Biotechnol. Prog. 1990, 6, 498-503
498
Intact Chemoreceptor-Based Biosensors Todd Q. Barker, R. Michael Buch, and Garry A. Rechnitz* Hawaii Biosensor Laboratory, Department of Chemistry, University of Hawaii, 2545 The Mall, Honolulu, Hawaii 96822
A receptrode biosensor is presented that uses intact chemoreceptor-based molecular recognition from antennular structures of the Hawaiian swimming crab species Portunis sunguinolentus. The sensor is coupled to a learning, pattern recognition calculation for performing analytical chemistry. Action potential waveforms are used to establish the identity of individual action potential types that can be associated t o particular analytes. T h e pattern recognition calculations used are referred t o as cluster analysis (CA) and principal component analysis (PCA). Action potential similarities are determined by using a dendrogram plot of the cluster analysis results and further elucidated by using principal component scores plots. Quantitative analysis was performed after classification of analyte and background responses. Chemoresponses to salinity and trimethylamine N-oxide, two chemical constituents that are found in the crustacean living environment, were investigated and gave analytic responses over several orders of magnitude.
Introduction Intact chemical recognition elements found in nature, such as the chemoresponsive elements found in certain crustacea, provide a novel approach to performing chemical analyses. The neural network in the crustacean antennular sensing structure, during times of neuronal activity from chemical stimulation, gives rise to data that can be used to perform chemical analysis. Recently, a body of literature has accrued in the field of neurophysiology regarding chemoreceptor chemical selectivity (Atema et al., 1982). Studies have focused on the role of chemoreception in olfaction and gustation of aquatic species, investigating exquisite sensitivity and selectivity for different chemical substances (Bauer and Hatt, 1980; Fuzessery et al., 1978). The studies used certain classes of chemicals as chemoreception probes, usually chemicals or substances of direct relevance for the survival and success of the species of interest. Chemoreception phenomena originate from the afferent innervation of sensory seta. In the crab, the aesthetasc seta of the antennules are responsible for olfaction, with at least 20 neurons per aesthetasc setum in decapods (Ache, 1982), and in the case of Panulirus argus, as many as approximately 350 neurons populate each aesthetasc sensillum (Ache, 1985). These arrays of sensing sites can be monitored with standard electrophysiologic techniques, potentially providing a wealth of information about the total chemical environment of the sensor. In our laboratory, a biosensor probe is being developed, which uses the molecular recognition from olfaction in swimming crabs. The nature of the neuronal connectivity (Ache and Derby, 1985; Sandeman, 1982) of the chemoreceptor sites has prompted our view that the chemoreceptive structures in the antennules can be utilized as array biosensors, consisting of several chemically selective, molecular recognition elements. This new type of biosensor, referred to as a “receptrode”, has been used to analyze amino acid solutions (Buch and Rechnitz, 1989b) and other substances, which are perhaps not as chemically relevant to a crab’s existence: insect molting hormone and kainic acid (Buch and Rechnitz, 1989a), a powerful chemostimulant used to induce seizure activity. In the present 8756-7938/90/3006-0498$02.50/0
work, the first stage of developing a receptrode as an array detection device is reported. Automatic learning calculations, which are widely used in the discipline of pattern recognition (Kittler, 1988), are used to separate analytical responses from multiunit background signals.
Experimental Section Reagents and Materials. All solutions were prepared from reagent-grade chemicals without further purification. 18 Mi2 purity water was used for all solution preparations. Artificial seawater was prepared by using the Woods Hole formula, and a nerve bathing solution was prepared by using the Panulirus saline formulation (Cavanaugh, 1964; Mulloney and Selverston, 1974). Fresh trimethylamine N-oxide, kainic acid, and glutamate stock solutions were prepared a t millimolar concentrations before each experiment. All receptrode-based measurements employed the antennular preparations from the Hawaiian swimming crab Portunis sanquinolentus. This species was obtained from Kaneohe Bay, on the island of Oahu, Hawaii. Experimental Equipment. Synthetic data were generated to test the data analysis procedures. A waveform generator (WaveTek, San Diego, CA) was used to generate simulated action potentials that were used as probes for developing noise reduction measures. The sinusoids were digitized in the same manner as actual electrophysiological data. Actual data were collected by making extracellular neurophysiological measurements of antennular nerve fibers with glass capillary electrodes after excising the antennule from the crab. The transmembrane flux of ions was measured relative to a reference electrode, which is immersed in a neurophysiological bathing solution. The details of making the extracellular measurement have been discussed at length in previous work from this laboratory (Buch and Rechnitz, 1989b). In those studies a small electrode tip was desired since it provided a measure of selectivity, allowing connection with few neurons, increasing the probability of obtaining a single unit signal, which is easily interpreted. A 10-km electrode was well suited to
0 1990 American Chemical Society and American
Institute of Chemical Engineers
Biotechnol. Prog., 1990, Vol. 6,No. 6
499 Artificial S e i Water (Carrier Stream)
I
Injection
Valve
Flow Cell Mounted on
Oscilloscope
Microscope Stage
I
3
20
30
Time (s)
Figure 1. Multiunit data showing a salinity-based chemore-
sponse. The data are shown sampled at two different sampling rates: at 15 ms per sample and a single action potential sampled at 50 ks per sample, which is plotted in the inset figure.
accomplish this. It is important to emphasize in this report that the glass capillaries were sized to allow measurement of several neurons in the crab antennule. Moderately large electrodes were fabricated that allow measurement of several units simultaneously. Outer diameters of these electrode tips were on the order of 100 pm. The larger electrodes used here are positioned in the same way as in the previous studies, with light suction applied to the electrode filling solution, allowing for a seal between neural tissue and the electrode tip. Multiunit responses result when several neurons are within range of the measurement electrode. Figure 1shows a plot of multiunit neuronal activity during a typical receptrode measurement. In the experiment, artificial sea water continuously flows over the antennular tip. The neuronal activity during this time period, when no chemostimulant is present, is primarily from mechanoreception. An injection of analyte into the flowing system results in a period of chemostimulus neuronal activity, which is mixed with the background mechanoresponse. As the analyte is swept past the antennule, the chemoresponse neuronal activity rises and then drops off. The data plotted in Figure 1show this effect, plotting the increase in neuronal activity, which is margined by background activity. The data set in Figure 1 is a time sequence of rapidly fluctuating electrochemical potentials, consisting of a number of different classes of waveforms. Each class of waveform is associated with the connection of the pickup electrode to a particular neuron. The inset in Figure 1shows the waveform shape of a typical action potential, illustrating t h a t , by increasing the data acquisition sampling rate, each action potential can be carefully described by both the amplitude as well as the shape of the action potential. An accurate record of the waveform shape can be obtained by using an analog to digital (A/D) conversion with computer storage of the data. In Figure 2, the experimental configuration is summarized. In these studies, a 12-bit AID conversion of the raw data was accomplished by using a MacAdios-SE (GW Instruments Inc., Somerville, MA) and data-spooling to a MacIntosh hard disk (Apple Computer Co., Cupertino, CA). The data were stored on a large capacity, high-speed hard disk so that several alternative data analysis strategies could be tested on large amounts of data. The data could be digitized directly from the experiment or be recorded on magnetic tape for later conversion. After converting and storing the data, the data were transferred to a work-
I
d
Microphone For Experimental Dictation
IOU I =i, I
Computer For Data Analysis
Figure 2. Experimental configuration.
station for data analysis using high-speed data transfer. Data analysis programs were written for use with an Intel 80386-based microcomputer (Intel Corp., Santa Clara, CA) equipped with an Intel 80387 floating point processer or a Spark work-station (SUN Microsystems Inc., Mountain View, CA). The sampling rate of the analog to digital converter is a critical consideration. An appropriate sampling rate is needed to insure high waveform accuracy. The mathematical analysis of the data will not produce meaningful results unless the sampling speed is high enough (Glaser and Ruchkin, 1976). As a general operating principle, the A/D conversion time was set at 50 NS per sample. By using this sampling rate, an action potential having a period of approximately 1ms is described by 20 data points. This rate allows for collection of sufficient information t o describe the action potentials, which are typically in the range of 500-1500 Hz. Numerical Preprocessing. The first stage in the data preprocessing involved peak identification and waveform sampling. A numerical first derivative of the original data was used to identify action potential peak locations, followed by sampling the data points that occur before and after each peak maximum. By using this approach, an abbreviated waveform is retained for each action potential. These contain sufficient information about the amplitude and temporal shape characteristics of the action potentials. In this way, a data compression takes place by extracting the high information content data and placing it in a new data matrix, which can be analyzed by multivariate data analysis methods. The data matrix constructed from sampled action potentials is n-dimensional with respect to the number of data points used to define a sampling window. The data matrix
X = x ( ~ , ~ )for i = 1, ..., m; k
= 1, ...,1
(1) states that m sampled objects, the action potentials, are
Biotechnol. Prog., 1990,Vol. 6,No. 6
500
described by 1 variables, the electrochemical measurements that are used in the sampled time envelope. It can be said that these variables are time-resolution elements, which are realized by the electrochemical potential at each point in time. The results of multivariate data analysis are affected by the presence of noise. Two types of noise are readily identified: interobject noise, or low-frequency noise, and intraobject noise, relatively high-frequency noise. Interobject noise results if the action potential units, occurring at frequencies between 500 and 1500 Hz, are superimposed on a low-frequency (approximately 60-Hz) background. The result is that the aperiodic sampled action potentials contain a random offset component. The power contribution of this noise component is described by lower frequencies relative to those describing the action potential and, because of the random occurrence of each action potential, is distributed randomly between objects (rows of the data matrix). In contrast, intraobject noise is described by higher frequencies than those describing action potentials. Both noise components can be subdued by using Fourier spectral analysis. A Fourier power spectrum analysis of an action potential sequence yields a frequency domain representation of the action potential (Oppenheim, 1975). This is done by first transforming the data as follows: N-1
F ( n ) = ~ C f ( k ) e - j ~for~ n~=~0,1, / ~...,N
- 1 (2)
k=O
O
A
0.50
I i - 7 1.00 r
Figure 3; Cluster analysis dendrogram. The centroid algorithm was applied in a cluster analysis of synthetic data consisting of the following units: B, 1200 Hz, 1.5V; C, 1200 Hz, 0.7 V; A, 600 Hz. 1.5 V.
This results in formation of a similarity matrix, which is then evaluated by any of a number of cluster analysis algorithms. A general description of membership to a cluster begins by considering the data matrix as a set of U objects. Calculation of the similarity matrix begins by first determining d by the following: U X U R , which assigns a real number, R, for each pair of elements of U. A well-known calculation of distance is the Euclidean distance:
-
where F ( n ) is the Fourier representation of the action potential f ( k ) . The power spectrum is then calculated by
P = IF(n)I2
(3) In this way, the power spectrum frequencies used to describe an action potential are realized by the power density for each frequency of the power spectrum. In the studies that follow, a Fourier power spectral analysis was used to form a transformed, relatively noise-free data matrix. This was done by transforming t h e action potentials to their associated power spectra, followed by forming a multivariate data matrix from a selected frequency range. Frequencies that are subject to interand intraobject noise are rejected while frequencies that describe the action potential waveform are retained. In this way, a data matrix can be formed that clarifies the description of the action potentials by disregarding noise contributions. In these studies, a waveform-frequency envelope was chosen over the range from 300 to 1700 Hz, frequencies from this range were used in building a data matrix since a typical action potential waveform period ranged between 500 and 1500 Hz. Pattern Recognition Methods. Pattern recognition is a numerical method developed in the field of artificial intelligence (Jurs et al., 1969) and has grown into a widely practiced chemometric technique (Brown et al., 1988).The use of pattern recognition in the present paper can be divided into two numerical procedures: cluster analysis (CAI and principal component analysis (PCA). Both are used as off-line learning calculations to identify the presence of action potential classes, providing information that leads to object classification. Once a class of objects is identified, it is possible to associate a particular analyte to that object, which provides a way for attaining analyte selectivity. Analyte quantitation is obtained by determining the number of members in a given set of classified objects. CA begins by calculating an appropriate metric of similarity for all objects in the data space (Spath, 1980).
for the vectors x and y. When applying cluster analysis algorithms, it is not generally known which one will be most appropriate. A good approach is to have several cluster analysis algorithms available. T h e one t h a t provides t h e most clearly distinguishable results is considered most meaningful. In this study, general agreement was obtained by applying several different clustering algorithms. Farthestneighbor, nearest-neighbor, centroid, median, and Lance and Williams algorithms gave good agreement in the classification assignment of action potentials. The results of a cluster analysis are summarized in a dendrogram plot, showing object similarity through the linkage of the objects with respect to a similarity metric. In the plot, action potentials (data objects) that have similar waveforms are linked a t a level of similarity approaching 1. Relatively dissimilar action potentials are linked a t a similarity index value approaching 0. As an example, a synthetic data set was created that consisted of three different action potential units. Figure 3 shows the results of applying the cluster analysis to this data set, showing that the data consists of three different units, as expected. The three units can be identified by noting the vertical linkage lines that are present at different levels of similarity. In this example, each unit consisted of 10 members, represented by the short branches at the extremities of the tree plot, at a similarity index = 1.0. Three units are clearly identified at levels of similarity ranging from approximately 0.25 to 0.75. These are denoted by the longer vertical lines in the midportion of the plot. From such a result it would be clear t h a t sufficient data structure exists, allowing for classification into groups of similar objects and cueing the analyst that further analysis by PCA may be informative. PCA is a method that has been used extensively in chemistry (Massart et al., 1987),providing a way to reduce multidimensional data to fewer dimensions. It is applied
501
Biotechnol. Prog, 1990, Vol. 6, No. 6
Table I. Eigenvalues from Analyzing Synthetic Neuronal Data % cumulative eigenvector eigenvalue variance variance 1
2 3 4 5
0.116 626 7 0.027 063 6 0.004 388 7 0.002 801 8 0.001 016 8
76.23 17.69 2.86 1.83 0.05
76.23 93.92 96.79 98.62 99.28
by using a numerical procedure, such as multiple linear regression, Eigen analysis, or singular value decomposition, to solve for a description of the data variance of old, correlated variables (measured quantities) by using a minimum set of new, uncorrelated variables (linear combinations of the old variables). In the previous section, cluster analysis was described to provide information about classes of particular objects, the action potential units. By identifying the frequency and amplitude character of each action potential unit, PCA is used to further elucidate action potential classes. In this study, Eigen analysis was employed to perform PCA. By using this approach, PCA can be summarized by RA = XA (5) In eq 5, a solution is found for the characteristic vector, or eigenvector, A, such that premultiplication by the correlation matrix, R,is equal to the product of A and the characteristic root, A. The largest characteristic root is assigned the variance of the first principal component. The associated characteristic vector, or loading, gives the relative weights of each of the variables. The first principal component describes the bulk of the data variance. Decreasing amounts of data variance are described by subsequent calculation of lesser components. I t is convenient if the analytically useful information is contained in the first two or three principal components, leaving the noise information in the latter components. This has the effect of concentrating relatively noise-free information in fewer variables. When this is possible, a plot of the principal component scores provides a new way of graphically visualizing the data. To demonstrate the use of PCA, the previously mentioned synthetic data was also analyzed with PCA. Table I shows the results of performing the Eigen analysis, showing that 93.92 % of the data variance can be attributed to the first two principal components. In this way, 93.92% of the data information from a 15-dimensionaldata matrix (the number of columns) is preserved in two dimensions, which can be graphically visualized. Figure 4 shows the scores plot by using the first two principal components. As in the cluster analysis, the classification of action potential units is possible, showing three identifiable swarms of points, corresponding to the three different synthetic units. However, additional information about the amplitude and frequency characteristics of the action potential units is now available. The orientation of swarms shows t h a t they are differentiated with respect t o amplitude by principal component 1and with respect to frequency (the period of the waveform) by principal component 2. The plot shows that the two units which only differ in amplitude are distributed on the principal component 1axis. Similarly, the units that differ only in waveform frequency are differentiated with respect to the second principal component.
Results and Discussion Calibration of a Salinity-BasedChemoresponse. An analyte-specific response was observed for solution salinity
IO
V
eV 6-
v v V 1-
N
k?
2-
-.
,
,
-8
-6
,
,
,
,
,
0
2
4
~
-6 -10
-1
-2
6
8
Kl
F i g u r e 4. Principal component analysis scores plot. The first two principal component scores were used to evaluate the synthetic data set. PCA was applied to a synthetic data set that consisted of the following units: A, 1200 Hz, 1.5 V; 0, 1200 Hz, 0.7 V; V, 600 Hz, 1.5 V.
by using receptrodes constructed with crab antennules from the species P. sanquinolentus. Early behavioral studies used this species to report that such a sensitivity existed (Lockwood, 1967). The chemoresponse was only observed when salinity changes were introduced and not observed for the other chemicals used as chemoresponse screening analytes. Kainic acid, glutamate, betaine, and trimethylamine N-oxide gave no response. A report of the salinity response is presented here as part of a continued investigation of the analytical capabilities of receptrode-based measurements. As described previously, CA was performed to evaluate the data structure of the chemoresponse data. A multivariate data matrix was constructed for analysis by assembling a representative set of objects from background and chemostimulus signals. By sampling both the background neuronal activity and the neuronal activity during an observed chemostimulus epoch, the data matrix includes information that is needed in classifying the background and chemostimulus responses. In this treatment, the analysis employed the centroid clusteranalysis algorithm (Spath, 1980), and is summarized in the dendrogram plot in Figure 5. Inspection of the plot shows that three primary classes of action potentials are identifiable at a similarity index threshold of approximately 0.6. Other subclasses are feasible if higher degrees of similarity are considered. From this, a further analysis is warranted by using PCA. The salinity chemoresponse data was analyzed by using PCA. Classification of chemostimulus action potentials was achieved by identifying action potentials that appear in a unique region of principal component scores space, apart from the appearance of background action potentials. Table I1 reports the results of Eigen analysis, showing that 98.88% of the data variance can be assigned to the first two principal components, allowing for a visually descriptive plot of the principal component scores. In Figure 6, the scores plot from the first two principal components is shown. The scores plot indicates that there are three action potential units, which is in agreement with the interpretation of the cluster analysis dendrogram. In the plot, a swarm of action potentials are located close to the plot origin, corresponding to cluster B in Figure 5. Cluster C is represented by the action potentials that extend “upward” with respect to the second principal component axis. Cluster A is associated with the swarm centered a t
Biotechnol. Prog., 1990, Vol. 6,No. 6
502 120
1 I
\\ -
01 0.00
I
I
I
I
I
I
0.05
0.10
0.15
0.20
0.25
0.30
lOO-1
Ly
Figure 5. Cluster analysis dendrogram from analyzing salinity chemoresponse data. Three units specified as A, B, and C are noted by using a similarity index threshold of approximately 0.6. Table 11. Eigenvalues from Analyzing Salinity-Based Chemoresponse 9% cumulative eigenvector
eigenvalue
variance
variance
9.327 636 8 0.560 456 1 0.085 635 7 0.014 386 5 0.005 699 4
93.27 5.60 0.85 0.14 0.05
93.27 98.88 99.73 99.88 99.93
8-
A
A 6-
k 4-
A A
N U
a
2-
A
-2 -0
-b
-4
-2
0
I
4
I
8
io
12
Y
K l
Figure 6. Principal component scoresplot from analyzing salinity chemoresponse data. The first two principal component scores were used to generate the plot of the background, 0,and chemoresponse, A. the coordinates 12 and -1 of principal components 1 and 2, respectively. A calibration curve was generated by analyzing measurements of a series of dilutions of artificial seawater injected into a flowing artificial seawater carrier. The data analysis results indicate that the action potentials associated with cluster B were from background neuronal activity, since the population of this group remains constant over a range of concentration of salinity. The other action potential units, however, show a concentration dependency. Figure 7 shows the result of plotting the frequency of firing of the concentration-dependent regions of cluster A. A concentration-dependent trend in the amount of chemoresponse neuronal activity is obvious from the plot. Calibration of Responses to Trimethylamine N-Oxide. Trimethylamine N-oxide (TMO) is a major constituent as a component of the cellular sap of marine organisms (Lockwood, 1967). As a result, crabs may have
,Oi
Concentration (ppt NoCI)
Figure 7. Calibration plot of chemoresponseto salinity. A series of diluted solutions of artificial seawater solutions were used to generate the calibration plot. A triplicate injection waa performed to calculate a confidence limit for an intermediate concentration in the calibration curve. The error bar shows the spread based on one standard deviation. the ability to sense this substance as a way of detecting food sources. A data set was collected of chemoresponses for TMO by using the P. sanquinolentus based receptrode. As with the salinity-based chemoresponse, the data involving TMO were analyzed with CA and PCA to generate a calibration response curve. The analysis of this data set was identical with that presented for the salinity responses, starting with an evaluation of the data structure by using CA and then further elucidation with PCA. To summarize the data analysis, the principal component scores plot is shown in Figure 8. In this case, 85.34% of the data variance is described by the first two principal components. This reflects that less data structure was available to facilitate the interpretation of the data. Like the salinity response data, a dense compaction of action potentials is observed near the origin of the scores plot. These do not appear to be associated with chemoresponsive action potentials, as they do not vary with respect t o concentration. However, the presence of a well-separated swarm of analytical action potentials greatly aids the interpretation of the chemoresponse neuronal activity. PCA was performed on a series of dilutions of TMO, and the calibration curve was generated by counting the number of action potentials from the previously indicated analytical region of scores space. The calibration curve, shown in Figure 9, demonstrates that the response is linear over a wide concentration range. The frequency of firing increased linearly as more concentrated solutions of TMO were passed over the antennular tip. I n the studies involving TMO calibration, it was sometimes possible to observe chemoresponse activity at much lower concentrations than are shown in the calibration plot of Figure 9, ranging as low as 1.0 X
Conclusions In this study, a naturally occurring chemical sensor array is explored that uses multiunit measurements of neuronal activity in the antennules of a swimming crab. A calibration method has been outlined t h a t relies on identification of analyte-specific chemoresponse regions in t h e plotted results of t h e unsupervised pattern recognition procedures CA and PCA. This was done by analyzing a series of calibration data sets relative to a
503
Biotechnol. Prog., 1990, Vol. 6 , No. 6
0
8
study has made use of single analyte responses, which were in the presence of nonchemoresponse background. A logical extension of this work is in progress that makes use of several simultaneous chemoresponse signals, thus taking advantage of the multiplexed information from receptrode measurements.
@ '
0% 0
Acknowledgment This work was supported by National Science Foundation Grant CHE-8921156 Much thanks goes to the Hawaiian Institute of Marine Biology for assisting in obtaining materials for this work.
0
A A 0
-6 -8
-2
-6
PCl
Figure 8. Principal component scores from analyzing a TMO chemoresponse. The training set action potentials are shown as follows: A,background; 0,chemoresponse. The training action potentials were referenced to all subsequent calibration data sets. Also plotted is a chemoresponse from a 1.0 X lo-* M TMO solution, 0. 25
p
0
1
m
/i/
154
m
i
Log Concentration
Figure 9. Calibration plot of chemoresponse to TMO. The standard deviation from triplicate analyses was used to calculate the error bar shown.
constant training set. By monitoring a chemoresponsespecific region for changes in the amount of neuronal activity during chemostimulation, a calibration response curve can be constructed. It should be noted that this
Literature Cited Ache, B. W. In The Biology of Crustacea Vol. 3, Neurobiology Structure and Function; Atwood, H. L., Sandeman, D. C., E&.; Academic Press: New York, 1982; p 376. Ache, B. W.; Derby, C. D. Trends Neurosci. 1985, 356-360. Atema, J.; Fay, R. R.; Popper, A. N.; Tavolga, W. N. Sensory Biology of Aquatic Animals; Springer-Verlag: New York, 1988. Bauer, U.; Hatt, H. Neurosci. Lett. 1980, 17, 209-214. Brown, S. D.; Barker, T. Q.; Monfre, S. L.; Larivee, R. L.; Wilk, H. R. Anal. Chem. 1988,60,252-294. Buch, R. M.; Rechnitz, G. A. Anal. Lett. 1989a,22 (13-14), 26852702. Buch, R. M.; Rechnitz, G. A. Biosensors 1989b, 4, 215-230. Cavanaugh, G . In Formulae and Methods V; Marine Biological Laboratory: Woods Hole, MA, 1964. Fuzessery, 2.; Carr, W.; Ache, B. Biol. Bull. 1978,154,224-240. Glaser, E. M.; Ruchkin, D. S. Principles of Neurobiological Signal Analysis; Academic Press: New York, 1976. Jurs, P. C.; Kowalski, B. R.; Isenhour, T. L. Anal. Chem. 1969, 41, 21. Kittler, J., Ed. Pattern Recognition: 4th International Conference; Springer-Verlag: New York, 1988. Lockwood, A. P. M. Aspects of the Physiology of Crustacea; W. H. Freeman: San Francisco, CA, 1967; pp 250-253. Massart, D. L.; Vandeginste, B. G. M.; Deming, S. N.; Mochotte, Y.; Kauffman, L. Chemometrics: A Textbook; Elsevier: Amsterdam, 1987. Mulloney, B.; Selverston, A. J . Comp. Physiol. 1974, 91, 1. Oppenheim, A. V.; Schafer, R. W. Digital Signal Processing; Prentice-Hall: Englewood Cliffs, NJ, 1975. w Sandeman, D. C. In The Biology of Crustacea Vol. 3, Neurobiology Structure and Function; Atwood, H. L., Sandeman, D. C., Eds.; Academic Press: New York, 1982, pp 1-61. Spath, H. Clllster Analysis Algorithms: Ellis-Honvood: West Sussex, U. K., 1980. Accepted September 14, 1990.
Registry No. Trimethylamine N-oxide, 1184-78-7.
498
Intact Chemoreceptor-Based Biosensors Todd Q. Barker, R. Michael Buch, and Garry A. Rechnitz* Hawaii Biosensor Laboratory, Department of Chemistry, University of Hawaii, 2545 The Mall, Honolulu, Hawaii 96822
A receptrode biosensor is presented that uses intact chemoreceptor-based molecular recognition from antennular structures of the Hawaiian swimming crab species Portunis sunguinolentus. The sensor is coupled to a learning, pattern recognition calculation for performing analytical chemistry. Action potential waveforms are used to establish the identity of individual action potential types that can be associated t o particular analytes. T h e pattern recognition calculations used are referred t o as cluster analysis (CA) and principal component analysis (PCA). Action potential similarities are determined by using a dendrogram plot of the cluster analysis results and further elucidated by using principal component scores plots. Quantitative analysis was performed after classification of analyte and background responses. Chemoresponses to salinity and trimethylamine N-oxide, two chemical constituents that are found in the crustacean living environment, were investigated and gave analytic responses over several orders of magnitude.
Introduction Intact chemical recognition elements found in nature, such as the chemoresponsive elements found in certain crustacea, provide a novel approach to performing chemical analyses. The neural network in the crustacean antennular sensing structure, during times of neuronal activity from chemical stimulation, gives rise to data that can be used to perform chemical analysis. Recently, a body of literature has accrued in the field of neurophysiology regarding chemoreceptor chemical selectivity (Atema et al., 1982). Studies have focused on the role of chemoreception in olfaction and gustation of aquatic species, investigating exquisite sensitivity and selectivity for different chemical substances (Bauer and Hatt, 1980; Fuzessery et al., 1978). The studies used certain classes of chemicals as chemoreception probes, usually chemicals or substances of direct relevance for the survival and success of the species of interest. Chemoreception phenomena originate from the afferent innervation of sensory seta. In the crab, the aesthetasc seta of the antennules are responsible for olfaction, with at least 20 neurons per aesthetasc setum in decapods (Ache, 1982), and in the case of Panulirus argus, as many as approximately 350 neurons populate each aesthetasc sensillum (Ache, 1985). These arrays of sensing sites can be monitored with standard electrophysiologic techniques, potentially providing a wealth of information about the total chemical environment of the sensor. In our laboratory, a biosensor probe is being developed, which uses the molecular recognition from olfaction in swimming crabs. The nature of the neuronal connectivity (Ache and Derby, 1985; Sandeman, 1982) of the chemoreceptor sites has prompted our view that the chemoreceptive structures in the antennules can be utilized as array biosensors, consisting of several chemically selective, molecular recognition elements. This new type of biosensor, referred to as a “receptrode”, has been used to analyze amino acid solutions (Buch and Rechnitz, 1989b) and other substances, which are perhaps not as chemically relevant to a crab’s existence: insect molting hormone and kainic acid (Buch and Rechnitz, 1989a), a powerful chemostimulant used to induce seizure activity. In the present 8756-7938/90/3006-0498$02.50/0
work, the first stage of developing a receptrode as an array detection device is reported. Automatic learning calculations, which are widely used in the discipline of pattern recognition (Kittler, 1988), are used to separate analytical responses from multiunit background signals.
Experimental Section Reagents and Materials. All solutions were prepared from reagent-grade chemicals without further purification. 18 Mi2 purity water was used for all solution preparations. Artificial seawater was prepared by using the Woods Hole formula, and a nerve bathing solution was prepared by using the Panulirus saline formulation (Cavanaugh, 1964; Mulloney and Selverston, 1974). Fresh trimethylamine N-oxide, kainic acid, and glutamate stock solutions were prepared a t millimolar concentrations before each experiment. All receptrode-based measurements employed the antennular preparations from the Hawaiian swimming crab Portunis sanquinolentus. This species was obtained from Kaneohe Bay, on the island of Oahu, Hawaii. Experimental Equipment. Synthetic data were generated to test the data analysis procedures. A waveform generator (WaveTek, San Diego, CA) was used to generate simulated action potentials that were used as probes for developing noise reduction measures. The sinusoids were digitized in the same manner as actual electrophysiological data. Actual data were collected by making extracellular neurophysiological measurements of antennular nerve fibers with glass capillary electrodes after excising the antennule from the crab. The transmembrane flux of ions was measured relative to a reference electrode, which is immersed in a neurophysiological bathing solution. The details of making the extracellular measurement have been discussed at length in previous work from this laboratory (Buch and Rechnitz, 1989b). In those studies a small electrode tip was desired since it provided a measure of selectivity, allowing connection with few neurons, increasing the probability of obtaining a single unit signal, which is easily interpreted. A 10-km electrode was well suited to
0 1990 American Chemical Society and American
Institute of Chemical Engineers
Biotechnol. Prog., 1990, Vol. 6,No. 6
499 Artificial S e i Water (Carrier Stream)
I
Injection
Valve
Flow Cell Mounted on
Oscilloscope
Microscope Stage
I
3
20
30
Time (s)
Figure 1. Multiunit data showing a salinity-based chemore-
sponse. The data are shown sampled at two different sampling rates: at 15 ms per sample and a single action potential sampled at 50 ks per sample, which is plotted in the inset figure.
accomplish this. It is important to emphasize in this report that the glass capillaries were sized to allow measurement of several neurons in the crab antennule. Moderately large electrodes were fabricated that allow measurement of several units simultaneously. Outer diameters of these electrode tips were on the order of 100 pm. The larger electrodes used here are positioned in the same way as in the previous studies, with light suction applied to the electrode filling solution, allowing for a seal between neural tissue and the electrode tip. Multiunit responses result when several neurons are within range of the measurement electrode. Figure 1shows a plot of multiunit neuronal activity during a typical receptrode measurement. In the experiment, artificial sea water continuously flows over the antennular tip. The neuronal activity during this time period, when no chemostimulant is present, is primarily from mechanoreception. An injection of analyte into the flowing system results in a period of chemostimulus neuronal activity, which is mixed with the background mechanoresponse. As the analyte is swept past the antennule, the chemoresponse neuronal activity rises and then drops off. The data plotted in Figure 1show this effect, plotting the increase in neuronal activity, which is margined by background activity. The data set in Figure 1 is a time sequence of rapidly fluctuating electrochemical potentials, consisting of a number of different classes of waveforms. Each class of waveform is associated with the connection of the pickup electrode to a particular neuron. The inset in Figure 1shows the waveform shape of a typical action potential, illustrating t h a t , by increasing the data acquisition sampling rate, each action potential can be carefully described by both the amplitude as well as the shape of the action potential. An accurate record of the waveform shape can be obtained by using an analog to digital (A/D) conversion with computer storage of the data. In Figure 2, the experimental configuration is summarized. In these studies, a 12-bit AID conversion of the raw data was accomplished by using a MacAdios-SE (GW Instruments Inc., Somerville, MA) and data-spooling to a MacIntosh hard disk (Apple Computer Co., Cupertino, CA). The data were stored on a large capacity, high-speed hard disk so that several alternative data analysis strategies could be tested on large amounts of data. The data could be digitized directly from the experiment or be recorded on magnetic tape for later conversion. After converting and storing the data, the data were transferred to a work-
I
d
Microphone For Experimental Dictation
IOU I =i, I
Computer For Data Analysis
Figure 2. Experimental configuration.
station for data analysis using high-speed data transfer. Data analysis programs were written for use with an Intel 80386-based microcomputer (Intel Corp., Santa Clara, CA) equipped with an Intel 80387 floating point processer or a Spark work-station (SUN Microsystems Inc., Mountain View, CA). The sampling rate of the analog to digital converter is a critical consideration. An appropriate sampling rate is needed to insure high waveform accuracy. The mathematical analysis of the data will not produce meaningful results unless the sampling speed is high enough (Glaser and Ruchkin, 1976). As a general operating principle, the A/D conversion time was set at 50 NS per sample. By using this sampling rate, an action potential having a period of approximately 1ms is described by 20 data points. This rate allows for collection of sufficient information t o describe the action potentials, which are typically in the range of 500-1500 Hz. Numerical Preprocessing. The first stage in the data preprocessing involved peak identification and waveform sampling. A numerical first derivative of the original data was used to identify action potential peak locations, followed by sampling the data points that occur before and after each peak maximum. By using this approach, an abbreviated waveform is retained for each action potential. These contain sufficient information about the amplitude and temporal shape characteristics of the action potentials. In this way, a data compression takes place by extracting the high information content data and placing it in a new data matrix, which can be analyzed by multivariate data analysis methods. The data matrix constructed from sampled action potentials is n-dimensional with respect to the number of data points used to define a sampling window. The data matrix
X = x ( ~ , ~ )for i = 1, ..., m; k
= 1, ...,1
(1) states that m sampled objects, the action potentials, are
Biotechnol. Prog., 1990,Vol. 6,No. 6
500
described by 1 variables, the electrochemical measurements that are used in the sampled time envelope. It can be said that these variables are time-resolution elements, which are realized by the electrochemical potential at each point in time. The results of multivariate data analysis are affected by the presence of noise. Two types of noise are readily identified: interobject noise, or low-frequency noise, and intraobject noise, relatively high-frequency noise. Interobject noise results if the action potential units, occurring at frequencies between 500 and 1500 Hz, are superimposed on a low-frequency (approximately 60-Hz) background. The result is that the aperiodic sampled action potentials contain a random offset component. The power contribution of this noise component is described by lower frequencies relative to those describing the action potential and, because of the random occurrence of each action potential, is distributed randomly between objects (rows of the data matrix). In contrast, intraobject noise is described by higher frequencies than those describing action potentials. Both noise components can be subdued by using Fourier spectral analysis. A Fourier power spectrum analysis of an action potential sequence yields a frequency domain representation of the action potential (Oppenheim, 1975). This is done by first transforming the data as follows: N-1
F ( n ) = ~ C f ( k ) e - j ~for~ n~=~0,1, / ~...,N
- 1 (2)
k=O
O
A
0.50
I i - 7 1.00 r
Figure 3; Cluster analysis dendrogram. The centroid algorithm was applied in a cluster analysis of synthetic data consisting of the following units: B, 1200 Hz, 1.5V; C, 1200 Hz, 0.7 V; A, 600 Hz. 1.5 V.
This results in formation of a similarity matrix, which is then evaluated by any of a number of cluster analysis algorithms. A general description of membership to a cluster begins by considering the data matrix as a set of U objects. Calculation of the similarity matrix begins by first determining d by the following: U X U R , which assigns a real number, R, for each pair of elements of U. A well-known calculation of distance is the Euclidean distance:
-
where F ( n ) is the Fourier representation of the action potential f ( k ) . The power spectrum is then calculated by
P = IF(n)I2
(3) In this way, the power spectrum frequencies used to describe an action potential are realized by the power density for each frequency of the power spectrum. In the studies that follow, a Fourier power spectral analysis was used to form a transformed, relatively noise-free data matrix. This was done by transforming t h e action potentials to their associated power spectra, followed by forming a multivariate data matrix from a selected frequency range. Frequencies that are subject to interand intraobject noise are rejected while frequencies that describe the action potential waveform are retained. In this way, a data matrix can be formed that clarifies the description of the action potentials by disregarding noise contributions. In these studies, a waveform-frequency envelope was chosen over the range from 300 to 1700 Hz, frequencies from this range were used in building a data matrix since a typical action potential waveform period ranged between 500 and 1500 Hz. Pattern Recognition Methods. Pattern recognition is a numerical method developed in the field of artificial intelligence (Jurs et al., 1969) and has grown into a widely practiced chemometric technique (Brown et al., 1988).The use of pattern recognition in the present paper can be divided into two numerical procedures: cluster analysis (CAI and principal component analysis (PCA). Both are used as off-line learning calculations to identify the presence of action potential classes, providing information that leads to object classification. Once a class of objects is identified, it is possible to associate a particular analyte to that object, which provides a way for attaining analyte selectivity. Analyte quantitation is obtained by determining the number of members in a given set of classified objects. CA begins by calculating an appropriate metric of similarity for all objects in the data space (Spath, 1980).
for the vectors x and y. When applying cluster analysis algorithms, it is not generally known which one will be most appropriate. A good approach is to have several cluster analysis algorithms available. T h e one t h a t provides t h e most clearly distinguishable results is considered most meaningful. In this study, general agreement was obtained by applying several different clustering algorithms. Farthestneighbor, nearest-neighbor, centroid, median, and Lance and Williams algorithms gave good agreement in the classification assignment of action potentials. The results of a cluster analysis are summarized in a dendrogram plot, showing object similarity through the linkage of the objects with respect to a similarity metric. In the plot, action potentials (data objects) that have similar waveforms are linked a t a level of similarity approaching 1. Relatively dissimilar action potentials are linked a t a similarity index value approaching 0. As an example, a synthetic data set was created that consisted of three different action potential units. Figure 3 shows the results of applying the cluster analysis to this data set, showing that the data consists of three different units, as expected. The three units can be identified by noting the vertical linkage lines that are present at different levels of similarity. In this example, each unit consisted of 10 members, represented by the short branches at the extremities of the tree plot, at a similarity index = 1.0. Three units are clearly identified at levels of similarity ranging from approximately 0.25 to 0.75. These are denoted by the longer vertical lines in the midportion of the plot. From such a result it would be clear t h a t sufficient data structure exists, allowing for classification into groups of similar objects and cueing the analyst that further analysis by PCA may be informative. PCA is a method that has been used extensively in chemistry (Massart et al., 1987),providing a way to reduce multidimensional data to fewer dimensions. It is applied
501
Biotechnol. Prog, 1990, Vol. 6, No. 6
Table I. Eigenvalues from Analyzing Synthetic Neuronal Data % cumulative eigenvector eigenvalue variance variance 1
2 3 4 5
0.116 626 7 0.027 063 6 0.004 388 7 0.002 801 8 0.001 016 8
76.23 17.69 2.86 1.83 0.05
76.23 93.92 96.79 98.62 99.28
by using a numerical procedure, such as multiple linear regression, Eigen analysis, or singular value decomposition, to solve for a description of the data variance of old, correlated variables (measured quantities) by using a minimum set of new, uncorrelated variables (linear combinations of the old variables). In the previous section, cluster analysis was described to provide information about classes of particular objects, the action potential units. By identifying the frequency and amplitude character of each action potential unit, PCA is used to further elucidate action potential classes. In this study, Eigen analysis was employed to perform PCA. By using this approach, PCA can be summarized by RA = XA (5) In eq 5, a solution is found for the characteristic vector, or eigenvector, A, such that premultiplication by the correlation matrix, R,is equal to the product of A and the characteristic root, A. The largest characteristic root is assigned the variance of the first principal component. The associated characteristic vector, or loading, gives the relative weights of each of the variables. The first principal component describes the bulk of the data variance. Decreasing amounts of data variance are described by subsequent calculation of lesser components. I t is convenient if the analytically useful information is contained in the first two or three principal components, leaving the noise information in the latter components. This has the effect of concentrating relatively noise-free information in fewer variables. When this is possible, a plot of the principal component scores provides a new way of graphically visualizing the data. To demonstrate the use of PCA, the previously mentioned synthetic data was also analyzed with PCA. Table I shows the results of performing the Eigen analysis, showing that 93.92 % of the data variance can be attributed to the first two principal components. In this way, 93.92% of the data information from a 15-dimensionaldata matrix (the number of columns) is preserved in two dimensions, which can be graphically visualized. Figure 4 shows the scores plot by using the first two principal components. As in the cluster analysis, the classification of action potential units is possible, showing three identifiable swarms of points, corresponding to the three different synthetic units. However, additional information about the amplitude and frequency characteristics of the action potential units is now available. The orientation of swarms shows t h a t they are differentiated with respect t o amplitude by principal component 1and with respect to frequency (the period of the waveform) by principal component 2. The plot shows that the two units which only differ in amplitude are distributed on the principal component 1axis. Similarly, the units that differ only in waveform frequency are differentiated with respect to the second principal component.
Results and Discussion Calibration of a Salinity-BasedChemoresponse. An analyte-specific response was observed for solution salinity
IO
V
eV 6-
v v V 1-
N
k?
2-
-.
,
,
-8
-6
,
,
,
,
,
0
2
4
~
-6 -10
-1
-2
6
8
Kl
F i g u r e 4. Principal component analysis scores plot. The first two principal component scores were used to evaluate the synthetic data set. PCA was applied to a synthetic data set that consisted of the following units: A, 1200 Hz, 1.5 V; 0, 1200 Hz, 0.7 V; V, 600 Hz, 1.5 V.
by using receptrodes constructed with crab antennules from the species P. sanquinolentus. Early behavioral studies used this species to report that such a sensitivity existed (Lockwood, 1967). The chemoresponse was only observed when salinity changes were introduced and not observed for the other chemicals used as chemoresponse screening analytes. Kainic acid, glutamate, betaine, and trimethylamine N-oxide gave no response. A report of the salinity response is presented here as part of a continued investigation of the analytical capabilities of receptrode-based measurements. As described previously, CA was performed to evaluate the data structure of the chemoresponse data. A multivariate data matrix was constructed for analysis by assembling a representative set of objects from background and chemostimulus signals. By sampling both the background neuronal activity and the neuronal activity during an observed chemostimulus epoch, the data matrix includes information that is needed in classifying the background and chemostimulus responses. In this treatment, the analysis employed the centroid clusteranalysis algorithm (Spath, 1980), and is summarized in the dendrogram plot in Figure 5. Inspection of the plot shows that three primary classes of action potentials are identifiable at a similarity index threshold of approximately 0.6. Other subclasses are feasible if higher degrees of similarity are considered. From this, a further analysis is warranted by using PCA. The salinity chemoresponse data was analyzed by using PCA. Classification of chemostimulus action potentials was achieved by identifying action potentials that appear in a unique region of principal component scores space, apart from the appearance of background action potentials. Table I1 reports the results of Eigen analysis, showing that 98.88% of the data variance can be assigned to the first two principal components, allowing for a visually descriptive plot of the principal component scores. In Figure 6, the scores plot from the first two principal components is shown. The scores plot indicates that there are three action potential units, which is in agreement with the interpretation of the cluster analysis dendrogram. In the plot, a swarm of action potentials are located close to the plot origin, corresponding to cluster B in Figure 5. Cluster C is represented by the action potentials that extend “upward” with respect to the second principal component axis. Cluster A is associated with the swarm centered a t
Biotechnol. Prog., 1990, Vol. 6,No. 6
502 120
1 I
\\ -
01 0.00
I
I
I
I
I
I
0.05
0.10
0.15
0.20
0.25
0.30
lOO-1
Ly
Figure 5. Cluster analysis dendrogram from analyzing salinity chemoresponse data. Three units specified as A, B, and C are noted by using a similarity index threshold of approximately 0.6. Table 11. Eigenvalues from Analyzing Salinity-Based Chemoresponse 9% cumulative eigenvector
eigenvalue
variance
variance
9.327 636 8 0.560 456 1 0.085 635 7 0.014 386 5 0.005 699 4
93.27 5.60 0.85 0.14 0.05
93.27 98.88 99.73 99.88 99.93
8-
A
A 6-
k 4-
A A
N U
a
2-
A
-2 -0
-b
-4
-2
0
I
4
I
8
io
12
Y
K l
Figure 6. Principal component scoresplot from analyzing salinity chemoresponse data. The first two principal component scores were used to generate the plot of the background, 0,and chemoresponse, A. the coordinates 12 and -1 of principal components 1 and 2, respectively. A calibration curve was generated by analyzing measurements of a series of dilutions of artificial seawater injected into a flowing artificial seawater carrier. The data analysis results indicate that the action potentials associated with cluster B were from background neuronal activity, since the population of this group remains constant over a range of concentration of salinity. The other action potential units, however, show a concentration dependency. Figure 7 shows the result of plotting the frequency of firing of the concentration-dependent regions of cluster A. A concentration-dependent trend in the amount of chemoresponse neuronal activity is obvious from the plot. Calibration of Responses to Trimethylamine N-Oxide. Trimethylamine N-oxide (TMO) is a major constituent as a component of the cellular sap of marine organisms (Lockwood, 1967). As a result, crabs may have
,Oi
Concentration (ppt NoCI)
Figure 7. Calibration plot of chemoresponseto salinity. A series of diluted solutions of artificial seawater solutions were used to generate the calibration plot. A triplicate injection waa performed to calculate a confidence limit for an intermediate concentration in the calibration curve. The error bar shows the spread based on one standard deviation. the ability to sense this substance as a way of detecting food sources. A data set was collected of chemoresponses for TMO by using the P. sanquinolentus based receptrode. As with the salinity-based chemoresponse, the data involving TMO were analyzed with CA and PCA to generate a calibration response curve. The analysis of this data set was identical with that presented for the salinity responses, starting with an evaluation of the data structure by using CA and then further elucidation with PCA. To summarize the data analysis, the principal component scores plot is shown in Figure 8. In this case, 85.34% of the data variance is described by the first two principal components. This reflects that less data structure was available to facilitate the interpretation of the data. Like the salinity response data, a dense compaction of action potentials is observed near the origin of the scores plot. These do not appear to be associated with chemoresponsive action potentials, as they do not vary with respect t o concentration. However, the presence of a well-separated swarm of analytical action potentials greatly aids the interpretation of the chemoresponse neuronal activity. PCA was performed on a series of dilutions of TMO, and the calibration curve was generated by counting the number of action potentials from the previously indicated analytical region of scores space. The calibration curve, shown in Figure 9, demonstrates that the response is linear over a wide concentration range. The frequency of firing increased linearly as more concentrated solutions of TMO were passed over the antennular tip. I n the studies involving TMO calibration, it was sometimes possible to observe chemoresponse activity at much lower concentrations than are shown in the calibration plot of Figure 9, ranging as low as 1.0 X
Conclusions In this study, a naturally occurring chemical sensor array is explored that uses multiunit measurements of neuronal activity in the antennules of a swimming crab. A calibration method has been outlined t h a t relies on identification of analyte-specific chemoresponse regions in t h e plotted results of t h e unsupervised pattern recognition procedures CA and PCA. This was done by analyzing a series of calibration data sets relative to a
503
Biotechnol. Prog., 1990, Vol. 6 , No. 6
0
8
study has made use of single analyte responses, which were in the presence of nonchemoresponse background. A logical extension of this work is in progress that makes use of several simultaneous chemoresponse signals, thus taking advantage of the multiplexed information from receptrode measurements.
@ '
0% 0
Acknowledgment This work was supported by National Science Foundation Grant CHE-8921156 Much thanks goes to the Hawaiian Institute of Marine Biology for assisting in obtaining materials for this work.
0
A A 0
-6 -8
-2
-6
PCl
Figure 8. Principal component scores from analyzing a TMO chemoresponse. The training set action potentials are shown as follows: A,background; 0,chemoresponse. The training action potentials were referenced to all subsequent calibration data sets. Also plotted is a chemoresponse from a 1.0 X lo-* M TMO solution, 0. 25
p
0
1
m
/i/
154
m
i
Log Concentration
Figure 9. Calibration plot of chemoresponse to TMO. The standard deviation from triplicate analyses was used to calculate the error bar shown.
constant training set. By monitoring a chemoresponsespecific region for changes in the amount of neuronal activity during chemostimulation, a calibration response curve can be constructed. It should be noted that this
Literature Cited Ache, B. W. In The Biology of Crustacea Vol. 3, Neurobiology Structure and Function; Atwood, H. L., Sandeman, D. C., E&.; Academic Press: New York, 1982; p 376. Ache, B. W.; Derby, C. D. Trends Neurosci. 1985, 356-360. Atema, J.; Fay, R. R.; Popper, A. N.; Tavolga, W. N. Sensory Biology of Aquatic Animals; Springer-Verlag: New York, 1988. Bauer, U.; Hatt, H. Neurosci. Lett. 1980, 17, 209-214. Brown, S. D.; Barker, T. Q.; Monfre, S. L.; Larivee, R. L.; Wilk, H. R. Anal. Chem. 1988,60,252-294. Buch, R. M.; Rechnitz, G. A. Anal. Lett. 1989a,22 (13-14), 26852702. Buch, R. M.; Rechnitz, G. A. Biosensors 1989b, 4, 215-230. Cavanaugh, G . In Formulae and Methods V; Marine Biological Laboratory: Woods Hole, MA, 1964. Fuzessery, 2.; Carr, W.; Ache, B. Biol. Bull. 1978,154,224-240. Glaser, E. M.; Ruchkin, D. S. Principles of Neurobiological Signal Analysis; Academic Press: New York, 1976. Jurs, P. C.; Kowalski, B. R.; Isenhour, T. L. Anal. Chem. 1969, 41, 21. Kittler, J., Ed. Pattern Recognition: 4th International Conference; Springer-Verlag: New York, 1988. Lockwood, A. P. M. Aspects of the Physiology of Crustacea; W. H. Freeman: San Francisco, CA, 1967; pp 250-253. Massart, D. L.; Vandeginste, B. G. M.; Deming, S. N.; Mochotte, Y.; Kauffman, L. Chemometrics: A Textbook; Elsevier: Amsterdam, 1987. Mulloney, B.; Selverston, A. J . Comp. Physiol. 1974, 91, 1. Oppenheim, A. V.; Schafer, R. W. Digital Signal Processing; Prentice-Hall: Englewood Cliffs, NJ, 1975. w Sandeman, D. C. In The Biology of Crustacea Vol. 3, Neurobiology Structure and Function; Atwood, H. L., Sandeman, D. C., Eds.; Academic Press: New York, 1982, pp 1-61. Spath, H. Clllster Analysis Algorithms: Ellis-Honvood: West Sussex, U. K., 1980. Accepted September 14, 1990.
Registry No. Trimethylamine N-oxide, 1184-78-7.
