WO2002033400A1 - Method for characterizing, identifying and designating microbial mixtures by comparing ir, ft-ir, raman, ft-raman and gc analysis data with data from a reference database - Google Patents

Abstract

In order to rapidly characterize, identify and designate microbial mixtures, vibrational spectra of samples are measured based on the IR, FT-IR, Raman and FT-Raman spectroscopy or gas chromatograms are recorded. The measurements are conducted in known experimental devices. The characterization and identification of microbial mixtures ensues by comparing the infrared, Raman spectra or the chromatograms of the gas chromatography or gas chromatography/mass spectroscopy of the sample to be examined with a reference database consisting of IR, FT-IR, Raman or FT-Raman spectra or of GC chromatograms/GC mass spectra that are classified into two or more classes according to the composition, to the property or to the state of the measured reference sample. The comparison preferably ensues using methods of pattern recognition such as artificial neural networks, multi-variant statistical models, machine learning, case-based classification and genetic algorithms.

Description

 METHOD FOR CHARACTERIZING, IDENTIFYING AND LABELING MICROBIAL MIXTURES BY COMPARING IR, FT-IR, RAMAN, FT-RAMAN AND GC ANALYSIS DATA WITH DATA FROM A REFERENCE DATABASE

The invention relates to a ner driving for characterizing, identifying and labeling microbial mixtures by means of ER. (Infrared), FT-IR (Fourier transform infrared), Raman, FT Raman (Fourier transform Raman) or GC analysis (gas chromatography).

10

Microbial mixtures are e.g. Bacterial mixtures, mixtures of yeasts or fungi. Such properties can be assigned to such microbial mixtures whose composition is unknown in detail. There are microbial mixtures that contain, for example, a certain natural substance

15 synthesize. A change in the composition of the microbial mixture has a direct influence on the amount of organic matter synthesized. Another application of microbial mixtures is in the area of displacement of undesirable bacterial cultures. For example, a characteristic mixture of bacteria in the intestine of chickens prevents Salmonella from settling in the intestine

20 of the chickens.

So far, there is no reliable analytical method to predict the presence of a particular property in a given microbial mixture, or even to characterize or identify the different composition of microbial mixtures 25. The only option is direct

Property tests. Such tests are lengthy and may require in vivo testing as in the chicken case described above.

Such tests are unsuitable for quality control during production and the associated standardization of the manufacturing process. The first approaches for the application of FT-IR spectroscopy to identify microorganisms for biomedical, biotechnological or diagnostic applications have been published for Actinomycetes (Haag H., Gremlich H. -U., Bergmann R. and Sanglier J. -J. (1996 ) Characterization and identification of actinomycetes by FT-IR spectroscopy. J Microbiol. Meth. 27 (2-3), 157-163), yeast in food (Kümmerle M., Scherer S. and Seiler H. (1998) Rapid and reliable identification of food-borne yeasts by Fourier-transform infrared spectroscopy. App. Env. Microbiol. 64 (6), 2207-2214.), fungal and algae pathogens of animals (Schmalreck AF, Tränkle P., Nanca E. and Blaschke- Hellmessen R. (1998) Differentiation and characterization of Candida albicans, Exophiala dermatitidis and Prototheca spp. By Fourier-transform infrared spectroscopy (FT-IR) in comparison with conventional methods. Mycoses. 41 (1), 71-77.) And some Species of the families Lactobacillus (Curk M. C, Peladan F. and Hubert JC (1994) Fourier-transform infrared (FT-IR) spectroscopy for identifying Lactobacillus species. FEMS Microbiol. Lett. 123, 241-248.), Listeria (Holt C, Hirst D.,

Sutherland A. and MacDonald F. (1995) Discrimination of species in the genus Listeria by Fourier transform infrared spectroscopy and canonical variate analysis. Appl. Env. Microbiol. 61, 377-378.), Streptococcus, Enterococcus (Goodacre R., Timmins E.M., Rooney P.J., Rowland J.J. and Kell D.B. (1996) Rapid identification of Streptococcus and Enterococcus species. FEMS Microbiol. Lett. 140

(2-3), 233-239.), Clostridium (Franz M. (1994) Identification of clostridia by FT-IR spectroscopy. Dtsch. Milchwirtsch. 3, 130-132.) Bacillus, Pseudomonas and Taphylococcus, Candida (Schmitt j ., T. Udeüioven (2000): Use of Artificial neural networks in biomedical applications. In: Gremlich H.-U., B. Yan: Infrared and Raman Spectroscopy of biologic materials. Marcel Dekker, New York, p. 379-420 .). Applications in clinical microbiology and biomedical applications are described in H. Mantsch, M. Jackson (Infrared Spectroscopy: A New Tool in Medicine. Washington DC, 1998) and in A. Mahadevans-Jansen, G. Puppeis (Biomedical Spectroscopy: Vibrational spectroscopy and other novel techniques, Washington DC, Bellingham, SPIE Vol. 3918, 2000). The invention has for its object to develop a ner driving for fast, reliable and economical detection of ner changes in microbial mixtures or the detection of a specific state in microbial mixtures. The ner driving according to the invention should also be able to be used effectively under the conditions of production in routine operation.

According to the invention, the object is achieved in that the following method steps are carried out in succession

a. the spectra / chromatograms of the microbial mixture are recorded, b. a reference database is provided which contains spectra / chromatograms of microbial mixtures classified into two or more predetermined classes, c. After comparison with the spectra / chromatograms of the reference database, the spectra / chromatograms of the microbial mixture are assigned to one of two or more predetermined classes.

The functional groups of all biochemical components of a microbial mixture, such as peptides, proteins, polysaccharides, phospholipids and nucleic acids from intact cells, contribute to a spectrum / chromatogram of this microbial mixture.

The spectra are very complex due to the high number of contributions and include many different vibration modes of the biomolecules of the cell wall

Membrane, the cytoplasm and the extracellular polymeric substances.

Despite their complexity, the spectra / chromatograms are very specific for a composition, property or state of a microbial mixture and can therefore be used to identify, characterize or label the microbial mixture. The method can be used for the classification of IR, FT-IR, Raman and FT-Raman spectra as well as for gas chromatograms and for chromatograms obtained from coupling methods such as GC-MS (gas chromatography and mass spectroscopy).

Any of the spectroscopic measuring arrangements known per se, such as transmission / absorption, attenuated total reflection, direct or diffuse reflection, can be used for IR or Raman spectroscopy. The measurement in transmission / absorption is particularly advantageous.

The IR spectra are typically recorded in the spectral range of the so-called middle infrared between 500 - 4000 cm ^{"1. In} principle, the near infrared of 1000 - 4000 cm ^{" 1} can also be used.

The samples are preferably solid or liquid. When measuring, the use of multi-cuvettes for measuring several samples or the use of microspectrometric techniques is advantageous. This minimizes the sample quantity and the use of automated sample preprocessing and measurement takes place in order to increase the sample throughput. It can too

Sample carriers as used for microtiter plates or flow-through cells are used. It is also possible to use infrared light guides for location-independent measurement automation.

All water-insoluble optical materials commonly used in IR spectroscopy, such as Ge, ZnSe, CaF2, BaF ₂ , can be used as the material for cuvettes and sample carriers of the preparation variants described above, ZnSe having proven successful for a multi-sample element. However, roughened metal plates or micrometallic grids are also suitable as sample holders, especially if they are designed on the scale of microtiter plates for a large number of samples and as disposable materials. The amount of sample for the acquisition of IR spectra can be kept very small and only a few μl (2-5 μl). Depending on the specified conditions with or without beam focusing, amounts of substance from μg to ng can be used. The diameters of the irradiated sample areas vary between 1-6 mm and 10-50 μm when microfused.

A possible classification can be that an investigated microbial mixture is suitable (“pass”) or unsuitable (“fail”) for a specific application or that it corresponds to different states (state A, state B or state C. etc.).

The spectra / chromatograms are classified by comparing the spectra / chromatograms of the sample to be examined with appropriately classified reference spectra / chromatograms of samples, their composition,

Property or state is known, the reference spectra. The reference spectra are in a reference database. This reference database can also have the form of an artificial neural network in which the spectral information is stored in the form of neuronal weights and can be used for the evaluation.

The creation of the reference database is only required once for each characterization, identification or labeling with regard to a composition, property or condition.

For the construction of the reference database, spectra / chromatograms of samples of microbial mixtures are measured and these samples are characterized by further tests with regard to their composition, their property or their state and then for example the classes "passed" and "failed" or one of several classes of states assigned. The sample preparation and the recording of the spectra / chromatograms are carried out in the same way for the reference samples and the unknown samples to be classified. It is crucial that all parameters for the reference sample measurement and the subsequent actual sample measurement are selected the same within narrow limits.

The physical measurement parameters of the measurements of the LR, FT-IR, Raman and FT-Raman spectra, such as the spectral resolution or the number of averaged spectra, etc., can be varied within the ranges customary in vibration spectroscopy without in practice as critical to the success of the

Classification or evaluation to prove. It is only important when specifying the parameters for spectra collection and sample preparation and sample preparation that the same parameters are selected within narrow limits for all measurements of a sample type. A sample type is a mixture that is always used for the same purpose as e.g. a commercial product based on a bacterial mixture.

For the chromatographic methods, the measurement parameters such as sample preparation, extraction, temperature program of the chromatograph, choice of separation columns can also be varied. It is only important when defining the parameters for the chromatogram extraction as well as sample preparation and sample preparation that the same parameters are selected within narrow limits for all measurements for a sample type.

The comparison of spectra / chromatograms of unknown samples with the

The reference database and the assignment to one of two or more predetermined classes is preferably carried out by means of computer-aided pattern recognition, such as, for example, multivariate statistics (Devijer PA, and Kittler J., Pattern Recognition: A statistical approach, Prentice Hall; London 1982) using discriminant analysis, cluster analysis, distance measurements, multiple regression (Otto M. Chemometrie, VCH Publisher, Weinheim, 1997), or through artificial neural networks (Zupan J., and Gasteiger J., Neural Networks for Chemists. VCH-Publisher, 1993), machine learning, case-based classification (Schank RC, Dynamic Memory: A Theory of Reminding and Learning in Computers and People. Cambridge University Press 1982), genetic algorithms (Holland JH, Adaptation in Natural and Artificial Systems 2nd edition MIT Press, Boston, 1992) or evolutionary programming (Fogel LJ, Owens AJ, Walsh MJ: Artificial Intelligence through Simulated Evolution, John Wiley 1966 and Fogel DB, Evolving Artificial Intelligence, Ph.D. thesis, University of California at San Diego, 1992).

The description of a class x is based on a finite set of features x (such as wavelength) and is defined as

An artificial neural network is preferably used for the comparison and the assignment of spectra / chromatograms of unknown samples to one of two or more predetermined classes.

Such an artificial neural network consisting of 3 layers is shown schematically in

Fig. 3 shown. The input layer consists of input neurons that represent selected wavelengths. This information is passed on from the input layer to a hidden layer, where calculations take place in the hidden neurons. The information flow is either increased or decreased by the hidden neurons and the result of the calculation is passed on to the output layer. The output layer corresponds to the specified classes, the number of output Neurons for the number of given classes. The activation of one of the output neurons is the result of the classification procedure.

A neuron is an information processing cell that produces an output when the cumulative effect of all input information reaches a particular one

Threshold exceeded. The flow of information on a neuron is illustrated in FIG.

The basic elements of the artificial neural network are defined as follows:

Input neurons are characterized by weight or strength. An input variable Xj is multiplied by the synaptic weight W _y . The first index indicates the corresponding neuron and the second index the input end of the synapse to which the weight belongs. The weights model the synaptic neural connections in their capacity to either amplify or attenuate an input signal. The weights are real positive or negative numbers.

The propagation function (netj) is used to sum up all input signals on a neuron, weighted by the corresponding connections of the neuron. If necessary, an external bias θj is added, which amplifies or weakens the sum of the input signals, depending on whether θj is positive or negative.

net _j =, w _v + θ _y equation 1

1 = 0

The activation function (f (net _j )) serves to limit the amplitude of the output of a neuron.

a _j = f (net _j ) equation 2 The most frequently used activation function is:

log (*) = 7 - equation 3

A neuron can use the output layer to calculate the assignment to one or more classes (FIG. 4). The components of the input vectors x are the wavelengths for spectra and the peaks for chromatograms. In the case of an artificial neural network with 3 layers, the output vector corresponds to the hidden neurons. In an example with two classes, the output can correspond to the classes "fail" and "pass" of a classification system.

The artificial neural networks can be divided into two types, the multilayer perceptrons (multilayer perceptrons (MLPs)) and the networks with radial basic functions (RBFs). The structures of these two types of neural networks

(Bishop CM. (1995). Neural networks for pattern recognition, Clarendon press, Oxford) are different. While MLPs assign the different classes through the hidden neurons and represent the hyperplanes in the input space, so to speak, the RBFs represent a class assignment using local kernel functions (FIG. 5).

An explicit expression for a fully connected 3-layer MLP is:

M _k (x) = f (ß _k + ∑ w _kj f (∑ w _ß x, + θ _j ) + ∑ w _h x _t ) Equation 4

7 = 1; = 1 (= 1 k = 1, 2 ..., K where:

- = activation of the output neuron k by input vector xf = xι, x ₂ , _k x)

Xn M> _J , = weight running in the hidden layer, from input neuron i to hidden neuron j W _k , = weight in the output layer, running from hidden neuron neuron j to output neuron k

W _h = weight in the output layer, running from input neuron i to

Output neuron k (shortcut connection) θ _{/ cj} = biases of output layer and hidden layer / '= activation functions for output layer and hidden layer n, M, K = number of input, hidden and output neurons

Artificial neural networks with radial basis functions show a more local modeling behavior than the MLPs and thus result in a better detection of outliers. A fully connected RBF network including direct connections (shortcut connections) between input and output neurons is represented by the following approximation:

M _k (x) = f (ß _k + ∑W Afl x -t _j \, θ _j ) + ∑w *, *,) Equation 5

J = l = 1 k = 1, 2 ..., K

With:

I ^x / 1 = Euclidean distance between the input vector and the centers of the RBFs h = basic function

The radial basis functions h are based on the distance (mostly Euclidean) between the input vector and a so-called prototype vector

(Equation 5) applied. The prototype vectors, the so-called centers, represent the different classes in the training set. The most common activation The function in RBF applications is the Gauss function, with _gauss -> 0 if | x | - • oo. The last term in Equation 5 takes into account the direct connections between input and output neurons.

An artificial neural feed-forward network with 3 layers is preferred, which uses a gradient descent method such as resililient propagation, backpropagation, quick-prop and variants of these methods as a learning algorithm.

Pre-processing of the spectra is preferably carried out before the classification.

This preprocessing can consist of a wavelength selection in the case of the vibration spectra or a feature selection in the case of the chromatograms. The number of parameters for the classification model can thus be reduced and only extend to the optimal spectral ranges for the classification.

Redundant information or spectral ranges or chromatogram components without essential information can thereby be eliminated, which improves the quality and stability of the classification model.

The selection of one or more spectral ranges can be done by visual inspection of the spectra or, sensibly, by multivariate methods for the selection of spectral features. The multivariate methods used are, in particular, the co ariancy analysis and the simple univariate analysis of variance (Otto M., chemometry, VCH-Puplisher, Weinheim 1997) or genetic algorithms (Holland J.H., Adaption in Natural and Artificial Systems 2nd edition

MIT Press, Boston, 1992) or discriminant methods (e.g. a distance measure, discriminant analysis).

When selecting wavelengths or features using multivariate methods, three different elements have to be taken into account: the selection criterion, the search strategy and the termination condition. The differences between the classes can serve as a selection criterion. These differences between the classes can be analyzed using statistical methods such as the probability of errors, distance between classes (inter-class distance), probabilistic distance (probabilistic distance), probabilistic dependence and entropy. Univariate and multivariate selection criteria can be derived from this. A univariate selection criterion is the F-value, which describes the ratio of the variance between and within the classes. The search strategy relates to a ranking of the F-values. The determination of a predetermined number of wavelengths with the highest F values in each case or a predetermined threshold value for the F values can serve as a termination condition.

In order to avoid the problem of intercorrelation of univariate methods and to further reduce the number of wavelengths (features) without losing their distinctive character, a multivariate approach such as the multiple can also be chosen

Analysis of covariance.

The principle in multiple covariance analysis is to iteratively link all unselected wavelengths (features) with already selected wavelengths (features) and calculate a partial F-value. The wavelength with the highest

Classification is taken over in the set of the already selected wavelengths (features) (the so-called control variables) and accordingly taken from the set of the unselected wavelengths (features). This procedure is repeated until the desired number of selected wavelengths (features) has been reached or the maximum of the calculated F values falls below a predetermined threshold value.

The preprocessing of the spectra can also consist of data reduction e.g. by methods of multivariate statistics or the transformation of the data either with the help of the wavelet transformation or by main component

Disassembly (e.g. main component analysis, factor analysis). Before the preprocessing, a step of processing the vibration spectra can also be carried out. Possible methods are, for example, calculation of the first or second derivative, spectra deconvolution or other methods for increasing the spectral contrast, which facilitate band detection and permit minimization of any existing baseline problems.

Overall, the method of the invention allows a characterization of _"identity fying or labeling of a microbial mixture after receipt of

Spectrum / chromatogram within fractions of a second after receipt of the sample with conventional personal computers (e.g. Pentium III, 400 Mhz, 64 MB RAM).

The method can be integrated very well into a routine process, since spectra recording and processing or the chromatography and the subsequent classification are completely computer-controlled and can be automated very easily. As a result, there is little need for personnel only for the simple sample preparation.

The method according to the invention can be carried out without the involvement of highly specialized experts, since the IR spectra / GC chromatograms are classified by the computer-aided pattern recognition methods optimized for the purpose. The evaluation based on strictly mathematical criteria simultaneously leads to a high level of certainty in the classification and does not require subjective experience. Figures and examples

Show it

Fig. 1: FT-IR absorption spectrum from Aviguard® Fig. 2: 1. Derivation of the absorption spectrum from Aviguard®

3: Schematic representation of an artificial neural network with 3

Layers Fig. 4: Schematic representation of a neuron, for example from the hidden layer Fig. 5: Classification using RBF networks with local kernel functions (a.) And with MLPs with hyperplanes (b.)

example

FT-IR spectroscopy of the Aviguard® bacterial mixture

The Aviguard® product was characterized using the method according to the invention as an example.

Aviguard® is a product that is produced by industrial fermentation and consists of a wide range of aerobic and anaerobic bacteria. The bacterial mixture in the intestine of healthy adult chickens was simulated.

Based on FT-IR spectra from Aviguard®, an IQassification scheme was created that corresponds to the classification as “pass” or “fail” according to the in-vivo quality test of a product batch from the Aviguard® manufacturing process. This quality test measures the effectiveness of Aviguard® in reducing colonization by Salmonella in young chickens.

The FT-IR spectra of Aviguard® samples from a standardized fermentation process were compared with FT-IR spectra of samples that had not successfully gone through the same fermentation process. The differences in the spectra between these two groups were extracted and used for classification by means of chemometry and pattern recognition.

The FT-IR spectra from Aviguard® were recorded with a Bruker JJFS 28B spectrometer with a ZnSe sample holder with 15 sample positions. The spectra were recorded with a DTGS detector and 64 scans in the wavelength range from 4000 to 500 cm ^"1. The Fourier transformation was carried out with a Blackman-Harris 3-term apodization function (apodization function), a zero-filling factor of 4 to achieve a spectral resolution of 6 cm ^"1 and a data point interval of approximately 1 point per wavenumber. The spectrometer was flooded with 500-1000 1 / h of dry air with the help of an air dryer during the recordings in order to avoid artefacts caused by steam lines minimize. The water vapor content was measured during the recording of the spectra in the range from 1837 to 1847 cm ^"1 and was not more than 0.0003 AU. The signal-to-noise ratio was not more than 0.0003 from peak to peak in the range from 2000 up to 2100 cm ^"1 . A quality control test of the measured FT-IR spectra was applied to the spectra with minimum thresholds

Absorption (0.3245 AU) and maximum absorption (1.24 AU) to determine the linearity range of the detector. A background spectrum was recorded before each measurement on a sample in order to compensate for the background. 7 independent measurements were made for each sample to include measurement-to-measurement variances for each sample.

A spectrum of Aviguard® is shown as an example in Fig. 1.

The biochemical components of Aviguard® can be assigned to 5 broad IR absorption bands: 1 - 3000-2800 cm ^"1 fatty acid range, 2- 1700-1500 cm ^{" 1}

Protein range, 3 - 1450-1200 cm ^"1 mixing range with carboxyl groups of proteins, free amino acids and polysacchiarids (1450-1400 cm ^{" 1} ) and RNA / DNA and phospholopid content (1250-1200 cm ^"1 ), 4 - 1200 -900 cm ^"1 polysaccharides, 5 - 900-700 cm ^{" 1} finger area with significant contribution to the specific distinction.

In a first processing step, the first derivative of the recorded spectra was calculated with a Savitzky-Golay algorithm with 9 points (A. Savitzky, MJ Golay, Anal. Chem. 36: 1627-1638, 1964) and then with a vector transformation normalized the entire recorded spectral range.

FIG. 2 shows the first derivative of the absorption spectrum from FIG. 1 over the entire wavelength range of 4000-500 cm ^"1 .

The spectra of the Aviguard® samples were each assigned to one of two classes, namely the classes "passed" and "failed". Based on this assignment A comparison was made between the spectra in the two classes in order to determine the variances of the spectra within the classes and between the classes (wavelength selection). Procedures were used to minimize the variance of spectra within a class and to maximize it between classes. This procedure was carried out data point by data point over the entire measured spectral range. The resulting spectral features with maximum distinctive character between the spectra of the two classes were then used for the classification.

The method used for this wavelength selection was the covariance analysis.

A 3-layer MLP consisting of an input layer, a hidden layer and an output layer, which were completely connected to each other, including direct connections between the input and output neurons, was used as an artificial neural network for classification , used. Sigmoid transfer functions were used as transfer functions, both in the hidden layer and on the output layer.

As a training method for the artificial neural network, a modification of the gradient descent method (Rumelhart D. and McClelland JL, Parallel Distributed Processing: Exploration in the Microstructure of Cognition Vol. 1,2, Cambridge, MA, MIT Press 1986) was called " Resilient Propagation "(Rprop) (Riedmiller M. and Braun H. RPROP. A almost an robust back-propagation learning strategy. Proceedings of the IEEE International Conference on

Neural Networks (ICNN) San Francisco 1993, p. 586-591) is used. The weights were not adjusted during the training according to the gradient of the error function, but according to the sign of this gradient. The slopes of the error function of the current and the previous point in time were used. The amount of weight change did not depend on the amount of

Slope, but only on its sign. If the signs were the same, so if the amount could be increased, the signs were different, the amount was reduced. The advantage of the process at RPROP is that it is superior to other processes in training time and has good generalization properties.

For the description of the classes, a 1-out-of-n coding was used in the output vector of the artificial neural network, with each class being represented by an output unit. Exactly one position of the output vector was coded with "1" (corresponding to the position of the class to which the input pattern was assigned), and all other positions were coded with "0".

The results of the classification by the trained neural network were then evaluated using an analysis tool using the functions “winner takes all” (WTA) and “402040”. The WTA function accepts an output pattern for classification if one output neuron is activated higher than the others. The

402040 rule requires that with a possible output range from 0 to 1, the activation of one output neuron is greater than 0.6 (the upper 40% of the output range), with the other output neurons less than 0.4 ( the lower 40% of the output range). The output patterns remain unclassified in all other cases. Both methods were used to describe the result, with the 402040 rule being the stricter criterion for evaluating the result.

The data set consisted of 10 samples marked "failed" and 26 Aviguard® samples marked "passed". Each sample consisted of

7 repeat measurements, so that 70 FT-LR spectra "failed" and 182 FT-IR spectra "passed" result. The validation was carried out using a cross-validation based on the leave-one-out method (k = 1) in order to check the quality, stability and robustness of the method according to the invention. The result of the classification with a 3-layer MLP with cross validation (k = l) and 5 input wavelengths after the wavelength selection resulted in 86.82% correctly assigned spectra, no incorrectly assigned spectra and 13.18% unassignable spectra. The error, as the mean square of error (SSE) related to the validation, was 13.77 for the "402040" validation method. The "WTA" validation method resulted in 92.25% correctly assigned spectra, 7.75% incorrectly assigned spectra and no unassignable spectra with a mean square of error (SSE) of 13.77.

Claims

Patentansprtiche 1. A method for characterizing, identifying or labeling microbial mixtures, characterized in that a. Spectra / chromatograms of the microbial mixture are recorded, b. a reference database is provided which contains spectra / chromatograms of microbial mixtures classified into two or more predetermined classes, c. the spectra / chromatograms of the microbial mixture are assigned to one of two or more predetermined classes after comparison with the spectra / chromatograms of the reference database. 2. The method according to claim 1, characterized in that the spectra / chromatograms before step c. in a step b2. be preprocessed. 3. The method according to claim 2, characterized in that the preprocessing of the spectra / chromatograms consists in a data reduction. 4. The method according to claim 2, characterized in that the preprocessing of the spectra / chromatograms consists in a wavelength selection in the case of the spectra or a feature selection in the case of the chromatograms. 5. The method according to claim 2, characterized in that the preprocessing of the spectra / chromatograms in a data reduction and a subsequent wavelength selection in the case of the spectra or Feature selection in the case of chromatograms exists. 6. The method according to any one of claims 1 to 5, characterized in that the spectra / chromatograms before step c. or possibly before step b2. in one step bl. be processed and this processing is in the formation of the first or second derivative, a spectral deconvolution or another method to increase the spectral contrast. 7. The method according to any one of claims 1 to 6, characterized in that the reference database in the form of spectra / chromatograms. or in the form of neural weights of a neural network. 8. The method according to any one of claims 1 to 7, characterized in that the classification by pattern recognition, preferably by means of artificial neural networks, methods of machine learning, algorithms of multivariate statistics, case-based classification, genetic algorithms or evolutionary programming. 9. The method of claim 8, characterized in that the algorithms of the multivariate statistics consist in a discrimination analysis, cluster analysis, distance measurements or multiple regression or combinations thereof. 10. The method according to claim 8, characterized in that a feed-forward network with 3 layers with a gradient descent method as a learning algorithm or an artificial neural network based on radial basic functions is used as the artificial neural network. 11. The method according to claim 10, characterized in that resililient propagation (RProp), backpropagation, quickprop and variants of these methods are used as the learning algorithm. 12. The method according to any one of claims 4 to 11, characterized in that the wavelength selection in the case of spectra or feature selection in the case of chromatograms with genetic algorithms, covariance analysis, simple univariate analysis of variance, a statistical distance measure or discriminant method. 13. The method according to any one of claims 4 to 11, characterized in that the wavelength selection in the case of spectra or feature selection in the case of chromatograms takes place visually. 14. The method according to any one of claims 3, 5 to 13, characterized in that the data reduction by methods of multivariate statistics or the transformation of the data using the wavelet transformation such as the main component decomposition (e.g. main component analysis, factor analysis). 15. The method according to any one of claims 1 to 14, characterized in that the microbial mixtures are bacterial mixtures or mixtures of yeasts or fungi or all three groups. 16. ^"The method according to any one of claims 1 to 15, characterized in that the spectra are IR, FT-IR, Raman or FT-Raman spectra and the chromatograms by gas chromatography (GC), or by coupling GC can be created with mass spectroscopy (GC / MS). 17. The method according to any one of claims 1 to 16, characterized in that the spectrum in one or more regions of the middle infrared range from 500 cm ^"1 to 4000 cm ^{" 1} or the near infrared range from 1000 cm ^"1 to 4000 cm ^{" 1} or measured in both regions.