USING ARTIFICIAL NEURAL NETWORKS FOR PLANT TAXONOMIC DETERMINATION BASED ON CHLOROPHYLL FLUORESCENCE INDUCTION CURVES

Transcription

1 USING ARTIFICIAL NEURAL NETWORKS FOR PLANT TAXONOMIC DETERMINATION BASED ON CHLOROPHYLL FLUORESCENCE INDUCTION CURVES Mariya Kirova 1, 3, Georgina Ceppi 2, Petko Chernev 1, Vasilij Goltsev 1*, Reto Strasser 2, 1 Dept Biophysics and Radiobiology, Biological Faculty, Sofia University, Sofia, Bulgaria; 2 Lab. Bioenergetics, Science Faculty, University of Geneva, Geneva, Switzerland. 3 Biomed future Ltd., Sofia, Bulgaria (current address) Correspondence to Vasilij Goltsev 1 goltsev@biofac.uni-sofia.bg ABSTRACT Photosynthetic apparatus is a highly conservative element of the plant cell and its characteristics are comparatively similar no matter where a plant stands in the taxonomic classification. A problem of great interest is whether a highly informative method can be applied for taxonomic classification of plants based on their photosynthetic characteristics. We used as such one method artificial neural networks and as a source of information about the photosynthetic apparatus, chlorophyll fluorescence induction curves. We used feedforward neural network with back-propagation of error signals through the connections of the network. The designed network has two layers, five hidden neurons, and logarithmic sigmoidal transfer function. As input the network receives the numerical equivalent of the induction curves. In our studies we used induction curves from 35 different species which belonged to 19 different families from 17 orders divided into 11 superorders, all of which were from one of the two major groups of flowering plants (Monocotyledones, Dicotyledones). All together we used 2016 curves (61 curves per plant for most of the species). When a neural network is trained with the whole data set, the error is very high (over 50 %). That s why we developed a system of neural networks each of which divides a taxonomic rank to its subgroups. This step-bystep approach starts with analysis of the whole data set and ends with classification of each family into its species used in the studies. This system allows taxonomic classification with high accuracy (the error is under 5 %). Considering the written above, we can conclude that photosynthetic apparatus contains information about the genotype of the plants and can be used for their taxonomic classification. Keywords: Artificial neural networks, JIP-test, Mathematical modeling, Taxonomy, Variable chlorophyll fluorescence Introduction Photosynthetic apparatus is the main structural and functional element of the plant cell. Its reaction centers are highly conservative with low species specific characteristics. On the other hand, the antenna complexes characteristics are highly variable and specific for varies groups of plant species (1). This gives us the possibility to use chlorophyll fluorescence, which depends on the reaction centers and the electron transport chains but which is emitted by the antenna complexes, for taxonomic classification. This emission doesn t contain obvious species-specific characteristics. That s why we are suggesting the application of highly informative method for division of plant objects into systematical groups based on their photosynthetic characteristics manifested trough chlorophyll fluorescence transients (2). These characteristics could be analyzed through artificial neural networks which give us the opportunity to check whether such possibility for plant classification exists or not. Materials and methods In the current work we have used chlorophyll fluorescence induction curves measured with a fluorimeter Handy PEA (Plant Efficiency Analyzer, Hansatech Instruments Ltd.,King s Lynn, Norfolk, UK) (3). All the used plants were in normal physiological state. These were 35 different species which belonged to 19 different families from 17 orders divided into 11 superorders, all of which were from one of 941

2 the two major groups of flowering plants (Monocotyledones, Dicotyledones). All together we used 2016 curves (61 curves per plant for most of the species). The JIP test is a method for in vivo description of the photosynthetic apparatus behaviour (3, 4). This test could be described as a set of mathematical operations that allows the user to transform information derived from the measured OJIP-transients into estimates of energy fluxes (absorption, trapping, electron transport and dissipation) per reaction center or per cross section. The most popular parameter of the JIP-test which is used to characterize the total yield of the system is the so-called performance index, PI. We created, trained and tested the artificial neural networks in a standard software product MATLAB. We created a feedforward neural network with backpropagation. In order to increase the network efficiency the input data underwent preliminary principal component analysis, PCA. The designed network had two layers, five hidden neurons, logarithmic sigmoidal transfer function and 640 epochs at training. As input the network received the numerical equivalent of the induction curves. Three quarters of the input data were used for network training. To avoid possible overfitting we used Bayesian regularization (5, 6). Additionally, for better generalization, the network was trained with early stopping. The last quarter of the input data was used to examine the ability of the trained network to recognize correctly unknown curves. Results and Discussion In order to examine whether there is a connection between the test parameters and the taxonomic identity of a particular species/group, we applied the JIP test to the used set of chlorophyll fluorescence induction curves. The JIP test was made for a chosen set of the input curves and then the derived parameters were analyzed in two different manners (Fig. 1; Fig. 2). Fig 1. Three dimensional parametric visualization of the JIP test results for six plant species from two families. A. Three parametric representation of the JIP test results for the plant species Geranium endressi, Geranium himayense, Pelargonium peltatum, Pelargonium hotatum (Geraniaceae). B. Three parametric representation of the JIP test results for the plant species Ludisia discolor, Oncidium marshallianum (Orchidaceae). C. Three parametric representation of the JIP test results for plant objects from two different families (Geraniaceae, Orchidaceae) 942

3 To check whether they had species species-specific behaviour we chose three photosynthetically important JIP test parameters (PI ABS,, ) and presented them in a three parametric diagram (Fig. 1). This comparative analysis was made for two pairs of plant groups, species from one family (Fig. 1A), and for two taxonomically distant (meaning genetically distant as well) families (Fig 1.B). Fig. 1A shows the existence of areas in the three dimensional space which characterize the values of the three parameters for the four Geraniaceae species used in the studies and their intersection is insufficient. When this approach was applied to the species from the two families simultaneously in the three parameter space there is a sufficient intersection between the values of the parameters (Fig. 1C). Considering these results, we concluded that the three-parameter test could be used to distinguish clearly enough between the objects from one taxonomic group. At the same time, if this test was applied to objects with greater genome difference (e.g. species from two different families), the species specificity disappears. Therefore the threeparametric description of a particular object is not enough for its clear taxonomic classification. A Δ(Ro) B ϕ(po) Δ(Ro) 1.2 ϕ(po) C Δ(Ro) Liliopsida Lilianae Commelinanae D ϕ(po) Δ(Ro) Lilianae Aspargales Liliales ϕ(po) Aspargales Orchidaceae Iridaceae Orchidaceae Ludisia discolor Oncidium marshalianum Fig. 2. Variations in the JIP test parameters for the representatives of the different taxonomic levels. On each of the axes is plotted one of the chosen JIP test parameters. The results are presented as a relative ratio of the average values of the same parameters for the previous taxonomic group (one level higher). A. Spider plot for superorders Lilianae and Commelinanae, the parameters are presented as relative parts from the ones for Liliopsida. B. Spider plot for orders Aspargales and Liliales, the parameters are presented as relative parts from the ones for Liliopsida. C. Spider plot for families Orchidaceae and Iridaceae, the parameters are presented as relative parts from the ones for Aspargales. D. Spider plot for Oncidium discolor and Iris germanica, the parameters are presented as relative parts from the ones for Orchidaceae. Yet, could we use analysis and comparison of greater number of JIP test parameters for taxonomic classification? For presentation of the multiparametric JIP test results we used a spider plot graph (Fig. 2). Eight JIP test parameters were chosen (φ Ро, Δ Ео,, RC/CS o, Δ Ro,,, PI(total)). Each axis of the plot corresponded to one of the eight parameters. The results were represented as a relative part of the average values for the particular taxonomic level. 943

4 The data for the different taxonomic levels were plotted, from superorder to species. This approach of presentation allowed us to follow the dependence of the different JIP test parameters on the taxonomic classification of the object for which the parameters had been evaluated. The average value of each parameter for a particular taxonomic level was used as the comparative value ( 1). Our results showed (Fig. 2 A, B, C, D) that the different JIP test parameters had different dependence on the taxonomic classification of the test object. This created a specific profile of the alteration of these chlorophyll fluorescence characteristics for a particular test object. This profile could be considered species specific and could contribute to the taxonomic classification of a plant. The information from the chlorophyll fluorescence induction curves could be derived by mathematical approaches without preliminary calculation of the JIP test parameters. This could be achieved by artificial neural networks. For the neural network training we used the induction curves described in the Materials and methods section. When the network was trained with the whole data set, the training wasn t efficient enough and the error was quite high (55.4 % with Bayesian regularization; 53.3% with early stopping). To increase the recognition efficiency we developed a system of several neural networks which consequently processed the input data set. Each of these networks divided the data set for a particular taxonomic group to its subgroups used in the experiment. On the next stage each of these subgroups was divided into its sub-subgroups. This step-by-step method started with processing of the whole input data set and ended with classification of each family to its species used in these studies. The training results for the system of neural networks are summarized in Table 1. Table 1 Summary of the results from the neural networks training. Above each line it is noted the error made by the network trained with Bayesian regularization and under each line it is noted the error of the neural network working with application of early stopping. 944

5 Conclusion The described results show that the constellation of the photosynthetic process functional characteristics contains enough information about the taxonomic classification of the studied plants. Application of trained artificial neural networks allows this species-specific information to be derived successfully for an automation screening analysis of plant populations to be achieved. Acknowledgment We thank Bulgarian National Science Fund, Project No DO / for financial support. R.J.S acknowledges support by the Swiss National Science Foundation, Project Nr: REFERENCES 1. Ke B. (2001) Photosynthesis. Photobiochemisry and Photobiophysics. Kluwer Acad. Publ. Dordrecht, p Strasser RJ, Srivastava A, Govindjee (1995) Photochem Photobiol 61: Strasser R., Srivastava A., Tsimilli-Michael M. (2000) Probing Photosynthesis. CRC, p Tsimilli-Michael M., Strasser R. (2008) Mycorrhiza. Springer, Berlin Heidelberg, p McQuarrie A., Tsai C. (1998) Regression and Time Series Model Selection. World Scientific, Singapore. 6. Schwarz G. (1978) Annals of Statistics, 6,