- Email: [email protected]
Artificial neural networks as a supporting tool for compatibility study based on thermogravimetric data
Thermochimica Acta 659 (2018) 222–231
Contents lists available at ScienceDirect
Thermochimica Acta journal homepage: www.elsevier.com/locate/tca
Artificial neural networks as a supporting tool for compatibility study based on thermogravimetric data Barbara Rojek, Bogdan Suchacz, Marek Wesolowski
T
⁎
Medical University of Gdansk, Department of Analytical Chemistry, Gen. J. Hallera 107, 80-416, Gdansk, Poland
A R T I C L E I N F O
A B S T R A C T
Keywords: Caffeine Compatibility/incompatibility Thermogravimetry Multivariate statistical analysis Artificial neural networks Self-organizing maps
This paper focuses on applying of artificial neural networks (ANNs) for the incompatibility detection between an active pharmaceutical ingredient (API) and excipients on the basis of thermogravimetric data. In binary model mixtures, caffeine was used as API mixed with selected excipients. The interpretation of ANN results was based on the assumption that the model mixtures with relatively high content of API (70 and 90%), should undergo similar course of thermal decomposition and be placed in nearby neurons, as opposed to mixtures where excipient predominated. When different behaviour of model mixtures was observed, the incompatibility between mixture ingredients was determined. The results indicate ANNs combined with thermogravimetry to be a simple diagnostic tool that visualizes the behaviour of ingredients in binary mixtures by placing them in different neurons of the topological map so as to determine incompatibility occurrence. The findings were confirmed with complementary techniques − DSC, FTIR and XRPD.
1. Introduction Preformulation is a critical stage in drug development where the physicochemical profiling of an active pharmaceutical ingredient (API) and excipients are determined and model formulation is prepared. The choice of excipients is associated with an exhaustive evaluation of drug-excipient compatibility or interaction [1]. Thus, compatibility screening of an API with excipients or other active ingredients is recognized as one of the mandatory factors that is at the forefront of drug product research and technology. In addition, a complete understanding of the physicochemical interactions in dosage forms is expected under quality by design prototype of drug development and is encouraged by various regulatory bodies worldwide. The introduction of thermoanalytical methods into the initial steps of pre-formulation studies have contributed significantly to early prediction, monitoring and characterization of the API incompatibility to avoid costly material wastage and considerably reduce the time required to arrive at an appropriate product formulation [2,3]. The most frequently used thermal methods for prospective compatibility screening studies include differential scanning calorimetry (DSC), differential thermal analysis (DTA), and thermogravimetry (TG) [2,4]. Since no generally accepted procedures to evaluate compatibility/incompatibility between drug and excipients exist, there is still ongoing search for more effective methods for incompatibility detection.
⁎
The TG by its nature is a quantitative technique and can frequently be used to determine the amount of a substance in a mixture or a purity of a substance. TG has been particularly fruitful in determination of reaction rates. Data obtained by this method are often more accurate than those from other techniques but usually require the support of complementary chemical and structural characterizations to identify and confirm the reaction stoichiometry [5]. To detect incompatibility in drug-excipient mixtures, the TG profiles of API and excipients are being compared with those of their mixtures. Unfortunately, these comparisons provide equivocal data, because crucial information that can be extracted from the shape of TG traces is the change of mass (loss or gain) during thermal decomposition and the temperature range within which this process occurs. Hence, full information about the character of thermal processes cannot be derived from the TG traces and a plain detection of incompatibility of API with excipients can be deceptive [6]. In addition, TG by itself is not a distinctive technique and the identification of incompatibility solely on TG data may lead to misinterpretation and unreliable conclusions. Therefore, TG data may require the application of multivariate methods [7,8]. Among multivariate methods, special meaning has been given to artificial neural networks (ANNs) [9,10]. The most characteristic feature of ANNs is the ability to learn complex nonlinear input-output relationships, use sequential training procedures, and adapt themselves to the data. The most commonly used ANNs for pattern classification tasks are feed-forward networks, i.e. multilayer perceptron and radial-
Corresponding author. E-mail address: [email protected] (M. Wesolowski).
https://doi.org/10.1016/j.tca.2017.12.015 Received 17 October 2017; Received in revised form 7 December 2017; Accepted 8 December 2017 Available online 11 December 2017 0040-6031/ © 2017 Elsevier B.V. All rights reserved.
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
basis function. Another popular networks are self-organizing maps (SOMs), also called Kohonen networks, used mainly for data clustering and feature mapping [11]. The SOMs are probably the best known of the unsupervised neural network methods and has been used in varied applications [9]. The SOM is often chosen because of its suitability to visualize data, otherwise difficult to interpret [12]. These ANNs are composed of only two layers: the input and the output layer of radial neurons. Accordingly, the formation of SOM is determined by the size of the output layer, where the samples are to be mapped and by the size of the neighbourhood (the distance between neurons where the samples are to be considered similar). SOM leads to clustering of data and relating similar classes to each other [13]. However, it must be noted that if the number of neurons is too small, the samples will be placed in the same or adjacent neuron regardless of their characteristics. If the number is too big, the samples will be scattered all over the map making identification of any vital similarities (differences) impossible to discern. Additionally, when two or more input variables are highly correlated, the SOM model can depend entirely either on one variable or on the other or on some arbitrary combination of them. In such situation it is advisable to eliminate some variables, because ANN will not recognize the data structure properly. The above mentioned properties of SOM underlay the search for effective and fast method for detecting drug-excipient incompatibility based on the data from TG traces. The additional incentive was also the fact that up to now, the use of ANNs as a solution to the problem of thermoanalytical results has rarely been reported [14–17]. As the study material the binary mixtures of caffeine and some excipients were used. The findings of ANN studies were to be confirmed by complementary techniques − DSC, Fourier transform infrared spectroscopy (FTIR) and powder X-ray diffraction (PXRD).
Fig. 1. Scheme of neural and pattern recognition techniques employed in the study.
were measured every 5% of mass loss, giving 16 variables (T5, T10, T15, T20, T25, T30, T35, T40, T45, T50, T55, T60, T65, T70, T75) in all. Through correlation analysis, only some temperature values were selected as input variables for ANN calculations. 2.3. Calculations Chemometric calculations were performed using Statistica 10 software (StatSoft Inc., Tulsa, OK, USA). A schematic of the computational procedure is shown in Fig. 1. Firstly, the correlation matrix between temperatures of particular mass loss was computed to identify strong linear correlations between variables. The SOM is made up of numerous processing units called neurons placed in two different layers − input and output (Fig. 2). The input layer is built of linear neurons, and the output layer of radial ones, which are arranged in two-dimensional grid often referred to as topological map. Each neuron uses an aggregation function to collect the weighted input values, which is passed through a transfer function to give their output. In linear neuron it is a weighted sum of its inputs, while in radial ones it is a scaled squared distance of weight vector from input vector. It must be also noted that the transfer functions differ in both types of neurons. In linear neurons the transfer function is typically sigmoidal, but in radial neurons it is Gaussian. The SOM operates using unsupervised training, in which the weights of neurons (w1, w2, …, wk) in the topological map are entirely adjusted in response to the input pattern (x1, x2, …, xm). A single pass through the entire input pattern to update the weights of neurons is called an epoch. The network forces the neurons to compete against each other to determine which one is to be activated. The result of the
2. Material and methods 2.1. Materials Caffeine, (C8H10N4O2); HPLC purity ≥ 99%, m.p. 232–236 °C, was supplied by Fluka (Siegen, Germany). Glicocol, (C2H5NO2); sorbitol, (C6H14O6); sucrose, (C12H22O11); and arabic gum were purchased from POCh (Gliwice, Poland). Glucose, (C6H12O6) was provided by CentroChem (Lublin, Poland), whereas microcrystalline cellulose (Avicel PH101), (C6H10O5)n was obtained from the FMC Corp. Europe NV (Brussels, Belgium). Caffeine and excipients were used as obtained without further purification. Binary physical mixtures of caffeine with selected excipient at molar or mass ratios of 9:1, 7:3, 1:1, 3:7 and 1:9 were prepared by gentle mixing of both ingredients in agate mortal for 5 min. Ingredients with similar molar masses were mixed at the molar ratios (caffeine with glicocol, glucose, sorbitol and sucrose), whereas those that differed significantly in molar masses were mixed at mass ratios (caffeine with arabic gum and microcrystalline cellulose). 2.2. Thermogravimetry The TG measurements were carried out with TGA Discovery device (TA Instruments, New Castle, DE, USA). The samples of approximately 10 mg were placed in platinum pans and heated from 25 to 700 °C at the heating rate 10 °C/min. The investigations were realized in air atmosphere (purity 99,999%, Air Products, Warsaw, Poland) with a flow rate of 25 mL/min. Mass-temperature diagrams were analysed using Trios software. The interpretation of TG traces was based on the determination of the temperatures of mass loss taken as an arithmetic mean of three measurements, starting with the temperature of 5% mass loss (T5) and ending on the temperature of 75% mass loss (T75). The temperatures
Fig. 2. The general architecture of SOM.
223
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
2.4. Other methods
competition between neurons is decided either on the largest output for a given input or the highest similarity of the weight vector to the input vector. When the neuron which best satisfies the selected criterion is found, it is activated and declared the winner. The criterion can be demonstrated by the following equation: out ⟵ min{Σ(x − w)2}, where out is the output of the activated neuron, x denotes the training vector and w indicates the weight vector. After choosing the winning neuron, its weights are modified so that its response can be closer to that desired. The weights of adjacent neurons are also modified, but with increasing distance from the winner the modifications get smaller. The update of the weight vector of the winner in response to the input pattern is applied according to the following formula: wnew = wold + η (x − wold), in which x is the training case and η is the learning rate [13,18]. This competitive learning leads to clustering of data and relating samples of similar qualities to each other, so the samples that went through similar stages of thermal decomposition shared similar values of mass loss temperatures and were localized in the neighbouring neurons in the map. To indicate fluctuations in mass losses of analysed mixtures it was important to establish the right number of neurons in the map. The optimization of this number was performed by trial and error method. The reason for such strategy is that it cannot be established in advance how many neurons in the map are sufficient to determine proper relationships between data. The networks of different sizes are trained and the outcomes analysed. It is easy to notice that if the number of neurons is not adequate, the samples are placed in the same or adjacent neuron regardless of their characteristics. Contrary to the situation in which there are too many neurons and the samples are scattered all over the map. In both scenarios the identification of any similarities is infeasible, therefore the SOMs of different dimensions (3 × 3, 4 × 4, 5 × 5 and 6 × 6) were trained to establish proper dimension of the map. The training of SOM was run in two phases. In the first phase high learning rate was gradually reduced from 0.9 to 0.1 with a neighbourhood equalled 2 over 200 epochs. In the second phase a low learning rate was set to 0.02 with a neighbourhood 0 run for 3000 epochs. After the training was accomplished the caffeine mixtures with selected excipients such as glicocol, glucose, sorbitol, sucrose, microcrystalline cellulose and arabic gum were projected into the topological map. To discover the temperatures of mass loss which can be regarded as a potential indicator of incompatibility, two unsupervised chemometric methods: cluster analysis (CA) and principal component analysis (PCA) were applied. To accomplish the task, the weight values of all incoming connections of SOM (denoted as wT) were selected as variables. Afterwards, CA and PCA were to arrange them into groups with similar characteristics. In CA, the distance between weight values was calculated using Euclidean distances and to join together comparable clusters the single linkage rule was applied. The purpose of PCA was to identify a smaller number of orthogonal variables from a large set of data to make the interpretation of the data and the detection of its structure simplified. The new variables called principal components (PCs) are linear combinations of the original variables, and it is important to explain the maximum amount of variance with as few number of PCs as possible. The PC1 is described by maximum variance among all linear combinations, whereas the PC2 accounts for as much of the remaining variation as possible. All subsequent PCs have this same property. The scatterplot of the principal components shows the role of the variables and their contribution to the explanation of the variance of original variables. However, the most important thing in the interpretation of PCA scatterplot is the position of variables in relation to each other. Therefore, if two variables are placed near each other it means that they are highly correlated. The variables that are positioned on the other side of the plot with high value of PC are correlated negatively.
DSC scans were obtained using a heat-flux Mettler Toledo instrument (Model DSC 822e, Schwerzenbach, Switzerland) with samples of approximately 5 mg weighed into flat-bottomed aluminium pans. Scans were obtained at a rate of 10 °C/min in the range from 20 to 300 °C, using nitrogen (purity 99,9997%, Air Products, Warsaw, Poland) as a purging gas at a flow rate of 70 mL/min. The indium and zinc standards were used to calibrate the DSC cell. A STARe software was used for the analysis of the DSC scans. FTIR spectra were recorded on a Nicolet 380 FTIR spectrometer (Thermo Fischer Scientific, Madison, USA) with a DTGS KBr detector in the spectral range of 4000–400 cm−1 with resolution of 4 cm−1. To prepare pellets, which contained 1 mg of a sample and 100 mg of KBr (Merck, Darmstadt, Germany), hydraulic press (Specac, Orpington, UK) was used. The background spectra for each measurement, were taken with average 16 scans. An OMNIC software was applied to collect IR spectra of the samples. PXRD patterns were obtained on a D2 Phaser equipment (Bruker, Karlsruhe, Germany) with a CuKα tube (k = 0.154060 nm), voltage of 30 kV and current of 10 mA, over the diffraction angle range 7−55° (2θ), using a step size of 0.02° under an exposure time of 0.10 s. The Diffrac.suite software was used for the analysis of diffraction data. 3. Results and discussion 3.1. Thermal behaviour and compatibility study The TG trace of caffeine (Fig. 3A, Table 1) reveals its decomposition in a single step between 150 and 275 °C [19] with a total mass loss. By the comparison of the TG traces of caffeine and selected excipient alone (Table 1) with traces of their physical mixtures, the differences or similarities in thermal runs are noticeable and can be attributed to any compatibility/incompatibility occurrence between ingredients. Two opposite thermal profiles of caffeine mixtures with glucose and sorbitol are presented in Figs. 3 and 4. According to the TG traces of mixtures of caffeine and glucose (Fig. 3), two thermal evens were found in the temperature ranges of 150–267.5 °C and 267.5–400 °C, due to thermal degradation of both ingredients. The first curved segment of TG traces of binary mixtures (Fig. 3B–F) in the region of 150–267.5 °C corresponds to decomposition of API, whereas the second segment in the range of 267.5–400 °C is due to thermal decomposition of glucose. The thermal behaviour of glucose (Fig. 3G) shows that no other thermal phenomena are observed before the beginning of decomposition between 250 and 400 °C with mass loss of 95%. If the excipient affects on API, it becomes reflected in TG traces of their mixtures by displacing the temperature of decomposition of API. The changes in thermal shapes of TG traces of binary mixtures, e.g. caffeine mixtures with sorbitol (Fig. 4B–F) in comparison to those of both ingredients can suggest potential incompatibility. TG traces of caffeine-sorbitol mixtures went through several degradation steps, whereas thermal degradation of the individual components transpired in a single step. The TG trace of sorbitol (Fig. 4G) showed its thermal degradation in the range of 160–350 °C with total mass loss. The visual examination of the TG traces of mixtures in contrast to those of the single ingredients is difficult, because APIs and the most of excipients undergo similar thermal events and the shapes of TG traces are not so distinctive. Furthermore, caffeine and excipients are organic compounds, so during the heating the same or similar thermal processes may occur. Thus, the thermal profiles of caffeine and excipient on TG traces of their mixtures can overlap and then to distinguish between the profiles of both ingredients becomes complicated. It suggests that the technique is not distinctive enough to detect directly incompatibility and requires the participation of other methods. Therefore, the present study was focused on the application of SOMs to support the interpretation of TG data. 224
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 3. TG traces of (a) caffeine, (g) glucose and their mixtures at the mass ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
The interpretation of SOM results was based on the assumption that the thermal decomposition of single substances and their mixtures, in which the content of API is relatively high i.e. 70 and 90%, should run similar course. As a result, caffeine and its mixtures (7:3, 9:1) should be placed in nearby neurons. Analogous situation was expected to happen in the case of mixtures where excipient predominates and caffeine is in the amount of 10 and 30% (3:7, 1:9). Those mixtures should be positioned in adjacent neurons, but far from caffeine and mixtures with its high content. The last issue was the mixture with equal quantities of
3.2. Compatibility study by SOM On the basis of the correlation analysis, 5 temperatures of mass loss were chosen as inputs to SOM, and they were as follows: T5, T20, T35, T55, T70. As it is shown in Fig. 5, the optimal size of the topological map of the SOM was established to be 5 × 5. The map consisted of enough neurons for proper projection of the samples, where the ones which went through similar heating decomposition course tended to localize in close neurons, others were placed farther away on the map. Table 1 Thermal characteristics of caffeine and excipients under study. Caffeine and excipients
Number of thermal stages
Temperature range (°C)
Mass loss (%)
DSC data (peak temperature, Tp) (°C)
Caffeine
1
125–250
100
Sorbitol Glicocol
1 3
Glucose
5
Sucrose
3
Microcrystalline cellulose
2
Arabic gum
5
160–360 180–288 288–425 425–632 50–98 98–149 149–247 247–386 386–597 170–233 233–400 400–585 225–362 362–520 41–181 181–275 275–400 400–525 525–611
98 60 14 26 7 2 14 53 24 21 50 29 90 10 10 22 34 31 1
159.4 (endo), polymorphism 236.2 (endo), melting point 99.7 (endo), melting point 254.7 (endo), melting point
225
162.1 (endo), melting point 219.4 (endo), caramelization
190.9 212.8 227.2 341.9
(endo), (endo), (endo), (endo),
melting point caramelization caramelization decomposition
128.2 (endo), dehydration 308.7 (exo), decomposition
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 4. TG traces of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
excipient (3:7, 1:9) were separated from single substances. Similar situation was observed for sorbitol (Fig. 5C) and arabic gum (Fig. 5E) as well. However, it must be noted that a small amount of sorbitol in 9:1 mixture does not strongly affect the thermal decomposition of caffeine, which is indicated by the fact that sorbitol and the mixture are located in neighbouring neurons on the map (Fig. 5C). The behaviour of the mixtures containing arabic gum is quite the opposite. The arrangement of samples in Fig. 5E shows that even a small addition of arabic gum to caffeine influences its thermal decomposition, so the mixtures are drifted away from the individual ingredients. Nevertheless, the increasing amount of arabic gum in different mixtures is not vividly reflected by their thermal profile, and in consequence three mixtures 9:1, 7:3 and 1:1 can be found in nearby neurons. Taken into regard the arrangement of those mixtures in the SOM map, it was specified that glicocol, sorbitol and arabic gum are incompatible with caffeine.
caffeine and excipient. The most reasonable assumption was that thermal decomposition of this mixture should be intermediate between the decomposition of single substances, provided that no unexpected reactions take place. Such arrangement of samples in the topological map would clearly indicate compatibility between mixture ingredients. The spatial arrangements of caffeine mixtures with microcrystalline cellulose, glucose and sucrose (Fig. 5A, D and F) show that the SOM grouped the samples containing those compounds according to the assumption made above. Caffeine and the mixtures 9:1 and 7:3 are gathered on one side of the map (dark grey circles), while the mixtures with a dominant quantity of excipient 3:7, 1:9 (light grey circles) on the other. The mixture 1:1 (black circle) is located in the middle, and its position seems to be determined by the thermal decomposition of either caffeine or excipient depending on which process is more distinctive. Consequently, the location of 1:1 mixture shifts either in the direction of 7:3 or 3:7. A slight deviation from this observation was demonstrated by the mixtures with sucrose, where 1:1 mixture was moved towards the upper part of the map (Fig. 5F). The way the samples were allocated all over the topological map evidently signified the lack of incompatibility between the excipients (microcrystalline cellulose, glucose, sucrose) and API. As regards glicocol, sorbitol and arabic gum (Fig. 5B, C and E), the thermal decomposition proceeded differently than it was presumed. The samples, which were expected to be located in near vicinity, were scattered away from each other. This means that the thermal profiles of those substances in combination with caffeine diverge substantially from those performed using single ingredients. It is particularly distinct in the case of caffeine-glicocol mixtures (Fig. 5B), where samples with dominant caffeine content (7:3, 9:1) and those containing more
3.3. Selection of key temperature values The SOM detection of incompatibility between API and excipient was performed on the basis of only selected temperatures of mass loss (T5, T20, T35, T55, T70) due to high linear correlations (above 0.9) between variables. In further step, therefore, it was important to establish the temperature values of key importance in such detection. Taking into consideration that in dissimilar study, specified single values of mass loss temperatures would not properly detect incompatibility, this search was also focused on finding the temperature range of potential detectability. The CA dendrograms presented in Fig. 6 indicate that the temperatures of the small (T5) and large mass loss (T70) do not contribute 226
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 5. The arrangement of caffeine and its mixtures at the ratios of 9:1, 7:3, 1:1, 3:7, 1:9 on the 5 × 5 topological maps with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
with glicocol, sorbitol and arabic gum, for which the incompatibilities were established. It may point out to T20 that it is a distinctive feature of some interactions happening between constituents of mixtures. For the rest of excipients i.e. microcrystalline cellulose, glucose and sucrose, the clusters of wT20, wT35 and wT55 are separated from the cluster
much to the detection of incompatibility. In all tree diagrams the weights for temperatures of 5 and 70% mass loss (wT5, wT70) are linked together. However, in three cases (Fig. 6B, C and E) the value of weights for the temperature of 20% mass loss (wT20) was connected to them, which suggested similar thermal characteristics of caffeine mixtures
Fig. 6. The CA dendrogram of weight values for selected temperatures for caffeine and its mixtures with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
227
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 7. The PCA plot of weight values for selected temperatures for caffeine and its mixtures with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
Table 2 The prediction of compatibility/incompatibility between ingredients of pharmaceutical mixtures using other analytical approaches. Analytical approaches
TG-SOMs DSC FTIR PXRD
Excipients Glicocol
Sorbitol
Glucose
Sucrose
Microcrystalline cellulose
Arabic gum
× × o o
× × × o
o o o o
o o × o
o o o o
× × × o
o − compatibility, × –incompatibility.
they are positively correlated and influence the PCA model in similar ways. The variable wT35 is negatively correlated to wT5 and wT70, because it is positioned opposite to the place where wT5 and wT70 are. Moreover, this denotes that wT35 contributes to the PC model in ways different from wT5 and wT70. Similar situation takes place in the case of wT20 and wT55, which are located on opposite sides of the plot on PC2 axis for glicocol, sorbitol and arabic gum (Fig. 7B, C and E). Due to high linear correlations (above 0.9) between certain pairs of temperatures, the results of PCA and CA results can be expanded by including the temperatures not used as input vector to SOMs. Consequently, the temperature range for the mass loss from 20 to 55% can be regarded as the one with potential ability to detect incompatibilities.
composed only of wT5 and wT70. The PCA plot is shown in Fig. 7. The first two PCs explain 72.0790.79% of variance depending on the mixture, so it was safe to assume that two-dimensional plots PC1 and PC2 were sufficient for data exploration. Although, the PCA results of weight values of some temperatures of mass loss were slightly problematic to interpret, some regularities were identified. For all analysed mixtures wT5 and wT70 were positioned together and were characterized by high values of PC1 (either positive or negative). The weight values for 20% mass loss of glicocol, sorbitol and arabic gum included in mixtures with caffeine (Fig. 7B, C and E) assumed the value of PC1 close to 0 and high values of PC2 close to 1. Another consistency was the fact that wT55 for microcrystalline cellulose, glucose and sucrose (Fig. 7A, D and F) was described by intermediate values of both PC1 and PC2 (about ± 0,5) with the tendency to move towards wT5 and wT70. The opposite situation was observed in the case of glicocol, sorbitol and arabic gum. In view of the fact that the values of wT5 and wT70 are around 0 in the PC2 axes means that they contribute very little to the variance explained by PC2. At the same time both variables greatly contribute to the variance explained by PC1, which is by definition the highest. The variables wT5 and wT70 are closely positioned, which indicates that
3.4. Confirmation of SOM outcome The incompatibility detection performed by SOM was verified by DSC, FTIR and XRPD (Table 2). Due to the similarity of the results for the mixtures, in which interaction with API was established, it was decided to present them on the example of sorbitol. The DSC scans, FTIR spectra and PXRD diffractograms of binary 228
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 8. DSC traces of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9. Fig. 9. FTIR spectra of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
referred to sorbitol, while the second small broader peak around 159 °C was correlated with polymorphic transition of caffeine. There are some published reports on the compatibility study between API and mannitol [22], probably thermal behaviour of mannitol in its mixtures with API is similar as sorbitol. Sorbitol and its isomer − mannitol are sugar alcohols belonging to polyols. The mechanism of reaction of mannitol with API is probably similar to mechanism of
mixtures were compared to those of individual ingredients. DSC scan of caffeine (Fig. 8) displays an endothermic peak due to a polymorphic transition at 159.42 °C [19,20], before the final melting at 236.32 °C [19–21]. Through the comparison of DSC scans of caffeine, sorbitol with those of binary mixtures, it was observed the absence of melting peak of caffeine at around 236 °C, suggesting incompatibility between API and excipient. The first sharp peak found at about 102 °C was 229
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
increased. Also, the absorption bands were encountered in the region of wavenumbers of 1200–1000 cm−1, corresponding to stretching vibrations of CeC and CeO of sorbitol and bands of caffeine (bending of CH, rocking of CH3, rocking of CH3 out of plane [25]) were changed. Any changes found in absorption bands of API in FTIR spectra of mixtures suggest incompatibility between API and excipient. Those changes were also found in the case of glicocol and arabic gum. The spectra of caffeine mixtures with microcrystalline cellulose, glucose and sucrose were a sum of all bands of both ingredients. The application of PXRD in the evaluation of interactions between API and excipients is sustained by the fact that a crystalline compound exhibit a unique set of diffraction peaks. Any interaction, which occurs between two crystalline ingredients or between a crystalline and an amorphous one affects the aspect of the PXRD pattern. As shown in Fig. 10, the peaks of caffeine with higher intensity are observed at 11.96°, 24.03°, 26.38°, 27.00°, revealing its high crystallinity. The diffractograms of mixtures are simple superimposition of patterns observed for single ingredients, which informs that the presence of excipient in mixtures of caffeine does not influence its crystallinity and nature. No incompatibility can be established from those diffractograms. Taking above into consideration, it must be stated that only DSC and FTIR data are consistent with those obtained using SOMs. It revealed that glicocol, sorbitol and arabic gum are incompatible with caffeine.
reaction of sorbitol with API. These excipients are generally used as tablet and capsule diluents, sweetening agents, plasticizers [23]. Bharate et al. [22] reported that mannitol was found to be incompatible with omeprazole, primaquine, and quinapril, which are all aminocontaining drugs. Stulzer et al. [24] indicated that interactions between API as piroxicam due to a possible hydrogen bond formed by nitrogen molecules of piroxicam and CeH and CH2 groups of mannitol. The FTIR spectra of caffeine mixtures with sorbitol (Fig. 9) show that characteristic bands of caffeine in the spectral range of 1800–1500 cm−1 and 1200–1000 cm−1 were altered. It can be observed that two overlapped bands (at 1700 cm−1 and 1660 cm−1 due to stretching of C]O) are split into two distinct ones, which indicates a possible physico-chemical interaction between these ingredients, i.e., hydrogen bond formation. Furthermore, the intensity of absorption band of caffeine at 1548 cm−1, attributable to bending vibrations of HeC]N and stretching vibrations of rings of imidazole and pyrimidine
4. Conclusions The application of ANNs for compatibility/incompatibility detection based on TG data provided an effective approach for resolving such issues. The SOM specified glicocol, sorbitol and arabic gum as the substances which interfere with caffeine. The outcomes of SOM studies were confirmed by complementary techniques − DSC and FTIR, proving that the application of TG combined with ANNs to be a promising alternative. The CA and PCA calculations revealed that the initial temperatures when the mass loss is slim (T5, T10, T15) and the ending temperatures when the mass loss is high (T60, T70, T75) cannot be regarded as a potential indicator of incompatibility. The range of temperatures, which can be the most valuable for such detection, was established for the mass loss from 20 to 55%. Acknowledgment The investigations were financially supported by a statutory research, Grant No. 02-0015/07/505, from the Ministry of Science and Higher Education, Poland. References [1] M. Ch. Adeyeye, H.G. Brittain, Preformulation in Solid Dosage Form Development, Informa Healthcare Inc., New York, 2008. [2] R. Chadha, S. Bhandari, Drug-excipient compatibility screening −role of thermoanalytical and spectroscopic techniques, J. Pharm. Biomed. Anal. 87 (2014) 82–97. [3] A. Marini, V. Berbenni, S. Moioli, G. Bruni, P. Cofrancesco, C. Margheritis, M. Villa, Drug-excipient compatibility study by physicochemical techniques; The case of indomethacin, J. Therm. Anal. Calorim. 73 (2003) 529–545. [4] C.M.S. de Mendonça, I.P. de Barros Lima, C.F.S. Aragão, A.P.B. Gomes, Thermal compatibility between hydroquinone and retinoic acid in pharmaceutical formulations, J. Therm. Anal. Calorim. 115 (2014) 2277–2285. [5] D.Q.M. Craig, M. Reading, Thermal Analysis of Pharmaceuticals, Taylor & Francis Group, LLC, Boca Raton, 2006. [6] B. Rojek, M. Wesolowski, Compatibility studies of hydrocortisone with excipients using thermogravimetric analysis supported by multivariate statistical analysis, J. Therm. Anal. Calorim. 127 (2017) 543–553. [7] I. Singh, P. Juneja, B. Kaur, P. Kumar, Pharmaceutical applications of chemometric techniques, ISRN Anal. Chem. 2013 (2013) 13. [8] S. Matero, Chemometrics Methods in Pharmaceutical Tablet Development and Manufacturing Unit Operations, Publications of the University of Eastern Finland Dissertations in Health Sciences, Kuopio, 2010. [9] M. Wesolowski, B. Suchacz, Artificial neural networks: theoretical background and pharmaceutical applications: a review, J. AOAC Int. 95 (2012) 652–668.
Fig. 10. PXRD patterns of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
230
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
[18] J. Zupan, J. Gasteiger, Neural Networks in Chemistry and Drug Design: An Introduction, second ed., VCH, Weinheim, 1999. [19] R.A. Carlton, Pharmaceutical Microscopy, Springer Science + Business Media, LLC, New York, 2011. [20] W. Lund, The Pharmaceutical Codex, twelfth ed., The Pharmaceutical Press, London, 2004. [21] Polish Pharmacopoeia VIII, Ministry of Health, Warsaw, 2008. [22] S.S. Bharate, S.B. Bharate, A.N. Bajaj, Interactions and incompatibilities of pharmaceutical excipients with active pharmaceutical ingredients: a comprehensive review, J. Excipients Food Chem. 1 (2010) 3–26. [23] R.C. Rowe, P.J. Sheskey, M.E. Quinn, Handbook of Pharmaceutical Excipients, sixth ed., American Pharmaceutical Association, Pharmaceutical Press, London, 2009. [24] H.K. Stulzer, M.P. Tagliari, A.P. Cruz, M.A.S. Silva, M.C.M. Laranjeira, Compatibility studies between piroxicam and pharmaceutical excipients used in solid dosage forms, Pharm. Chem. J. 42 (2008) 215–219. [25] F. Ucun, A. Sağlam, V. Güçlü, Molecular structures and vibrational frequencies of xanthine and its methyl derivatives (caffeine and theobromine) by ab initio Hartree −Fock and density functional theory calculations, Spectrochim. Acta A 67 (2007) 342–349.
[10] J. Djuris, Computer-Aided Applications in Pharmaceutical Technology, Woodhead Publishing Cambridge, 2013. [11] J.K. Basu, D. Bhattacharyya, T. Kim, Use of artificial neural network in pattern recognition, Int. J. Software Eng. Appl. 4 (2010) 23–34. [12] J. Malone, K. McGarry, S. Wermter, C. Bowerman, Data mining using rule extraction from Kohonen self-organising maps, Neural Comput. Appl. 15 (2006) 9–17. [13] T. Kohonen, Self-Organising Maps, Springer, Berlin, 2000. [14] N. Sbirrazzuoli, D. Brunel, Computational neural networks for mapping calorimetric data: application of feed-forward neural networks to kinetic parameters determination and signal filtering, Neural Comput. Appl. 5 (1997) 20–32. [15] N. Sbirrazzuoli, D. Brunel, L. Élégant, Neural networks for kinetic parameters determination, signal filtering and deconvolution in thermal analysis, J. Therm. Anal. Calorim. 49 (1997) 1553–1564. [16] M. Wesolowski, B. Suchacz, Classification of rapeseed and soybean oils by use of unsupervised pattern-recognition methods and neural networks, Fresh J. Anal. Chem. 371 (2001) 323–330. [17] M. Wesolowski, B. Suchacz, Neural networks as a tool for quality assessment of soybean and rapeseed oils analyzed by thermogravimetry, J. Therm. Anal. Calorim. 68 (2002) 893–899.
231
Contents lists available at ScienceDirect
Thermochimica Acta journal homepage: www.elsevier.com/locate/tca
Artificial neural networks as a supporting tool for compatibility study based on thermogravimetric data Barbara Rojek, Bogdan Suchacz, Marek Wesolowski
T
⁎
Medical University of Gdansk, Department of Analytical Chemistry, Gen. J. Hallera 107, 80-416, Gdansk, Poland
A R T I C L E I N F O
A B S T R A C T
Keywords: Caffeine Compatibility/incompatibility Thermogravimetry Multivariate statistical analysis Artificial neural networks Self-organizing maps
This paper focuses on applying of artificial neural networks (ANNs) for the incompatibility detection between an active pharmaceutical ingredient (API) and excipients on the basis of thermogravimetric data. In binary model mixtures, caffeine was used as API mixed with selected excipients. The interpretation of ANN results was based on the assumption that the model mixtures with relatively high content of API (70 and 90%), should undergo similar course of thermal decomposition and be placed in nearby neurons, as opposed to mixtures where excipient predominated. When different behaviour of model mixtures was observed, the incompatibility between mixture ingredients was determined. The results indicate ANNs combined with thermogravimetry to be a simple diagnostic tool that visualizes the behaviour of ingredients in binary mixtures by placing them in different neurons of the topological map so as to determine incompatibility occurrence. The findings were confirmed with complementary techniques − DSC, FTIR and XRPD.
1. Introduction Preformulation is a critical stage in drug development where the physicochemical profiling of an active pharmaceutical ingredient (API) and excipients are determined and model formulation is prepared. The choice of excipients is associated with an exhaustive evaluation of drug-excipient compatibility or interaction [1]. Thus, compatibility screening of an API with excipients or other active ingredients is recognized as one of the mandatory factors that is at the forefront of drug product research and technology. In addition, a complete understanding of the physicochemical interactions in dosage forms is expected under quality by design prototype of drug development and is encouraged by various regulatory bodies worldwide. The introduction of thermoanalytical methods into the initial steps of pre-formulation studies have contributed significantly to early prediction, monitoring and characterization of the API incompatibility to avoid costly material wastage and considerably reduce the time required to arrive at an appropriate product formulation [2,3]. The most frequently used thermal methods for prospective compatibility screening studies include differential scanning calorimetry (DSC), differential thermal analysis (DTA), and thermogravimetry (TG) [2,4]. Since no generally accepted procedures to evaluate compatibility/incompatibility between drug and excipients exist, there is still ongoing search for more effective methods for incompatibility detection.
⁎
The TG by its nature is a quantitative technique and can frequently be used to determine the amount of a substance in a mixture or a purity of a substance. TG has been particularly fruitful in determination of reaction rates. Data obtained by this method are often more accurate than those from other techniques but usually require the support of complementary chemical and structural characterizations to identify and confirm the reaction stoichiometry [5]. To detect incompatibility in drug-excipient mixtures, the TG profiles of API and excipients are being compared with those of their mixtures. Unfortunately, these comparisons provide equivocal data, because crucial information that can be extracted from the shape of TG traces is the change of mass (loss or gain) during thermal decomposition and the temperature range within which this process occurs. Hence, full information about the character of thermal processes cannot be derived from the TG traces and a plain detection of incompatibility of API with excipients can be deceptive [6]. In addition, TG by itself is not a distinctive technique and the identification of incompatibility solely on TG data may lead to misinterpretation and unreliable conclusions. Therefore, TG data may require the application of multivariate methods [7,8]. Among multivariate methods, special meaning has been given to artificial neural networks (ANNs) [9,10]. The most characteristic feature of ANNs is the ability to learn complex nonlinear input-output relationships, use sequential training procedures, and adapt themselves to the data. The most commonly used ANNs for pattern classification tasks are feed-forward networks, i.e. multilayer perceptron and radial-
Corresponding author. E-mail address: [email protected] (M. Wesolowski).
https://doi.org/10.1016/j.tca.2017.12.015 Received 17 October 2017; Received in revised form 7 December 2017; Accepted 8 December 2017 Available online 11 December 2017 0040-6031/ © 2017 Elsevier B.V. All rights reserved.
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
basis function. Another popular networks are self-organizing maps (SOMs), also called Kohonen networks, used mainly for data clustering and feature mapping [11]. The SOMs are probably the best known of the unsupervised neural network methods and has been used in varied applications [9]. The SOM is often chosen because of its suitability to visualize data, otherwise difficult to interpret [12]. These ANNs are composed of only two layers: the input and the output layer of radial neurons. Accordingly, the formation of SOM is determined by the size of the output layer, where the samples are to be mapped and by the size of the neighbourhood (the distance between neurons where the samples are to be considered similar). SOM leads to clustering of data and relating similar classes to each other [13]. However, it must be noted that if the number of neurons is too small, the samples will be placed in the same or adjacent neuron regardless of their characteristics. If the number is too big, the samples will be scattered all over the map making identification of any vital similarities (differences) impossible to discern. Additionally, when two or more input variables are highly correlated, the SOM model can depend entirely either on one variable or on the other or on some arbitrary combination of them. In such situation it is advisable to eliminate some variables, because ANN will not recognize the data structure properly. The above mentioned properties of SOM underlay the search for effective and fast method for detecting drug-excipient incompatibility based on the data from TG traces. The additional incentive was also the fact that up to now, the use of ANNs as a solution to the problem of thermoanalytical results has rarely been reported [14–17]. As the study material the binary mixtures of caffeine and some excipients were used. The findings of ANN studies were to be confirmed by complementary techniques − DSC, Fourier transform infrared spectroscopy (FTIR) and powder X-ray diffraction (PXRD).
Fig. 1. Scheme of neural and pattern recognition techniques employed in the study.
were measured every 5% of mass loss, giving 16 variables (T5, T10, T15, T20, T25, T30, T35, T40, T45, T50, T55, T60, T65, T70, T75) in all. Through correlation analysis, only some temperature values were selected as input variables for ANN calculations. 2.3. Calculations Chemometric calculations were performed using Statistica 10 software (StatSoft Inc., Tulsa, OK, USA). A schematic of the computational procedure is shown in Fig. 1. Firstly, the correlation matrix between temperatures of particular mass loss was computed to identify strong linear correlations between variables. The SOM is made up of numerous processing units called neurons placed in two different layers − input and output (Fig. 2). The input layer is built of linear neurons, and the output layer of radial ones, which are arranged in two-dimensional grid often referred to as topological map. Each neuron uses an aggregation function to collect the weighted input values, which is passed through a transfer function to give their output. In linear neuron it is a weighted sum of its inputs, while in radial ones it is a scaled squared distance of weight vector from input vector. It must be also noted that the transfer functions differ in both types of neurons. In linear neurons the transfer function is typically sigmoidal, but in radial neurons it is Gaussian. The SOM operates using unsupervised training, in which the weights of neurons (w1, w2, …, wk) in the topological map are entirely adjusted in response to the input pattern (x1, x2, …, xm). A single pass through the entire input pattern to update the weights of neurons is called an epoch. The network forces the neurons to compete against each other to determine which one is to be activated. The result of the
2. Material and methods 2.1. Materials Caffeine, (C8H10N4O2); HPLC purity ≥ 99%, m.p. 232–236 °C, was supplied by Fluka (Siegen, Germany). Glicocol, (C2H5NO2); sorbitol, (C6H14O6); sucrose, (C12H22O11); and arabic gum were purchased from POCh (Gliwice, Poland). Glucose, (C6H12O6) was provided by CentroChem (Lublin, Poland), whereas microcrystalline cellulose (Avicel PH101), (C6H10O5)n was obtained from the FMC Corp. Europe NV (Brussels, Belgium). Caffeine and excipients were used as obtained without further purification. Binary physical mixtures of caffeine with selected excipient at molar or mass ratios of 9:1, 7:3, 1:1, 3:7 and 1:9 were prepared by gentle mixing of both ingredients in agate mortal for 5 min. Ingredients with similar molar masses were mixed at the molar ratios (caffeine with glicocol, glucose, sorbitol and sucrose), whereas those that differed significantly in molar masses were mixed at mass ratios (caffeine with arabic gum and microcrystalline cellulose). 2.2. Thermogravimetry The TG measurements were carried out with TGA Discovery device (TA Instruments, New Castle, DE, USA). The samples of approximately 10 mg were placed in platinum pans and heated from 25 to 700 °C at the heating rate 10 °C/min. The investigations were realized in air atmosphere (purity 99,999%, Air Products, Warsaw, Poland) with a flow rate of 25 mL/min. Mass-temperature diagrams were analysed using Trios software. The interpretation of TG traces was based on the determination of the temperatures of mass loss taken as an arithmetic mean of three measurements, starting with the temperature of 5% mass loss (T5) and ending on the temperature of 75% mass loss (T75). The temperatures
Fig. 2. The general architecture of SOM.
223
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
2.4. Other methods
competition between neurons is decided either on the largest output for a given input or the highest similarity of the weight vector to the input vector. When the neuron which best satisfies the selected criterion is found, it is activated and declared the winner. The criterion can be demonstrated by the following equation: out ⟵ min{Σ(x − w)2}, where out is the output of the activated neuron, x denotes the training vector and w indicates the weight vector. After choosing the winning neuron, its weights are modified so that its response can be closer to that desired. The weights of adjacent neurons are also modified, but with increasing distance from the winner the modifications get smaller. The update of the weight vector of the winner in response to the input pattern is applied according to the following formula: wnew = wold + η (x − wold), in which x is the training case and η is the learning rate [13,18]. This competitive learning leads to clustering of data and relating samples of similar qualities to each other, so the samples that went through similar stages of thermal decomposition shared similar values of mass loss temperatures and were localized in the neighbouring neurons in the map. To indicate fluctuations in mass losses of analysed mixtures it was important to establish the right number of neurons in the map. The optimization of this number was performed by trial and error method. The reason for such strategy is that it cannot be established in advance how many neurons in the map are sufficient to determine proper relationships between data. The networks of different sizes are trained and the outcomes analysed. It is easy to notice that if the number of neurons is not adequate, the samples are placed in the same or adjacent neuron regardless of their characteristics. Contrary to the situation in which there are too many neurons and the samples are scattered all over the map. In both scenarios the identification of any similarities is infeasible, therefore the SOMs of different dimensions (3 × 3, 4 × 4, 5 × 5 and 6 × 6) were trained to establish proper dimension of the map. The training of SOM was run in two phases. In the first phase high learning rate was gradually reduced from 0.9 to 0.1 with a neighbourhood equalled 2 over 200 epochs. In the second phase a low learning rate was set to 0.02 with a neighbourhood 0 run for 3000 epochs. After the training was accomplished the caffeine mixtures with selected excipients such as glicocol, glucose, sorbitol, sucrose, microcrystalline cellulose and arabic gum were projected into the topological map. To discover the temperatures of mass loss which can be regarded as a potential indicator of incompatibility, two unsupervised chemometric methods: cluster analysis (CA) and principal component analysis (PCA) were applied. To accomplish the task, the weight values of all incoming connections of SOM (denoted as wT) were selected as variables. Afterwards, CA and PCA were to arrange them into groups with similar characteristics. In CA, the distance between weight values was calculated using Euclidean distances and to join together comparable clusters the single linkage rule was applied. The purpose of PCA was to identify a smaller number of orthogonal variables from a large set of data to make the interpretation of the data and the detection of its structure simplified. The new variables called principal components (PCs) are linear combinations of the original variables, and it is important to explain the maximum amount of variance with as few number of PCs as possible. The PC1 is described by maximum variance among all linear combinations, whereas the PC2 accounts for as much of the remaining variation as possible. All subsequent PCs have this same property. The scatterplot of the principal components shows the role of the variables and their contribution to the explanation of the variance of original variables. However, the most important thing in the interpretation of PCA scatterplot is the position of variables in relation to each other. Therefore, if two variables are placed near each other it means that they are highly correlated. The variables that are positioned on the other side of the plot with high value of PC are correlated negatively.
DSC scans were obtained using a heat-flux Mettler Toledo instrument (Model DSC 822e, Schwerzenbach, Switzerland) with samples of approximately 5 mg weighed into flat-bottomed aluminium pans. Scans were obtained at a rate of 10 °C/min in the range from 20 to 300 °C, using nitrogen (purity 99,9997%, Air Products, Warsaw, Poland) as a purging gas at a flow rate of 70 mL/min. The indium and zinc standards were used to calibrate the DSC cell. A STARe software was used for the analysis of the DSC scans. FTIR spectra were recorded on a Nicolet 380 FTIR spectrometer (Thermo Fischer Scientific, Madison, USA) with a DTGS KBr detector in the spectral range of 4000–400 cm−1 with resolution of 4 cm−1. To prepare pellets, which contained 1 mg of a sample and 100 mg of KBr (Merck, Darmstadt, Germany), hydraulic press (Specac, Orpington, UK) was used. The background spectra for each measurement, were taken with average 16 scans. An OMNIC software was applied to collect IR spectra of the samples. PXRD patterns were obtained on a D2 Phaser equipment (Bruker, Karlsruhe, Germany) with a CuKα tube (k = 0.154060 nm), voltage of 30 kV and current of 10 mA, over the diffraction angle range 7−55° (2θ), using a step size of 0.02° under an exposure time of 0.10 s. The Diffrac.suite software was used for the analysis of diffraction data. 3. Results and discussion 3.1. Thermal behaviour and compatibility study The TG trace of caffeine (Fig. 3A, Table 1) reveals its decomposition in a single step between 150 and 275 °C [19] with a total mass loss. By the comparison of the TG traces of caffeine and selected excipient alone (Table 1) with traces of their physical mixtures, the differences or similarities in thermal runs are noticeable and can be attributed to any compatibility/incompatibility occurrence between ingredients. Two opposite thermal profiles of caffeine mixtures with glucose and sorbitol are presented in Figs. 3 and 4. According to the TG traces of mixtures of caffeine and glucose (Fig. 3), two thermal evens were found in the temperature ranges of 150–267.5 °C and 267.5–400 °C, due to thermal degradation of both ingredients. The first curved segment of TG traces of binary mixtures (Fig. 3B–F) in the region of 150–267.5 °C corresponds to decomposition of API, whereas the second segment in the range of 267.5–400 °C is due to thermal decomposition of glucose. The thermal behaviour of glucose (Fig. 3G) shows that no other thermal phenomena are observed before the beginning of decomposition between 250 and 400 °C with mass loss of 95%. If the excipient affects on API, it becomes reflected in TG traces of their mixtures by displacing the temperature of decomposition of API. The changes in thermal shapes of TG traces of binary mixtures, e.g. caffeine mixtures with sorbitol (Fig. 4B–F) in comparison to those of both ingredients can suggest potential incompatibility. TG traces of caffeine-sorbitol mixtures went through several degradation steps, whereas thermal degradation of the individual components transpired in a single step. The TG trace of sorbitol (Fig. 4G) showed its thermal degradation in the range of 160–350 °C with total mass loss. The visual examination of the TG traces of mixtures in contrast to those of the single ingredients is difficult, because APIs and the most of excipients undergo similar thermal events and the shapes of TG traces are not so distinctive. Furthermore, caffeine and excipients are organic compounds, so during the heating the same or similar thermal processes may occur. Thus, the thermal profiles of caffeine and excipient on TG traces of their mixtures can overlap and then to distinguish between the profiles of both ingredients becomes complicated. It suggests that the technique is not distinctive enough to detect directly incompatibility and requires the participation of other methods. Therefore, the present study was focused on the application of SOMs to support the interpretation of TG data. 224
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 3. TG traces of (a) caffeine, (g) glucose and their mixtures at the mass ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
The interpretation of SOM results was based on the assumption that the thermal decomposition of single substances and their mixtures, in which the content of API is relatively high i.e. 70 and 90%, should run similar course. As a result, caffeine and its mixtures (7:3, 9:1) should be placed in nearby neurons. Analogous situation was expected to happen in the case of mixtures where excipient predominates and caffeine is in the amount of 10 and 30% (3:7, 1:9). Those mixtures should be positioned in adjacent neurons, but far from caffeine and mixtures with its high content. The last issue was the mixture with equal quantities of
3.2. Compatibility study by SOM On the basis of the correlation analysis, 5 temperatures of mass loss were chosen as inputs to SOM, and they were as follows: T5, T20, T35, T55, T70. As it is shown in Fig. 5, the optimal size of the topological map of the SOM was established to be 5 × 5. The map consisted of enough neurons for proper projection of the samples, where the ones which went through similar heating decomposition course tended to localize in close neurons, others were placed farther away on the map. Table 1 Thermal characteristics of caffeine and excipients under study. Caffeine and excipients
Number of thermal stages
Temperature range (°C)
Mass loss (%)
DSC data (peak temperature, Tp) (°C)
Caffeine
1
125–250
100
Sorbitol Glicocol
1 3
Glucose
5
Sucrose
3
Microcrystalline cellulose
2
Arabic gum
5
160–360 180–288 288–425 425–632 50–98 98–149 149–247 247–386 386–597 170–233 233–400 400–585 225–362 362–520 41–181 181–275 275–400 400–525 525–611
98 60 14 26 7 2 14 53 24 21 50 29 90 10 10 22 34 31 1
159.4 (endo), polymorphism 236.2 (endo), melting point 99.7 (endo), melting point 254.7 (endo), melting point
225
162.1 (endo), melting point 219.4 (endo), caramelization
190.9 212.8 227.2 341.9
(endo), (endo), (endo), (endo),
melting point caramelization caramelization decomposition
128.2 (endo), dehydration 308.7 (exo), decomposition
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 4. TG traces of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
excipient (3:7, 1:9) were separated from single substances. Similar situation was observed for sorbitol (Fig. 5C) and arabic gum (Fig. 5E) as well. However, it must be noted that a small amount of sorbitol in 9:1 mixture does not strongly affect the thermal decomposition of caffeine, which is indicated by the fact that sorbitol and the mixture are located in neighbouring neurons on the map (Fig. 5C). The behaviour of the mixtures containing arabic gum is quite the opposite. The arrangement of samples in Fig. 5E shows that even a small addition of arabic gum to caffeine influences its thermal decomposition, so the mixtures are drifted away from the individual ingredients. Nevertheless, the increasing amount of arabic gum in different mixtures is not vividly reflected by their thermal profile, and in consequence three mixtures 9:1, 7:3 and 1:1 can be found in nearby neurons. Taken into regard the arrangement of those mixtures in the SOM map, it was specified that glicocol, sorbitol and arabic gum are incompatible with caffeine.
caffeine and excipient. The most reasonable assumption was that thermal decomposition of this mixture should be intermediate between the decomposition of single substances, provided that no unexpected reactions take place. Such arrangement of samples in the topological map would clearly indicate compatibility between mixture ingredients. The spatial arrangements of caffeine mixtures with microcrystalline cellulose, glucose and sucrose (Fig. 5A, D and F) show that the SOM grouped the samples containing those compounds according to the assumption made above. Caffeine and the mixtures 9:1 and 7:3 are gathered on one side of the map (dark grey circles), while the mixtures with a dominant quantity of excipient 3:7, 1:9 (light grey circles) on the other. The mixture 1:1 (black circle) is located in the middle, and its position seems to be determined by the thermal decomposition of either caffeine or excipient depending on which process is more distinctive. Consequently, the location of 1:1 mixture shifts either in the direction of 7:3 or 3:7. A slight deviation from this observation was demonstrated by the mixtures with sucrose, where 1:1 mixture was moved towards the upper part of the map (Fig. 5F). The way the samples were allocated all over the topological map evidently signified the lack of incompatibility between the excipients (microcrystalline cellulose, glucose, sucrose) and API. As regards glicocol, sorbitol and arabic gum (Fig. 5B, C and E), the thermal decomposition proceeded differently than it was presumed. The samples, which were expected to be located in near vicinity, were scattered away from each other. This means that the thermal profiles of those substances in combination with caffeine diverge substantially from those performed using single ingredients. It is particularly distinct in the case of caffeine-glicocol mixtures (Fig. 5B), where samples with dominant caffeine content (7:3, 9:1) and those containing more
3.3. Selection of key temperature values The SOM detection of incompatibility between API and excipient was performed on the basis of only selected temperatures of mass loss (T5, T20, T35, T55, T70) due to high linear correlations (above 0.9) between variables. In further step, therefore, it was important to establish the temperature values of key importance in such detection. Taking into consideration that in dissimilar study, specified single values of mass loss temperatures would not properly detect incompatibility, this search was also focused on finding the temperature range of potential detectability. The CA dendrograms presented in Fig. 6 indicate that the temperatures of the small (T5) and large mass loss (T70) do not contribute 226
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 5. The arrangement of caffeine and its mixtures at the ratios of 9:1, 7:3, 1:1, 3:7, 1:9 on the 5 × 5 topological maps with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
with glicocol, sorbitol and arabic gum, for which the incompatibilities were established. It may point out to T20 that it is a distinctive feature of some interactions happening between constituents of mixtures. For the rest of excipients i.e. microcrystalline cellulose, glucose and sucrose, the clusters of wT20, wT35 and wT55 are separated from the cluster
much to the detection of incompatibility. In all tree diagrams the weights for temperatures of 5 and 70% mass loss (wT5, wT70) are linked together. However, in three cases (Fig. 6B, C and E) the value of weights for the temperature of 20% mass loss (wT20) was connected to them, which suggested similar thermal characteristics of caffeine mixtures
Fig. 6. The CA dendrogram of weight values for selected temperatures for caffeine and its mixtures with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
227
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 7. The PCA plot of weight values for selected temperatures for caffeine and its mixtures with (a) microcrystalline cellulose, (b) glicocol, (c) sorbitol, (d) glucose, (e) arabic gum and (f) sucrose.
Table 2 The prediction of compatibility/incompatibility between ingredients of pharmaceutical mixtures using other analytical approaches. Analytical approaches
TG-SOMs DSC FTIR PXRD
Excipients Glicocol
Sorbitol
Glucose
Sucrose
Microcrystalline cellulose
Arabic gum
× × o o
× × × o
o o o o
o o × o
o o o o
× × × o
o − compatibility, × –incompatibility.
they are positively correlated and influence the PCA model in similar ways. The variable wT35 is negatively correlated to wT5 and wT70, because it is positioned opposite to the place where wT5 and wT70 are. Moreover, this denotes that wT35 contributes to the PC model in ways different from wT5 and wT70. Similar situation takes place in the case of wT20 and wT55, which are located on opposite sides of the plot on PC2 axis for glicocol, sorbitol and arabic gum (Fig. 7B, C and E). Due to high linear correlations (above 0.9) between certain pairs of temperatures, the results of PCA and CA results can be expanded by including the temperatures not used as input vector to SOMs. Consequently, the temperature range for the mass loss from 20 to 55% can be regarded as the one with potential ability to detect incompatibilities.
composed only of wT5 and wT70. The PCA plot is shown in Fig. 7. The first two PCs explain 72.0790.79% of variance depending on the mixture, so it was safe to assume that two-dimensional plots PC1 and PC2 were sufficient for data exploration. Although, the PCA results of weight values of some temperatures of mass loss were slightly problematic to interpret, some regularities were identified. For all analysed mixtures wT5 and wT70 were positioned together and were characterized by high values of PC1 (either positive or negative). The weight values for 20% mass loss of glicocol, sorbitol and arabic gum included in mixtures with caffeine (Fig. 7B, C and E) assumed the value of PC1 close to 0 and high values of PC2 close to 1. Another consistency was the fact that wT55 for microcrystalline cellulose, glucose and sucrose (Fig. 7A, D and F) was described by intermediate values of both PC1 and PC2 (about ± 0,5) with the tendency to move towards wT5 and wT70. The opposite situation was observed in the case of glicocol, sorbitol and arabic gum. In view of the fact that the values of wT5 and wT70 are around 0 in the PC2 axes means that they contribute very little to the variance explained by PC2. At the same time both variables greatly contribute to the variance explained by PC1, which is by definition the highest. The variables wT5 and wT70 are closely positioned, which indicates that
3.4. Confirmation of SOM outcome The incompatibility detection performed by SOM was verified by DSC, FTIR and XRPD (Table 2). Due to the similarity of the results for the mixtures, in which interaction with API was established, it was decided to present them on the example of sorbitol. The DSC scans, FTIR spectra and PXRD diffractograms of binary 228
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
Fig. 8. DSC traces of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9. Fig. 9. FTIR spectra of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
referred to sorbitol, while the second small broader peak around 159 °C was correlated with polymorphic transition of caffeine. There are some published reports on the compatibility study between API and mannitol [22], probably thermal behaviour of mannitol in its mixtures with API is similar as sorbitol. Sorbitol and its isomer − mannitol are sugar alcohols belonging to polyols. The mechanism of reaction of mannitol with API is probably similar to mechanism of
mixtures were compared to those of individual ingredients. DSC scan of caffeine (Fig. 8) displays an endothermic peak due to a polymorphic transition at 159.42 °C [19,20], before the final melting at 236.32 °C [19–21]. Through the comparison of DSC scans of caffeine, sorbitol with those of binary mixtures, it was observed the absence of melting peak of caffeine at around 236 °C, suggesting incompatibility between API and excipient. The first sharp peak found at about 102 °C was 229
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
increased. Also, the absorption bands were encountered in the region of wavenumbers of 1200–1000 cm−1, corresponding to stretching vibrations of CeC and CeO of sorbitol and bands of caffeine (bending of CH, rocking of CH3, rocking of CH3 out of plane [25]) were changed. Any changes found in absorption bands of API in FTIR spectra of mixtures suggest incompatibility between API and excipient. Those changes were also found in the case of glicocol and arabic gum. The spectra of caffeine mixtures with microcrystalline cellulose, glucose and sucrose were a sum of all bands of both ingredients. The application of PXRD in the evaluation of interactions between API and excipients is sustained by the fact that a crystalline compound exhibit a unique set of diffraction peaks. Any interaction, which occurs between two crystalline ingredients or between a crystalline and an amorphous one affects the aspect of the PXRD pattern. As shown in Fig. 10, the peaks of caffeine with higher intensity are observed at 11.96°, 24.03°, 26.38°, 27.00°, revealing its high crystallinity. The diffractograms of mixtures are simple superimposition of patterns observed for single ingredients, which informs that the presence of excipient in mixtures of caffeine does not influence its crystallinity and nature. No incompatibility can be established from those diffractograms. Taking above into consideration, it must be stated that only DSC and FTIR data are consistent with those obtained using SOMs. It revealed that glicocol, sorbitol and arabic gum are incompatible with caffeine.
reaction of sorbitol with API. These excipients are generally used as tablet and capsule diluents, sweetening agents, plasticizers [23]. Bharate et al. [22] reported that mannitol was found to be incompatible with omeprazole, primaquine, and quinapril, which are all aminocontaining drugs. Stulzer et al. [24] indicated that interactions between API as piroxicam due to a possible hydrogen bond formed by nitrogen molecules of piroxicam and CeH and CH2 groups of mannitol. The FTIR spectra of caffeine mixtures with sorbitol (Fig. 9) show that characteristic bands of caffeine in the spectral range of 1800–1500 cm−1 and 1200–1000 cm−1 were altered. It can be observed that two overlapped bands (at 1700 cm−1 and 1660 cm−1 due to stretching of C]O) are split into two distinct ones, which indicates a possible physico-chemical interaction between these ingredients, i.e., hydrogen bond formation. Furthermore, the intensity of absorption band of caffeine at 1548 cm−1, attributable to bending vibrations of HeC]N and stretching vibrations of rings of imidazole and pyrimidine
4. Conclusions The application of ANNs for compatibility/incompatibility detection based on TG data provided an effective approach for resolving such issues. The SOM specified glicocol, sorbitol and arabic gum as the substances which interfere with caffeine. The outcomes of SOM studies were confirmed by complementary techniques − DSC and FTIR, proving that the application of TG combined with ANNs to be a promising alternative. The CA and PCA calculations revealed that the initial temperatures when the mass loss is slim (T5, T10, T15) and the ending temperatures when the mass loss is high (T60, T70, T75) cannot be regarded as a potential indicator of incompatibility. The range of temperatures, which can be the most valuable for such detection, was established for the mass loss from 20 to 55%. Acknowledgment The investigations were financially supported by a statutory research, Grant No. 02-0015/07/505, from the Ministry of Science and Higher Education, Poland. References [1] M. Ch. Adeyeye, H.G. Brittain, Preformulation in Solid Dosage Form Development, Informa Healthcare Inc., New York, 2008. [2] R. Chadha, S. Bhandari, Drug-excipient compatibility screening −role of thermoanalytical and spectroscopic techniques, J. Pharm. Biomed. Anal. 87 (2014) 82–97. [3] A. Marini, V. Berbenni, S. Moioli, G. Bruni, P. Cofrancesco, C. Margheritis, M. Villa, Drug-excipient compatibility study by physicochemical techniques; The case of indomethacin, J. Therm. Anal. Calorim. 73 (2003) 529–545. [4] C.M.S. de Mendonça, I.P. de Barros Lima, C.F.S. Aragão, A.P.B. Gomes, Thermal compatibility between hydroquinone and retinoic acid in pharmaceutical formulations, J. Therm. Anal. Calorim. 115 (2014) 2277–2285. [5] D.Q.M. Craig, M. Reading, Thermal Analysis of Pharmaceuticals, Taylor & Francis Group, LLC, Boca Raton, 2006. [6] B. Rojek, M. Wesolowski, Compatibility studies of hydrocortisone with excipients using thermogravimetric analysis supported by multivariate statistical analysis, J. Therm. Anal. Calorim. 127 (2017) 543–553. [7] I. Singh, P. Juneja, B. Kaur, P. Kumar, Pharmaceutical applications of chemometric techniques, ISRN Anal. Chem. 2013 (2013) 13. [8] S. Matero, Chemometrics Methods in Pharmaceutical Tablet Development and Manufacturing Unit Operations, Publications of the University of Eastern Finland Dissertations in Health Sciences, Kuopio, 2010. [9] M. Wesolowski, B. Suchacz, Artificial neural networks: theoretical background and pharmaceutical applications: a review, J. AOAC Int. 95 (2012) 652–668.
Fig. 10. PXRD patterns of (a) caffeine, (g) sorbitol and their mixtures at the molar ratios: (b) 9:1, (c) 7:3, (d) 1:1, (e) 3:7, (f) 1:9.
230
Thermochimica Acta 659 (2018) 222–231
B. Rojek et al.
[18] J. Zupan, J. Gasteiger, Neural Networks in Chemistry and Drug Design: An Introduction, second ed., VCH, Weinheim, 1999. [19] R.A. Carlton, Pharmaceutical Microscopy, Springer Science + Business Media, LLC, New York, 2011. [20] W. Lund, The Pharmaceutical Codex, twelfth ed., The Pharmaceutical Press, London, 2004. [21] Polish Pharmacopoeia VIII, Ministry of Health, Warsaw, 2008. [22] S.S. Bharate, S.B. Bharate, A.N. Bajaj, Interactions and incompatibilities of pharmaceutical excipients with active pharmaceutical ingredients: a comprehensive review, J. Excipients Food Chem. 1 (2010) 3–26. [23] R.C. Rowe, P.J. Sheskey, M.E. Quinn, Handbook of Pharmaceutical Excipients, sixth ed., American Pharmaceutical Association, Pharmaceutical Press, London, 2009. [24] H.K. Stulzer, M.P. Tagliari, A.P. Cruz, M.A.S. Silva, M.C.M. Laranjeira, Compatibility studies between piroxicam and pharmaceutical excipients used in solid dosage forms, Pharm. Chem. J. 42 (2008) 215–219. [25] F. Ucun, A. Sağlam, V. Güçlü, Molecular structures and vibrational frequencies of xanthine and its methyl derivatives (caffeine and theobromine) by ab initio Hartree −Fock and density functional theory calculations, Spectrochim. Acta A 67 (2007) 342–349.
[10] J. Djuris, Computer-Aided Applications in Pharmaceutical Technology, Woodhead Publishing Cambridge, 2013. [11] J.K. Basu, D. Bhattacharyya, T. Kim, Use of artificial neural network in pattern recognition, Int. J. Software Eng. Appl. 4 (2010) 23–34. [12] J. Malone, K. McGarry, S. Wermter, C. Bowerman, Data mining using rule extraction from Kohonen self-organising maps, Neural Comput. Appl. 15 (2006) 9–17. [13] T. Kohonen, Self-Organising Maps, Springer, Berlin, 2000. [14] N. Sbirrazzuoli, D. Brunel, Computational neural networks for mapping calorimetric data: application of feed-forward neural networks to kinetic parameters determination and signal filtering, Neural Comput. Appl. 5 (1997) 20–32. [15] N. Sbirrazzuoli, D. Brunel, L. Élégant, Neural networks for kinetic parameters determination, signal filtering and deconvolution in thermal analysis, J. Therm. Anal. Calorim. 49 (1997) 1553–1564. [16] M. Wesolowski, B. Suchacz, Classification of rapeseed and soybean oils by use of unsupervised pattern-recognition methods and neural networks, Fresh J. Anal. Chem. 371 (2001) 323–330. [17] M. Wesolowski, B. Suchacz, Neural networks as a tool for quality assessment of soybean and rapeseed oils analyzed by thermogravimetry, J. Therm. Anal. Calorim. 68 (2002) 893–899.
231