Periodic Reporting for period 1 - CENTR-ALL (Clarifying role of the CENTROSOME in abnormal mitotic processes featuring the most commonchildhood malignancy: paediatric High Hyperdiploid Acute Lymphoblastic Leukaemia (HHDpALL))
Okres sprawozdawczy: 2019-01-24 do 2021-01-23
Paediatric Acute Lymphoblastic Leukaemia (pALL) still accounts for most cancer-related deaths in children. Hyperdiploidy (HHD) is the most frequent subtype, having an elevated chromosome number compared to other patients. It is still not clear how this disease develops, but recently it was shown that chromosomal instability (CIN) plays a key role. The most common cause of CIN is centrosome amplification (CA). Centrosomes coordinate cell division and having access number of them (CA), can result in an unequal distribution of the chromosomes to the daughter cells and eventually cancer. If a cell has too many centrosomes, the cell division can get out of hand, which can result in cell death (Fig. 1). Preventing centrosome clustering (by which tumor cells keep CA under control) seems a promising therapeutic option for patients with CA. HHD-pALL is also known to often have mutations in RAS signalling genes, which, in turn, might effect centrosome development; the exact relation, however, is unclear. To reveal if CA plays a role in the development of HHD-pALL, we analyzed patient samples and investigated, in an experimental setting, the effect that the most frequently mutated RAS gene (KRAS) has on the development of these organelles.
Centrosomes are mostly visualized by immunofluorescence (IF) microscopy, but precise evaluation is error-prone and not always straightforward. A diagnostic and high-throughput screening method would thus be of great help. To develop a fully automated fluorescence light microscopy application, determining the amount of cells with CA, we combined a deep neural network (DNN) classifier with an automated slide scanning system.
Research was carried out primarily by Dr Gabor Pajor in a field-leading group (headed by Prof. Alwin Krämer) at the German Cancer Research Center (DKFZ, Germany) and in a cross-sector collaboration with a non-academic world leader in automated image analysis (MetaSystems, Germany).
Centrosomes are mostly visualized by immunofluorescence (IF) microscopy, but precise evaluation is error-prone and not always straightforward. A diagnostic and high-throughput screening method would thus be of great help. To develop a fully automated fluorescence light microscopy application, determining the amount of cells with CA, we combined a deep neural network (DNN) classifier with an automated slide scanning system.
Research was carried out primarily by Dr Gabor Pajor in a field-leading group (headed by Prof. Alwin Krämer) at the German Cancer Research Center (DKFZ, Germany) and in a cross-sector collaboration with a non-academic world leader in automated image analysis (MetaSystems, Germany).
For clarifying KRAS’s role in CA, mutant versions of this gene were introduced into the mammary cell line MCF10A with the help of viruses that were manipulated so that upon infecting the cell, they delivered the mutated gene into the nucleus, integrating it into the DNA (lentiviral transduction). Centrosomal status was analyzed at various time points after transduction by IF microscopy. During IF, proteins of the centrioles and also its surrounding pericentriolar matrix were stained using fluorescent antibodies (CP110, centrin, pericentrin and γ-tubulin) that attach to them and can be visualized under the microscope.
The IF experiments showed that the cells that successfully carried (i.e. expressed) a KRAS mutation, were featured with a modest but statistically significant level of CA (Fig. 1). It was also observed that the rate of CA, in general, increased over the course of two weeks. Our results suggest that point mutations of KRAS are able to induce moderate levels of centrosome amplification, and – in the light of previous findings - might therefore be responsible for the elevated chromosome number seen in HHDpALL. The pilot investigation using patient samples showed signs of CA in about ¼ of the cells analyzed. Further experiments are needed to confirm our findings, investigating more patient-samples and experimenting with further genes and cells more similar to the leukemic cells of HHDpALL, to gain more disease-specific conclusions. (Results had been submitted for conference presentation.)
For creating an automated centrosome detection workflow, cells with different levels of CA were used in order to train the DNN. Centrosomes were stained using IF, whereas the nucleus was counterstained with a DNA stain. After scanning slides from 5 different cell-lines, cells were identified based on counterstain-related appearance. In the following, more than 35 000 cells were categorized (labelled) into seven classes manually, for training the DNN. The 7 classes were (1) interphase cell with CA; (2) CA with (pseudo)bipolar mitosis; (3) CA with multipolar mitosis; (4) non-amplified centrosomes in G1-phase of the cell cycle; (5) non-amplified centrosomes in S/G2-phases; (6) non-amplified centrosomes in mitosis; (7) other (Fig. 2). The above also means that classes 1-3 belong to the bigger group of ‘amplified centrosomes’ whereas classes 4-6 represent cells with ‘non-amplified centrosomes’. Slides were scanned using Metafer Slide-Scanning System.
The data set generated upon manual labelling was divided into two subsets (training set and validation set) to train a custom deep neural net (DNN). To increase robustness and avoid pure memorization of images by the DNN, the training images were continuously changed using various techniques such as image flipping, rotation, zooming and the addition of random noise.
Currently, the DNN’s overall accuracy in predicting all of the classes correctly is 80.5%, whereas its ability to distinguish ‘amplified’ from ‘non-amplified’, is close to 90% (88.3%). Further experiments are needed to fine-tune performance, before integrating the DNN into the automated slide scanning system. This will result in a workflow which fully automatically analyses a sample, predicting the normal to aberrant ratio of centrosomes with the only mandatory user interactions being the initial loading of the slide feeder and the final review of results. (Results had been submitted for conference presentation; journal submission is expected within 3-6 months.)
The IF experiments showed that the cells that successfully carried (i.e. expressed) a KRAS mutation, were featured with a modest but statistically significant level of CA (Fig. 1). It was also observed that the rate of CA, in general, increased over the course of two weeks. Our results suggest that point mutations of KRAS are able to induce moderate levels of centrosome amplification, and – in the light of previous findings - might therefore be responsible for the elevated chromosome number seen in HHDpALL. The pilot investigation using patient samples showed signs of CA in about ¼ of the cells analyzed. Further experiments are needed to confirm our findings, investigating more patient-samples and experimenting with further genes and cells more similar to the leukemic cells of HHDpALL, to gain more disease-specific conclusions. (Results had been submitted for conference presentation.)
For creating an automated centrosome detection workflow, cells with different levels of CA were used in order to train the DNN. Centrosomes were stained using IF, whereas the nucleus was counterstained with a DNA stain. After scanning slides from 5 different cell-lines, cells were identified based on counterstain-related appearance. In the following, more than 35 000 cells were categorized (labelled) into seven classes manually, for training the DNN. The 7 classes were (1) interphase cell with CA; (2) CA with (pseudo)bipolar mitosis; (3) CA with multipolar mitosis; (4) non-amplified centrosomes in G1-phase of the cell cycle; (5) non-amplified centrosomes in S/G2-phases; (6) non-amplified centrosomes in mitosis; (7) other (Fig. 2). The above also means that classes 1-3 belong to the bigger group of ‘amplified centrosomes’ whereas classes 4-6 represent cells with ‘non-amplified centrosomes’. Slides were scanned using Metafer Slide-Scanning System.
The data set generated upon manual labelling was divided into two subsets (training set and validation set) to train a custom deep neural net (DNN). To increase robustness and avoid pure memorization of images by the DNN, the training images were continuously changed using various techniques such as image flipping, rotation, zooming and the addition of random noise.
Currently, the DNN’s overall accuracy in predicting all of the classes correctly is 80.5%, whereas its ability to distinguish ‘amplified’ from ‘non-amplified’, is close to 90% (88.3%). Further experiments are needed to fine-tune performance, before integrating the DNN into the automated slide scanning system. This will result in a workflow which fully automatically analyses a sample, predicting the normal to aberrant ratio of centrosomes with the only mandatory user interactions being the initial loading of the slide feeder and the final review of results. (Results had been submitted for conference presentation; journal submission is expected within 3-6 months.)
Our results suggest that KRAS might indeed play a role in generating centrosome amplification (CA) and also showed some indication that CA is a characteristic of HHDpALL. Further experiments are needed to confirm these findings. However, should the causal relationship between 'KRAS' - 'CA' - 'hyperdiploidy in pALL' gain further confirmation, that could help to understand the true pathomechanism of this childhood malignancy, which, in turn, could result in novel therapeutic solutions.
Currently, results from the project aiming to automatize evaluation of centrosome-related immunofluorescence (IF) show positive signs of feasibility. Upon completion, we will have a truly automated pipe-line in our hands, that requires minimal human interaction to analyze centrosomes by IF, providing a great opportunity for developing screening modules for testing the effect of different genetic modifications on centrosome biology. The notion that this classifier has been trained to differentiate between different cell-cycle phases further raises its research potential. In addition, current work will not only provide an effective fully automated analytical pipeline for one particular centrosome-related IF analysis, but stand as the ground for countless similar applications, using different cell lines and/or different antibodies. It should even be possible to re-train the same DNN for the analysis of human patient samples. Centrosome related immunofluorescence has never truly reached a state where it could be integrated into routine diagnostics. The main reasons for this are the difficulty of optimizing IF on the varying quality of patient samples, but also the laborious nature of evaluation. Maybe an artificial intelligence based automated screening tool is all that is needed to elevate IF from being “only” a research tool to becoming a routine application for detecting numerical centrosome abnormalities.
Currently, results from the project aiming to automatize evaluation of centrosome-related immunofluorescence (IF) show positive signs of feasibility. Upon completion, we will have a truly automated pipe-line in our hands, that requires minimal human interaction to analyze centrosomes by IF, providing a great opportunity for developing screening modules for testing the effect of different genetic modifications on centrosome biology. The notion that this classifier has been trained to differentiate between different cell-cycle phases further raises its research potential. In addition, current work will not only provide an effective fully automated analytical pipeline for one particular centrosome-related IF analysis, but stand as the ground for countless similar applications, using different cell lines and/or different antibodies. It should even be possible to re-train the same DNN for the analysis of human patient samples. Centrosome related immunofluorescence has never truly reached a state where it could be integrated into routine diagnostics. The main reasons for this are the difficulty of optimizing IF on the varying quality of patient samples, but also the laborious nature of evaluation. Maybe an artificial intelligence based automated screening tool is all that is needed to elevate IF from being “only” a research tool to becoming a routine application for detecting numerical centrosome abnormalities.