On-Line Process Control of Gradient Elution Liquid Chromatography

Jun 30, 2004 - The data are analyzed in real time with chemometrics to produce easily interpreted control charts. Gradient elution LC is extensively e...
0 downloads 0 Views 352KB Size
Anal. Chem. 2004, 76, 4875-4880

Technical Notes

On-Line Process Control of Gradient Elution Liquid Chromatography Leif Schweitz,*,† Magnus Fransson,† Lars Karlsson,† Arne Torstensson,† and Erik Johansson‡

Analytical Development, AstraZeneca R&D Mo¨lndal, S-431 83 Mo¨lndal, Sweden, and Umetrics AB, S-907 19 Umeå, Sweden

The typical measure of the stability of analytical HPLC methods in the pharmaceutical laboratory is standards injected repeatedly throughout the sample sequence. To obtain improved control of the analysis and reduction of the number of standards and replicates, a novel approach to treat the analytical run as a process, where the chromatographic data is the product, is proposed. Thus, an alternative and continuous system suitability test procedure is described. This is obtained by continuous monitoring of several parameters of the chromatographic system such as pressure, temperature, and conductivity. The data are analyzed in real time with chemometrics to produce easily interpreted control charts. Gradient elution LC is extensively employed in pharmaceutical analysis. A gradient elution system is inherently dynamic due to the mobile-phase composition being changed during the chromatographic run. To handle the dynamics, suitable chemometric tools are needed. In this report, we extend the use of liquid chromatography process control (LCPC) to gradient elution LC by creating partial least-squares regression batch models of the data collected. The gradient elution LCPC system was evaluated by inducing disturbances, and it was shown to easily detect any real or simulated deviation. In the pharmaceutical industry, liquid chromatography (LC) is most often the technique of choice for the analysis of drugs and related substances. The technique is used extensively in discovery, development, and production. Especially when employed in a GLP or GMP area, the requirements on method ruggedness and system stability are very strict as set in directives issued by regulatory authorities. To conform to international regulations and ensure analytical quality, various measures are taken in the respective analytical laboratories. First, in terms of instrumentation, the various units of the chromatographic system are checked at regular intervals. Second, the developed LC methods are thoroughly validated. Third, prior to initiation of an analytical sequence, a system suitability test is performed to * Corresponding author. Phone: +46 31 7762997. Fax: + 46 31 7763768. E-mail: [email protected]. † AstraZeneca R&D Mo ¨lndal. ‡ Umetrics AB. 10.1021/ac0497050 CCC: $27.50 Published on Web 06/30/2004

© 2004 American Chemical Society

confirm that the equipment is operating properly in order to reach (most often) separation criteria as defined in the analytical method. Finally, and of primary importance to this paper, during an actual run, system stability is only indirectly tested through reference solutions analyzed repeatedly throughout the sample sequence. As a consequence, the ratio of references to actual samples is often very high. It is not uncommon that every second injection is a reference solution. In a previous report,1 we presented an alternative approach to monitor LC system stability during the actual run. The liquid chromatography process control (LCPC) system entails monitoring of certain key apparatus-related physical parameters, such as pressures, temperature, and conductivity of the mobile phase, in combination with multivariate statistical process control (MSPC). Although most commonly used for monitoring of industrial processes,2-7 there have been earlier reports on the use of MSPC for the surveillance of analytical systems.8-10 These were, however, all based on monitoring the analytical signals of either the analyte itself8,9 or of a reference, a so-called “deferred standard”.10 To the best of our knowledge, our setup was the first monitoring approach of an analytical system where instrument-related parameters were modeled. The LCPC system treats the chromatographic analysis as a process, with the chromatographic data as the product. To ensure high quality of the product, focus is thus not only put on already existing postrun checks of retention time, plate numbers, resolution, and peak areas but rather on monitoring of the process itself. The parameters of potential interest are continuously logged throughout the sample (1) Fransson, M.; Spare´n, A.; Lagerholm, B.; Karlsson, L. Anal. Chem. 2001, 73, 1502-1508. (2) Nomikos, P.; MacGregor, J. F. Chemom. Intell. Lab. Syst. 1995, 30, 97108. (3) Harnett, M.; Lightbody, G.; Irwin, G. W. Analyst (Cambridge, U. K.) 1996, 121, 749-754. (4) Skagerberg, B.; MacGregor, J. F.; Kiparissides, C. Chemom. Intell. Lab. Syst. 1992, 14, 341-356. (5) Lepeniotis, S. S.; Vigezzi, M. J. Chemom. Intell. Lab. Syst. 1995, 29, 133139. (6) Hodouin, D.; MacGregor, J. F.; Hou, M.; Franklin, M. CIM Bull. Miner. Proc. 1993, 86, 23-34. (7) MacGregor, J. F. Adv. Control Chem. Processes 1994, 427-437. (8) Nijhuis, A.; de Jong, S.; Vandeginste, B. G. M. Chemom. Intell. Lab. Syst. 1997, 38, 51-62. (9) Nijhuis, A.; de Jong, S.; Vandeginste, B. G. M. Chemom. Intell. Lab. Syst. 1999, 47, 107-125. (10) Guillemin, C. L. Chemom. Intell. Lab. Syst. 1992, 17, 201-211.

Analytical Chemistry, Vol. 76, No. 16, August 15, 2004 4875

Figure 1. Liquid chromatography process control instrument setup with the sensors employed for data collection indicated. Details described in the Experimental Section.

sequence, thus creating a continuous system suitability test. It was argued1 that not only can the number of reference solution injections potentially be reduced, process monitoring of the analytical system also enables an improved quality of the data obtained. However, the approach was at the time limited to isocratic runs, whereas many LC methods are performed in the gradient mode. Hence, the LCPC approach needs to be further developed to also include gradient runs. In this paper, we present a model example on how a continuous system suitability test can be implemented on a gradient LC run using the LCPC system. EXPERIMENTAL SECTION Apparatus. For a schematic description of the LCPC setup, see Figure 1. The LC system consisted of a PerkinElmer Series 200 pump (PerkinElmer Instruments, Norwalk, CT), a Waters 717+ autosampler, a Waters Symmetry C18 precolumn, a Waters Symmetry C18, 3.5-µm, 75 × 4.6 mm column (all from Waters Co, Milford, MA), and a Kratos Spectroflow 783 UV detector (Kratos Analytical, Ramsey, NJ). Chromatographic data were collected with a Waters BusSAT/IN module connected to a Millennium32 acquisition server. The pressure sensors were acquired from Entran Devices, Inc. (Fairfield, NJ): two pressure transducers model EPXO-X22-3KP (P1 and P2) and one transducer model EPXO-X22-150P (P3). The temperature sensor was an IPAQ-2 from INOR Process AB (Malmo¨, Sweden). Conductivity data were collected with a Crison Conductimeter 525 (Crison Instruments S.A., Barcelona, Spain). The reference lamp energy was accessed via the remote/test connector on the back of the UV detector. The A/D converter was a NI-DAQ PC-LPM-16/PnP with a CB-50 cable accessory, all from National Instruments (Austin, TX). The PC was a Compaq Deskpro EN series 6400 (Compaq Computer Corp., Houston, TX). 4876

Analytical Chemistry, Vol. 76, No. 16, August 15, 2004

Software. Millennium32 version 3.05.01 was from Waters. LabVIEW version 5.0 was from National Instruments. SIMCA-P+ v 10.0 was from Umetrics (Umeå, Sweden). PLS_Toolbox 2.0 was from Eigenvector Research Inc.. Chromatography. The 0.05 M ammonium dihydrogen phosphate (Merck) adjusted to pH 3.1 with 1 mol/L phosphoric acid was used as the buffer component and acetonitrile (Merck) as the organic modifier in the mobile phases used. The gradient program was as follows (if not stated otherwise): (A) phosphate buffer/acetonitrile (92/8 v/v); (B) phosphate buffer/acetonitrile (55/45 v/v). Start: 100% A, after 3 min linearly changed to 100% B during 9 min and held at 100% B for additional 3 min. Then, during 0.5 min, the mobile phase was linearly changed back to 100% A and held for 11.5 min. The flow was constant at 1.5 mL/min. A 20-µL aliquot of a model sample containing a drug substance and known related substances was injected. Procedure. To monitor and collect data, the software developed previously1 was used. Data were only collected during the elution phase (thus none during the injection phase) of the chromatography process. Data from 50 chromatographic runs were collected and constitute the data set for an unfolded PLS batch model of the gradient LCPC system. The data collected from the lamp reference voltage were, prior to chemometric treatment, batch wise smoothed using the savgol m-file in PLS_Toolbox 2.0 (Eigenvector Research Inc.). This m-file uses the well-known Savitsky-Golay algorithm, here with a window setting of 51. Extraordinary events such as modifications of the gradient, the mobile-phase flow, temperature changes, and mobile-phase leakage were induced to evaluate the performance of the gradient LCPC system. All chemometric treatment was performed using the SIMCA-P+ v 10.0 software (Umetrics AB).

Table 1. Summary of Models

observation level batch level

model

number of components

R2X

R2Y

PLS (phase 1) PLS (phase 2) PLS (phase 3) PLS (phase 4) PCA

2 2 2 2 5

0.936 0.969 0.932 0.938 0.914

0.643 0.978 0.977 0.563

Q2 (sum) 0.643 0.978 0.976 0.563 0.88

RESULTS AND DISCUSSION Gradient LCPC System. To study and evaluate a gradient LC system using LCPC, a number of sensors were connected to

the LC to monitor a variety of parameters. In the present example, three pressure sensors were placed as follows: after the pump and prior to the injector, after the injector and prior to the column, and after the analytical column and prior to the UV detector (Figure 1). Furthermore, a temperature sensor was placed on the analytical column since temperature changes will affect the solute partitioning to the column’s stationary phase and thus affect the separation. The reference voltage of the UV detector lamp was collected in order to detect lamp disturbances or failure. A conductivity probe was placed in the flow stream after the UV detector to monitor changes in mobile-phase composition. It should be noted that this model system could easily be expanded by implementing any number of additional sensors.

Figure 2. Batch-level multivariate statistical control charts revealing chromatographic runs (batches) that exceed the limit set by the user. Any run exceeding the limit will be regarded as outside the model and will alert the analyst that the run is deviating from the normal, indicating a problem or error during the run. Several induced disturbances were performed including the following; a positive and negative 30-s shift of the gradient start (early gradient and late gradient, respectively), gradient evolving nonlinearly with gradient curve values of +1.5 (positive curve) and -1.5 (negative curve), flow rate changed with a 7% increase or decrease until gradient starts to evolve (high flow and early low flow, respectively), temperature change of the column by holding it in one’s hand for 30 s (temperature disturbance), a minute leakage induced by loosening a nut at a connection between the pump and the column (leakage), and a chromatographic run with no gradient (isocratic elution). All chromatographic runs with induced disturbances are indicated as outliers in the DModX plot (component 5) (A) as well as in the Hotellings T2 plot (components 1-5) (B) (data not shown for the runs with the leakage and isocratic elution, respectively, since these exceed the limits extensively in both plots). A batch-level score plot, in this case of components 1 and 4 (C) may alternatively illustrate the designated chromatographic runs with deliberate disturbances induced as compared to the model batches. The ellipse is the Hotelling’s 95% critical limit for the batch level model. It should be noted that runs situated outside the ellipse are outliers, but this does not imply that these runs are by default erroneous. Statistically, 5% of all batches should be outside the ellipse. Thus, all plots indicate the runs with induced disturbances as outliers and strongly suggest them to be erroneous. In all cases, however, to determine whether a batch is a true outlier or not, the data need a closer examination.

Analytical Chemistry, Vol. 76, No. 16, August 15, 2004

4877

Figure 3. Observed vs predicted time plot (component 2, average removed) of all phases corresponding to a whole chromatographic run in which a deliberate shift to an earlier time point of the gradient was introduced. The deviation from a “normal” run is noticeable at the end of phase 2 where the graph exceeds the (3 standard deviations control limit (red lines). The graph remains outside the control limit throughout phase 3 until the system returns to normal again in the beginning of phase 4.

Chemometric Modeling. In the previously reported isocratic LCPC system,1 the elution phase was considered a fairly static process and continuously treated with principal component analysis (PCA). In this example, however, the elution phase is not static but dynamic since the mobile-phase composition changes during the run. The data acquired were put into a threeway matrix with sensor data (K) in the X-direction, run time (J) in the Y-direction, and the different chromatographic runs or batches (N) in the Z-direction. The three-way matrix (K × J × N) was then unfolded along the batch direction to get a two-way matrix with N × J rows and K columns. Thus, each row, hereafter called observation, is the sensor data acquired at a certain time point j for a specific batch n.11 This unfolded two-way matrix was then subjected to partial least-squares (PLS) regression batch modeling in Simca-P+, where batch time was used as the Y-vector in the regression. Each batch is represented by 359 observations, 1 observation for each sampled time point, stacked on top of each other. The observations were divided into four classes, where the points of division were derived from the different phases of the gradient curve. A separate PLS class model was fitted for each phase. The modeling was done at two different levels: the observation level and the batch level. At the observation level, after fitting a PLS model to the data from each phase, the scores reflect the change of sensor data over time. At the batch level, each batch is represented by one observation. The variables are the scores from the models at the observation level, put side by side. For example, the model for phase 1 has two PLS components and is 120 time points long. Thus, the number of variables from this phase is 120 from component 1 + 120 from component 2 ) 240. The sum total of variables at the batch level will then be as follows: (120 + 120) + (130 + 130) + (20 + 20) + (89 + 89) ) 718. A PCA model is then fitted (Table 1). Using these models, new observations are predicted and the results presented in standard MSPC charts, e.g., DModX (distance to model plane) and Hotelling’s T2 (distance to model center within the model plane). (11) Wold, S.; Geladi, P.; Esbensen, K.; Oehman, J. J. Chemom. 1987, 1, 4156.

4878

Analytical Chemistry, Vol. 76, No. 16, August 15, 2004

LCPC System Evaluation. A series of experiments to evaluate the gradient LCPC system was carried out. During these experiments, measures were taken to simulate different kinds of disturbances. All runs with simulated disturbances were immediately detected as outliers in the batch level control charts (Figure 2). A set of experiments to simulate disturbances in the time point where the gradient starts to evolve were conducted. Here the B solvent starts delivering at 0.5 min too early and too late, respectively, as compared to the primary gradient program (see Experimental Section above). This deviation results in a shifted gradient, which obviously will alter the retention of components in the sample. This is readily recognized in the observed versus predicted time plot (Figure 3). To further elucidate the LCPC system’s ability to detect deviations in the gradient performance of the chromatographic process, experiments were performed where the flow rate was altered prior to the evolvement of the gradient. A 7% increase and decrease in the flow rate, respectively, was introduced to simulate poor pump performance. The flow rate was thus changed from 1.5 mL/min to 1.4 and 1.6 mL/min, respectively, and restored to 1.5 mL/min after 3 min. The deviation from the original flow rate was evident in the DModX plots (Figure 4). When a change in the evolvement of the gradient was induced, the change was detected as a deviation in the t1 scores plot (Figure 5) at the observation level. For example, when the linear evolvement of the gradient was changed to nonlinear (with a curve of -1.5), it could be noted as a deviation outside the 3σ limit of the model in phases 2-4 (Figure 5). This change was not as obvious in the DModX plot. In one instance, the pump nondeliberately failed to deliver the gradient and, thus, performed an isocratic elution; the LCPC system clearly recognized the deviation (Figure 2). To assess the sensitivity of the system, the temperature of the column was raised ∼2 °C when the analyst held the column in his hand for a few seconds. The response was intense, and the deviation was clearly seen in the Hotelling’s T2 plot (Figure 6A). When the temperature returns to normal, the signal, again, indicates that the system performs normally. If a temperature

Figure 4. DModX plot (component 2) of all the phases of the gradient showing a clear discrepancy from the model (green line, model average; red line, +3 standard deviations) the first 35 time points, corresponding to the first 3 min of the chromatographic run in which a deliberate increase (A) and decrease (B) in the flow rate was induced. This error will result in erroneous retention times of the separated compounds, which will normally only be evident for the analyst when the run is complete and the chromatographic peaks are evaluated.

Figure 5. t1 scores plot at the observation level over four phases of a chromatographic run in which the primary linear gradient was altered to evolve nonlinearly (curve -1.5). The deviation is first noted in phase two where the run progress outside the limit ((3 standard deviations) of the model.

disturbance like this was to occur during a routine run, it could be concluded that it was short in time and arguably only had minute influence of the overall performance of the chromatographic run. A major and disastrous event such as a leakage was clearly recognized instantaneously in all MSPC charts. In the batch level t1/t4 scores plot, for example, the run is noted as an extreme

outlier (Figure 2). It should be noted that the leakage was so small that it could only be detected by the analyst as a liquid droplet after 0.5 h of leaking. Thus, the analyst was alerted rapidly when employing the LCPC system. The LCPC approach presented here also offer troubleshooting facilities as exemplified in Figure 6. The reason for the deviation in Figure 6A can be elucidated in terms of observation contribuAnalytical Chemistry, Vol. 76, No. 16, August 15, 2004

4879

Figure 6. Small temperature increase of the column indicated at the observation level in the Hotelling’s T2 plot (A). The temperature deviation is easily recognized as a peak-shaped event in the middle of phase 2. To elucidate the reason for the deviation from the “normal”, a contribution plot at time point 78 (peak apex in A) was obtained (B). It is clearly shown that the contribution for the deviating point is almost solely from the response of the temperature sensor. Thus, it can be concluded from these two plots that a rather transient temperature deviation occurred during this chromatographic run.

tion. The contribution plot (Figure 6B) derived from the peak apex in Figure 6A reveals that the temperature sensor is mainly responsible for the deviation in this case. Thus, it can be concluded that the temperature deviated from normal during a short period of time. CONCLUSIONS By regarding the chromatographic system as a process and monitoring signals from a number of sensors simultaneously, the LCPC approach allows for an efficient, real-time, continuous system suitability test. This report shows that the LCPC approach also can be adopted to the widely employed gradient elution mode of liquid chromatography. This is indeed a prerequisite for this novel methodology to be widely implemented into the routine analytical separation laboratory. To deal with the dynamic situation a gradient elution by definition provides, PLS regression batch modeling was utilized. The observations (i.e., the sensor data) were divided into phases, following the elution gradient, and a PLS regression model was fitted for each individual phase. A PCA model was then fitted to the scores from the phase models, thus creating a batch-level model where each run is reduced to one observation. This offers the attractive possibility to monitor the chromatographic system at both the observation level and the batch level. High efficiency of the gradient LCPC system was demonstrated by deliberate alterations of the system such as disturbances in

4880

Analytical Chemistry, Vol. 76, No. 16, August 15, 2004

temperature, flow, mobile-phase composition, gradient evolvement, or leakage. All these disturbances cold be detected in the MSPC charts. We believe that the LCPC concept, which is further emphasized in this report, constitutes a novel, attractive monitoring system that ensures, proves, and documents that an analysis is performed with a stable system. Potentially, this will allow a reduction of the number of chromatographic reference solutions included in an sample sequence while increasing the documented quality of the analysis. A more direct range of application would be to utilize the LCPC system, especially with the system’s monitoring ability in mind, in trouble shooting and optimization of robust LC methods. In conclusion, if appropriate sensors can be installed, the LCPC principle can be applied to any chromatographic or chromatography-like separation system which is further emphasized in this report for gradient elution liquid chromatography. Provided calculations and warning limits are adequate, the advantages of the LCPC approach can be fully exploited.

Received for review February 23, 2004. Accepted May 19, 2004. AC0497050