Abstract
The invention relates to a method for screening of at least one chemical substance by treatment of plant material, comprising the following process steps: a) Applying the plant material into a cavity; b) Treatment of the plant material with the chemical substance; c) Creating at least one dataset showing at least one phenotypical characteristic of the plant material after treatment with the chemical substance; and d) Assigning the chemical substance based on the dataset to at least one site of action (SoA) and/or at least one mode of action (MoA) of a multitude of stored SoA and/or MoA by using a SoA- and/or MoA-compendium containing data regarding dependencies between phenotypical characteristics of at least one plant material treated by at least one reference substance of a known SoA and/or MoA.
Claims
1. A method for screening of at least one chemical substance by treatment of plant material, comprising the steps of: a. applying the plant material into a cavity; b. treatment of the plant material with the chemical substance; c. creating at least one dataset showing at least one phenotypical characteristic of the plant material after treatment with the chemical substance; and d. assigning the chemical substance based on the dataset to at least one site of action (SoA) and/or at least one mode of action (MoA) of a multitude of stored SoA and/or MoA by using a SoA- and/or MoA-compendium containing data regarding dependencies between phenotypical characteristics of at least one plant material treated by at least one reference substance of a known SoA and/or MoA.
2. The method according to claim 1, wherein the dataset is obtained by use of a sensor unit.
3. The method according to claim 2, wherein the sensor unit is used to obtain the dataset in a non-destructive and non-intrusive matter.
4. The method according to claim 2, wherein the sensor unit comprises at least one of hyperspectral VIS, hyperspectral NIR, chlorophyll fluorescence or RGB sensors.
5. The method according to claim 4, wherein a light source with a circular arrangement of lamps and reflective surface is used for the hyperspectral VIS and hyperspectral NIR sensors to homogeneously illuminate the plant material.
6. The method according to claim 1, wherein the cavity is a well of a multi-well plate.
7. The method according to claim 6, wherein each well contains one piece of plant material.
8. The method according to claim 1, wherein the chemical substance is applied at different concentrations.
9. The method according to claim 1, wherein one collective dataset is taken for a multitude of plant material pieces and subsequently decomposed into single datasets per single piece of plant material.
10. The method according to claim 4, wherein multiple datasets of the plant material are obtained after treatment at different predefined times or by the use of different sensors.
11. The method according to claim 1, wherein the dataset of the plant material can subsequently be decomposed into single datasets of different parts of the plant material.
12. The method according to claim 1, wherein at least one dataset showing at least one phenotypical characteristic of the plant material before treatment with the chemical substance is obtained.
13. The method according to claim 1, wherein the assigning of the chemical substance to at least one SoA or at least one MoA is carried out by using an adapted program performing a machine learning process.
14. The ethod according to claim 1, wherein the SoA- and/or MoA-compendium is augmented by recording of data of at least one reference substance of a further SoA and/or MoA.
15. The method according to claim 1, wherein an uncharacterized SoA and/or MoA is identified for each chemical substance not assignable to any recorded SoA and/or MoA in the SoA- and/or MoA-compendium.
16. The method according to claim 1, wherein the plant material is in a seed or seedling stage at step a.
17. The method according to claim 1, wherein the plant material belongs to the plant species Arabidopsis thaliana.
18. The method according to claim 1, wherein the chemical substance is a plant growth regulator.
Description
[0055] FIG. 1 shows an overview of different stages of data acquisition in the method according to the invention;
[0056] FIG. 2 shows an overview of the method according to the invention;
[0057] FIG. 3 shows an exemplary cutout of a multi-well plate used for the method according to the invention;
[0058] FIG. 4a shows an exemplary plot displaying raw data used in the method according to the invention;
[0059] FIG. 4b shows an exemplary plot displaying data of FIG. 4a in a scaled and smoothed way;
[0060] FIG. 5a-d show mean spectra of plant material treated with different known chemical substances obtained in the method according to the invention;
[0061] FIG. 6 shows an exemplary code workflow for image processing using hyperspectral imaging for obtaining datasets;
[0062] FIG. 7 shows an exemplary code workflow for image processing using ImagingPAM for obtaining datasets; and
[0063] FIG. 8 shows an exemplary code workflow for SoA/MoA classification.
[0064] FIG. 1 shows an overview of different stages of data acquisition in the method according to the invention. In step 1, plant material is applied into a cavity, for example a well of a multi-well plate. Afterwards, data is obtained in step 2 of the method. In this embodiment, datasets of the plant material are taken via digital optical sensors. Data acquisition may be carried out by one or several different sensors, as for example by a hyperspectral VIS sensor 2.1, a hyperspectral NIR sensor 2.2, a chlorophyll fluorescence sensor 2.3, an RGB sensor 2.4 or a hyperspectral UV sensor 2.5. A combination of different sensors of step 2 is advantageous, for instance data acquisition by use of the hyperspectral VIS sensor 2.1, the hyperspectral NIR sensor 2.2 and the chlorophyll fluorescence sensor 2.3. The use of these sensors is preferred, whereby the RGB sensor 2.4 and the hyperspectral UV sensor 2.5 may also be used optionally. In step 3, the raw data sets taken in step 2 are collected. The image processing pipeline 4 acts as a feature extractor representing all obtained data for plant material (different sensors/spatial and spectral information) in a feature vector. All datasets are merged in step 5. At the end of the method, a universal dataset is gathered (step 6).
[0065] FIG. 2 shows an overview of the method according to the invention. At the beginning, a selection of compounds 7 takes place. In this step, one or more chemical substances to be screened, plant material and data acquisition sensors are selected. The chemical substances can be manually selected covering established herbicidal SoA and/or MoA or uncharacterized ones. In step 8, one cavity per piece of plant material is prepared. Each cavity is equipped with growth medium for the plant material. In this specific example, Arabidopsis thaliana seeds are used as plant material. One seed is put into one cavity. This ensures that phenotypic data and derived parameters of one plant can be extracted by the image analysis pipeline. The cavity in this example is a well of a multi-well plate, as the use of multi-well plates facilitates handling. Before the multi-well plate is put into a growth chamber for plant growth 9, stratification is carried out to break seed dormancy. Standard growth conditions in this embodiment include 5 days of pre-growth on solid growth media with sucrose in a growth chamber with 16 h light per day (120 mol m.sup.2 s.sup.1) and 8 h of darkness. Temperature inside the growth chamber is constantly kept at 22 C. Different growth conditions can be adapted in the screening method to enhance specific SoA and/or MoA, which affect plant materials marginal or not at all under standard growth conditions. Growth conditions can also be adapted for different plant species. After the pre-growth, a first dataset of the plant material is obtained just before treatment 11. In this embodiment, this is carried out via image acquisition (pre-treatment) 10 with different sensors, preferably by use of a combination of the hyperspectral VIS sensor, the hyperspectral NIR sensor and the chlorophyll fluorescence sensor. The use of these sensors enables the recording of digital imaging data with a broad spectrum and high resolution to characterize the phenotype of the plant material. The combination of these sensors with customized acquisition software for full parameter control and automation of image acquisition and processing is preferred. One collective dataset per sensor is taken of the whole multi-well plate. Afterwards, treatment 11 can be carried out by pipetting or foliar application via spraying of an aerosol. Each multi-well plate contains control wells with untreated plant materials. This helps to identify process errors. In this embodiment, image acquisition (after treatment) 12 is carried out 24, 48, 72 and 96 h after treatment 11. In the next step (imaging raw data 3) the dataset taken for the multi-well plate is divided into single datasets for each single well containing one seedling, further analyzed by an image processing pipeline 4 and merged to a processed dataset 13. Data filtering 14 helps to identify ungerminated seeds, contaminated wells or plant material that has not grown optimal before treatment. These datasets are excluded from further analysis in step 15. Data normalization 16 is carried out by a mixed-effects model for phenotypic data of the treated plant material to remove confounding effects. This step is crucial to compare data over time (for weeks, months and years). After data normalization 16, datasets are summarized in step 17 for further processing. The automated feature extraction 18 serves to extract the most relevant phenotypic characteristics. SoA and/or MoA signatures gathered in the previous steps are assigned to the data in step 19. An algorithm is trained subsequently on recognition of SoA and/or MoA. After training it, the model is validated. In case of non-validated data, the algorithm has to be trained further and the data is reviewed again. The use of chemical substances with known SoA and/or MoA is essential for the proper training and validation of the algorithm. When the data is validated, a SoA- and/or MoA-classifier 20 is trained based on this data consisting of SoA and/or MoA of known chemical substances. The data obtained in the training phase is used to build this mathematical classification model trained to predict the correct SoA and/or MoA of the chemical substance by supervised machine learning (including for example random forest or support vector machines). Classical machine as well as deep learning is used on all data points to accurately classify SoA and/or MoA. Data assigned by the SoA- and/or MoA-classifier 20 is then stored in a SoA- and/MoA-compendium 21. The outlier check 23 is used to categorize the SoA and/or MoA: if the data matches data patterns stored in the SoA- and/MoA-compendium 21, classification 25 takes place. In case the data does not match any of the stored data, an uncharacterized (potential novel) SoA and/or MoA is detected (step 24). The detection of an uncharacterized MoA and/or SoA requires further tasks and opens up the possibility of adding further data to the MoA- and/or SoA-compendium.
[0066] FIG. 3 shows a cutout of a multi-well plate 26 comprising 96 cavities 27, which are automatically detected by an image handling process. Each cavity 27 contains one plant material 28. In further analysis the pixels associated with plant material 28 can be separated from a background 29 pictured in a zoom-in of a single cavity 27. This way the background comprising growth media and/or the cavity can be excluded from further data processing.
[0067] In FIGS. 4a and 4b, two different exemplary plots of spectral imaging data are shown: FIG. 4a displays raw data, while FIG. 4b shows scaled and smoothed data. The sensitivity of hyperspectral camera sensors is different depending on the spectral range. To normalize and scale this effect, it is necessary to have white and dark references. In both plots, FIG. 4a and FIG. 4b, the dash line 30 represents a dark reference, whereas the dotted line 31 displays the white reference (imaging of highly reflecting material: Polytetrafluorethylen) and the solid line 32 the plant material. On the outer regions of a here shown VIS spectrum the sensitivity is lower than in the center (illustrated in FIG. 4a) and the plant material data is between the white and dark references. The raw data displayed in FIG. 4a is edited in the image processing pipeline. FIG. 4b shows the scaled and smoothed data. The scaling ensures to consider the dynamic range over the whole spectra whereas the smoothing of the data planish small fluctuations and therefore strengthen relevant changes in the spectra.
[0068] FIG. 5a-d show mean spectra of plant material (Arabidopsis thaliana) treated with different known chemical substances to illustrate the effect of the treatment on the spectra of the plant material. The different spectra each represent the mean value of datasets obtained from a multitude of plant material 96 h after application of the chemical substances. Applied concentrations per chemical substance are mentioned in brackets: FIG. 5a shows the spectra of a control treatment (mock) and three treatments with different SoA/MoA: Norflurazone (5 g/ha)phytoene desaturase (PDS) inhibitor; 2,4-DB (100 g/ha)synthetic Auxins; and Sulcotrione (2 g/ha)4-hydroxyphenyl-pyruvate-dioxygenase (HPPD) inhibitor. These spectra exemplify the effect of treatments with different SoA and/or MoA, which add up to different spectra of the plant material. FIGS. 5b-d show spectra of plant material treated with chemical substances which act by the same SoA and/or MoA. Plant material of these spectra cluster together. The results shown in FIG. 5b are obtained by use of synthetic Auxins: 2,4-DB (100 g/ha); MCBA (10 g/ha); and Dicamba (50 g/ha). The results shown in FIG. 5c are obtained by use of phytoene desaturase (PDS) inhibitors: Beflubutamid (10 g/ha); Norflurazon (5 g/ha); and Picolinafen (1 g/ha). Results shown in FIG. 5d are obtained by use of 4-hydroxyphenyl-pyruvate-dioxygenase (HPPD) inhibitors: Sulcotrione (2 g/ha); Mesotrione (1 g/ha); and Topramezone (100 g/ha). For the results shown in FIGS. 5a-d, Arabidopsis thaliana (accession Col-0) plants are grown as described in EXAMPLE 1. The chemical substances are applied as described in EXAMPLE 2. 96 plants per treatment and concentration are analyzed. Image acquisition is carried out before the treatment as well as 24, 48, 72 and 96 h after the treatment as described in EXAMPLE 3. The data created is processed as described in EXAMPLE 4 and datasets are generated for SoA and/or MoA classification. Assigning of chemical substances with known SoA and or MoA is conducted as described in EXAMPLE 5 with each of the chemical substances listed in table 1 used as test data. On average, the phenotypical characteristics of plant materials treated with five chemical substances describe one SoA and/or MoA. Classification performance is tested by preferably leaving each chemical substance out of the training dataset and testing the assignment of individual plant materials to the correct SoA and/or MoA (table 2a). For SoA and/or MoA with less than two chemical substances available, accuracy is calculated based on classical cross validation (table 2b). Classification performance is evaluated using the prediction accuracy as a statistical measure of how well a classification model correctly identifies the correct MoA and/or SoA. Here, accuracy is the proportion of correct predictions (both true positives and true negatives) among the total number of cases tested.
TABLE-US-00002 TABLE 2a SoA/MoA Accuracy ACCase 0.791176 ALS 0.770563 DXR/DXS 0.454936 FAT 0.458333 HPPD 0.864198 Microtubule assembly 0.701847 PDS 0.682927 PPO 0.705989 PSII 0.932347 Synthetic auxins 0.60452 VLCFA 0.721925
TABLE-US-00003 TABLE 2b SoA/MoA Accuracy ACCase 0.842324 ALS 0.92579 Cellulose biosynthesis 0.735509 DXR/DXS 0.688025 EPSP synthase 0.429589 FAT 0.468932 Glutamine synthethase 0.885473 HPPD 0.92269 Microtubule assembly 0.83791 PDS 0.839713 PPO 0.845077 PSII 0.954997 SPS 0.648352 Synthetic auxins 0.839989 VLCFA 0.796927
[0069] FIG. 6 shows an exemplary code workflow for image processing using hyperspectral imaging to obtain datasets. In step 3, raw data imaging is carried out for one multi-well plate at once. Each dataset obtained may later be divided into single datasets per piece of plant material, but at this point of the workflow, one dataset per multi-well plate is used. Image correction 35 and pixel dropout correction 36 takes place for the dataset as well as for white and dark reference images 33,34 used for normalization. The white and dark reference images 33,34 undergo deadspot correction 37 before building a base of processed references 38 that can later be used for spectral range scaling 39. Spatial destriping 40 may optionally be carried out before spectral smoothing 41, which is leading to step 42, where optionally spectral band removal is conducted to avoid lower accuracy of the later-built classifier. The processed data cube 45 is then used for feature extraction 18. Step 46 describes an image thresholding process wherein together with step 47, the plate template matching, the data points of the dataset are categorized as background or foreground data. The plate template matching 47 is still carried out on multi-well plate template level. Afterwards, the thresholding is repeated in step 48 on plate well level. If the optional step 43, the synthesizing of an RGB image, was previously carried out, the RGB/color data 44 that is obtained on multi-well plate level, can be used at this point as well. To acquire datasets per single piece of plant material, object masking 49 is conducted. In this step, data belonging to each single piece of plant material in the multi-well plate is masked out individually creating multiple single datasets. After spatial summarization 17, a feature vector 50 per single piece of plant material is consecutively built before stopping the workflow shown in FIG. 6.
[0070] FIG. 7 shows an exemplary code workflow for image processing using ImagingPAM for obtaining datasets. In a first step, the imaging of the raw data 3 per multi-well plate and afterwards image thresholding 46 is carried out. The background-corrected dataset is then compared to a multi-well plate template for plate template matching 47 before the thresholding takes places for the plate wells in step 48. As previously shown in FIG. 6, the single pieces of plant material are masked out individually in step 49. In the last step before stopping the workflow shown in FIG. 7, photosynthesis parameters 51 for each dataset of a single piece of plant material are determined based on the data processed.
[0071] FIG. 8 shows a code workflow for SoA and/or MoA classification. After starting the workflow, photosynthesis parameters 51 and the feature vector 50, both per single piece of plant material, are combined. Afterwards, data filtering 14 is conducted. If the data is not sufficient, these data points are discarded in step 17. Data passing the filtering 14 is processed in step 52, by scaling the dataset in step 53 and augmented in step 54. The dataset created this way is then used for the assigning of phenotypical characteristics of the plant material to one or more SoA and/or MoA (step 55). A support vector machine (SVM) SoA- and/or MoA-classifier 20 based on the SoA- and/or MoA-compendium 21 is used for the SoA and/or MoA prediction 56 per single piece of plant material. By majority voting 57, SoA and/or MoA prediction 56 is carried out per chemical substance before stopping the workflow exemplified in FIG. 8.
