Data fusion of fourier transform infrared spectra and powder x-ray diffraction patterns for pharmaceutical mixtures

Contents lists available at Journal of Pharmaceutical and Biomedical Analysis Data fusion of Fourier transform infrared spectra and powder X-ray diffraction patterns for pharmaceutical mixtures Rahul V. Haware , Patrick R. Wright , Kenneth R. Morris , Mazen L. Hamad a University of Hawaii at Hilo, College of Pharmacy, 34 Rainbow Drive, Hilo, HI 96720, USA b Department of Chemistry, University of Hawaii at Hilo, 200 W Kawili St., Hilo, HI 96720, USA Fusing complex data from two disparate sources has been demonstrated to improve the accuracy in quan- Received 18 May 2011 tifying active ingredients in mixtures of pharmaceutical powders. A four-component simplex-centroid Received in revised form 9 August 2011 design was used to prepare blended powder mixtures of acetaminophen, caffeine, aspirin and ibuprofen.
Accepted 10 August 2011 The blends were analyzed by Fourier transform infra-red spectroscopy (FTIR) and powder X-ray diffrac- Available online 17 August 2011 tion (PXRD). The FTIR and PXRD data were preprocessed and combined using two different data fusion methods: fusion of preprocessed data (FPD) and fusion of principal component scores (FPCS). A partial least square (PLS) model built on the FPD did not improve the root mean square error of prediction.
However, a PLS model built on the FPCS yielded better accuracy prediction than PLS models built on indi- Multivariate analysis Pharmaceutical powder mixtures vidual FTIR and PXRD data sets. The improvement in prediction accuracy of the FPCS may be attributed to Fourier transform infrared spectroscopy the removal of noise and data reduction associated with using PCA as a preprocessing tool. The present Powder X-ray diffraction approach demonstrates the usefulness of data fusion for the information management of large data sets from disparate sources.
2011 Elsevier B.V. All rights reserved.
A second obstacle associated with using high data density tools arises when these techniques are used to make measurements on The need to understand the critical material and process samples that are considered non-ideal. An example of a non-ideal attributes on the end product quality of pharmaceutical products sample for FTIR is a heterogeneous, multi-component, solid state is now an imperative with respect to the ICH Q8 guideline issued pharmaceutical mixture. Ideally, the absorbance of infrared light by FDA Consequently, quantitative and qualitative applica- at a particular wavenumber will be directly proportional to the tions of sophisticated high data density analytical tools like Fourier concentration of each absorbing species; however, the variation transform infrared spectroscopy (FTIR), powder X-ray diffrac- in the extent of light scattering at particulate interfaces tends to tion (PXRD), Raman spectroscopy and near infrared spectroscopy significantly increase the error in the measurement. This artefact, have gained wider acceptance in characterizing pharmaceutical caused by variation in the physical aspects of the sample matrix, processes, intermediates and products major obstacle asso- may prevent the use of FTIR as a technique for the quantitative ciated with these analytical techniques is the generation of large characterization of multi-component solid state pharmaceutical data matrices which may be complex and difficult to interpret.
Thus, it is critical that the end users of these tools have appropriate PXRD, on the other hand, is a technique known for structural methods of data reconciliation in order to extract the sought after characterization, not chemical characterization, of single compo- information for subsequent prediction of the process outcomes.
nent solid-state material samples. Thus, neither method is ideal Multivariate analysis, also called chemometrics when applied to for the quantitative chemical characterization of multi-component, chemical-specific applications, has been offered as one key to pharmaceutical samples. However, since the two techniques yield extracting critical information from large data sets generated by different kinds of information, their individual data sets can be com- a single high data density tool.
bined into a single data set which provides more information than either technique by itself. What is not clear is how data from the two techniques can be combined to better perform a single task than either technique could perform on its own.
An important technique emerging from the informatics domain Corresponding author. Tel.: +1 808 933 2194; fax: +1 808 974 7693.
is data fusion. The aim of data fusion is to facilitate the faultless E-mail address: (M.L. Hamad).
1 Student undertaking summer internship from Albert-Ludwigs-University integration of information from various sources to develop a single Freiburg, Freiburg, Germany.
model or decision is hypothesized that data fusion may be a 0731-7085/$ – see front matter 2011 Elsevier B.V. All rights reserved.
R.V. Haware et al. / Journal of Pharmaceutical and Biomedical Analysis 56 (2011) 944–949 equal proportions of the selected two components, 3 face centers Four component simplex-centroid experimental design in units of percentage con- of the ternary mixtures with equal proportions of the selected three centration [acetaminophen (APAP), caffeine (CAF), ibuprofen (IBU) and aspirin components, 4 axes of the quaternary mixtures with varying pro- portions of the selected four components and 3 center experiments of the quaternary mixtures with equal proportions of the selected four components to check both the linearity and repeatability of the experimental results.
2.3. Fourier transform infrared spectroscopy (FTIR) The spectra of the 21 calibration samples and 4 unknown samples were collected using a Thermo Nicolet NEXUS 670 FTIR instrument equipped with a Nicolet Smart MIRacle accessory (Thermo Fisher Scientific, Waltham, MA). The MIRacle accessory uses a glassy material known as AMTIR (Amorphous Material Trans- mitting Infrared Radiation – composed of Ge, As, and Se) to measure the absorbance in the attenuated total reflectance (ATR) mode.
The samples were measured by inserting approximately 25 mg of mixed sample powder into the trough insert and supplying suffi- cient pressure using the micrometer pressure clamp to compress the sample against the AMTIR glass. For each sample, 32 scans in the wavenumber range from 650 cm−1 to 4000 cm−1 (at a resolution of 4 cm−1) were averaged to produce a single spectrum. The resulting spectral data vectors contained 1738 data points. The spectral data useful strategy to integrate data from FTIR and PXRD for the char- were acquired in absorbance mode using OMNIC software (Thermo acterization of multi-component, pharmaceutical samples. Various Fisher Scientific, Waltham, MA) and exported to MATLAB® (Math- scientific and engineering disciplines, such as robotics, remote works, Natick, MA) and the Unscrambler® (Unscrambler® 10.0.1, sensing, image analysis, and analytical chemistry, are employing CAMO AS, Norway) for data processing.
data fusion concepts and realizing better information management data fusion coupled with multivariate analysis (MVA) is a new and potentially very powerful approach to modeling 2.4. Powder X-ray diffraction (PXRD) The goal of the present work was to investigate the suitability of The PXRD data were collected for all experiments that were data fusion combination with MVA methods to build more conducted on a Bruker D8 Advanced system in Bragg-Brentano accurate predictive models. Principal component analysis (PCA) geometry using a Cu K␣ radiation point source ( = 1.5406 ˚A) at was used for exploratory data analysis and it was also used as a data an operating voltage and amperage of 40.0 kV and 40.0 mA, respec- reduction technique prior to data fusion. Partial least square (PLS) tively. The powdered samples were analyzed in a low background regression was ultimately used to build predictive models based on cell. The samples (approx. 25 mg) were scanned at a rate of 0.005◦ the FTIR data set, the PXRD data set, the data set prepared by fusion per minute at step size of 0.01◦ from 5◦ to 35◦ 2, resulting in row of preprocessed data (FPD) and the data set prepared by fusion vectors of 2894 data points for each sample. The obtained PXRD of principal component scores (FPCS). The quantitative prediction data was exported to Unscrambler® prior to MVA modeling and accuracy of fractions of acetylsalicylic acid, caffeine, ibuprofen, and data fusion.
acetaminophen in blended powder samples was compared using leave-one-out cross validation. The models were also used to pre- dict fractions of the four components in blind, unknown powder The data from FTIR and PXRD were combined using two differ- ent fusion methods: fusion of preprocessed data (FPD) and fusion 2. Experimental and methods
of principal component scores (FPCS). It was necessary to prepro- cess each set of data individually prior to data fusion. Without preprocessing, the scales for each data set would have been dra- matically different and this would have caused inappropriate and Acetylsalicylic acid (ASA) was purchased from Alfa Aesar unequal weighting in the models. Therefore, the FTIR data set and (Ward Hill, MA). Acetaminophen (APAP) was purchased from the PXRD data set were preprocessed using the standard normal Ortho-McNeil Pharmaceuticals (Titusville, NJ). Ibuprofen (IBU) and variate (SNV) function in the PLS Toolbox (Eigenvector Research, caffeine (CAF) were purchased from Spectrum Chemical (Gardena, Inc., Wenatchee, WA, USA). The SNV function standardizes the row vectors to mean zero and unit variance. Additionally, the CO2 peak, including data from 2268 cm−1 to 2402 cm−1 was removed 2.2. Experimental design from each FTIR spectrum and the FTIR data above 3377 cm−1 were removed since they did not contain any useful information. After Four active ingredients (APAP, ASA, CAF and IBU) were tested removal of these data points, the FTIR spectra contained 1346 data by a four-component simplex-centroid design (SCD). The four- points. After preprocessing, the data were considered normalized, component SCD was used to achieve better predictability with a allowing the multivariate models to apply appropriate weightings high accuracy of unknown fractions of subjected active ingredients to each variable to yield the most descriptive models. For FPD, the total, 21 combinations of subjected active ingredients were normalized data for each sample was fused by concatenating its tested by both FTIR and PXRD techniques (4 vertices for FTIR row vector with its PXRD row vector, resulting in row vectors 4 pure components, 6 edge centers of the binary mixtures with with 4240 data points.

R.V. Haware et al. / Journal of Pharmaceutical and Biomedical Analysis 56 (2011) 944–949 Fig. 1. Schematic flow of data processing in the PLS prediction of the fusion of principal component scores (FPCS) data set. Blocks in the diagram are not intended to represent
matrix size. Matrix sizes are represented as M rows × N columns = M samples × N variables.
In FPCS, it was the PCA score values that were fused. A PCA was the 21 calibration samples is complex due to the large quantity performed on each individual data set. A PCA was performed on of overlapping data; however, it is instructive to view an overlay the normalized FTIR spectra of the 21 calibration samples and the plot of the fingerprint spectral region of the 4 vertices (i.e. pure 4 unknown samples (a total of 25 samples). The intention of this components). that each of the pure components has at procedure was to extract as much variation from these samples as least one absorption peak with little overlap from the peaks of other possible so the score values of the first 20 principal components pure components. Vertical lines are included in indicate the were saved. Next, a PCA was performed on the PXRD data set and locations of these unique peaks. APAP has a peak at 1562.1 cm−1, the score values for its first 20 principal components were saved.
CAF has a peak at 744.4 cm−1, IBU has a peak at 779.1 cm−1, and The score values from each technique were concatenated into one ASA has a peak at 1562.1 cm−1. This feasibility check shows that fused matrix of 25 rows (samples) by 40 columns (score values).
there is sufficient variation in the FTIR absorbance spectra to enable Finally, this matrix was separated into a calibration matrix (21 rows multivariate quantitative analysis of these 4 components. A simi- by 40 columns) and an unknown matrix (4 rows by 40 columns).
lar analysis was performed on the functional group region of the Thus, the fused data set was condensed from the respectively large FTIR spectra (2400–3450 cm−1). It was found that APAP contained number of variables (4240) to a matrix containing only 40 variables.
a unique peak at 3223 cm−1 representing the N–H vibration of its secondary amine, but the rest of peaks in the functional group 2.6. Multivariate analysis (MVA) region were broader than those in the fingerprint region. There- fore, most of the peaks from a single component in the functional Principal component analysis (PCA) followed by partial least square regressions (PLS-2) were performed on the individual FTIR and PXRD data, as well as on the fused data from both FTIR and PXRD analysis, to check the three dimensional spatial distribution of the score values (Unscrambler® 10.0.1, CAMO AS, Norway). Optimized PLS-2 models were used to predict the unknown concentrations of the ingredients in the mixtures. Leave-one-out cross validation was used to calculate the PLS-2 models best PLS-2 model for the prediction of the concentrations of the ingredients was selected on the basis of yielding the lowest root mean square of cross vali- dation (RMSECV) values.
Finally, the ability of the optimized PLS models based on an individual FTIR data, PXRD data and the fused data were tested by subjecting the data of four unknown samples mixtures of varying compositions of active ingredients. A schematic of the data pro- cessing steps involved in the PLS prediction of the FPCS is shown in 3. Results and discussion
00 12 00 1100 1000 900 3.1. Analysis of calibration samples by FTIR and PXRD Fig. 2. Fingerprint region of the FTIR spectra of APAP, CAF, IBU, and ASA. Vertical
The FTIR spectra of the 21 calibration samples were obtained, lines show locations of peaks for single components with little overlap from other as indicated in Section overlay plot of the FTIR spectra of R.V. Haware et al. / Journal of Pharmaceutical and Biomedical Analysis 56 (2011) 944–949 Normalized Intensity Fig. 4. The fused FTIR spectra and PXRD patterns. The first 1346 data points rep-
Fig. 3. Powder X-ray diffraction patterns of the 4 pure components: APAP, CAF, IBU,
resent the FTIR spectra and data points 1347–4240 represent the PXRD patterns.
and ASA. Vertical lines show locations of peaks for single components with little SNV was applied separately to each data set prior to concatenation. The tallest ASA overlap from other components. The ASA peak at 2 = 15.5◦ was cut off to provide peak (18.5 normalized intensity units) was cut off to provide a better view of the an enlarged view of the majority of peaks in the pattern.
remaining peaks.
group region shared overlap with peaks from at least one of the set is compressed into 3 principal components (PCs). Second, the other pure components.
pure component samples (labeled as 1-APAP, 2-CAF, 3-IBU, and 4- A similar feasibility check was performed on the PXRD patterns ASA) align themselves at the vertices of a trigonal pyramid. Third, of the four pure components. the PXRD patterns for each of the calibration samples aligns itself at the expected location each of the pure components. Vertical lines are included in within the three dimensional structure. For example, sample #8 is to point out that there is at least one peak for each component 50% (2-CAF) and 50% (3-IBU); thus aligning itself halfway between with little overlap from the peaks of other components. APAP has samples 2 and 3 while not showing any orientation with samples 1 a peak at 2 = 24.2067◦, CAF has a peak at 2 = 11.6477◦, IBU has and 4. Furthermore, the center-points (samples 19, 20, and 21) are a peak at 2 = 18.565◦, and ASA has a peak at 2 = 7.6964◦. Again, precisely overlapping one another and located at the center of the this feasibility check shows there is sufficient variation in the PXRD pattern to enable multivariate quantitative analysis of the four pure shows the resulting three-dimensional PCA plots for the FTIR data sets for the 21 calibrations samples. A similar trigonal Since it was found that each technique could measure variation pyramidal structure emerges, however, the error in the measure- in each of the four components, the next step was to fuse the data.
ment is evident as the calibration samples do not align themselves In FPD, the data sets were concatenated after preprocessing each at the exact locations of the pyramid that would be expected. For individual data set with SNV. The resulting overlay plots of the four example, sample 7, which is 50% (1-APAP) and 50% (4-ASA), is found pure components for the normalized and fused data vectors are somewhat half way between samples 1 and 4, but it is now located shown in Data points 1–1346 represented the FTIR spectra off of the line connecting vertex points 1 and 4. shows the and data points 1347–4240 represented the PXRD patterns. While corresponding PCA plot for the PXRD patterns for the 21 calibration there were more data points in the PXRD data, there was a similar samples. Again, the same pattern emerges and it appears there is degree of variation in each of the data sets due to preprocessing less error in the PXRD data set than in the FTIR data set. Finally, Fig with SNV. This is ensured because the SNV subtracts the mean of 5D shows the corresponding PCA plot for the FPD. Again, the same each data set from each data vector, then divides each data vector pattern emerges and from a qualitative point of view, it appears by the standard deviation each data set. Thus, the FTIR spectra and there is less error in the FPD set than in the FTIR data set, but the PXRD patterns have been placed onto similar normalized scales approximately the same amount of error as in the PXRD data set.
and the fused data is now ready for multivariate analysis.
shows that the FTIR, PXRD and fused data sets, respec- tively, track the trends in the concentration variance in the samples (as elucidated in The next step is to use PLS to determine 3.2. PCA patterns and trends the extent to which each of the methods can quantitatively pre- dict the concentrations of each of the components in each of the The purpose of using PCA is usually to elucidate trends or classify samples within data sets. The trends in the 21 calibration samples were clear as these trends were deliberately introduced into the four-component simplex-centroid experimental design. The ques- 3.3. PLS quantitative predictions for calibration samples A tion is whether the analytical tools were sensitive to those trends and whether the fused data would provide better sensitivity to the PLS was performed on each of the data sets (FTIR, PXRD, and trends. To examine these questions, a PCA was first performed on fused data) to determine the accuracy of prediction for each of the the y-data block (the concentration matrix) of the experimentally components. an example of the PLS results for the designed data set (i.e. the pure component percentages listed in prediction of APAP. Similar trends in results were seen for CAF, IBU, The resulting PCA plot is given in showing sev- and ASA, although each of the plots is not included here. eral interesting features. First, 100% of the variation in the data shows the PLS prediction based on the FTIR data set for each of the R.V. Haware et al. / Journal of Pharmaceutical and Biomedical Analysis 56 (2011) 944–949 Fig. 5. PCA of (A) the percentages of the pure components in the 21 calibrations samples, (B) the preprocessed FTIR spectra of the 21 calibration samples, (C) the preprocessed
PXRD patterns of the 21 calibration samples, and (D) the fusion of preprocessed data (FPD) set representing the 21 calibration samples. The percent variance captured by each PC is shown in parenthesis along each axis.
21 calibration samples. shows the corresponding plot for the FPCS. The RMSEC and RMSECV values included in PXRD patterns. As can be seen from the decrease in the scatter in that the PXRD prediction (is better than the FTIR prediction the plot as well as the decrease in RMSEC and RMSECV values, the (but the FPCS prediction (is better than both the PLS model for the PXRD patterns did a better job of prediction than FTIR and PXRD data sets when they are used alone. The RMSECV the PLS model built on the FTIR data. A PLS model was also built on values were calculated using the "leave one out" cross validation the FPD set and the results were very similar to those obtained from method. Each of the RMSEC and RMSECV values for each of the the PXRD PLS model. However, using the FPCS approach, there was four components is detailed in results indicate that a an improvement in the accuracy of the prediction. shows FPCS model based on the concatenation of PCA scores can be used the corresponding plot for the PLS model that was applied to the to improve prediction accuracy.
once ntratio n (%) al APAP Co nce ntratio n (%) AP Concentratio n (%) Fig. 6. Representative predicted vs. measured models obtained by partial least square regression (PLS-2) of the (A) FTIR spectra, (B) PXRD patterns, and (C) fusion of principal
component scores (FPCS) data for the quantitative prediction of APAP.
R.V. Haware et al. / Journal of Pharmaceutical and Biomedical Analysis 56 (2011) 944–949 analysis, can improve the prediction outcome as compared with Summary of partial least square regression (PLS-2) models (RMSEC, root means the single instrument PLS prediction outcomes. Furthermore, it square error of calibration; RMSECV, root mean square error of prediction; PCs, was found that the FPCS PLS prediction outperformed the FPD PLS Partial least square regressions FTIR
Optimum no. of PCs used
The present work demonstrates the ability of data fusion to combine the information in the FTIR and PXRD data for bet- ter quantification and prediction of the amounts of APAP, CAF, IBU and ASA from blended, powder mixtures. PCA of FTIR data, Optimum no. of PCs used PXRD data, and the FPD indicated that, with some error, each of the data sets showed similar trends to the concentration Fusion of principal component scores (FPCS)
variation intrinsic in the 21 calibration samples. The use of Optimum no. of PCs used FPCS was the key to improving the PLS prediction. A compar- ison of the PLS regression analysis of the FTIR data, the PXRD data, the FPD and the FPCS demonstrated that the FPCS pro- duced the best prediction accuracies of unknown amounts of active pharmaceutical ingredients. The improvement in predic- Actual and predicted values of unknown samples of APAP, CAF, IBU and ASA by tion accuracy of the FPCS method over the FPD method may be partial least square regressions (PLS-2) method.
attributed to the noise removal and the data reduction associated with extracting the principal component scores prior to build- ing the PLS regression model. The implications for developing distinctive signatures for unknown mixtures as in, e.g., natural products, are an exciting direction under investigation in the UHH [1] Guidance for Industry: Q8 Pharmaceutical Development, US Department of Health and Human Services, Food and Drug Administration, Center for Drug Evaluation and Research, Center for Biologics Evaluation and Research, Rockville, 2006.
[2] A. Gupta, G.E. Peck, R.W. Miller, K.R. Morris, Real-time near-infrared monitoring of content uniformity, moisture content, compact density, tensile strength, and Young's modulus of roller compacted powder blends, J. Pharm. Sci. 94 (2005) [3] S. Romero-Torres, J.D. Perez-Ramos, K.R. Morris, E.R. Grant, Raman spec- troscopy for tablet coating thickness quantification and coating characteriza- tion in the presence of strong fluorescent interference, J. Pharm. Biomed. Anal.
41 (2006) 811–819.
Fusion of principle component scores (FPCS) [4] W. Cao, M.P. Mullarney, B.C. Hancock, S. Bates, K.R. Morris, Modeling of trans- mitted X-ray intensity variation with sample thickness and solid fraction in glycine compacts, J. Pharm. Sci. 92 (2003) 2345–2353.
[5] P.M. Ramos, I. Ruisanchez, K.S. Andrikopoulos, Micro-Raman and X-ray fluores- cence spectroscopy data fusion for the classification of ochre pigments, Talanta 75 (2008) 926–936.
[6] V. Steinmetz, F. Sevila, V. Bellon-Maurel, A methodology for sensor fusion design: application to fruit quality assessment, J. Agric. Eng. Res. 74 (1999) 3.4. PLS quantitative predictions for blind, unknown samples [7] P. Boilot, E.L. Hines, M.A. Gongora, R.S. Folland, Electronic noses inter- comparison, data fusion and sensor selection in discrimination of standard fruit The PLS models built in Section PLS model of the solutions, Sens. Actuators B 88 (2003) 80–88.
[8] S. Roussel, W. Bellon-Maurel, J.M. Roger, Authenticating white grape must FTIR data, the PLS model of the PXRD data, and the PLS model variety with classification models based on aroma sensors, FT-IR and UV spec- of the FPCS) were applied to unknown samples. The actual and trometry, J. Food Eng. 60 (2003) 407–419.
predicted values for each of the four components in each of the [9] J. Esteban, A. Starr, R. Willetts, P. Hannah, P. Bryanston-Cross, A review of data fusion models and architectures: towards engineering guidelines, Neural four unknowns are given in The results indicate that the Comput. Appl. 14 (2005) 273–281.
PXRD PLS prediction is more accurate than the FTIR PLS predic- [10] J. Llinas, D.L. Hall, An introduction to multi-sensor data fusion, Proc. IEEE Int.
tion, but the FPCS PLS prediction is more accurate than either of Symp. Circuit Syst. 6 (1998) 537–540.
[11] K. Esbensen, S. Schoenkopf, T. Midtgaard, D. Guyot, Multivariate Data Analysis the single instrument PLS predictions. Again, the FPD model did – In Practice, Camo ASA, Trondheim, Norway, 1994.
not improve the accuracy prediction over the single instrument [12] H. Martens, M. Martens, Modified Jack-knife estimation of parameter uncer- PLS predictions. Therefore, these results indicate that FPCS, com- tainty in bilinear modelling by partial least squares regression (PLSR), Food Qual. Pref. 11 (2000) 5–16.
bined with appropriate preprocessing and multivariate statistical


Thermo Scientific iCAP 7000 Series ICP-OES The new Thermo Scientific™ iCAP™ 7000 Series ICP-OES provides the lowest cost multi-element analysis for measuring trace elements in a diverse sample range, combining advanced performance with high productivity and ease of use. The instrument provides your environmental, pharmaceutical, industrial or food safety facility with the lowest cost of analysis per sample, producing consistently reliable data, whilst ensuring compliance to global regulations and standards.


Management of acute and chronic migraine Gianluca Cand Jean Sc Purpose of reviewWe highlight the recent clinical trials for the management of acute and chronic migraine. Recent findingsIn women with menstrual migraine, triptans seem to be well tolerated irrespective of whether or not patientsare taking oestrogen-containing contraceptives or have comorbidities that indicate increased cardiovascularrisk. The new acute drug, telcagepant, a calcitonin gene-related peptide (CGRP) antagonist, is safe forlong-term use (up to 18 months) in migraine patients with stable coronary artery disease in whom the useof triptans is not advisable. From the pooled analysis of the two Phase III Research Evaluating MigraineProphylaxis Therapy studies of onabotulinumtoxinA (BOTOX) in chronic migraineurs, it clearly emerged thatefficacy increases overtime (up to 56 weeks) and paralleled self-perceived improvement in quality of life.Effectiveness was also observed in patients with severely disabling headaches, who met criteria for triptanabuse and were refractory to several prophylactic treatments. Finally, combination of preventivepharmacological agents with different action mechanisms may be the next frontier in therapeuticadvancements for treating migraine.