress (Fig. 6i to j and k to l). Further investigation into the raw data making up PF demonstrated the presence of two trials where the combination of tDF and [NfF ]/[N0F ] could be considered as outliers, i.e., resulting in PF values of 0.11 and 0.61 compared with the narrow range recorded for the controlled trials 0.31 < PF < 0.47. However separate investigation of how tDF and [NfF ] as independent variables affect the degree of fit of the ANN model gave insufficient improvement to merit further consideration in this study. The holding time of the harvested cells after the final passage, HT again has little effect on the resistance of the cells to hydrodynamic stress (Fig. 6i to k and j to l). The effect of cell and processing history on the resistance of cells to change as measured by the presence of surface markers is also explored. With reference back to Fig. 6, CD9 shows the same behavior as membrane integrity for increase in cell age (Fig. 7a–d) but to a far less extent. Otherwise CD9 is little affected by all the other changes studied, e.g., independent variation of GNF and PNF (Fig. 7e–h) and of PF and HT (Fig. 7i–l). Hence CD9 appears to be a robust surface marker unaffected by any of the conditions used for cell preparation prior to processing. However on exposure to hydrodynamic stress the CD9 marker is lost to a marked extent but not at quite the same level as for membrane integrity, i.e., alterations in the cell membrane seems to cause the CD9 marker to be less accessible for analysis or even lost. The effect of hydrodynamic stress on CD147 is complicated by the trend leading to greater marker expression or accessibility to analysis for greater extents of exposure to stress. Increased cell age leads to greater changes in CD147 detection (Fig. 8a–d). Otherwise the changes arising from alterations in the operating line for cell preparation (Fig. 8e–h) and for different cell handling procedures (Fig. 8i–l) are relatively insignificant. Hence the overall behavior is similar to that for CD9 except that the effect of hydrodynamic stress is to lead to increase in marker access. The effect of cell preparation method on the response to hydrodynamic stress of HLA A-C is relatively similar to that for CD147 (Fig. 9a–d and e–h) but with less certainty of this observation. Again the effect of PF is low, but there is a clear
Fig. 6. Interrogation of trained optimal ANN model for the effect of the cell preparation and processing conditions on the fractional change in the cell membrane integrity. 0 W/mL; 14 W/mL; - - - 52 W/mL. (a–d) show the effect of Each graph shows the predicted response over ST = 180 min for exposure to hydrodynamic stress, ε of: increasing combinations of GNF and PNF for fixed PF and HT. The GNF and PNF combinations follow the relationship from Fig. 5, GNF = 0.85 PNF . (e and f) show the effect of increasing GNF for the same PNF and (g and h) the effect of increasing PNF for the same GNF ((a) is at the midpoint of these sequences). (i, j and k, l) show the effect of increasing PF for the same HT and (i, k and j, l) show the effect of increasing HT for the same PF ; these are all for the same GNF and PNF combination as also used for (a).
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9 6
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Fig. 7. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, CD9 – the figure layout follows the same description as for Fig. 6.
Fig. 8. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, CD147 – the figure layout follows that for Fig. 6. To note that in some cases increases in surface markers are shown.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
7
Fig. 9. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, HLA A-C – the figure layout follows that for Fig. 6. effect of greater holding time with the marker becoming increasingly susceptible to hydrodynamic stress (Fig. 9i–l).
5. Discussion The effects of shear on surface marker levels may be explained by changes in cellular protein metabolism or the physical presentation of the protein to analysis. Protein metabolism refers to the dynamic states of continuous protein synthesis, drawing upon intracellular protein pools, and degradation. Stimuli such as disease states and chemical and physical stress can affect the protein levels. For example the stimuli may cause up-regulation in protein turnover and synthesis leading to increased expression, or to cause up-regulation of protein degradation pathways (proteolysis) leading to a decrease [34]. As a result of changes in the intracellular protein pools, surface protein expression levels may alter rapidly (few hours). Cells try to achieve a state of homeostasis (a balance between synthesis and degradation). However, following insults such as exposure to hydrodynamic stress, this may not be possible. Hydrodynamic stress may also bring about alterations in the physical presentation of proteins to analysis. This may take several forms such as the reduced or enhanced epitope-antibody accessibility due to blocking or epitope structural alteration or loss, or due to denaturation or movement of neighboring proteins [8,20]. The three surface markers studied in this paper demonstrate different responses to hydrodynamic shear stress. Both increases and decreases in surface marker presence are reported in the literature. The exposure of murine bone cell lines (MC3T3-E1 and MLO-Y4) led to examples of slower turnover and synthesis, increased synthesis with greater turnover rate and possible denaturation or removal by hydrodynamic stress [35]. Exposure to laminar shear stress of rabbit carotid artery led to changes in various adhesion molecules with decreased and increased presence of intracellular and vascular molecules respectively [36]. HLA A-C (human leukocyte antigens) are proteins of approximately 40 kDa that make up a large
subset of the human major histocompatibility complex (MHC). These molecules are involved in presentation of foreign antigens to the host immune system in order to elicit an immune response [37]. A proportion of HLA subtypes that reside on the cell surface internalize over time and a steady state recycling of molecules occurs naturally [38]. Hence the shear stress-related decrease identified in this paper HLA A-C at the cell surface may be due to an increase in the rate of internalization and/or a decrease in the rate of replacement. The same may be said for CD9 (also known as Motility Related Protein 1) which shows similar, although to a greater extent. This is a 24–27 kDa member of the tetraspanin family of proteins that play a role in the regulation and modulation of cell development, activation, growth, aggregation, adhesion and motility [39–41]. CD147 (also known as Collagenase Stimulatory Factor) is an approximately 42 kDa cell surface transmembrane protein involved in the co-ordination of cell adhesion with proteolysis, cell communication and signal transduction [42–44]. Here the effects of shear stress appear to be much reduced and even lead to up-regulation or increased accessibility to the marker for analysis. The results reported in this paper can only provide a limited biological understanding of how hydrodynamic stress affects any particular marker protein while the cells are suspended in formulation medium in the presence of DMSO, i.e., whether the changes are differences in metabolism or location/accessibility for analysis. However they do provide insight of how the cells and surface markers respond to shear stress and indicate regions of interest for further analysis. For example CD9 is significantly affected by the level of hydrodynamic stress and time of exposure. This almost parallels loss of membrane integrity (Fig. 6) except that the responsiveness at low levels of stress is still high for CD9 whereas it is low for membrane integrity. Hence considerable care is needed in process design if CD9 is a critical marker with respect to clinical efficacy. In addition to the effect of stress and time, the effect of holding time in formulation medium before processing (HT) is also significant with more than a doubling of loss observed on increasing HT
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
8
from 1:45 to 3:00 h. This is in contrast to the relatively low effect of HT on the susceptibility of the cells to lose membrane integrity (Fig. 6). The relative instability of CD9 for cells held in suspension may be a reflection of the cell down-regulating a protein which functions to control cell adhesion and its motility while surface attached. Interestingly, HLA A-C (Fig. 9) shows the same effect of holding time response as for CD9. The development of future cell lines for therapy will require in each case an understanding of the relationship between the operating conditions used and the resulting properties of the cells. Most importantly, to ensure process robustness is secured, it is necessary to understand degree of control needed for each operating condition to ensure a manufacturing process meets necessary regulatory standards. This paper demonstrates that information captured from individual trials may be combined with the use of ultra scale-down methods to study the processibility of the cells and the impact on cell quality. Each cell processing stage will impose its own process stress on the cells, e.g., mixing, vialling, centrifugation, filtration etc. Individual cell lines will demonstrate different responses to process stress making the application of numerical methods such as ANN to draw out data trends an important tool toward discovering robust routes for routine manufacture of cells for therapy. The effect of cell age, i.e., increase of PNF and GNF , is to make cells more robust to the effect of processing in the formulation medium which really means a reduced tendency to lose membrane integrity or surface marker, although for CD147 the effect is an increased level of marker present or available for assay. 6. Conclusions The use of ANNs has been demonstrated to help gain insight into the effect of cell preparation conditions on their processibility during formulation stages. Increased confidence of data fitting is achieved by taking into account cell passage and generation numbers, the time used for the final passage stage and the hold time in formulation medium. The trends observed provide insight into the design constraints of how cells are prepared for bioprocessing as determined by the extent and duration of hydrodynamic stress in a formulation medium. Such insight information gained at the small scale can help in understanding the contribution of operating variables in cell processing procedures and also in design consideration for full-scale manufacture. List of symbols
DT ε GN HT N [N] P PN ST tD TE
doubling time maximum power dissipation generation number holding time number of cells cell density or concentration proportion of passage time for cells in stationary phase passage number time of exposure to hydrodynamic stress time between cell detachments/at start and end of passage extended time in exponential phase
Subscripts f final for a particular passage stage F final passage stage i passage stage initial for a particular passage stage 0 working cell bank WCB
Acknowledgments Funding provided by the UK Technology Strategy Board High Value Manufacturing Initiative (TP/7/B10/6/I/M0011 (P550)) for Gopal Krishna Dhondalay and Katherine Lawrence is gratefully acknowledged as is also resource provided by the EPSRC Innovative Manufacturing Research Center for Bioprocessing (EP/I033270/1) hosted at University College London. We acknowledge Ryan McCoy (funded by EPSRC Industrial Doctoral Training Center in Bioprocessing (EP/G034656/1)) who supplied the original data [8] for analysis in this paper. References [1] Brandenberger R, Burger S, Campbell A, Fong T, Lapinskas E, Rowley JA. Cell therapy bioprocessing. BioProcess Int 2011:30–7. [2] Eaton JD, Perry MJ, Nicholson S, Guckian M, Russell N, Whelan M, et al. Allogeneic whole-cell vaccine: a phase I/II study in men with hormone-refractory prostate cancer. BJU Int 2002;89:19–26. [3] Staveley-O’Carroll K, Sotomayor E, Montgomery J, Borrello I, Hwang L, Fein S, et al. Induction of antigen-specific T cell anergy: an early event in the course of tumor progression. Proc Natl Acad Sci U S A 1998;95:1178–83. [4] Michael A, Ball G, Quatan N, Wushishi F, Russell N, Whelan J, et al. Delayed disease progression after allogeneic cell vaccination in hormone-resistant prostate cancer and correlation with immunologic variables. Clin Cancer Res 2005;11:4469–78. [5] Copier J, Ward S, Dalgleish A. Cell based cancer vaccines: regulatory and commercial development. Vaccine 2007;25(Suppl. 2):B35–46. [6] Ochsenbein AF, Klenerman P, Karrer U, Ludewig B, Pericin M, Hengartner H, et al. Immune surveillance against a solid tumor fails because of immunological ignorance. Proc Natl Acad Sci U S A 1999;96:2233–8. [7] Acosta-Martinez JP, Papantoniou I, Lawrence K, Ward S, Hoare M. Ultra scaledown stress analysis of the bioprocessing of whole human cells as a basis for cancer vaccines. Biotechnol Bioeng 2010;107:953–63. [8] McCoy R, Hoare M, Ward S. Ultra scale-down studies of the effect of shear on cell quality; processing of a human cell line for cancer vaccine therapy. Biotechnol Progr 2009;25:1448–58. [9] McCoy R, Ward S, Hoare M. Sub-population analysis of human cancer vaccine cells – ultra scale-down characterization of response to shear. Biotechnol Bioeng 2010;106:584–97. [10] Stadnyk AW. Cytokine production by epithelial cells. FASEB J 1994;8: 1041–7. [11] Petricciani J, Egan W, Vicari G, Furesz J, Schild G. Potency assays for therapeutic live whole cell cancer vaccines. Biologicals 2007;35:107–13. [12] Barry FP, Murphy JM. Mesenchymal stem cells: clinical applications and biological characterization. Int J Biochem Cell Biol 2004;36: 568–84. [13] Bubela T, Reshef A, Li MD, Atkins H, Caulfield T, Culme-Seymour E, et al. Enabling advanced cell therapies (EnACT): invitation to an online forum on resolving barriers to clinical translation. Regen Med 2012;7:735–40. [14] Carmen J, Burger SR, McCaman M, Rowley JA. Developing assays to address identity, potency, purity and safety: cell characterization in cell therapy process development. Regen Med 2012;7:85–100. [15] Diogo MM, da Silva CL, Cabral JM. Separation technologies for stem cell bioprocessing. Biotechnol Bioeng 2012;109:2699–709. [16] Mason C, Hoare M. Regenerative medicine bioprocessing: the need to learn from the experience of other fields. Regen Med 2006;1:615–23. [17] Papadaki M, McIntire LV, Eskin SG. Effects of shear stress on the growth kinetics of human aortic smooth muscle cells in vitro. Biotechnol Bioeng 1996;50:555–61. [18] Esquenet M, Swinnen JV, Heyns W, Verhoeven G. LNCaP prostatic adenocarcinoma cells derived from low and high passage numbers display divergent responses not only to androgens but also to retinoids. J Steroid Biochem Mol Biol 1997;62:391–9. [19] Wall ME, Bernacki SH, Loboa EG. Effects of serial passaging on the adipogenic and osteogenic differentiation potential of adipose-derived human mesenchymal stem cells. Tissue Eng 2007;13:1291–8. [20] Al-Rubeai M, Emery AN, Chalder S, Goldman MH. A flow cytometric study of hydrodynamic damage to mammalian cells. J Biotechnol 1993;31:161–77. [21] Born C, Zhang Z, Al-Rubeai M, Thomas CR. Estimation of disruption of animal cells by laminar shear stress. Biotechnol Bioeng 1992;40:1004–10. [22] Tanzeglock T, Soos M, Stephanopoulos G, Morbidelli M. Induction of mammalian cell death by simple shear and extensional flows. Biotechnol Bioeng 2009;104:360–70. [23] Godoy-Silva R, Mollet M, Chalmers JJ. Evaluation of the effect of chronic hydrodynamical stresses on cultures of suspensed CHO-6E6 cells. Biotechnol Bioeng 2009;102:1119–30. [24] Goodacre R, Kell DB. Pyrolysis mass spectrometry and its applications in biotechnology. Curr Opin Biotechnol 1996;7:20–8. [25] Khan J, Wei JS, Ringner M, Saal LH, Ladanyi M, Westermann F, et al. Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med 2001;7:673–9.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
[26] Ball G, Mian S, Holding F, Allibone RO, Lowe J, Ali S, et al. An integrated approach utilizing artificial neural networks and SELDI mass spectrometry for the classification of human tumours and rapid identification of potential biomarkers. Bioinformatics 2002;18:395–404. [27] Mian S, Ball G, Hornbuckle J, Holding F, Carmichael J, Ellis I, et al. A prototype methodology combining surface-enhanced laser desorption/ionization protein chip technology and artificial neural network algorithms to predict the chemoresponsiveness of breast cancer cell lines exposed to paclitaxel and doxorubicin under in vitro conditions. Proteomics 2003;3:1725–37. [28] Kothari SC, Oh H. Neural networks for pattern recognition. Adv Comput 1993;37:119–66. [29] McCartney S, Weltin M, Burchiel KJ. Use of an artificial neural network for diagnosis of facial pain syndromes: an update. Stereotact Funct Neurosurg 2014;92:44–52. [30] Hu X, Cammann H, Meyer HA, Miller K, Jung K, Stephan C. Artificial neural networks and prostate cancer – tools for diagnosis and management. Nat Rev Urol 2013;10:174–82. [31] Nakajima K, Nakajima Y, Horikoshi H, Ueno M, Wakabayashi H, Shiga T, et al. Enhanced diagnostic accuracy for quantitative bone scan using an artificial neural network system: a Japanese multi-center database project. EJNMMI Res 2013;3:83. [32] Lancashire LJ, Powe DG, Reis-Filho JS, Rakha E, Lemetre C, Weigelt B, et al. A validated gene expression profile for detecting clinical outcome in breast cancer using artificial neural networks. Breast Cancer Res Treat 2010;120:83–93. [33] Lowery AJ, Miller N, Devaney A, McNeill RE, Davoren PA, Lemetre C, et al. MicroRNA signatures predict oestrogen receptor, progesterone receptor and HER2/neu receptor status in breast cancer. Breast Cancer Res 2009;11: R27. [34] Fonseca R, Vabulas RM, Hartl FU, Bonhoeffer T, Nagerl UV. A balance of protein synthesis and proteasome-dependent degradation determines the maintenance of LTP. Neuron 2006;52:239–45.
9
[35] Thi MM, Kojima T, Cowin SC, Weinbaum S, Spray DC. Fluid shear stress remodels expression and function of junctional proteins in cultured bone cells. Am J Physiol Cell Physiol 2003;284:C389–403. [36] Walpola PL, Gotlieb AI, Cybulsky MI, Langille BL. Expression of ICAM-1 and VCAM-1 and monocyte adherence in arteries exposed to altered shear stress. Arterioscler Thromb Vasc Biol 1995;15:2–10. [37] Gruen JR, Weissman SM. Evolving views of the major histocompatibility complex. Blood 1997;90:4252–65. [38] Vega MA, Strominger JL. Constitutive endocytosis of HLA class I antigens requires a specific portion of the intracytoplasmic tail that shares structural features with other endocytosed molecules. Proc Natl Acad Sci U S A 1989;86:2688–92. [39] Higginbottom A, Takahashi Y, Bolling L, Coonrod SA, White JM, Partridge LJ, et al. Structural requirements for the inhibitory action of the CD9 large extracellular domain in sperm/oocyte binding and fusion. Biochem Biophys Res Commun 2003;311:208–14. [40] Ikeyama S, Koyama M, Yamaoko M, Sasada R, Miyake M. Suppression of cell motility and metastasis by transfection with human motility-related protein (MRP-1/CD9) DNA. J Exp Med 1993;177:1231–7. [41] Masellis-Smith A, Shaw AR. CD9-regulated adhesion. Anti-CD9 monoclonal antibody induce pre-B cell adhesion to bone marrow fibroblasts through de novo recognition of fibronectin. J Immunol 1994;152:2768–77. [42] Guo H, Majmudar G, Jensen TC, Biswas C, Toole BP, Gordon MK. Characterization of the gene for human EMMPRIN, a tumor cell surface inducer of matrix metalloproteinases. Gene 1998;220:99–108. [43] Muramatsu T, Miyauchi T, Basigin. (CD147): a multifunctional transmembrane protein involved in reproduction, neural function, inflammation and tumor invasion. Histol Histopathol 2003;18:981–7. [44] Nabeshima K, Lane WS, Biswas C. Partial sequencing and characterization of the tumor cell-derived collagenase stimulatory factor. Arch Biochem Biophys 1991;285:90–6.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Contents lists available at ScienceDirect
Artificial Intelligence in Medicine journal homepage: www.elsevier.com/locate/aiim
Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis Gopal Krishna Dhondalay a,b,1 , Katherine Lawrence c,1 , Stephen Ward d,2 , Graham Ball b,∗ , Michael Hoare c,∗ a
National Heart and Lung Institute, Imperial Centre for Translational and Experimental Medicine, Imperial College of London, London W120NN, UK School of Science and Technology, Nottingham Trent University, Clifton Campus, Nottingham, Nottinghamshire NG11 8NS, UK c Department for Biochemical Engineering, Advanced Centre of Biochemical Engineering, University College London, Torrington Place, London WC1E 7JE, UK d Cell Therapy Catapult, Biomedical Research Centre, Research and Development, 16th Floor Tower Wing, Guy’s Hospital, Great Maze Pond, London SE1 9RT, UK b
a r t i c l e
i n f o
Article history: Received 28 February 2014 Received in revised form 29 April 2014 Accepted 12 July 2014 Keywords: Artificial neural networks Cells for therapy Bioprocessing Membrane integrity Surface markers Ultra scale-down
a b s t r a c t Objective: The successful preparation of cells for therapy depends on the characterization of causal factors affecting cell quality. Ultra scale-down methods are used to characterize cells in terms of their response to process engineering causal factors of hydrodynamic shear stress and time. This response is in turn characterized in terms of causal factors relating to variations as may naturally occur during cell preparation, i.e., passage number, generation number, time of the final passage stage and hold time in formulation medium. Methods: To investigate the influence of all of these causal factors we have adopted a non-linear, multivariate predictive artificial neural network (ANN) based modeling approach to help create clearer insights into their effect on cell membrane integrity and surface marker content. A prostate cancer cell line candidate for cancer therapy (P4E6) was used and cell surface markers CD9, CD147 and HLA A-C were investigated. Results: All causal factors studied were found to be significant in establishing an ANN model for the prediction of cell quality parameters with the extent of exposure to shear stress being the most significant and then passage number (range 57–66) and generation number (range 10–19) determining most strongly the cells’ resistance to shear stress. Both the operation of the final cell passage and the hold time of the cells in a formulation buffer also determine the cells’ resistance to shear stress. The processing parameters related to cell handling after preparation, i.e., shear stress and time of exposure were found to be the most influential affecting cell quality. Conclusion: CD9 surface marker loss was the most sensitive indicator of the effects of shear stress followed by loss of membrane integrity and then HLA A-C, while CD147 remained unaffected by shear stress or even prone to increase. Also greater stability of cell surface marker presence was noted for cells generated at greater passage numbers or generation numbers or for reduction in hold time in formulation buffer. © 2014 Elsevier B.V. All rights reserved.
1. Introduction The biopharmaceutical industry has taken the first steps into production of ‘therapeutic cells’ where cells are introduced to the patient that are capable of restoring function and helping to combat disease. This type of therapy is particularly complex and the
∗ Corresponding authors. E-mail addresses: [email protected] (G. Ball), [email protected] (M. Hoare). 1 Co-lead authors. 2 Previously at: Onyvax Limited, St Georges Hospital Medical School, Cranmer Terrace, Tooting, London SW17 0RE, UK.
industry faces the challenge to manufacture cells at different scales in a reproducible and cost effective manner. All phases from initial cell sourcing to cell expansion and processing to formulation, storage and administration require a great deal of research and development to achieve the required cell quality [1]. Allogeneic cell therapies offer the opportunity to manufacture at scales of up to10s and even 100s L of culture. At such scales processing considerations become even more of a key issue than for bench scale, e.g., for cell transfer, hold stages, etc. The example studied in this paper is a cell candidate for a therapeutic whole cell vaccine [2] with cells selected to promote a potent immune stimulus [3]. For example, several prostate cell lines representing different stages of prostate tumor progression may be incorporated
http://dx.doi.org/10.1016/j.artmed.2014.07.003 0933-3657/© 2014 Elsevier B.V. All rights reserved.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9 2
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
into an allogeneic whole cell tumor vaccine [2,4] to target multiple tumor specific antigens [5]. An overriding important factor is to ensure the required cell quality is maintained throughout the culture and manufacturing processes. Methods of cell characterization range from cell membrane integrity analysis which is of importance for the responsiveness of the immune system [6], surface marker analysis [7–9] the cells’ cytokine release profile [10] and their potency in creating an appropriate immune response [11]. Similar processing and characterization challenges are faced in the production of cells for regenerative medicine therapies [12–16] as well as their ability to act as an inoculum for continued growth [17]. A record of cell history is important when producing a therapeutic whole cell product [18,19]. The most common record is the passage number assigned to the culture which increases by one each time sub-culturing occurs. Other information such as cell generation number, the extent of cell expansion especially at the final stage, time between passages, time of exposure to a dissociative enzyme (at passage) and confluency at passage all provide the user with further quantitative information about the history of the cells prior to bioprocessing and subsequent administration to the patient. Cell preparation for patient administration involves a sequence of operations during which the cells are exposed to a wide range of hydrodynamic forces. Such forces can often lead to membrane rupture and death [20–22]. Physical conditions that must be considered include the extent and duration of shear stress (or intensity of power dissipation), e.g., during mixing and vialling and the use of holding stages, e.g., during preparation for centrifugation or for formulation. Ultra scale-down devices of rotating disk [8,9] and capillary [7,23] configurations have been developed to help inform of the effects on cells of the process environments which may exist during large-scale preparation. The nature of the interaction of biological systems with physical bioprocess conditions is inherently nonlinear, making their effects difficult to interpret. This is compounded when the effects of multiple parameters are considered. Such nonlinearity and cocorrelation rule out the use of linear based statistical methods such as multivariate regression. Nonlinear predictive approaches such as artificial neural networks (ANNs) can overcome these problems; for example for the identification of “biologically relevant” molecules, in pyrolysis mass spectrometry [24]; genomic microarraying of tumor tissue [25]; discrimination between multiple classes for blind data for MALDI mass spectrometry [26,27]; or SELDI-MS-derived data containing a high level of background noise [28]. Generalized models may be produced without relying on predetermined relationships (e.g., in medical diagnostics [29], prostate cancer [30]) and can be subjected to further development through continued interpretation to improve the accuracy of outputs [31].
The research presented here uses an ANN to mine sets of experimental data parameters comprising cell culture and cell bioprocessing operating variables and resulting cell properties, e.g., membrane integrity or surface marker content, and help identify relationships between the parameters studied. 2. Materials and methods The input and output data analyzed in this paper are based on the study of bioprocessing of a prostate cancer cell line [8]. A brief description is given of the cell preparation and processing along with the nature of the data recorded during experimentation. 2.1. Cell preparation and processing A prostate cancer cell line, P4E6 (provided by Onyvax Limited, London, UK) was grown at 37 ◦ C in a humidified atmosphere containing 5% CO2 using keratinocyte serum free medium (Invitrogen, Paisley, UK) supplemented with 2% (v/v) fetal calf serum (PAA laboratories, Linz, Austria) and 5 g/L epidermal growth factor (Invitrogen, Paisley, UK) and sub-cultured when the cells reached approximately 70% confluency [8]. Cells were harvested by removal of spent medium, detachment in Accutase (Sigma–Aldrich, Poole, UK), quenching in complete growth medium, separation by centrifugation at 500 g for 3 min followed by washing twice in Hank’s balanced salt solution (HBSS; Sigma–Aldrich, Poole, UK), and final resuspension in HBSS containing 8% (v/v) dimethyl sulfoxide (DMSO; Sigma, Poole, UK) to the required concentration range of 2–5 × 107 cells/mL [8]. In all cases the cell passage number and the number of cells at the start and end of each passage stage was recorded. The cells were held in static suspension at 21 ◦ C for 1.5–3.5 h, and then gently resuspended and then exposed to a defined hydrodynamic stress environment in a 20 mL rotating disk shear device, disk diameter 40 mm operating at 4000 rpm (equivalent maximum shear rate 90 × 103 s−1 or maximum power dissipation, ε, 14 W/mL), 5000 rpm (130 × 103 s−1 or 31 W/mL) or 6000 rpm (175 × 103 s−1 or 52 W/mL) for up to 3 h [8,9]. Cells held in static suspension were used as a non-stressed control. Cells were sampled every 45 min in triplicate and each sample analyzed in triplicate for cell membrane integrity (trypan blue exclusion method using a haemocytometer) or for cell surface marker analysis, CD9, CD147 and HLA A-C using flow cytometry (FACScan; Becton Dickinson, New Jersey, USA with data analyzed using FCS Express software, De Novo Software, Ontario, Canada). 3. Data sets for analysis The process sequence studied is described in Fig. 1 along with the classification of the various input factors and resulting
Fig. 1. Process sequence studied and experimental data collected for the development of an ANN model linking output values to input factors related to cell preparation and cell testing. The flow sequence is related to factors and experimental data sets as described directly below the individual operating stages. The key to the factors used: Fixed values: passage number (PNWCB ) and generation number (GNWCB ) for working cell bank (see Eq. (1)). Input factors for cell manufacture: passage number (PNF ) and generation number (GNF ) (Eq. (1)). Proportion of time spent in stationary phase, time (tDF ) and concentration [NfF ] for final phase (Eq. (2)) and holding time after formulation (HT). Input factors and output values for cell testing: maximum power dissipation (ε) and time of exposure to shear (ST).
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
output values related to cell quality. A working cell bank is used as a source of cells of fixed history. A set number of these cells are cultured through a series of passage steps and generations with the final cells produced being defined by the resulting passage number, PNF , and generation number, GNF . The majority of the cells are produced on the final passage, the operation of which is characterized by the final cell density, [NfF ], and the time taken to complete the final passage, tDF . The cells are then formulated and subjected to a hold period, HT, until further processing is possible. The cells are then exposed to a pre-specified set of shear stress levels, ε, for set times, ST, to provide a measure of their robustness to processing conditions, e.g., centrifugation and resuspension, vialling, mixing, transport, etc. The range of stress conditions used encompasses those which may apply in production as well as some extreme conditions to help increase significance of changes in cell properties measured. The cell properties were originally characterized in relation to the testing of cell processibility, i.e., for a range of ε and ST values with the assumption that all cells produced for the study would be considered the same [8]. The considerable variations noted in the comparison of predicted and measured output data were attributed to potential differences in cell preparation and it is these differences which are explored here. 3.1. Input factors relating to cell preparation – passage number (PNF ) and generation number (GNF ) Cells are supplied for production from a working cell bank with a set passage number. The cells are inoculated, expanded and harvested for a number of additional passages until the number of cells required is achieved. In this study the working cell bank is at passage 50 (PNWCB ) resulting in: 57 ≤ PNF ≤ 66. The total number of cell doublings is given by GNF =
i=F ln(Nf /N0 )i i=1
ln 2 + GNWCB
(1)
where (Nf /N0 )i is the increase in cell number for a particular passage stage. In this study GNWCB = 10 and 10.9 ≤ GNF ≤ 18.7. The ratio PNF :GNF = (PNF − PNWCB ):(GNF − GNWCB ) reflects the history of cell preparation following recovery from the working cell bank. Greater values represent shorter times between passages or higher extents of dilution prior to the next passage stage. This ratio will be used as a record of procedures carried out, rather than as input data for modeling. 3.2. Input factors relating to cell processing decisions – time of final passage (tDF ), final cell number density ([NfF ]) and hold time after cell formulation (HT) The cells were seeded at constant cell density ([N0F ]) for the final passage stage (excess cells were discarded). Decisions when to harvest were moderated by the need to achieve a certain cell count. These two variables are analyzed together as a proportion of the total time spent in stationary phase, PF : PF =
tDF − DT ln ([Nf /N0 ])F / ln 2 tDF − TE = tDF tDF
(2)
where DT is the cell doubling time in exponential phase (=24 h) and TE is the estimated time in exponential phase. In this study PF ranges from 0.11 to 0.61. Evidently tDF and [NfF ] might be considered as separate variables, this is discussed during interpretation of results. Some manufacturing methods use an apparent doubling time to represent the final phase, i.e., tDF ln 2/ln([NfF ]/[N0F ]); this ranges from 27 to 61 h in this study (Onyvax Limited, private communication). Cells were resuspended immediately after final passage into
3
Fig. 2. Process of application of the ANN algorithm for identification of models to predict cell quality factors. The experimental data sets compared values of input data of ε, ST, GNF , DT, HT, PF and output data of cell membrane integrity and CD9 and CD147 and HLA A-C. For each partition a random selection of 60% of the data sets was used for “training”, i.e., development of a model. Various combinations of some or all of the input data were used in a multilayer perceptron approach to predict separately each of the cell quality parameters, i.e., predicted outputs. Back propagation (see Section 2) was used to minimize RMS; this procedure was stopped after 3000 epochs or sooner within 1000 epochs if, using the test data sets the mean squared error >0.1 (i.e., RMS > 0.32). The resultant model was further developed via validation and repartitioning to yield a final trained model for particular combinations of input data (e.g., ε and ST in Fig. 3 and ε, ST, GNF , DT, HT, PF in subsequent figures).
the formulation buffer (HBSS + 8% DMSO). The hold times at 21 ◦ C prior to exposure to hydrodynamic stress varied from 1.5 to 3.5 h. 3.3. Input factors relating to ultra scale-down testing of cell processibility: maximum power dissipation (ε) and time of exposure (ST) During processing cells will be exposed to a wide range of flow conditions ranging from static (hold stages) to gentle (mixing, membrane concentration) to harsh (vialling, sediment resuspension during centrifugation). An ultra scale-down shear device is used to mimic these conditions [8,9]. The applied hydrodynamic stress is noted in terms of the maximum power dissipation, ε, which exists at the tip of the rotating disk and for times of application of the stress (ST). ST is in addition to the hold time (HT) and ranges from 0 to 3 h reflecting rapid operations (e.g., vialling, sediment resuspension) to longer term operations (e.g., mixing, membrane separation). 3.4. ANN model development and interrogation The ANN model (Fig. 2) consisted an input, an output and a sandwiched hidden layer with two nodes, with feed forward flow of information and backpropagation of error algorithm for iterative learning from data [32]. The model development commenced with random data partitioning into three subsets: training, test and validation sets with partition ratio of 60:20:20, i.e., 60% of the data is used to train the network, 20% act as early stopping criteria for the network learning and the remaining 20% is used for validation of the network model. The whole process of random data partitioning is repeated for 50 predictive sub models with learning rate of 0.1 and momentum of 0.5 [32]. The sigmoid activation function was applied in the hidden nodes at hidden layer. The predictions will iterate until the termination criteria is met under each model
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model
ARTICLE IN PRESS
ARTMED-1352; No. of Pages 9
G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
4
Fig. 3. Correlations of cell quality parameters using maximum power dissipation (ε) and time of exposure (ST) only. Parity plots (—) and lines of best fit (—) comparing predicted and observed fractional changes of membrane integrity and surface markers using an ANN model. See Table 1, line 1 for details of RMS error of parity and R2 value for line of best fit. The line of best fit offset (intercept, y-axis) and skewedness (slope) values are for: membrane integrity, 0.106, 0.873; CD9, 0.384, 0.517; CD147, 0.613, 0.435; HLA A-C, 0.431, 0.525.
which is 3000 epochs or a threshold window of 1000 epochs without improvement of root mean square error of more than 10% [33]. The resulting ANN models were used to predict, as a function of particular combinations of values of operating parameters within the ranges studied, the properties of the cells with respect to their membrane integrity and three surface markers. 4. Results Fig. 3 presents parity plots obtained comparing predicted and observed cell properties resulting from ANN models derived using just exposure to hydrodynamic stress (ε and ST) as the input data. Reasonable correlation of fit (see line 1, Table 1) is gained for membrane integrity (R2 = 0.891) but poor correlation for the surface markers (R2 = 0.501 for CD9, 0.488 for CD147 and 0.586 for HLA A-C). Also the line of best fit markedly deviates from the parity line for all surface markers. For CD147 there is evidently little effect on predicted values except at exposure at high levels of ε and ST. For the other two markers some loss is predicted for cells exposed to these conditions. To study further the impact of cell preparation conditions the ANN model was used to explore how PNF , GNF , PF and HT impact on how the resultant cells respond to ε and ST. Table 1 relates the result of increasing the number of variables, added in a stepwise fashion, used for constructing the ANN model. Each variable was added in order of its ability to improve prediction of response. The subsequent reduction in the RMS error upon addition of parameters relates with the increasing R2 values for correlations for membrane integrity and for the cell surface markers. The most important parameters and the sequence in which they demonstrated increasing quality of fit were PNF > GNF > PF > HT (Table 1) to yield finally a good correlation for
Fig. 4. Correlations of cell quality parameters using ε, ST, PNF , GNF , PF and HT as variables for determination of the ANN model and the resultant parity plots (—) and lines of best fit (—) comparing predicted and observed fractional changes. See Table 1, line 5 for details of RMS error of parity and R2 value for line of best fit. The line of best fit offset and skewedness values (see Fig. 3 legend) are for: membrane integrity, 0.028, 0.965; CD9, 0.172, 0.793; CD147, 0.449, 0.580; HLA A-C, 0.328, 0.644.
membrane integrity (R2 = 0.978) and moderate correlations for CD9 (0.853), CD147 (0.746) and HLA A-C (0.760). A comparison between the equation of a best fit line and a parity line is also shown in Table 1 and Fig. 4. The presence of offset and skewedness indicates that any extrapolation of the model outside the envelope of values of the variables studied needs to be avoided as confidence in true prediction will be low. When using the ANN model to predict values of changes in membrane integrity and the surface markers the line of best fit rather than the parity fit is used to give the predicted relationships. This approach is to recognize that unaccounted variables are affecting the degree of fit achieved with parity. Evidently, for membrane integrity and CD9, the relatively close agreement of the parity line and the line of best fit (Fig. 4) gives greater confidence in the predicted responses to hydrodynamic stress. The effect of cell history, i.e., PNF and GNF , is studied first. The relationship between the number of cells to be produced from the given working cell bank stock is reflected in the overall number of passages to be used, i.e., greater numbers of cells will require an increase in PNF with the accompanying increase in GNF. This approximately linear relationship is described in Fig. 5 but with variations both of smaller or larger GNF values for the same PNF value. The first set of conditions studied are for combinations of GNF and PNF which best follow the linear relationship. This combined increase of PNF and GNF leads to greater resistance to change in membrane integrity on exposure to hydrodynamic stress (Fig. 6a–d). This result may be a consequence of selective pressure leading to preferred growth of the more robust cells, i.e., cells which better survive the interstage processing between passages including enzyme treatment for cell detachment followed by low RCF centrifugation, cell pellet resuspension and replating for further growth.
Table 1 Development of ANN model with addition of input data variables, i.e., bioprocessing variables determining the impact of cell preparation and processing on cell quality parameters. The input data variables are included in order of decreasing significance of impact of confidence of fit obtained. Input data variables
1. 2. 3. 4. 5.
ε + ST ε + ST + PNF ε + ST + PNF + GNF ε + ST + PNF + GNF + PF ε + ST + PNF + GNF + PF + HT
Membrane integrity
CD9
CD147
HLA A-C
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
RMS error of parity
R2 of best-fit
0.099 0.073 0.058 0.053 0.045
0.891 0.941 0.964 0.969 0.978
0.196 0.124 0.121 0.113 0.107
0.501 0.803 0.816 0.842 0.853
0.267 0.238 0.226 0.204 0.200
0.488 0.628 0.657 0.734 0.746
0.170 0.174 0.153 0.149 0.133
0.586 0.563 0.683 0.690 0.760
For completeness the offset and skewedness values are for: membrane integrity line 2 0.067, 0.919, line 3 0.055, 0.933, line 4 0.050, 0.944; CD9 line 2 0.219, 0.736, line 3 0.220, 0.729, line 4 0.201, 0.759; CD147 line 2 0.564, 0.469, line 3 0.519, 0.517, line 4 0.461, 0.567; HLA A-C line 2 0.395, 0.562, line 3 0.409, 0.555, line 4 0.381, 0.593. See Figs. 2 and 3 for lines 1 and 5 values, respectively.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Fig. 5. Combinations of final stage generation numbers (GNF ) and passage number (PNF ) studied in this paper. Both the experimentally studied cells and the ANN modeled cells are represented: experimental cell batches used; + combinations of GNF and PNF examined in the derived ANN models (with appropriate figure letter identified for reference to Figs. 6–9); — line of best fit for experimentally prepared cells, GNF :PNF = 0.85.
The way in which passages are carried out, e.g., seeding density or point of harvest, will affect the resultant relationship between PNF and GNF . This is explored in the sequence of Fig. 6e to a to f where for increased GNF for the same PNF value, cells are obtained which are more resistant to hydrodynamic stress. This combination of increased GNF and fixed PNF is achieved either by seeding cells at lower cell density and/or allowing cells to grow to higher proliferation density. Again it is likely that selective pressure is such that the more robust cells will be preferably recovered after extended growth. The converse effect is noted when keeping GNF the same and increasing PNF (Fig. 6g to a to h) where the cells become less resistant to hydrodynamic stress. (It is evident that GNF and PNF could be controlled to yield
5
cells of unchanged robustness to hydrodynamic stress. However this approach to achieving consistency in cell properties does have to be offset against the value of routine in carrying out passaging operations, e.g., fixed time of passage and/or fixed split numbers.) PF reflects decisions made relating to the operation of the final passage of cell preparation e.g., the point at which the operators will be ready to start the final harvest of the cells. The second is holding time (HT) in formulation media before processing. Within the scope of this study, this variable has little effect on the resistance of the cells to hydrodynamic stress (Fig. 6i to j and k to l). Further investigation into the raw data making up PF demonstrated the presence of two trials where the combination of tDF and [NfF ]/[N0F ] could be considered as outliers, i.e., resulting in PF values of 0.11 and 0.61 compared with the narrow range recorded for the controlled trials 0.31 < PF < 0.47. However separate investigation of how tDF and [NfF ] as independent variables affect the degree of fit of the ANN model gave insufficient improvement to merit further consideration in this study. The holding time of the harvested cells after the final passage, HT again has little effect on the resistance of the cells to hydrodynamic stress (Fig. 6i to k and j to l). The effect of cell and processing history on the resistance of cells to change as measured by the presence of surface markers is also explored. With reference back to Fig. 6, CD9 shows the same behavior as membrane integrity for increase in cell age (Fig. 7a–d) but to a far less extent. Otherwise CD9 is little affected by all the other changes studied, e.g., independent variation of GNF and PNF (Fig. 7e–h) and of PF and HT (Fig. 7i–l). Hence CD9 appears to be a robust surface marker unaffected by any of the conditions used for cell preparation prior to processing. However on exposure to hydrodynamic stress the CD9 marker is lost to a marked extent but not at quite the same level as for membrane integrity, i.e., alterations in the cell membrane seems to cause the CD9 marker to be less accessible for analysis or even lost. The effect of hydrodynamic stress on CD147 is complicated by the trend leading to greater marker expression or accessibility to analysis for greater extents of exposure to stress. Increased cell age leads to greater changes in CD147 detection (Fig. 8a–d). Otherwise the changes arising from alterations in the operating line for cell preparation (Fig. 8e–h) and for different cell handling procedures (Fig. 8i–l) are relatively insignificant. Hence the overall behavior is similar to that for CD9 except that the effect of hydrodynamic stress is to lead to increase in marker access. The effect of cell preparation method on the response to hydrodynamic stress of HLA A-C is relatively similar to that for CD147 (Fig. 9a–d and e–h) but with less certainty of this observation. Again the effect of PF is low, but there is a clear
Fig. 6. Interrogation of trained optimal ANN model for the effect of the cell preparation and processing conditions on the fractional change in the cell membrane integrity. 0 W/mL; 14 W/mL; - - - 52 W/mL. (a–d) show the effect of Each graph shows the predicted response over ST = 180 min for exposure to hydrodynamic stress, ε of: increasing combinations of GNF and PNF for fixed PF and HT. The GNF and PNF combinations follow the relationship from Fig. 5, GNF = 0.85 PNF . (e and f) show the effect of increasing GNF for the same PNF and (g and h) the effect of increasing PNF for the same GNF ((a) is at the midpoint of these sequences). (i, j and k, l) show the effect of increasing PF for the same HT and (i, k and j, l) show the effect of increasing HT for the same PF ; these are all for the same GNF and PNF combination as also used for (a).
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9 6
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
Fig. 7. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, CD9 – the figure layout follows the same description as for Fig. 6.
Fig. 8. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, CD147 – the figure layout follows that for Fig. 6. To note that in some cases increases in surface markers are shown.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
7
Fig. 9. Interrogation of trained optimal ANN models for the effect of processing conditions on the fractional change of cell surface marker level, HLA A-C – the figure layout follows that for Fig. 6. effect of greater holding time with the marker becoming increasingly susceptible to hydrodynamic stress (Fig. 9i–l).
5. Discussion The effects of shear on surface marker levels may be explained by changes in cellular protein metabolism or the physical presentation of the protein to analysis. Protein metabolism refers to the dynamic states of continuous protein synthesis, drawing upon intracellular protein pools, and degradation. Stimuli such as disease states and chemical and physical stress can affect the protein levels. For example the stimuli may cause up-regulation in protein turnover and synthesis leading to increased expression, or to cause up-regulation of protein degradation pathways (proteolysis) leading to a decrease [34]. As a result of changes in the intracellular protein pools, surface protein expression levels may alter rapidly (few hours). Cells try to achieve a state of homeostasis (a balance between synthesis and degradation). However, following insults such as exposure to hydrodynamic stress, this may not be possible. Hydrodynamic stress may also bring about alterations in the physical presentation of proteins to analysis. This may take several forms such as the reduced or enhanced epitope-antibody accessibility due to blocking or epitope structural alteration or loss, or due to denaturation or movement of neighboring proteins [8,20]. The three surface markers studied in this paper demonstrate different responses to hydrodynamic shear stress. Both increases and decreases in surface marker presence are reported in the literature. The exposure of murine bone cell lines (MC3T3-E1 and MLO-Y4) led to examples of slower turnover and synthesis, increased synthesis with greater turnover rate and possible denaturation or removal by hydrodynamic stress [35]. Exposure to laminar shear stress of rabbit carotid artery led to changes in various adhesion molecules with decreased and increased presence of intracellular and vascular molecules respectively [36]. HLA A-C (human leukocyte antigens) are proteins of approximately 40 kDa that make up a large
subset of the human major histocompatibility complex (MHC). These molecules are involved in presentation of foreign antigens to the host immune system in order to elicit an immune response [37]. A proportion of HLA subtypes that reside on the cell surface internalize over time and a steady state recycling of molecules occurs naturally [38]. Hence the shear stress-related decrease identified in this paper HLA A-C at the cell surface may be due to an increase in the rate of internalization and/or a decrease in the rate of replacement. The same may be said for CD9 (also known as Motility Related Protein 1) which shows similar, although to a greater extent. This is a 24–27 kDa member of the tetraspanin family of proteins that play a role in the regulation and modulation of cell development, activation, growth, aggregation, adhesion and motility [39–41]. CD147 (also known as Collagenase Stimulatory Factor) is an approximately 42 kDa cell surface transmembrane protein involved in the co-ordination of cell adhesion with proteolysis, cell communication and signal transduction [42–44]. Here the effects of shear stress appear to be much reduced and even lead to up-regulation or increased accessibility to the marker for analysis. The results reported in this paper can only provide a limited biological understanding of how hydrodynamic stress affects any particular marker protein while the cells are suspended in formulation medium in the presence of DMSO, i.e., whether the changes are differences in metabolism or location/accessibility for analysis. However they do provide insight of how the cells and surface markers respond to shear stress and indicate regions of interest for further analysis. For example CD9 is significantly affected by the level of hydrodynamic stress and time of exposure. This almost parallels loss of membrane integrity (Fig. 6) except that the responsiveness at low levels of stress is still high for CD9 whereas it is low for membrane integrity. Hence considerable care is needed in process design if CD9 is a critical marker with respect to clinical efficacy. In addition to the effect of stress and time, the effect of holding time in formulation medium before processing (HT) is also significant with more than a doubling of loss observed on increasing HT
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
8
from 1:45 to 3:00 h. This is in contrast to the relatively low effect of HT on the susceptibility of the cells to lose membrane integrity (Fig. 6). The relative instability of CD9 for cells held in suspension may be a reflection of the cell down-regulating a protein which functions to control cell adhesion and its motility while surface attached. Interestingly, HLA A-C (Fig. 9) shows the same effect of holding time response as for CD9. The development of future cell lines for therapy will require in each case an understanding of the relationship between the operating conditions used and the resulting properties of the cells. Most importantly, to ensure process robustness is secured, it is necessary to understand degree of control needed for each operating condition to ensure a manufacturing process meets necessary regulatory standards. This paper demonstrates that information captured from individual trials may be combined with the use of ultra scale-down methods to study the processibility of the cells and the impact on cell quality. Each cell processing stage will impose its own process stress on the cells, e.g., mixing, vialling, centrifugation, filtration etc. Individual cell lines will demonstrate different responses to process stress making the application of numerical methods such as ANN to draw out data trends an important tool toward discovering robust routes for routine manufacture of cells for therapy. The effect of cell age, i.e., increase of PNF and GNF , is to make cells more robust to the effect of processing in the formulation medium which really means a reduced tendency to lose membrane integrity or surface marker, although for CD147 the effect is an increased level of marker present or available for assay. 6. Conclusions The use of ANNs has been demonstrated to help gain insight into the effect of cell preparation conditions on their processibility during formulation stages. Increased confidence of data fitting is achieved by taking into account cell passage and generation numbers, the time used for the final passage stage and the hold time in formulation medium. The trends observed provide insight into the design constraints of how cells are prepared for bioprocessing as determined by the extent and duration of hydrodynamic stress in a formulation medium. Such insight information gained at the small scale can help in understanding the contribution of operating variables in cell processing procedures and also in design consideration for full-scale manufacture. List of symbols
DT ε GN HT N [N] P PN ST tD TE
doubling time maximum power dissipation generation number holding time number of cells cell density or concentration proportion of passage time for cells in stationary phase passage number time of exposure to hydrodynamic stress time between cell detachments/at start and end of passage extended time in exponential phase
Subscripts f final for a particular passage stage F final passage stage i passage stage initial for a particular passage stage 0 working cell bank WCB
Acknowledgments Funding provided by the UK Technology Strategy Board High Value Manufacturing Initiative (TP/7/B10/6/I/M0011 (P550)) for Gopal Krishna Dhondalay and Katherine Lawrence is gratefully acknowledged as is also resource provided by the EPSRC Innovative Manufacturing Research Center for Bioprocessing (EP/I033270/1) hosted at University College London. We acknowledge Ryan McCoy (funded by EPSRC Industrial Doctoral Training Center in Bioprocessing (EP/G034656/1)) who supplied the original data [8] for analysis in this paper. References [1] Brandenberger R, Burger S, Campbell A, Fong T, Lapinskas E, Rowley JA. Cell therapy bioprocessing. BioProcess Int 2011:30–7. [2] Eaton JD, Perry MJ, Nicholson S, Guckian M, Russell N, Whelan M, et al. Allogeneic whole-cell vaccine: a phase I/II study in men with hormone-refractory prostate cancer. BJU Int 2002;89:19–26. [3] Staveley-O’Carroll K, Sotomayor E, Montgomery J, Borrello I, Hwang L, Fein S, et al. Induction of antigen-specific T cell anergy: an early event in the course of tumor progression. Proc Natl Acad Sci U S A 1998;95:1178–83. [4] Michael A, Ball G, Quatan N, Wushishi F, Russell N, Whelan J, et al. Delayed disease progression after allogeneic cell vaccination in hormone-resistant prostate cancer and correlation with immunologic variables. Clin Cancer Res 2005;11:4469–78. [5] Copier J, Ward S, Dalgleish A. Cell based cancer vaccines: regulatory and commercial development. Vaccine 2007;25(Suppl. 2):B35–46. [6] Ochsenbein AF, Klenerman P, Karrer U, Ludewig B, Pericin M, Hengartner H, et al. Immune surveillance against a solid tumor fails because of immunological ignorance. Proc Natl Acad Sci U S A 1999;96:2233–8. [7] Acosta-Martinez JP, Papantoniou I, Lawrence K, Ward S, Hoare M. Ultra scaledown stress analysis of the bioprocessing of whole human cells as a basis for cancer vaccines. Biotechnol Bioeng 2010;107:953–63. [8] McCoy R, Hoare M, Ward S. Ultra scale-down studies of the effect of shear on cell quality; processing of a human cell line for cancer vaccine therapy. Biotechnol Progr 2009;25:1448–58. [9] McCoy R, Ward S, Hoare M. Sub-population analysis of human cancer vaccine cells – ultra scale-down characterization of response to shear. Biotechnol Bioeng 2010;106:584–97. [10] Stadnyk AW. Cytokine production by epithelial cells. FASEB J 1994;8: 1041–7. [11] Petricciani J, Egan W, Vicari G, Furesz J, Schild G. Potency assays for therapeutic live whole cell cancer vaccines. Biologicals 2007;35:107–13. [12] Barry FP, Murphy JM. Mesenchymal stem cells: clinical applications and biological characterization. Int J Biochem Cell Biol 2004;36: 568–84. [13] Bubela T, Reshef A, Li MD, Atkins H, Caulfield T, Culme-Seymour E, et al. Enabling advanced cell therapies (EnACT): invitation to an online forum on resolving barriers to clinical translation. Regen Med 2012;7:735–40. [14] Carmen J, Burger SR, McCaman M, Rowley JA. Developing assays to address identity, potency, purity and safety: cell characterization in cell therapy process development. Regen Med 2012;7:85–100. [15] Diogo MM, da Silva CL, Cabral JM. Separation technologies for stem cell bioprocessing. Biotechnol Bioeng 2012;109:2699–709. [16] Mason C, Hoare M. Regenerative medicine bioprocessing: the need to learn from the experience of other fields. Regen Med 2006;1:615–23. [17] Papadaki M, McIntire LV, Eskin SG. Effects of shear stress on the growth kinetics of human aortic smooth muscle cells in vitro. Biotechnol Bioeng 1996;50:555–61. [18] Esquenet M, Swinnen JV, Heyns W, Verhoeven G. LNCaP prostatic adenocarcinoma cells derived from low and high passage numbers display divergent responses not only to androgens but also to retinoids. J Steroid Biochem Mol Biol 1997;62:391–9. [19] Wall ME, Bernacki SH, Loboa EG. Effects of serial passaging on the adipogenic and osteogenic differentiation potential of adipose-derived human mesenchymal stem cells. Tissue Eng 2007;13:1291–8. [20] Al-Rubeai M, Emery AN, Chalder S, Goldman MH. A flow cytometric study of hydrodynamic damage to mammalian cells. J Biotechnol 1993;31:161–77. [21] Born C, Zhang Z, Al-Rubeai M, Thomas CR. Estimation of disruption of animal cells by laminar shear stress. Biotechnol Bioeng 1992;40:1004–10. [22] Tanzeglock T, Soos M, Stephanopoulos G, Morbidelli M. Induction of mammalian cell death by simple shear and extensional flows. Biotechnol Bioeng 2009;104:360–70. [23] Godoy-Silva R, Mollet M, Chalmers JJ. Evaluation of the effect of chronic hydrodynamical stresses on cultures of suspensed CHO-6E6 cells. Biotechnol Bioeng 2009;102:1119–30. [24] Goodacre R, Kell DB. Pyrolysis mass spectrometry and its applications in biotechnology. Curr Opin Biotechnol 1996;7:20–8. [25] Khan J, Wei JS, Ringner M, Saal LH, Ladanyi M, Westermann F, et al. Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med 2001;7:673–9.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
G Model ARTMED-1352; No. of Pages 9
ARTICLE IN PRESS G.K. Dhondalay et al. / Artificial Intelligence in Medicine xxx (2014) xxx–xxx
[26] Ball G, Mian S, Holding F, Allibone RO, Lowe J, Ali S, et al. An integrated approach utilizing artificial neural networks and SELDI mass spectrometry for the classification of human tumours and rapid identification of potential biomarkers. Bioinformatics 2002;18:395–404. [27] Mian S, Ball G, Hornbuckle J, Holding F, Carmichael J, Ellis I, et al. A prototype methodology combining surface-enhanced laser desorption/ionization protein chip technology and artificial neural network algorithms to predict the chemoresponsiveness of breast cancer cell lines exposed to paclitaxel and doxorubicin under in vitro conditions. Proteomics 2003;3:1725–37. [28] Kothari SC, Oh H. Neural networks for pattern recognition. Adv Comput 1993;37:119–66. [29] McCartney S, Weltin M, Burchiel KJ. Use of an artificial neural network for diagnosis of facial pain syndromes: an update. Stereotact Funct Neurosurg 2014;92:44–52. [30] Hu X, Cammann H, Meyer HA, Miller K, Jung K, Stephan C. Artificial neural networks and prostate cancer – tools for diagnosis and management. Nat Rev Urol 2013;10:174–82. [31] Nakajima K, Nakajima Y, Horikoshi H, Ueno M, Wakabayashi H, Shiga T, et al. Enhanced diagnostic accuracy for quantitative bone scan using an artificial neural network system: a Japanese multi-center database project. EJNMMI Res 2013;3:83. [32] Lancashire LJ, Powe DG, Reis-Filho JS, Rakha E, Lemetre C, Weigelt B, et al. A validated gene expression profile for detecting clinical outcome in breast cancer using artificial neural networks. Breast Cancer Res Treat 2010;120:83–93. [33] Lowery AJ, Miller N, Devaney A, McNeill RE, Davoren PA, Lemetre C, et al. MicroRNA signatures predict oestrogen receptor, progesterone receptor and HER2/neu receptor status in breast cancer. Breast Cancer Res 2009;11: R27. [34] Fonseca R, Vabulas RM, Hartl FU, Bonhoeffer T, Nagerl UV. A balance of protein synthesis and proteasome-dependent degradation determines the maintenance of LTP. Neuron 2006;52:239–45.
9
[35] Thi MM, Kojima T, Cowin SC, Weinbaum S, Spray DC. Fluid shear stress remodels expression and function of junctional proteins in cultured bone cells. Am J Physiol Cell Physiol 2003;284:C389–403. [36] Walpola PL, Gotlieb AI, Cybulsky MI, Langille BL. Expression of ICAM-1 and VCAM-1 and monocyte adherence in arteries exposed to altered shear stress. Arterioscler Thromb Vasc Biol 1995;15:2–10. [37] Gruen JR, Weissman SM. Evolving views of the major histocompatibility complex. Blood 1997;90:4252–65. [38] Vega MA, Strominger JL. Constitutive endocytosis of HLA class I antigens requires a specific portion of the intracytoplasmic tail that shares structural features with other endocytosed molecules. Proc Natl Acad Sci U S A 1989;86:2688–92. [39] Higginbottom A, Takahashi Y, Bolling L, Coonrod SA, White JM, Partridge LJ, et al. Structural requirements for the inhibitory action of the CD9 large extracellular domain in sperm/oocyte binding and fusion. Biochem Biophys Res Commun 2003;311:208–14. [40] Ikeyama S, Koyama M, Yamaoko M, Sasada R, Miyake M. Suppression of cell motility and metastasis by transfection with human motility-related protein (MRP-1/CD9) DNA. J Exp Med 1993;177:1231–7. [41] Masellis-Smith A, Shaw AR. CD9-regulated adhesion. Anti-CD9 monoclonal antibody induce pre-B cell adhesion to bone marrow fibroblasts through de novo recognition of fibronectin. J Immunol 1994;152:2768–77. [42] Guo H, Majmudar G, Jensen TC, Biswas C, Toole BP, Gordon MK. Characterization of the gene for human EMMPRIN, a tumor cell surface inducer of matrix metalloproteinases. Gene 1998;220:99–108. [43] Muramatsu T, Miyauchi T, Basigin. (CD147): a multifunctional transmembrane protein involved in reproduction, neural function, inflammation and tumor invasion. Histol Histopathol 2003;18:981–7. [44] Nabeshima K, Lane WS, Biswas C. Partial sequencing and characterization of the tumor cell-derived collagenase stimulatory factor. Arch Biochem Biophys 1991;285:90–6.
Please cite this article in press as: Dhondalay GK, et al. Relationship between preparation of cells for therapy and cell quality using artificial neural network analysis. Artif Intell Med (2014), http://dx.doi.org/10.1016/j.artmed.2014.07.003
