ftirforum (12745 Byte)
ftirdiag (22089 Byte)

Informationsbulletin
Ausgabe (Februar 1998)

1. Vorwort
2. Einige aktuelle Projekte

Identifizierung von Bazillen mit FTIR-Spektroskopie
Identifizierung von Hefen mit FTIR-Spektroskopie
EU-funded Biomed-II project "Rapid identification and antibiotic/antifungal-agent susceptibility testing by Fourier-Transform Infrared (FTIR) and Raman spectroscopy (MIDAS)"
Identification and characterization of micro-organisms by Raman spectroscopy
Stacked Spectral Data Processing and Artificial Neural Networks Applied to FT-IR and FT-Raman Spectra in Biomedical Applications
3. Literatur

Vorwort

    Nachdem sich vor mehr als einem Jahr im Oktober 1996 in Berlin der Arbeitskreis FTIR-Diagnostik konstituiert hat, ist der Kreis der Interessenten an den Anwendungen der FTIR-Diagnostik in der mikrobiologischen und medizinischen Diagnostik ständig angewachsen.

    Die vorliegende zweite Ausgabe des "Informationsbulletins FTIR-Diagnostik" möchte weiter über Neues auf diesem Gebiet berichten. So haben dankenswerterweise verschiedene Arbeitsgruppe Material zur Verfügung gestellt, das in diesem Rahmen einer breiteren Öffentlichkeit vorgestellt werden kann. Auch in Zukunft möchten wir solche Beiträge im Rahmen dieses Informationsbulletins veröffentlichen. Diesbezügliches Material kann (möglichst in digitaler Form) an die Herausgeber geschickt werden.

    Das Herausgeberkollegium
    Dieter Naumann:  Robert-Koch-Institut, FB 311, Nordufer 20, D-13353 Berlin
    Tel. 030-187542259; Fax: 030-187542606; naumannd@rki.de
    Frank Mertens: Bruker Analytik GmbH, Wikingerstr. 13, D-76189 Karlsruhe
    Tel: 0721-9528743, Fax: 0721-9528712; Frank.Mertens@bruker.de
    Jürgen Schmitt: Rheinisch-Westfälisches Institut für Wasserchemie und Wassertechnologie (IWW), Aquatische Mikrobiologie, Moritzstr. 28, D-187546 Mülheim/Ruhr
    Tel. 0208-40303435, Fax: 0208-4030384; 100740.1762@compuserve.com


Einige aktuelle Projekte
Identifizierung von Bazillen mit FTIR-Spektroskopie
H. Seiler
Institut für Mikrobiologie, FML Weihenstephan, TU München, Vöttingerstraße 45, 85350 Freising;
T: 08161/713519; FAX: 08161/714492; E-mail: seiler@lrz.tu-muenchen.de
Die Methode der FTIR-Spektroskopie wurde auf die Gruppe der Bazillen aus Lebensmitteln angewandt. Die Spektren dieser Organismen sind im Vergleich zu anderen Gruppen relativ schlecht reproduzierbar, da sie stark vom Grad der Sporulation und somit von der jeweiligen Vitalität eines Stamms abhängig sind. So zeigten z.B. B. cereus und B. megaterium im jeweils unversporten bzw. versporten Zustand starke bzw. fast keine Differenzen. Auch das Ausblenden von sporentypischen Spektralbereichen ergab keine besseren Ergebnisse.

Es wurde nach einer Anzuchtmethode gesucht, bei der die Bazillen einerseits ausreichend wachsen konnten, andererseits aber noch keine Sporen produzierten. Wir prüften folgende Modifikationen der Anzuchtbedingungen: Reduzierung von Bebrütungstemperatur und -zeit, Erhöhung der Glucosekonzentration und Zusatz von L-Alanin sowie Kombinationen dieser Parameter. Auf den Zusatz anderer Chemikalien oder Antibiotika wurde verzichtet.

Auf Plate-Count-Agar bei 30 °C waren die Stämme nach 24 h häufig schon stark versport; der Zusatz von bis zu 2% Glucose verzögerte die Sporulation erheblich; andererseits wurden aber teilweise große Mengen Schleim produziert, was zu Präparationsproblemen führte. L-Alanin, das die Auskeimung von Sporen verbessern soll, hatte keinen negativen Einfluß auf die Sporulation. Die Verkürzung der Bebrütungszeit ergab gute Resultate, aber diese Modifikation hätte den 24-stündigen Arbeitsrhythmus gestört.

Als beste Alternative erwies sich die Reduzierung der Bebrütungstemperatur auf 25 °C bei Beibehaltung von CASO-Agar. Nach 24 h Inkubation waren die meisten Bazillen erst am Ende der logarithmischen Wachstumsphase und zeigten somit noch keine Sporulation. Zwar waren manche Species (z.B. B. cereus und B. polymyxa) zu 1-3 % versport. Dieser geringe Sporulationsgrad beeinträchtigte zwar die Stammunterscheidung innerhalb der jeweiligen Art, nicht jedoch die Speciesunterscheidung.

Manche Stämme der Arten B. globisporus, B. haloalkalophilus, B. marinus und B. psychrophilus wuchsen unter diesen Bedingungen zu langsam für eine Analyse, zählen aber nicht zu den lebensmitteltypischen Bazillenarten. Für die extremophilen Bazillen werden eigenständige Bibliotheken mit jeweils angepaßten Kultivierungsbedingungen erarbeitet werden müssen.

Das Dendrogramm in Abb. 1 repräsentiert die Spektren von 56 definierten Stämmen aus der Institutssammlung (Weihenstephan Bacterial Collection = WS), die zu den Species alvei, badius, brevis, cereus, circulans, firmus, fusiformis, insolitus, laterosporus, lentus, licheniformis, macerans, megaterium, mycetoides, panthotenticus, pasteurii, polymyxa, psychrophilus, pumilus, sphaericus, subtilis, thuringiensis, t. ssp. berlin, t. ssp. galleriae, t. ssp. thuringiensis und weihenstephanensis gehören. Hier sind einige interessante Details zu erkennen.

Des weiteren sollten 3 Stämme aus Milchreis (Auftragsuntersuchungen) identifiziert werden. 70813 und 71101 wurden als B. subtilis bzw. B. amyloliquefaciens bestimmt. Der Stamm 71112-18 dagegen konnte nicht identifiziert werden, da er sehr spärlich wuchs (vermutlich war er durch Hitzeschädigung verändert).
 
 
seiler.jpg (42601 Byte)
       
      Abb. 1: Dendrogramm der Ähnlichkeitsbeziehungen von FTIR-Spektren mit 58 Bacillus-Stämmen aus 29 Species und Subspecies (55 Referenzstämme + 3 unbekannte Stämme). Average linkage; normalized to reprolevel (Division = 30); spektrale Fenster 3032 bis 2819 (Gewichtung = 1) + 1352 bis 1200 (1) + 902 bis  699 (1) Wellenzahlen; Abkürzungen und Einzelheiten siehe Text 
top of the page

 
 
 
 
 
 
 


Identifizierung von Hefen mit FTIR-Spektroskopie
H. Seiler, M. Kümmerle, S. Scherer
Institut für Mikrobiologie, FML Weihenstephan, TU München, Vöttingerstraße 45, 85350 Freising
T: 08161/713519; FAX: 08161/714492; E-mail: seiler@lrz.tu-muenchen.de
Die Spektrenbibliothek für fermentative Hefen aus Lebensmitteln wurde vervollständigt und validiert. Jede Art ist durch mehrere Stämme repräsentiert. In einer zweiten Bibliothek wurden Referenzspektren für die häufigsten oxidativen Hefen aus Milch und Milchprodukten sowie dem Produktionsumfeld bei der Lebensmittelherstellung und -bearbeitung gesammelt. Die beiden Dateien umfassen 74 Hefenarten von 18 Gattungen mit insgesamt 332 Stämmen. Die Stämme stammen teils (174) aus internationalen Sammlungen, teils sind es Isolate aus unseren Schadensfall- oder Umfeldanalysen. Die Identität aller Stämme wurden mit klassischen Methoden (ca. 100 Merkmale) gegengeprüft.

Die abgelegten Spektren sind Mittelwerte von Einzelspektren von Kulturen aus mehreren unabhängigen Präparationen auf Yeastextract-Glucose-Chloramphenicol-Agar (YGCA, Merck). YGCA wird in den IDF-Normen und im §35 LMBG als Nähragar für den Hefennachweis vorgeschrieben. Bei Verwendung dieses Agars für die Bibliothekserstellung kann man später in der Praxis u.U. direkt von den Gesamtkeimzahlbestimmungsansätzen oder nach höchstens einer weiteren Passage die Keimsuspensionen erstellen.

Die FTIR-Identifizierung wurde mit einer universellen Kombinatorik spektraler Fenster durchgeführt. Bei einzelnen Gattungen, beispielsweise Zygosaccharomyces, führte dies nicht zu einer ganz klaren Auftrennung der Arten. Hier gibt es die Möglichkeit, in einer zweiten Stufe die Identifizierung mit einer gattungsspezifischen Fensterkombination so zu erweitern, daß die Dendrogramme der spektralen Ähnlichkeiten jenen von phylogenetischen Analysen nahekommen.

Die Referenzdatei wurde mit ca. 450 bibliotheksunabhängigen Isolaten aus unserer Routinekontrolle geprüft, indem man sowohl mit der FTIR-spektroskopischen Methode als auch mit der klassischen Differenzierung die Identifizierung durchführte. Für die Güte der Identifizierung mit dem physiochemischen Verfahren wurden 6 Güteklassen bestimmt: sehr gut (100% zuverlässig), gut (³ 90%), befriedigend (³ 75%), ausreichend (³ 60%), schlecht (falsches Ergebnis) und fraglich (weder mit FTIR-Spektroskopie noch mit klassischen Methoden eine Entscheidung möglich). Von den 450 Teststämmen konnten 96% den ersten 4 Güteklassen und nur 1% (5 Stämme) der fünften Klasse zugeordnet werden; 15 Isolate waren grundsätzlich nicht eindeutig identifizierbar.

Dieses Ergebnis bestätigte endgültig, daß die FTIR-Spektroskopie eine konkurrenzlos einfache, schnelle und zuverlässige Identifzierungsmethode für Hefen ist, so daß wir heute ausschließlich nur noch dieses Verfahren anwenden. Lediglich in den seltenen fraglichen Fällen wird auf die klassische Methode zurückgegriffen. Hierbei zeigte es sich, daß diese Kulturen zumeist nicht rein waren oder bibliotheksfremde Arten repräsentierten. In Einzelfällen handelte es sich auch um wachstumsinhibierte Defektmutanten, um nicht registrierte Species oder um biotopspezifische Varianten. Die Bibliothek wird fortlaufend mit den ausgefallenen Varietäten ergänzt.

Diese Untersuchungen sollten vordringlich auf alle lebensmittelrelevanten Bakterien ausgedehnt werden. Wir erwarten, daß dieser Innovationsschub im Bereich der Mikroorganismenidentifizierung bei Lebensmittelbetrieben dann erhebliche Verbesserungen beim präventiven Qualitätsmangement, beim Monitoring von Rekontaminationskeimen, bei der hygienischen Qualität, bei der Kontrolle von Säureweckern und beim Studium von Populationsdynamiken bei Reifungsvorgängen bringen wird.

   top of the page

EU-funded Biomed-II project
"Rapid identification and antibiotic/antifungal-agent susceptibility testing by Fourier-Transform Infrared (FTIR) and Raman spectroscopy (MIDAS)"

This project comprises 4 partners from scientific institutes in Rotterdam/The Netherlands, Berlin/Germany, Reims/France, Milano/Italy, and from three partners in the industry (France, Germany, and United Kingdom). The project has started in September 1997 and will end in August 2000.

Objectives of the EC-funded project

The central objective of the MIDAS-project is to develop Raman and FT-IR spectroscopic methods for rapid identification and antibiotic/antifungal-agent susceptibility testing of clinically relevant micro-organisms. The methods should ultimately enable a microbiological analysis to be completed on the day that patient material is collected. Patient material needed should not exceed the amount of 103 cells per isolate.

Starting point:

Microbiological analysis of patient material, i.e. identification of micro-organisms and antibiotic/antifungal-agent susceptibility testing, usually takes 2 or more days depending on growth rate of micro-organisms. A minimum of 106-108 cells of pure cultures are needed. Most susceptibility tests are based on the determination of cell mass or cell numbers as a function of growth time, and therefore inherently slow.

The research and development tasks in the project are the following:

Industrial context:

This project enables the development of dedicated Raman and FT-IR instruments, software and methodology for a clinical application. It may thus serve to open the health care market to the industrial partners in this project. Pharmaceutical industry will obtain rapid techniques for fundamental studies on drug-cell interactions, crucial to the design and optimisation of new antibiotics and antifungal agents.
 

top of the page

Identification and characterization of micro-organisms by Raman spectroscopy
K. Maquelin, L.-P. Choo-Smith, H.A. Bruining, G.J. Puppels
Laboratory for Intensive Care Research and Optical Spectroscopy, Erasmus University Rotterdam & Dept. of General Surgery 10M, University Hospital Rotterdam ("Dijkzigt"), Dr. Molewaterplein 40, 3015 GD Rotterdam, The Netherlands, Tel. ++ 31 10/ 4635890 or 4633795, e-mail: Puppels@HLKD.AZR.NL
In clinical practice, micro-organisms are increasingly complicating the recovery of patients over the past few decades. Intensive surgery, AIDS and chemotherapy are examples of several factors compromising the immune status of patients, thereby facilitating all kinds of infectious diseases. At present, routine microbiological analysis of patient material takes up to 2 to 3 days before a result can be given to the clinician. In life threatening situations such as meningitis and sepsis, the treatment cannot be delayed and therapy is started with a broad spectrum antibiotic, where resistance to the drug and effects on the body’s normal flora can be complicating factors to the therapy.

Molecular biological techniques are now being developed to obtain a faster identification of micro-organisms and detection of genes coding for antibiotic resistance (1-3). These techniques often still include a culturing step and require highly trained staff. Vibrational spectroscopic techniques (Raman and FT-IR spectroscopies) have proved useful in medical research (4,5) including microbiology (6). These techniques have the advantage that they are fast, relatively easy to automate and hence, can be performed with minimal training. FT-IR spectroscopy has already been successfully used for the identification of micro-organisms (7,8).

In collaboration with several European laboratories, Raman and FT-IR spectroscopies are being optimised for the rapid identification and characterisation of clinically relevant micro-organisms. Confocal Raman spectroscopy is performed with near-IR monochromatic laser light of ~850 nm (9), at this wavelength sample fluorescence that can be present when exciting with visible light is avoided.

Raman spectra were obtained from a set of vancomycin sensitive and resistant Enterococcus faecium strains. The strains were cultured on Mueller-Hinton medium (MH) for 16 hours at 37°C.The medium containing a subMIC level of vancomycin for the sensitive strains was used to induce the translation of the resistance genes. After culturing bacteria were transferred onto a CaF2 window and samples were measured in a confocal setup with a total signal integration time of 5 minutes. The spectra displayed a high signal to noise ratio allowing the assignment of spectral features typically arising from biological molecules (Fig. 1).
 
 
Fig. 1: Average Raman spectrum (measured on CaF2) of E. faecium sensitive (S) and resistant (R) strains 
 and the corresponding difference spectrum (R-S).

The difference spectrum of resistant minus sensitive strains showed significant differences in carbohydrate, nucleic acid and protein content of the cells (Fig. 1), clearly indicating the underlying biochemical differences between the two groups. Upon induction the differences became more pronounced (data not shown for brevity). In addition Enterococcus faecium strains could be grouped into two clusters composed of vancomycin sensitive and vancomycin resistant strains, using principal component analysis (PCA) and hierarchical clustering as non-subjective grouping methods. Of the uninduced samples, 2 strains were misclassified, while no misclassification was observed for the induced strains (Fig. 2). These results demonstrate that Raman spectroscopy has much potential as a new rapid tool for the microbiologist.
 
Fig. 2: Dendrograms arising from hierarchical cluster analysis (using principal component analysis scores as input variables, squared Euclidean distance measure and Ward's clustering algorithm) of uninduced and vancomycin induced E. faecium strains. 
S=sensitive strains, R=resistant strains, *=misclassification.
References

1. Tenover FC, Huang MB, Rasheed JK, Persing DH. (1994) Development of PCR assays to detect ampicillin resistance genes in cerebrospinal fluid samples containing Haemophilus influenzae. J. Clin. Microbiol. 32(11):2729-37.

2. Relman DA. (1997) Emerging infections and newly-recognised pathogens. Netherlands J. Med. 50(5):216-20.

3. Schmitz FJ, Mackenzie CR, Hofmann B, et al. (1997) Specific information concerning taxonomy, pathogenicity and methicillin resistance of staphylococci obtained by a multiplex PCR. J. Med. Microbiol. 46(9):773-8.

4. Puppels GJ, de Mul FFM, Otto C, et al. (1990) Studying single living cells and chromosomes by confocal Raman microscopy. Nature 347:301-303.

5. Jackson M, Choo LP, Watson PH, Halliday WC, Mantsch HH. (1995) Beware of connective tissue proteins: assignment and implications of collagen absorptions in infrared spectra of human tissues. Biochim. Biophys. Acta 1270(1):1-6.

6. Helm D, Labischinski H, Naumann D. (1991) Elaboration of a procedure for identification of bacteria using Fourier-Transform IR spectral libraries: a stepwise correlation approach. J. Microbiol. Meth. 14:127-42.

7. Naumann D, Helm D, Labischinski H. (1991) Microbiological characterizations by FT-IR spectroscopy. Nature 351:81-82.

8. Helm D, Labischinski H, Schallehn G, Naumann D. (1991) Classification and identification of bacteria by Fourier-transform infrared spectroscopy. J. Gen. Microbiol. 137(Pt 1):69-79.

9. Puppels, G.J. (in preparation).
 

top of the page

Stacked Spectral Data Processing and Artificial Neural Networks Applied to FT-IR and FT-Raman Spectra in Biomedical Applications
J. Schmitt*, T. Udelhoven**
* Dept. Aquatic Microbiology, Spectroscopy Group, University of Duisburg, Moritzstr.26, 187546 Mülheim, Germany
**University of Trier FB VI, 54286 Trier, Germany

Biomedical applications of vibrational spectroscopy developed for routine analysis require reliable methods for data evalution. Artificial neural networks open a new perspective for the spectra differentiation and identification of biological samples with their small spectra variance. A stacked spectral data processing and the following use of neural networks for spectral identification was investigated. 6 different neural network architectures were tested in their capability to built spectral libraries for different bacterial genera and for yeasts, using FT-IR and FT-Raman spectra. After developing these libraries, they were connected to a large library, what we called "multilayered neural networks". This combines the advantages that the wavelength can be choosen more selective for a given differentiation problem and the network architecture and training function can be more adapted to a special task. The spectral identification procedure is performed within one library without the need for manual preselection and integrated in one software.

Introduction

Biological and biomedical applications of vibrational spectroscopy have witnessed an enormous progress in the past. The common spectral feature of all complex biological materials is their small spectral variance between the different objects. Biological samples therefore require more sophisticated methods, for differentiation and identification. Especially because the transfer of elaborated spectroscopic techniques into routine analysis in medicine, biodiagnostics and other fields of application depends on its simplicity and reliability, new methods for data evaluation are required. Self-learning systems like artificial neural networks with appropriate techniques of spectral feature selection and data pretreatment are new and promising techniques.

In the present study the applicability of neural networks in spectra identification of different bacteria genera, yeasts for practical purposes was tested.

The establishment of neural libraries can be divided into several steps, and the stacked data processing and pretreatment together with the appropriate network architecture, training-function and learning parameters decides about success and failure (Fig. 1). Our aim was to built several neural librairies, which can easily be connected by the software-user to form one large library. This has the advantage, that the parameters of each library can be individually optimized and adapted to the given problem. This can mean, that the wavelength for differentiation is more selective or the network architecture and training function is more appropriate for a special task. This scheme we call "multilayered neural networks". The user, working with such types of libraries can take advantage, that a given identification problem can be solved within one run of the unknown spectra through all layers. The information flux in such a combined network is controlled by a heading network, whose single task is to seperate between different e.g genera. In the following, the further networks are optimized for the species and subspecies specific separation .
 
 
Fig 1: Schematic diagramm of a "multilayered neural network". The spectral library was divided in subnets. Each net consists of a neural network, where the wavelength selection, the network architecture, 
 training function and training parameters were optimized.

The final library consisted of 6 FT-IR subnets to identifiy 6 bacterial genera with 182 classes (Pseudomonas, Bacillus, Staphylococcus, Streptococcus, Aeromonas, Mycobacterium) and 2 subnets for yeast identification (Candida, 37 classes) and testing their antibiotic resistence against flucanazol (Candida albicans, 2 classes). The yeast identification was based on a combination of FT-IR and FT-Raman spectra.

Wavelength selection

In order to provide the most significant wavelengths for spectral differentiation within a dataset, different techniques were tested. They can be based on the discriminant analysis (linear and quadratic). Another approach is using genetic algortihms in combination with the discriminant analysis, neural networks or euklidian distance methods. The latter methods are used for the optimisation criterion. The advantage of genetic algorithms is their ability to solve search and optimisation problems in large datasets efficiently. The disadvantage is still their relative complexity for routine analysis.

We decided to implement the variance analysis in our program, because of its ease, the ability to deal with many classes (> 10) and its computational speed.

Network architectures, training and testing.

At the beginning, training of the individual subnets was accomplished with different learning algorithms. The one with the best result was selected in each case. Besides standard backpropagation, backpropagation with flat spot elimination and weight decay, the Quickprop algorithm, R-Prop and Cascade correlation were used.

Results

In order to establish an neural database for the identificaion of bacterial genera on the species/strain level, the following species were to be classified: Pseudomonas (46 classes), Bacillus (34), Aeromonas (25), Staphylococcus (47), Streptococcus (10), Mycobacterium (20), Yeasts / Candida (37).

The selected wavelengths and the significane gained by the variance analysis have been utilised to decide, how many datapoints where selected as input neurons and at which significance level. If there are too many input neurons, redundant information is unneccessarily used, resulting in more weights and hence more training time and complexity of the network.We tested the amount of input neurons and assessed the results by the overall error (SSE) of the validation and the misclassifications. (R-Prop network).

The results are summarized in table 1.
# of datapoints
SSE
misclassifications [%]
10
17,4
27,27
20
11,03
21,2
50
11,32
24,24
100
8,79
12
200
4,05
3
300
0,0
0
400
0,0
0
In all examples, we found the best results using between 250 and 400 input neurons. The automatic selection by the variance analysis exhibited better results than manual selection and the discriminant analysis.

Tab.2: Topology and learning function
Bacteria (genus)
Hidden Units
Learning Function
# of Training Cycles (Validation Phase)
Correct Identification [%]
Pseudosomonas
40
R-Prop
111
100
Bacillus
100
Cascade-C.
1000
98
Aeromonas
40
R-Prop
83
100
Staphylococcus
40
R-Prop
150
100
Streptococcus
120
Cascade-C.
600
94
Mycobacterium
50
Quickprop
400
98
 As it can be seen from table 2 considerable numbers of hidden units were required to separate between the individual strains within one species. In general, one hidden layer was used in the subnets . The numbers of learning cycles differ severely between the methods. R-Prop solved the problem always very fast, e.g. with 111 training cycle and 46 classes. Quickprop was second in speed, followed by Cascade correlation. The Backpropagation algorithm showed generally bad results in comparision to the other methods. R-Prop was in most cases superior in its properties. In two cases, learning with Quickprop or Cascade-Correlation was significant faster and the generalisation properties of the networks better. It can also be seen from table 2 that most of the spectra in the validation pattern set could be classified correctly.

Combination of FT-Raman and FT-IR spectra

In order to get additional information about the cells, Raman spectra can be integrated into the analysis. This is demonstrated in table 3, where 37 Candida strains were seperated with neural networks.

Tab.3: Topology and training of 37 Candida strains using FT-IR and FT-Raman
 
Hidden Units
Learning Function
# of Training Cycles (Validation Phase)
Correct Identification [%]
FT-IR
40
R-Prop
60
100
FT-Raman
65
Quickprop
1500
82
Fuconazol resistance* 
30
R-Prop
45
97,5

*of Candida albicans by FT-IR It can be seen that an identification of the yeasts is possible on the basis of the IR spectra as well as with the FT-Raman data. Despite of a faster training and a better classification accuracy with FT-IR spectra, the Raman spectra can give valuable further and complementary information with more distinct and well resolved bands.

A medical application after the identification of the candida species would imply a further differentiation to answer the question for a efficient therapy . The resistance to the antibiotic fluconazol can also be tested by using FT-IR spectra. A neural network with R-Prop as learning algorihm and 30 hidden units was able to distinguish after 45 training cycles between the resistance and non-resistance of Candidaalbicans against fluconazol. This net was linked after the network for species identification.

Conclusion

All tackled spectral identfication problems could be solved with a high percentage of correct assignments (100% -94%) . This success depended on the data pretreatment as described, the wavelength selection and the appropriate network architecture. Whereas the data-pretreatment can be implemented into the software as a standard operation, the choice of the wavelenghts and the configuration of the training and validation datasets can strongly determine the succes of the network after training and its use in routine analysis. Our decision to use the variance analysis for automatic wavelength selection gained good results in combination with neural networks. Our results encouraged us, to use the variance analysis as a pragmatic wavelenght selection tool. In the case where no learning and hence no differentiation occurs and the test of different network architectures does not improve training, the dataset should be split into subgroups to create subnets. The decision, which classes should be combined in the subnets, can be determined by e.g. cluster analysis. This approach is more accurate and better results are obtained, than training large nets. Especially for practical applications in the routine analysis, where reliability has a high status. The subnets can linked later on to one big library by the software.

One of the positive features of neural networks libraries is their somewhat different information content. Whereas classical libraries often use average spectra or a small set to minimize the library size, the information of all spectra is stored in the neural networks. Hence, they have an advantage in the coverage of the variables space and the variance within. Especially biological samples always exhibit a certain variance, e.g. due to growth conditions. Despite this variance, the correct assignment can explicitly trained by the selected dataset, covering the variance range. But care should be taken during trainning. Overtraining is the most prominent fault why neural networks can fail their task in some instances. To test the properties of a trained network after the first validation phase during training, it is recommended to test it in a second validation phase, using a series of new and independant data e.g. of an other instrument or other laboratory.

In our examples, the training showed for some network architectures, that only 100 -500 training cycles are necessary, rareley more. This strongly reduces computation time, and enables to perform the training within a few minutes on personal computers of the new generation. A few years ago workstations dominated in this field. This circumstance can offer new perspectives for new applications. IR-spectroscopy in combination with neural networks offer an alternative approach to traditional methods with large potentials for a rapid and reliable identifiation of bacteria and in biodiagnostics.

These combined techniques of data treatment, wavelength selection and generation of libraries seem to open the perspective of a methodology that can be used as an effective tool in diagnosis and therapy .
 

to of the page
Literatur
Im folgenden sind einige neuere Literaturstellen aufgeführt, in denen Anwendungen der FT-IR- und Ramanspektroskopie in Mikrobiologie und Medizin beschrieben werden. Die Liste stellt lediglich eine Auswahl dar, die keinerlei Anspruch auf Vollständigkeit erhebt. Die Herausgeber wären für Hinweise auf weitere Literaturstellen zum Thema dankbar.

Benedetti et al. (1997) Determination of the relative amount of nucleic acids and proteins in leukemic and normal lymphocytes by means of Fourier transform infrared microspectroscopy. Applied Spectroscopy 51:792ff.

Bouhedja et al. (1997) ATR-FTIR spectroscopic investigation of E. coli transconjugants beta-lactamase resistance phenotype. FEBS Letters 412:39ff.

Brunner et al. (1996) FT-NIR spectroscopy and wood identification. Holzforschung 50:130ff.

Budínová et al. (1997) Application of molecular spectroscopy in the mid-infrared region to the determination of glucose and cholesterol in whole blood and in blood serum. Applied Spectroscopy 51:631ff.

Calonje et al. (1996) New contributions to the wall polysaccharide structure of vegetative mycelium and fruit body cell walls of Agaricusbisporus. Microbiologia 12:599ff.

Chiovitti et al. (1997) Cell-wall polysaccharides from australian red algae of the family Solieraceae (Gigartinales, Rhodophyta) - Novel, highly pyruvated carageenans from the genus Callophycus. Carbohydrate Research 299:229ff.

Cote (1997) Noninvasive optical glucose sensing - An overview. Journal of Clinical Engineering 22:253ff.

Fayolle et al. (1997) Monitoring of fermentation processes producing lactic acid bacteria by mid-infrared spectroscopy. Vibrational Spectroscopy 14:247ff.

Fournet et al. (1997) In situ measurements of cell wall components in the Red alga Solieria chordalis (Solieriaceae, Rhodophyta) by FTIR microspectrometry. Botanica Marina 40:45ff.

Gniadecka et al. (1997) Distinctive molecular abnormalities in benign and malignant skin lesions - Studies by Raman spectroscopy. Photochemistry & Photobiology 66:418ff.

Haaland et al. (1997) Multivariate classification of the infrared spectra of cell and tissue samples. Applied Spectroscopy 51:340ff.

Harthun et al. (1997) Determination of recombinant protein in animal cell culture supernatant by near-infrared spectroscopy. Analytical Biochemistry 251:73ff.

Heise (1996) Non-invasive monitoring of metabolites using near infrared spectroscopy: State of the art. Hormone & Metabolic Research 28:527ff.

Heise (1997) Medical applications of infrared spectroscopy. Mikrochim. Acta [Suppl.] 14:67ff.

Hellwig et al. (1996) Carboxyl group protonation upon reduction of the Paracoccus denitrificans cytochrome c oxidase: direct evidence by FTIR spectroscopy. FEBS Letters 385:53ff.

Imazato et al. (1997) Incorporation of antibacterial monomer MDPB into dentin primer. Journal of Dental Research 76:768ff.

Jung (1997) Analysis of forage fiber and cell walls in ruminant nutrition. Journal of Nutrition 127(Suppl):810Sff.

Lewis et al. (1997) Applications of Fourier transform infrared imaging microscopy in neurotoxicity. Annals of the New York Academy of Sciences 820:234ff.

Malins et al. (1996) Tumor progression to the metastatic state involves structural modifications in DNA markedly different from those associated with primary tumor formation. Proceedings of the National Academy of Sciences of the United States of America 93:14047ff.

Matsuhiro (1996)Vibrational spectroscopy of seaweed galactans. Hydrobiologia 327:481ff.

McCann et al. (1997) Infrared microspectroscopy - Sampling heterogeneity in plant cell wall composition and architecture. Physiologia Plantarum 100:729ff.

Miller et al. (1997) Identification of conformational substates involved in nitric oxide binding to ferric and ferrous myoglobin through difference Fourier transform infrared spectroscopy. Biophysical Journal 72:A425.

Moore et al. (1996) Fluctuations in IR spectral parameters detected in mixed acyl chain membranes of Acholeplasma laidlawii B. Biochimica et Biophysica Acta 1279:49ff.

Muguruma et al. (1997) A new diagnostic method using an infrared endoscope for digestive tract micro-lesions. Gastroenterology 112(Suppl.):A620.

Paschalis et al. (1997) Fourier transform infrared microspectroscopy reveals changes in newly deposited bone in fluoride-treated osteoporotic patients. Bone 20(Suppl.):8S.

Ridgway & Chambers (1996) Detection of external and internal insect infestation in wheat by near-infrared reflectance spectroscopy. Journal of the Science of Food & Agriculture 71:251ff.

Riley et al. (1997) Simultaneous measurement of glucose and glutamine in insect cell culture media by near infrared spectroscopy. Biotechnology & Bioengineering 55:11ff.

Sato et al. (1996) Lipid-protein interactions in human and bovine lens membranes by Fourier transform Raman and infrared spectroscopies. Experimental Eye Research 62:47ff.

Sockalingum et al. (1997) ATR-FTIR spectroscopic investigation of imipenem-susceptible and resistant Pseudomonas aeruginosa isogenic strains. Biochemical & Biophysical Research Communications 232:240ff.

Stewart (1996) Fourier Transform infrared microspectroscopy of plant tissues. Applied Spectroscopy 50:357ff.

Vrany et al. (1997) Comparison of recalcitrance to ciproflaxin and levofloxacin exhibited by Pseudomonas aeruginosa biofilms displaying rapid transport characteristics. Antimicrobial Agents & Chemotherapy 41:1352ff.

Wang et al. (1997) Microscopic FTIR studies of lung cancer cells in pleural fluid. Science of the Total Environment 204:283ff.

Watson et al. (1996) Visual and infrared assessment of root colonization of apple trees by Phymatotrichum omnivorum. Phytopathology 86(Suppl.):S22.

Yamamoto et al. (1997) Usefulness of an infrared spectrometer (IRS) for monitoring of Helicobacter pylori (HP) infection with 13C-urea breath test. Gastroenterology 112(Suppl.):A335.

Yano et al. (1996) Evaluation of glycogen level in human lung carcinoma tissues by an infrared spectroscopic method. Cancer Letters 110:29ff.
 
 
 
 


FT-IR Workshop Series, Berlin, Robert Koch-Institute Home

Any opinions, findings and conclusions or recommendations expressed in this publication are those of the workshop organizers and do not necessarily reflect the views of the Robert Koch-Institute. © 2017 Peter Lasch