- Research article
- Open Access
A high-throughput screening system for barley/powdery mildew interactions based on automated analysis of light micrographs
© Ihlow et al; licensee BioMed Central Ltd. 2008
- Received: 13 August 2007
- Accepted: 23 January 2008
- Published: 23 January 2008
To find candidate genes that potentially influence the susceptibility or resistance of crop plants to powdery mildew fungi, an assay system based on transient-induced gene silencing (TIGS) as well as transient over-expression in single epidermal cells of barley has been developed. However, this system relies on quantitative microscopic analysis of the barley/powdery mildew interaction and will only become a high-throughput tool of phenomics upon automation of the most time-consuming steps.
We have developed a high-throughput screening system based on a motorized microscope which evaluates the specimens fully automatically. A large-scale double-blind verification of the system showed an excellent agreement of manual and automated analysis and proved the system to work dependably. Furthermore, in a series of bombardment experiments an RNAi construct targeting the Mlo gene was included, which is expected to phenocopy resistance mediated by recessive loss-of-function alleles such as mlo5. In most cases, the automated analysis system recorded a shift towards resistance upon RNAi of Mlo, thus providing proof of concept for its usefulness in detecting gene-target effects.
Besides saving labor and enabling a screening of thousands of candidate genes, this system offers continuous operation of expensive laboratory equipment and provides a less subjective analysis as well as a complete and enduring documentation of the experimental raw data in terms of digital images. In general, it proves the concept of enabling available microscope hardware to handle challenging screening tasks fully automatically.
- Active Contour Model
- Moment Invariant
- Fourier Descriptor
- Susceptibility Index
- Powdery Mildew Fungus
Recent molecular methods have paved the way for a number of new experimental approaches in life science which were not available several years ago. As a matter of fact, these new techniques exceed the capacity of well-established manual or scantily automated analysis by far. Automated high-throughput analysis techniques not only solve this problem – they generally introduce a less subjective, more reproducible, and potentially more accurate data processing. However, competing with intuitive and trainable human skills, even though only for a rather specific problem, often turns out to be a difficult task.
This paper introduces a fully automated high-throughput screening system which has been developed for supporting a functional genomics approach in the field of plant-pathogen interactions.
A significant increase or decrease of this index indicates a relation of the test gene to the plant's defense mechanism.
Manual screening has been done for a large number of experiments and this proved to be a tedious and very time-consuming mission. The desired screening of thousands of candidate genes would require many person years without automation. Relegating this task to a fully automated high-throughput screening system offers a number of advantages: Besides saving labor, the subjective component of the human observer is replaced by deterministic image analysis algorithms. Due to the autonomous operation, continuous activity (24/7) becomes possible, leading to a higher utilization degree of expensive laboratory equipment. Last but not least, the intrinsic storage of the experimental raw data as digital images provides a complete and enduring documentation of the experiments for further reference.
Expenditure of time for image acquisition
1st pass: preview scan (5×-objective, 432 × 342 pixels)
2nd pass: detailed scan (10×-objective, 1300 × 1030 pixels)
≈ 0.5 s
≈ 1 s
negligible (once per slide)
≈ 3 s
≈ 0.7 s
≈ 1.5 s
several focal layers
5 layers per position
≈ 10 min per slide (scanning 15 × 38 = 570 positions)
≈ 20 min per slide (scanning ≈ 100 positions*)
System output and intermediate results
Benchmarking the system gives evidence about its reliability. One problem, the rather variable staining intensity of the transformed cells, can be dealt with by sophisticated image processing algorithms to detect these properly. But the generally low contrast of haustoria as well as the occurrence of salient discolorations in the stained cells preclude a naive haustoria detection solely on weak color differences to the staining. For a reliable classification, we applied machine-learning techniques which are trained on a reference data set previously labeled by an expert human observer. Based on this manually labeled data set, we investigated a classification accuracy of 95 ± 1%.
From the biological point of view, only large changes of the susceptibility index are significant and must be reliably detected by the system. Evidently, this goal is reached in practice.
Screening for gene discovery
The described high-throughput screening system enables a large-scale analysis of candidate genes regarding the resistance of crop plants against the powdery mildew fungus by automating very time-consuming screening tasks. Proved to work dependably and at operational stage now, it provides a novel tool of medium- to high-throughput phenomics in the crop plant barley allowing researchers to address gene function in host- or nonhost interactions for resistance. A single experimenter is expected to test up to 100 candidate genes per person month, which is approximately two orders of magnitude higher than whole-plant approaches in barley such as stable transgenic plants or TILLING mutants. Currently, the system is established in a number of projects at our research institute as well as at international cooperating partners.
As a general conclusion, the developed solution can be understood as a proof of concept of how to extend already available microscope hardware to handle challenging screening tasks fully automatically by bringing together research and development both from the fields of biology and engineering. Of course, this concept is neither limited to the described application nor to the currently used microscope hardware. In the future, the system will be adapted to further challenges and we will focus also on screening problems incorporating fluorescence microscopy. This paper should encourage other researchers to tackle analogous screening tasks in a similar way.
The screening was carried out in seven-day-old susceptible barley plants of cv. Golden Promise, as described in . Briefly, leaf segments were bombarded by gold particles that had been coated with a mixture of pUbiGUS (reporter plasmid) and pIPKTA30 Target (RNAi plasmid) using a PDS 1000/He system (Bio-Rad, Munich, Germany). Three days after the bombardment, leaf segments were inoculated with Blumeria graminis f. sp. hordei and incubated for another 40 h, followed by staining of transformed cells for beta-glucuronidase (GUS) activity and microscopy.
Having isolated the stained, genetically transformed cells, potential haustoria must be detected therein. Generally, both haustoria and other salient objects exhibit a slightly more saturated color than the remaining cell area. Exploiting this feature, we perform the following heuristic: First, a contrast enhancement via morphological top-hats  is applied on the color image of the cell cutout. Since morphological top-hats operate contrarily by extracting objects which cannot contain the structuring element, a rectangular or disk-shaped structuring element being somewhat larger than a haustorium is appropriate. (Actually, we use a disk-shaped structured element of 37 pixels diameter.) This leads to a content-adaptive contrast enhancement, increasing the saliency of potential haustoria regions while preserving the saliency of the remaining cell area. Second, the color saturation of the enhanced image cutout is taken as feature image for the segmentation. Salient regions are extracted by a region growing segmentation starting from seeds which exceed a certain high threshold until falling below a second, low threshold. This is efficiently realized by using a binary morphological reconstruction method . We describe this haustoria segmentation process in detail in .
At this stage, where the stained cells have been segmented and potential objects that might be haustoria have been marked, the image processing part is completed. Looking back at Figures 4 and 5 we now have obtained a sketch of the color images, representing the objects of interest. As a last milestone, the identified potential haustoria regions must be classified into true haustoria and false positive objects. In Figure 5 this is illustrated by the color, where haustoria are marked in red.
To transfer the expert knowledge, enabling the human observer to detect haustoria, to the machine, distinctive features of the considered objects are extracted and fed into an appropriate classifier. In general, the composition of an adequate feature vector is the most important step to obtain a good classification performance. Furthermore, the selected features have to be considered in relation to the subsequent classification technique. In  we tested several feature combinations on sophisticated classifiers and drew the conclusion that a classification accuracy of more than 90% is possible.
Due to an improved feature preprocessing we recently simplified the solution by enabling the use of a linear classifier. Confinement to an uncomplex classifier has the advantage of being independent of further specifications such as neural network topologies, training algorithms, learning rates, or other parameters. In addition, there is the least risk of overfitting. To reach this aim, the feature vector is adapted by nonlinear transformations in advance. As an example, consider incorporating the object's perimeter P and area A. For a given object, both features are related nonlinearly by A ~ P2. Hence, it can help a linear classifier to adapt to this feature vector when the object area is incorporated as , as the former nonlinear correlation of both features is thus linearized. Such a nonlinear preprocessing can also be done for more sophisticated features.
In order to train a classifier onto the extracted features, a representative data set is needed which contains samples of virtually all possible cases both of the classes "haustorium" and "other object". Therefore, we manually annotated a large set of digital images containing transformed cells with and without haustoria and stopped when the data set comprised 500 objects for each class.
Shape features for haustoria recognition
In order to complement the color features which have already been exploited during the segmentation, we must now focus on the object's shape. For haustoria recognition, the features must reflect a class of specifically shaped objects, consisting of a body with "fingers". Therefore, beside basic shape descriptors such as the object's area, its contour length, or principal axes, we incorporate two sophisticated approaches: moment invariants [17, 18] and Fourier descriptors .
Moment invariants can be derived both from a gray-level image and a binarized image. We incorporate the invariants of both the binary image and the color saturation image, leading to 22 features.
In the feature vector we use , which leads to 27 features due to omitting and . So far, composed of moment invariants and Fourier descriptors, the feature set is scale invariant, i.e., the information about the object's size is not contained. This changes through the incorporation of the square root of object area , the object perimeter P, and the major and minor axis length a and b, respectively. Additionally, we consider the normalized multiscale bending energy (NMBE) Ψ [20, 21] in terms of and the mean color saturation S as features. Together with the 22 moment invariants and the 27 Fourier descriptors this finally leads to a set of 55 features.
After normalizing the feature vector to zero mean and unit variance by applying the standard score (also called z-score or normal score) transformation , it is ready for the actual classification.
For estimating the expected classification performance on unknown data, the representative data set needs to be partitioned into a training subset and a disjoint test subset . We randomly partitioned the data set and used one half for training the classifier and the other half for testing. To obtain a stable informational value, this partitioning, training, and testing was performed in terms of 500 different realizations. As a result, we obtained a classification accuracy of 95 ± 1%.
At the end of this analysis pipeline, all necessary information is available to distinguish infected cells from uninfected ones. Cells are considered as infected if there is at least one object classified as a haustorium.
We thank Stefanie Lück, Manuela Knauft, Gabi Brantin, and Dimitar Douchkov for their support concerning the preparation of the specimens and the manual screening. Many thanks to Ralf Tautenhahn, Christian Schulze, Tobias Senst, Martin Kalev, and Burkhard Steuernagel for supporting the system development as well as to Cornelia Brüß, Felix Bollenbeck, Tobias Czauderna, Rainer Pielot, and Marc Strickert for fruitful discussions. This work was supported by the German Ministry of Education and Research (BMBF) under grant 0312706A.
- Shen QH, Saijo Y, Mauch S, Biskup C, Bieri S, Keller B, Seki H, Ülker B, Somssich IE, Schulze-Lefert P: Nuclear activity of MLA immune receptors links isolate-specific and basal disease-resistance responses. Science. 2007, 315 (5815): 1098-1103. 10.1126/science.1136372.PubMedView ArticleGoogle Scholar
- Zimmermann G, Bäumlein H, Mock HP, Himmelbach A, Schweizer P: The multigene family encoding germin-like proteins of barley. Regulation and function in basal host resistance. Plant Physiology. 2006, 142: 181-192. 10.1104/pp.106.083824.PubMedPubMed CentralView ArticleGoogle Scholar
- Dong W, Nowara D, Schweizer P: Protein polyubiquitination plays a role in basal host resistance of barley. The Plant Cell. 2006, 18: 3321-3331. 10.1105/tpc.106.046326.PubMedPubMed CentralView ArticleGoogle Scholar
- Panstruga R: A golden shot: How ballistic single cell transformation boosts the molecular analysis of cereal-mildew interactions. Molecular Plant Pathology. 2004, 5 (2): 141-148. 10.1111/j.1364-3703.2004.00208.x.PubMedView ArticleGoogle Scholar
- Douchkov D, Nowara D, Zierold U, Schweizer P: A high-throughput gene-silencing system for the functional assessment of defense-related genes in barley epidermal cells. Molecular Plant-Microbe Interactions (MPMI). 2005, 18 (8): 755-761. 10.1094/MPMI-18-0755.View ArticleGoogle Scholar
- Canny JF: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). 1986, 8 (6): 679-698.View ArticleGoogle Scholar
- Ihlow A, Seiffert U: Microscope color image segmentation for resistance analysis of barley cells against powdery mildew. 9. Workshop "Farbbildverarbeitung", ZBS Zentrum für Bild- und Signalverarbeitung e.V. Ilmenau, Report Nr. 3/2003 Ostfildern-Nellingen, Germany. 2003, 59-66.Google Scholar
- Ihlow A, Seiffert U: Adaptive color spaces based on multivariate Gaussian distributions for color image segmentation. 12. Workshop "Farbbildverarbeitung", ZBS Zentrum für Bild- und Signalverarbeitung e.V. Ilmenau, Ilmenau, Germany. 2006, 86-96.Google Scholar
- Bilmes JA: A gentle tutorial of the EM algorithm and its application to parameter estimation for Gaussian mixture and hidden Markov models. Tech Rep ICSI-TR-97-021. University of Berkeley; 1997.Google Scholar
- Ihlow A, Seiffert U: Automating microscope colour image analysis using the Expectation Maximisation algorithm. Pattern Recognition: 26th DAGM Symposium, Volume LNCS 3175 of Lecture Notes in Computer Science. Edited by: Rasmussen CE, Bülthoff HH, Giese MA, Schökopf B. 2004, Tübingen, Germany: Springer, 536-543. [http://0-www.springerlink.com.brum.beds.ac.uk/content/ufaq4c9gdma7d19q/]View ArticleGoogle Scholar
- Kass M, Witkin A, Terzopoulos D: Snakes: Active contour models. International Journal of Computer Vision. 1988, 1 (4): 321-331. 10.1007/BF00133570.View ArticleGoogle Scholar
- Ihlow A, Seiffert U: Snakes revisited – speeding up active contour models using the Fast Fourier Transform. Proceedings of the Eighth IASTED International Conference on Intelligent Systems and Control (ISC 2005). 2005, Cambridge, MA, USA, 416-420.Google Scholar
- Soille P: Morphological Image Analysis 2nd edition. Berlin: Springer; 2004.View ArticleGoogle Scholar
- Vincent L: Morphological grayscale reconstruction in image analysis: Applications and efficient algorithms. IEEE Transactions on Image Processing. 1993, 2 (2): 176-201. 10.1109/83.217222.PubMedView ArticleGoogle Scholar
- Ihlow A, Seiffert U: Haustoria segmentation in microscope colour images of barley cells. 10. Workshop "Farbbildverarbeitung". Edited by: Droege D, Paulus D. Koblenz: Der andere Verlag; 2004,119-126.Google Scholar
- Tautenhahn R, Ihlow A, Seiffert U: Adaptive feature selection for classification of microscope images. Fuzzy Logic and Applications: 6th International Workshop, WILF 2005, Crema, Italy, September 15–17, 2005, Revised Selected Papers, Volume LNAI 3849 of Lecture Notes in Artificial Intelligence. Edited by: Bloch I, Petrosino A, Tettamanzi AGB. 2006, Springer, 215-222. [http://0-www.springerlink.com.brum.beds.ac.uk/content/m441755p245l1521/]View ArticleGoogle Scholar
- Hu MK: Visual pattern recognition by moment invariants. IRE Transactions on Information Theory. 1962, 8 (2): 179-187. 10.1109/TIT.1962.1057692.View ArticleGoogle Scholar
- Flusser J: On the independence of rotation moment invariants. Pattern Recognition. 2000, 33 (9): 1405-1410. 10.1016/S0031-3203(99)00127-2.View ArticleGoogle Scholar
- van Otterloo PJ: A contour-oriented approach to digital shape analysis. PhD thesis. 1988, TU DelftGoogle Scholar
- Cesar RM, Costa LdF: Application and assessment of multiscale bending energy for morphometric characterization of neural cells. Review of Scientific Instruments. 1997, 68 (5): 2177-2186. 10.1063/1.1148112.View ArticleGoogle Scholar
- Young IT, Walker JE, Bowie JE: An analysis technique for biological shape I. Information and Control. 1974, 25 (4): 357-370. 10.1016/S0019-9958(74)91038-9.View ArticleGoogle Scholar
- Duda RO, Hart PE, Stork DG: Pattern Classification. New York: John Wiley & Sons; 2001.Google Scholar
- Fisher RA: The use of multiple measurements in taxonomic problems. Annals of Eugenics. 1936, 7 (Part II): 179-188.View ArticleGoogle Scholar
- Duchene J, Leclercq S: An optimal transformation for discriminant and principal component analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). 1988, 10 (6): 978-983. 10.1109/34.9121.View ArticleGoogle Scholar
- Bishop CM: Neural Networks for Pattern Recognition. Oxford: Oxford University Press; 1995.Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.