Methods and systems for imaging cells in culture

Abstract

The invention relates to a method for characterizing images of cell cultures in order to quantify cell parameters. The steps of the method comprise detecting at least one cell parameter or cell marker parameter in an image of cells in culture; constructing a mosaic of the image (squares or rectangles, also referred to as tiles); defining an image with smaller dimensions (also referred to as a density map) by integrating values of the cell parameters or cell marker parameters in each fraction of the mosaic; applying deep learning algorithms, which are implemented for example by a convolutional neural network, to the images generated in step c in order to determine a proximity score for the images of treated cells with respect to images of healthy cells or untreated diseased cells.

Description

[0001] METHODS AND SYSTEMS FOR IMAGING CULTURED CELLS

[0002] TECHNICAL FIELD OF THE INVENTION

[0003] The invention relates to the field of cell imaging and, in particular, a method for quantifying cellular parameters, notably using deep learning algorithms.

[0004] STATE OF THE ART

[0005] Cell culture is a laboratory technique that allows the in vitro growth of cells of various types and origins, the modification of their properties, or their increase in number, in order to obtain them in large quantities. Cell culture is an indispensable tool for biological research. Indeed, it is the very first stage of experimentation on living organisms and has become crucial for evaluating treatments.

[0006] Cell imaging involves taking images of cells or their components and analyzing them to extract useful information for researchers. The analysis of these images consists of measuring certain cellular parameters such as their shape, size, texture, etc.

[0007] Very often, cells or their various cellular components are labeled with dyes (called "cell markers"), usually fluorescent, to improve the analysis of cellular parameters. The labeled cells or their various components are then observed under a microscope, and certain parameters of these cell markers (such as their number, their intensity of staining, particularly fluorescence) are analyzed to extract information useful to researchers.

[0008] However, the analysis phase of cellular parameters, or parameters of cellular markers, is severely limited by its length and tedious nature, especially since it is now necessary to evaluate several parameters per image. This notably delays the development of new treatments. The need to automate and accelerate these analyses is therefore essential today.

[0009] Recent advances in artificial intelligence have improved the automation of image quantification in general, and cellular image quantification in particular, enabling the recognition of, for example, the precise outlines of features of interest within cells (Stephan Steigele, SLAS Discovery, 2020). However, deep learning on images is a lengthy and computationally intensive process. Some high-definition, high-resolution cell culture images are mosaics of images reconstructed using a confocal microscope. Consequently, they are very large and voluminous (e.g., 15,000 x 15,000) and can contain hundreds of millions of pixels.The usual methods using deep learning applied to images involve dividing large images into a succession of smaller image tiles, each of which can be processed by convolutional neural networks (deep learning). These image tiles rarely exceed one million pixels. These approaches require very long computation times, proportional to the number of image tiles, which can quickly become prohibitive if the number of images to be processed is large. A possible alternative is to perform random sampling on the slice to be processed. Unfortunately, this method is biased and, more importantly, fails to account for the spatial distribution within a cell culture.

[0010] US patent document US20240871 12 describes a system and method for the automated analysis of biological images, particularly those from cell cultures, using artificial intelligence techniques to detect, segment, and quantify structures of interest. Preferred cell cultures are of immune or tumor origin.

[0011] The state of the art therefore presents several limitations which the present invention makes it possible to solve.

[0012] DESCRIPTION OF THE INVENTION

[0013] The solution of the invention to the problem posed relates to a method for characterizing images of cell cultures comprising the steps of: a. detecting at least one cellular parameter and / or one parameter of cellular markers to be analyzed on an image of cell cultures; b. constructing a mosaic of said image; c. generating an image of reduced dimensions compared to the image of step a, by assigning to each square of the mosaic values ​​corresponding to the measured parameters; d. applying deep learning algorithms on the images generated in step c in order to determine a proximity score of the images of treated cells compared to images of healthy cells or untreated diseased cells.

[0014] The invention also relates to a computer program product, said computer program comprising code instructions enabling the steps of the process according to the invention to be carried out when said program is executed on a computer.

[0015] The invention consists of reducing the size of the images processed by a neural network, while preserving the measured parametric information and its location within the cell culture. In particular, in muscle cell cultures, cells differentiate and change shape. The spatial distribution of the measured cellular parameters can then also reflect the stage of differentiation.

[0016] One of the advantages of the invention is to take into account the spatial distribution of each parameter studied within the cell culture, allowing the different images to be classified and sorted more efficiently.

[0017] The method of the invention allows, in the case of high-throughput data acquisition, for a very rapid scoring, taking into account several or all of the analyzed and relevant parameters. Furthermore, the method of the invention can be used for spatial transcriptomic analyses.

[0018] The figures and examples in the detailed description illustrate the manipulation of proximity scores, different types of markers, and the use of various fluorescent dyes. System and software aspects are also described.

[0019] The density map creation step addresses the problem of excessively large images. It reduces the image size while retaining relevant information, namely the values ​​of the cellular parameters or cellular markers studied in each tile.

[0020] The density map creation step also has the advantage of not involving image compression upstream of the application of predictive artificial intelligence and, on the contrary, of selecting the relevant information that will be analyzed by the predictive artificial intelligence.

[0021] Another advantage of the method according to the invention lies in the fact that several cellular parameters or cellular marker parameters can be quantified at the same time, on the same cell culture image.

[0022] The method according to the invention preferably allows for the evaluation of treatments in research, particularly Gene Therapy (GT) treatments. The method according to the invention can be applied to all types of cultured cells. Muscle cells are, however, preferred.

[0023] PRESENTATION OF THE DRAWINGS

[0024] Other features and advantages of the invention will become apparent from the following description and the figures in the accompanying drawings, in which:

[0025] Figure 1 illustrates an example of an embodiment of a process according to the invention.

[0026] Figure 2 illustrates the steps of creating a mosaic of the cell culture image, producing the density map and obtaining the proximity score.

[0027] Figure 3 illustrates the learning process. A neural network was trained with the CTRL and U18 conditions, associating the CTRL condition with a score of 0 and the U18 condition with a score of 1. The neural network then evaluated each condition according to different criteria and assigned them a score between 0 and 1. Figure 4 describes the neural network used in the example of the invention.

[0028] Figure 5 shows the quantification of green fluorescence (corresponding to cholesterol loading) for the results of different doses of the treatment tested in the exemplary invention: a) Immortalized human myoblasts are differentiated into myotubes for 6 days in bright-field microscopy plates. The cells are then treated for 24 hours with U 18666A, followed by a possible drug to be tested (T1-T5), loaded with NBD-cholesterol and LysoTracker for 4 hours, fixed, and imaged using a confocal microscope. b) Confocal images showing the NBD-cholesterol signal in the myotubes. c) Quantification of corrected total green fluorescence in at least 180 myotubes per condition.

[0029] Figure 6 shows the results of assigning scores to the different treatments (T1-T5) of the example of the invention.

[0030] DETAILED DESCRIPTION

[0031] Definitions

[0032] Cell imaging, or imaging of cultured cells, refers to all the techniques used to visualize and capture images of cells. Analyzing these images allows researchers to study certain aspects of their function.

[0033] Cell culture is a set of biological techniques used to grow cells in vitro for scientific experimentation. The cells cultured can be free-living microorganisms (bacteria or yeast). They can also be healthy cells freshly taken from an organism (biopsy, etc.), in which case it is called a "primary culture." These cells cannot be maintained in culture indefinitely, mainly due to their limited number of divisions. They can also be cells with an unlimited capacity for division (these are called "cell lines"). Cell lines are either cancer cells, cells undergoing cancerization, healthy cells made artificially "immortal," or stem cells.

[0034] Cell labeling involves labeling different cell components. In a cell labeling assay, several dyes (usually fluorescent) are used to label various cell components, such as the nucleus, endoplasmic reticulum, lysosomes, mitochondria, cytoskeleton, Golgi apparatus, membranes, or specific nucleic acids, carbohydrates, lipids, or proteins. Many cellular parameters can also be measured based solely on cell morphology, thus eliminating the need for prior labeling. Examples include cell surface area, perimeter, diameter, circularity, and even optical thickness when using holographic microscopy technologies.

[0035] Confocal microscopy is a type of optical microscopy that produces images with a very shallow depth of field (approximately 400 nm) called "optical sections." By positioning the focal plane of the objective lens at different depths within the sample, it is possible to acquire a series of images from which a three-dimensional representation of the object can be obtained. The object is therefore not directly observed by the user; instead, the user sees a computer-generated image. Confocal microscopy operates using reflected light or fluorescence. Most often, a laser is used as the light source. This is then referred to as confocal laser scanning microscopy (CLSM).

[0036] Phase-contrast microscopy (or phase-contrast microscopy) is an optical microscopy technique that transforms the differences in refractive indices between two structures into contrast levels. These differences translate into phase differences for light waves passing through them. This allows the visualization of transparent structures (such as cells) when their refractive index differs from that of their surroundings.

[0037] Holographic microscopy is a recognized optical method for recording the "relief" of an object on a flat photosensitive plate and reconstructing that object in 3D. It uses the coherence of light to record the phase of the light wave scattered or transmitted by the object.

[0038] Fluorescence microscopy allows the visualization of naturally fluorescent objects (chlorophyll, etc.) or molecules made fluorescent for better observation (GFP-coupled proteins, DAPI for DNA, fluorochromes, etc.). Several techniques using this microscopy have been developed, such as immunofluorescence (labeling with an antibody coupled to a fluorochrome); FISH (fluorescence in situ hybridization) to label nucleotide sequences using oligonucleotides coupled to fluorochromes; FRET (fluorescent energy transfer) to visualize whether two molecules interact; BIFC (bimolecular fluorescence complementation) to visualize an interaction by reconstituting the integrity of a fluorochrome using two molecules; and so on.

[0039] In biology, the nucleus is a cellular structure present in almost all animal cells. It contains the cell's genetic material. In mature muscle cells (or muscle fibers), the nucleus is normally located at the periphery.

[0040] Centronucleation refers to the fact that the nucleus is located in the center of the cell. In a muscle fiber, centronucleation is an indicator of a regeneration phase. The end of the regeneration phase is characterized by the return of the nucleus to its peripheral position. Myopathies, in general, are associated with a greater number of centronucleated muscle fibers.

[0041] Muscle fiber size (or diameter) is a parameter widely used in the study of neuromuscular diseases. For example, the typical size of a mouse muscle fiber is approximately 30 µm. It is also considered an indicator of muscle status (whether or not it is in the regeneration phase). The end of the regeneration phase in this case is linked to the growth of these fibers and their return to the normal average size. On the other hand, many myopathies are associated with muscle atrophy and therefore a decrease in fiber size.

[0042] Gene therapy involves introducing genetic material (a therapeutic gene) into cells to treat a genetic disease caused by an error in a gene, resulting in a lack of protein production or the production of a defective protein. The corrected genes are delivered by a gene therapy vector (such as an adeno-associated virus or AAV) into the cells and restore the production of a functional protein.

[0043] A mosaic is a set of juxtaposed figures that cover a plane according to a defined arrangement rule. A mosaic can be described as multi-regular, non-homogeneous, regular, or semi-regular. A preferred mosaic of the invention is the division of images into fractions called tiles, which are in the shape of squares or rectangles.

[0044] Examples of steps in the process according to the invention

[0045] The method of the invention includes a first step consisting of detecting at least one cellular parameter and / or one parameter of the cellular markers to be analyzed on an image of cell cultures (for example by proprietary software or by algorithms accessible in the literature).

[0046] Non-exhaustive examples of cellular parameters or cellular marker parameters to be analyzed:

[0047] Nuclei (study of cell maturation)

[0048] Intensity and surface area of ​​cell markings (especially fluorescent)

[0049] Raw pixel values ​​and / or their statistics (Mean, Standard Deviation, Minimum, Maximum, Median, etc.) of the colors (especially fluorescent)

[0050] The said cellular parameters or parameters of cellular markers to be evaluated on cultured cells may include: cell components such as the nucleus, endoplasmic reticulum, lysosomes, mitochondria, cytoskeleton, Golgi apparatus, membranes or certain specific nucleic acids, carbohydrates, lipids or proteins.

[0051] The process of the invention includes a second step consisting of constructing a mosaic of the image (for example according to squares, also called tiles).

[0052] This image mosaic allows us to generate, in a third step, a smaller image (i.e., with a lower pixel count) than the one from the first step by assigning values ​​corresponding to the measured parameters (e.g., nuclear density, marking area, mean, sum, pixel values, etc.) to each square. This smaller image is also called a density map.

[0053] The tile size is determined by striking a balance between the resolution and precision of the cellular parameter being evaluated. If the tile is too small, the image tends towards a binary state (the extreme case where the tile equals the size of the nucleus). If the tile is too large, the image tends towards a homogeneous state (the extreme case where the tile equals the size of the original image).

[0054] A fourth step involves training a deep learning model to calculate a proximity score relative to the two reference categories: (i) the untreated cellular model of the disease under investigation - also called the negative control and (ii) the healthy cell - also called the positive control for each of the different cellular images.

[0055] This fourth step can advantageously use a convolutional neural network architecture, or convolutional neural network (CNN for Convolutional Neural Networks) based on supervised classification.

[0056] The CNN network determines a proximity score, which can be used to predict the effectiveness of a given gene therapy.

[0057] In a development, the mosaic is multi-regular or non-homogeneous, or regular or semi-regular.

[0058] In mathematics, a "mosaic" refers to a set of juxtaposed figures that cover the plane according to a defined arrangement rule. A mosaic can be described as multi-regular, non-homogeneous, regular, or semi-regular. An "adaptive" tiling can be advantageous for optimizing calculations.

[0059] In a development, the dimensions of the tiles are determined dynamically.

[0060] In a development, the dimensions of the tiles are determined dynamically, based on an external parameter received. It is indeed possible to define (for example) an irregular mosaic whose tile dimensions optimize, improve, or otherwise modify one or more constraints received from the outside (e.g., computation speed, spatial distribution, etc.).

[0061] In a development, the dimensions of the tiles are determined dynamically as a function of an external parameter received so as to optimize one or more of the following parameters: the number of tiles; computer memory consumption; calculation time of one or more steps of the process; optimization of the spatial distribution of biological parameters; one or more statistics of the marked biological parameters: presence, cardinality, mean, median, sum, standard deviation, nth moments, etc.); calculation speed of deep learning algorithms; optimization of proximity score; type of cell considered; presence of nucleus, centronucleation parameters, size (or diameter) of fibers, biological parameters to be evaluated on a histological section (which may include: the number of nuclei (indicator of the number of inflammatory cells), the number of muscle fibers, the diameter (size) of muscle fibers, the type of muscle fibers (slow or fast), connective tissue proteins, the position of nuclei within muscle fibers (Centronucleation), the presence of fibrosis, the presence of necrosis, the intensity of a labeling, etc.).

[0062] In particular, the tile size can be determined, influenced, moderated or weighted by a compromise between the definition and the precision of the cellular parameter to be evaluated.

[0063] One or more feedback loops can indeed be advantageously implemented.

[0064] Optimization can notably be multi-objective. Multi-objective optimization (also called multi-criteria optimization) is indeed a branch of mathematical optimization that specifically deals with optimization problems having several objective functions.

[0065] Figure 1 illustrates one embodiment of a process according to the invention.

[0066] The image mosaic is made up of squares. Each square is associated with a value for cellular parameters or cellular markers.

[0067] A density map is then generated, where each pixel corresponds to the number of detections. This map has a small number of pixels, but still retains the relevant information about the parameter(s) being studied. This type of image will serve as the input for the neural network.

[0068] After the parameter being studied has been detected and the image has been scaled, the data is transferred as input to a CNN neural network.

[0069] The CNN network performs supervised deep learning and produces results, enabling the determination of a proximity score relative to predefined reference categories, and then predicting the effectiveness of a treatment in the research phase based on this proximity score. The method of the invention can be implemented with any type of cell, with muscle cells being preferred.

[0070] In a particular mode, the process of the invention is implemented with cellular parameters detected by one or more fluorescent markers.

[0071] According to an advantageous embodiment, the method according to the invention does not include an image compression step upstream of the deep learning prediction algorithms (i.e. to determine the proximity score).

[0072] According to a preferred embodiment, the method according to the invention is lossless in terms of spatial information during the application of deep learning prediction algorithms (i.e., to determine the proximity score). In other words, the method according to the invention does not require the use of artificial intelligence to recover spatial information (reconstruct the two-dimensional image).

[0073] The invention is described in more detail in the following experimental examples.

[0074] These examples are provided for illustrative purposes only and are not intended to be exhaustive.

[0075] EXAMPLE 1: Evaluation of the effect of different drugs on restoring cholesterol transport to lysosomes in cultured cells

[0076] The study focused on the effect of different drugs on C25 myoblast cell cultures labeled with NBCholesterol (a fluorescent analog of cholesterol) and with Lysotracker Red DND (red fluorescent), a lysosome marker.

[0077] C25 myoblast cells were cultured in petri dishes under different conditions.

[0078] Some were left as they were; these will be called "Control" or CTRL cells. They represent the reference of healthy cells.

[0079] Others have been treated with "U 18666A," an inhibitor of NPC1, a lysosomal cholesterol transporter. These cells have inhibited cholesterol transport to lysosomes. Cholesterol therefore accumulates in these cells and is present in greater quantities than in CTRL cells. They represent the reference for pathological cells. When they receive no treatment, they are called "U 18" cells.

[0080] Some cells that have received "U 18666A" are exposed to one of five different treatments. These are called T1, T2, T3, T4, and T5 respectively (for Treatment 1, 2, 3, 4, and 5).

[0081] After fixation, the cells are labeled with DAPI (nuclear marker - blue fluorescent), with NBDCholesterol, a cholesterol marker (green fluorescent), and with lysotracker Red DND, a lysosome marker (red fluorescent). The cells are then observed using a confocal microscope.

[0082] For each cell culture, one acquisition corresponds to a mosaic of 2x2 images, each with 4096 pixels. The mosaic therefore measures approximately between 7000 and 8144 pixels (depending on the image overlap). Each mosaic contains four channels: three fluorescence channels (blue, green, and red) and one transmission channel (which produces a bright-field image).

[0083] For the CTRL and U 18 conditions, 6 mosaics were acquired and used as ground truth to train a neural network to score mosaics corresponding to treatments, (figure 3).

[0084] Figure 5 shows the quantifications of green fluorescence (corresponding to the cholesterol load).

[0085] Methodology

[0086] A neural network was trained with the CTRL and U 18 conditions, associating the CTRL condition with a score of 0 and the U 18 condition with a score of 1. The neural network then evaluated each treated condition according to different criteria and assigned them a score between 0 and 1 (Figure 3).

[0087] Different embodiments of the invention

[0088] The principle of the method is constant: it consists of creating a virtual mosaic with 20-meter square tiles on the image. Each tile is assigned a value based on cellular parameters or cell marker parameters. From this mosaic, a low-resolution image, in this example 80x80 pixels, is created for each measured parameter. All these images, corresponding to a single microscopy acquisition and therefore to a single condition, are compiled into a stack of images that will constitute the input for the neural network.

[0089] Example of image evaluation based on 3 parameters

[0090] In this case, the associated parameters in each tile are:

[0091] The number of nuclei per tile (blue fluorescence), based on the position of the center of mass of the nucleus.

[0092] Cholesterol staining area (green fluorescence) per tile

[0093] The lysosome labeling area (red fluorescence) per tile

[0094] Figure 6 shows the results of assigning scores to the different treatments. These results are very consistent and close to those found by simply quantifying green fluorescence (a marker of cholesterol levels) in Figure 5.

[0095] In this example, the measured parameters are those of the fluorescent markers used, namely a cholesterol marker and a lysosome marker. However, it is entirely feasible to consider different protein markers (labeled by fluorescence) and quantify them per tile, depending on the cell line being studied. For example, in our case, since C25 cells are muscle cells, it is quite possible to have a similar approach with labeling for dystrophin, the various alpha, beta, gamma, and delta subunits of the sarcoglycan complex, titin, actin, myosin and all its chains, and of course any type of protein relevant to muscle analysis in the context of a given scientific study.

[0096] In a cellular model of cancerous diseases, this approach can also be applied using known markers for these cancer cells; in particular, proliferation markers such as Kl 67, or apotosis markers, such as Annexin V, when seeking to evaluate the effectiveness of an anticancer drug.

[0097] EXAMPLE 2

[0098] The invention has several objectives:

[0099] Object 1: A method for characterizing cell culture images comprising the steps of: a) detecting at least one cellular parameter and / or one parameter of cellular markers to be analyzed on a cell culture image; b) constructing a mosaic of said image; c) generating an image of reduced dimensions compared to the image of step a, by assigning to each square of the mosaic values ​​corresponding to the measured parameters; d) applying deep learning algorithms on the images generated in step c in order to determine a proximity score of the treated cell images compared to images of healthy cells or untreated diseased cells.

[0100] Item 2: A process according to item 1, in which cell cultures are stained with a fluorescent and / or immunofluorescent marker.

[0101] Item 3: A method according to any one of the preceding items, wherein the image of cell cultures is an image of cultured muscle cells.

[0102] Item 4: A method according to one of the preceding items, in which the image of cell cultures is that of cultured myoblasts.

[0103] Item 5: A process according to any one of the preceding items, wherein the evaluated treatment of the treated cells is a gene therapy treatment.

[0104] Object 6: A method according to any of the preceding objects, in which the mosaic is multi-regular, non-homogeneous, regular, or semi-regular. Object 7: A method according to any of the preceding objects, in which the dimensions of the tiles or blocks of the mosaic are determined dynamically, for example, as a function of an external parameter received.

[0105] Object 8: A process based on one of the previous objects, without image compression upstream of deep learning algorithms to determine the proximity score.

[0106] Object 9: A method according to one of the preceding objects, without loss of spatial information when applying deep learning algorithms to determine the proximity score. Object 10: A computer program product, said computer program comprising code instructions for performing the steps of the method according to one of the objects 1 to 11, when said program is executed on a computer.

Claims

DEMANDS 1. A method for characterizing cell culture images comprising the steps of: a. detecting at least one cellular parameter and / or one parameter of cellular markers to be analyzed on a cell culture image; b. constructing a mosaic of said image; c. generating an image of reduced dimensions compared to the image of step a, by assigning to each square of the mosaic values ​​corresponding to the measured parameters; d. applying deep learning algorithms on the images generated in step c in order to determine a proximity score of the treated cell images compared to images of healthy cells or untreated diseased cells.

2. A method according to claim 1, wherein the cell cultures are stained with a fluorescent and / or immunofluorescent marker.

3. A method according to any one of the preceding claims, wherein the image of cell cultures is an image of cultured muscle cells.

4. A method according to any one of the preceding claims, wherein the image of cell cultures is that of cultured myoblasts.

5. A method according to any one of the preceding claims, wherein the evaluated treatment of the treated cells is a gene therapy treatment.

6. A method according to any one of the preceding claims, wherein said mosaic is multi-regular, non-homogeneous, regular or semi-regular.

7. A method according to any one of the preceding claims, wherein the dimensions of the tiles or mosaic pieces are determined dynamically, for example as a function of an external parameter received.

8. A method according to any one of the preceding claims, without image compression upstream of deep learning algorithms to determine the proximity score.

9. A method according to any one of the preceding claims, without loss of spatial information when applying deep learning algorithms to determine the proximity score.

10. A computer program product, said computer program comprising code instructions enabling the steps of the process according to any one of claims 1 to 9 to be carried out when said program is executed on a computer.

Images