Method for classifying a biological sample

Abstract

The invention relates to a method for classifying a biological sample, the method comprising the following steps implemented by data processing means (21) of an item of equipment (2): a) a step of obtaining data relating to the biological sample; b) a step of extracting a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, the extraction step being performed by applying an encoder of an encoder-decoder neural network taking as input the data relating to the biological sample, the encoder of the network comprising at least one self-attention block taking as input a context vector relating to the context of the biological sample and an encoder hidden state vector representative of the data relating to the biological sample and generating as output the updated context vector and the updated encoder hidden state vector; c) a step of classifying the biological sample by applying a decoder of the neural network to the final encoder hidden state vector.

Description

[0001] DESCRIPTION

[0002] TITLE: Method for classifying a biological sample. FIELD OF THE INVENTION

[0003] The present invention relates to the field of classification of a biological sample, more particularly a blood sample, by means of a neural network.

[0004] STATE OF THE ART

[0005] Flow cytometry allows for the analysis of a biological sample, particularly a blood sample. This makes it possible to identify abnormalities in a biological sample, thus enabling, for example, the detection of pathologies such as a hematological disorder.

[0006] Flow cytometry involves performing at least one hydrodynamic focusing and passing the blood cells of the sample one by one through a measuring device which, depending on what is implemented, produces a number of physical measurements for each cell.

[0007] For measurements to be clearly discernible, the cells must be separated and move at speeds that allow for measurement and data acquisition. Furthermore, the cell counts must be sufficient to allow for accurate statistical evaluations of each cell population. To achieve this, the blood sample is not analyzed undiluted but diluted. Moreover, given that red blood cells are a thousand times more numerous than white blood cells, the sample must remain in the flow cytometer for several minutes to perform an analysis that is statistically representative of the white blood cell populations.The other solution, to reduce the time the sample spends in the cytometer by a factor of ten, consists of performing, in addition to a dilution of the sample, a lysis which selectively destroys red blood cells to obtain a sample sufficiently concentrated in leukocytes and thus obtain statistically correct counts for these populations.

[0008] The measurements made by the cytometer can, as explained in patent application no. FR3078777, make it possible to determine for each biological cell of the biological sample to be analyzed, a point in an N-dimensional space whose coordinates are defined according to the cytometry parameters measured for the corresponding biological cell, where N is an integer greater than or equal to 2.

[0009] Then, an operator identifies one or more anomalies in the sample cluster file compared to normal samples.

[0010] This method is time-consuming, expensive, and prone to errors. A method for detecting cancer based on cytometric data using artificial intelligence is also known (see patent application W02020081582A1). However, this method lacks reliability and efficiency. Furthermore, it does not allow for the determination of parameters for a blood sample such as the complete blood count, differential, or the probability of hematological malignancies.

[0011] There is therefore a need for a process to analyze biological samples, in particular to detect any anomalies, that is more efficient, less expensive and more reliable.

[0012] DESCRIPTION OF THE INVENTION

[0013] One aim of the invention is to enable the analysis of a biological sample, and preferably also the prediction of the state of an individual from whom the biological sample is taken, in a way that is more efficient, less expensive and more reliable.

[0014] According to the first aspect, a method for classifying a biological sample is proposed, comprising the following steps implemented by data processing means of a piece of equipment:

[0015] a) obtaining data relating to the biological sample;

[0016] b) extraction of a final encoder hidden state vector representative of the data relating to the biological sample and of a final context vector relative to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relative to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0017] c) classification of the biological sample by applying a decoder of said neural network to the hidden state vector of the final encoder.

[0018] According to advantageous and non-limiting characteristics of this first aspect, taken alone or in any combination:

[0019] The encoder comprises a sequence of n>1 consecutive self-attentive blocks such that the i+1-th self-attentive block, 1 <i<n, prend en entrée des i+1-ème vecteur de contexte et i+1 -ème vecteur d’état caché d’encodeur correspondant aux i-ème vecteur de contexte mis à jour et i-ème vecteur d’état caché d’encodeur mis à jour générés par le i-ème blocauto-attentif, le vecteur d’état caché d’encodeur final étant le n+1 -ème vecteur d’état caché d’encodeur ;

[0020] the encoder includes a first self-attentive block, the first self-attentive block taking as input the data relating to the biological sample as the first encoder hidden state vector and a first predetermined context vector; the first context vector is learned with said neural network and / or relating to data relating to an individual from whom said biological sample originates; the encoder includes a sequence of at least three self-attentive blocks;

[0021] at least one self-attentive block of the encoder includes at least one multi-head attention layer and one forward-propagating neural network;

[0022] at least one self-attentive block of the encoder includes at least two multi-head attention layers;

[0023] the decoder includes a classifier block which generates as output a vector of classification scores for a plurality of possible classes of the biological sample and / or components of the biological sample;

[0024] at least one self-attentive block upstream of said classifier block;

[0025] the decoder's self-attentive block takes as input the final context vector and the final state vector and generates as output a hidden state vector of the decoder; the biological sample is or includes at least one of the following liquids: blood, plasma, serum, urine, cerebrospinal fluid, synovial fluid, cell suspension, lymphatic fluid, aspiration, lavage or bronchoalveolar lavage fluid and particle suspension in biological or industrial media; step c) of classification consists of determining a type of at least one cell of the biological sample;

[0026] step a) of obtaining includes the acquisition of data relating to the biological sample by means of at least one flow cytometer;

[0027] The data relating to the biological sample include measurements of the volume of the cells of the biological sample by impedance, measurements of optical absorption of the cells of the biological sample, diffraction measurements at different angles generated by the passage of the cells of the biological sample in a laser beam and / or fluorescence measurements of nucleic acids made fluorescent by one or more fluorochromes previously brought into contact with the cells of the biological sample;

[0028] the process includes a step d) of predicting a state of an individual from which said biological sample is taken by applying a second decoder of said neural network to the final context vector;

[0029] the process includes a step aO) of supervised learning of the neural network on the basis of a training database relating to reference biological samples associated with reference classes.

[0030] According to a second aspect, a biological sample classification system is proposed, comprising at least one piece of equipment including data processing means, characterized in that said data processing means are configured to implement:

[0031] - obtaining data relating to the biological sample;

[0032] - the extraction of a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0033] - the classification of the biological sample by applying a decoder of said neural network to the hidden state vector of the final encoder.

[0034] Advantageously, the system further includes at least one flow cytometer configured to acquire biological sample data from the biological sample.

[0035] According to a third aspect, a computer program product is proposed comprising code instructions for the execution of a process such as previously presented for the classification of a biological sample, when said program is executed on a computer.

[0036] According to a fourth aspect, a computer-readable storage method is proposed on which a computer program product includes code instructions for executing a process such as previously presented for classifying a biological sample.

[0037] According to another aspect, a method is proposed for predicting the state of an individual from which a biological sample is taken, comprising the following steps implemented by data processing means of a piece of equipment:

[0038] a) obtaining data relating to the biological sample;

[0039] b') extraction of a final encoder hidden state vector representative of the data relating to the biological sample and of a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0040] c') prediction of a state of an individual from which said biological sample is taken by applying a decoder of said neural network to the final context vector.

[0041] According to advantageous and non-limiting characteristics of this prediction method, taken alone or in any combination:

[0042] The encoder comprises a sequence of n>1 consecutive self-attentive blocks such that the i+1-th self-attentive block, 1 <i<n, prend en entrée des i+1-ème vecteur de contexte et i+1 -ème vecteur d’état caché d’encodeur correspondant aux i-ème vecteur de contexte mis à jour et i-ème vecteur d’état caché d’encodeur mis à jour générés par le i-ème bloc auto-attentif, le vecteur de contexte final étant le n+1 -ème vecteur de contexte ; l’encodeur comprend un premier bloc auto-attentif, le premier bloc auto-attentif prenant en entrée les données relatives à l’échantillon biologique en tant que premier vecteur d’état caché d’encodeur et un premier vecteur de contexte prédéterminé ; le premier vecteur de contexte est appris avec ledit réseau de neurone et / ou relatif à des données relatives à un individu dont provient ledit échantillon biologique ; l’encodeur comprend une séquence d’au moins trois blocs auto-attentifs ;

[0043] at least one self-attentive block of the encoder includes at least one multi-head attention layer and a forward-propagating neural network; at least one self-attentive block of the encoder includes at least two multi-head attention layers;

[0044] the decoder includes a predictor block generating as output a vector of classification scores for a plurality of possible classes of the individual's condition; the biological sample is or includes at least one of the following fluids: blood, plasma, serum, urine, cerebrospinal fluid, synovial fluid, cell suspension, lymphatic fluid, aspiration, lavage or bronchoalveolar lavage fluid and particle suspension in biological or industrial media; the prediction step c') consists of determining a pathology of the individual from whom the biological sample originates;

[0045] step a') of obtaining includes the acquisition of data relating to the biological sample by means of at least one flow cytometer;

[0046] The data relating to the biological sample include measurements of the volume of the cells of the biological sample by impedance, measurements of optical absorption of the cells of the biological sample, diffraction measurements at different angles generated by the passage of the cells of the biological sample in a laser beam and / or fluorescence measurements of nucleic acids made fluorescent by one or more fluorochromes previously brought into contact with the cells of the biological sample;

[0047] the prediction process includes a step d') of classifying the biological sample by applying another decoder of said neural network to the hidden state vector of the final encoder;

[0048] The prediction process includes a step a0') of supervised learning of the neural network based on a training dataset of reference biological samples associated with reference states; the biological sample is a first biological sample, step a) further includes obtaining data relating to a second biological sample, the final encoder hidden state vector is a first final encoder hidden state vector representative of the data relating to the first biological sample, step b) includes extracting a second final encoder hidden state vector representative of the data relating to the second biological sample, the final context vector is relative to the context of each biological sample, the encoder takes as input the data relating to each biological sample,The context vector taken as input by the self-attentive block is relative to the context of each biological sample; the encoder hidden state vector taken as input by the self-attentive block is representative of the data relating to the first biological sample; and the self-attentive block further takes as input an encoder hidden state vector representative of the data relating to the second biological sample. The self-attentive block then generates as output said updated context vector and each updated encoder hidden state vector. Step c) includes the classification of each biological sample by applying at least one decoder of said neural network to the final encoder hidden state vectors.

[0049] The encoder generates a single final context vector and the self-attentive block generates a single updated context vector.

[0050] Furthermore, a system is proposed for predicting the state of an individual from whom a biological sample is taken, comprising at least one piece of equipment including data processing means, characterized in that said data processing means are configured to implement:

[0051] - obtaining data relating to the biological sample;

[0052] - the extraction of a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0053] - the prediction of a state of an individual from which said biological sample is taken by applying a decoder of said neural network to the final context vector.

[0054] Advantageously, the system further includes at least one flow cytometer configured to acquire biological sample data from the biological sample.

[0055] Also proposed is a computer program product comprising code instructions for the execution of a process as previously described for predicting the state of an individual from whom a biological sample is taken, when said program is executed on a computer.

[0056] Also proposed is a computer-readable storage method in which a computer program product includes code instructions for executing a process, as previously described, of predicting the state of an individual from whom a biological sample is taken.

[0057] DESCRIPTION OF THE FIGURES

[0058] Other features and advantages of the present invention will become apparent from the following description of a preferred embodiment. This description will be given with reference to the accompanying figures, including:

[0059] - Figure 1 represents a classification system for a biological sample;

[0060] - Figure 2 represents the steps in the process of classifying a biological sample; - Figure 3 schematically represents, in landscape format, an encoder-decoder type neural network adapted to implement the classification;

[0061] - Figure 3bis schematically represents, in landscape format, an encoder-decoder type neural network adapted to implement classification for several samples;

[0062] - Figure 4 schematically represents, in portrait format, a self-attentive block of the encoder-decoder type neural network;

[0063] - Figure 4bis schematically represents, in portrait format, a self-attentive block of the encoder-decoder type neural network according to another embodiment;

[0064] - Figure 4ter schematically represents, in portrait format, a self-attentive block of the encoder-decoder type neural network according to an embodiment allowing the classification of several samples;

[0065] - Figure 5 schematically illustrates in detail, in portrait format, an encoder-decoder type neural network adapted to implement classification including self-attentive blocks such as in Figure 4;

[0066] - Figure 6 represents the steps in the process of predicting the state of an individual from which a biological sample is taken;

[0067] - Figure 7 schematically represents, in landscape format, an encoder-decoder type neural network adapted to implement prediction; - Figure 8 schematically represents, in landscape format, an encoder-decoder type neural network adapted to implement both classification and prediction.

[0068] DETAILED DESCRIPTION OF THE INVENTION

[0069] System

[0070] With reference to Figure 1, a 100 system for classifying a biological sample is proposed.

[0071] The biological sample contains, in suspension, biological cells, and advantageously blood cells.

[0072] The cells circulating in the blood include non-nucleated cells such as red blood cells or erythrocytes (approximately 5 million per mm³ of normal blood) and platelets (approximately 300,000 per mm³), as well as nucleated cells called leukocytes (approximately 10,000 per mm³). Blood may also contain other nucleated cells, such as erythroblasts, which are immature red blood cells, and other rarer cell types. Each cell type constitutes what is also called a population.

[0073] Other biological fluids contain blood cells, such as cerebrospinal fluid or urine.

[0074] Subsequently, we speak of a biological sample, but this is not limited to a blood sample.

[0075] The biological sample may be one of the following liquids:

[0076] - Blood and its derivatives:

[0077] - Whole blood: Blood cells (red blood cells, white blood cells, platelets) can be analyzed for studies of enumeration, differentiation, or presence of specific markers.

[0078] - Plasma and serum: These liquids can be used to study proteins, complexes, or lipoproteins. They are often used in analyses of particles such as exosomes or soluble markers.

[0079] - Bodily biological fluids:

[0080] - Urine: Flow cytometry can be used to analyze urinary cells (such as circulating epithelial cells or tumor cells) or particles present in the urine.

[0081] - Cerebrospinal fluid (CSF): This fluid can be analyzed for the detection of cells and other particles in neurological or infectious studies. - Synovial fluid: Used to analyze cells in inflammatory or infectious joint diseases.

[0082] - Cell culture: - Cell suspensions: Cultured cells (e.g., cancer cells, immune cells) can be analyzed for functional, proliferation, cell cycle, or apoptosis studies.

[0083] - Microbiological samples:

[0084] - Bacterial or fungal suspensions: Flow cytometry can be used to analyze bacteria or fungi in suspension, based on their physical characteristics or the expression of specific markers (e.g., size or the presence of certain proteins).

[0085] - Other biological or environmental fluids:

[0086] - Lymphatic fluids: Lymphatic fluid, originating from the lymph nodes, can be analyzed to detect specific immune cells, such as lymphocytes.

[0087] - Suction, lavage or bronchoalveolar lavage fluids: These samples can be used for flow cytometry analyses, for example to detect respiratory cells in lung diseases.

[0088] - Other non-biological particles or media:

[0089] - Suspended particles (i.e., particle suspensions) in biological or industrial media: Flow cytometry can also be used to analyze particles of any size and nature, such as microparticles, nanoparticles, or particles suspended in industrial or environmental liquids. The "classification" of a biological sample consists of determining at least one class from a set of possible descriptive classes for the biological sample. For example, in the case of a blood sample, a binary classification is possible, i.e., two possible classes: "presence of an anomaly" or "absence of an anomaly," indicating, respectively, the presence or absence of an anomaly in the blood sample.

[0090] The classes can be pathologies, for example different types of blood disorders, indicating whether the individual from whom the sample is taken is affected by one or more of these diseases or not.

[0091] Classes can be cell types. For example, one class might be "red blood cell" and another class "white blood cell".

[0092] In one embodiment, the biological sample comprises only white blood cells or only red blood cells (the sample being obtained by extracting white or red blood cells from a base sample). The sample may be associated with several cell types in varying proportions. For example, a sample containing white blood cells could contain 30% lymphocytes, 10% monocytes, and 60% neutrophils. The cell types thus include, in particular, the "lymphocyte," "monocyte," and "neutrophil" classes.

[0093] In practice, to study white blood cells, a blood sample is lysed to remove red blood cells, as white blood cells are far less numerous than red blood cells. This results in a biological sample that is notably devoid of red blood cells. Therefore, this biological sample may contain, in addition to white blood cells, nucleated cells that are not white blood cells, such as erythroblasts, and other cells like platelets or platelet aggregates, or other types of particles.

[0094] To study red blood cells, the biological sample can be an unlysed blood sample and therefore include red blood cells but also platelets, white blood cells and all cells possibly present in the blood.

[0095] The classes may, for example, include platelets, erythrocytes, lymphocytes, monocytes, and the various populations of polymorphonuclear leukocytes. Advantageously, the classes include the following: basophil (a type of white blood cell), erythroblast (Nucleated Red Blood Cell), lymphocyte (a type of white blood cell), monocyte (a type of white blood cell), blast (which is an immature hematopoietic cell), neutrophil (a type of white blood cell), immature granulocyte, eosinophil (a type of white blood cell), red blood cell (RBC), reticulocytes (a type of red blood cell), platelets, and platelet aggregate.

[0096] The classes can be adapted to allow the determination of the complete blood count (CBC) (which can allow alarms to be issued on levels that are too high or too low of certain cells) and / or the probability of the presence of a hematological disorder, etc.

[0097] Advantageously, classification allows us to obtain a cell type for some cells, or even for each cell, in the biological sample. In other words, the process makes it possible to determine the type of each cell in the biological sample, thus determining the sample's composition.

[0098] Thus, classification allows us to assign one or more classes to the biological sample and / or to assign one or more classes to its components. These components include, for example, cells, particles, debris, and molecules (e.g., proteins). Examples of components are also detailed previously in the sample example list.

[0099] Classification can consist of classifying one or more, or even each, of the components of the biological sample and then deducing, from all the classes assigned to the components, one or more classes for the biological sample as a whole. Classification of the biological sample therefore refers to the assignment of one or more classes to the biological sample as a whole and / or the assignment of one or more classes to components of the biological sample.

[0100] The present invention is not limited to any particular type of classification. The classification is implemented within a system 100 as represented in Figure 1, by means of a second piece of equipment 2 and optionally by means of a first piece of equipment 1. The first piece of equipment 1 is preferably the learning equipment (implementing the learning process) and the second piece of equipment 2 is a user equipment (implementing the classification process), for example a terminal in a biological analysis laboratory.

[0101] It is entirely possible that the two devices 1, 2 are confused, but preferably the first device 1 is a remote device, and the second device 2 is a consumer device, in particular a desktop computer, a laptop, etc.

[0102] Advantageously, the second device 2 is connected to the first device 1 solely for training purposes. Once the training is complete, the second device 2 is disconnected from the first device 1 and the second device 2 implements the classification process independently.

[0103] The second piece of equipment 2 is advantageously connected to one or more flow cytometers 3, so that data relating to the biological sample can be acquired directly, typically for real-time processing; alternatively, the biological sample data can be loaded onto the second piece of equipment 2. A flow cytometer 3 includes measuring means configured to measure cytometry parameters of biological cells, and in particular to measure the optical and electrical properties of biological cells. The measurements obtained are referred to as the biological sample data. It should be noted that the optical and electrical properties of biological cells can be properties of the cells themselves or properties derived from markers (e.g., fluorochromes) attached directly or indirectly (via antibodies, for example) to the cells.

[0104] Each device 1, 2 is typically a remote computing device connected to a local area network or a wide area network such as the internet for data exchange. Each includes data processing means 11, 21 such as a processor and data storage means 12, 22 such as computer memory, for example flash memory or a hard drive.

[0105] The data processing means 21 of the second equipment 2 are configured to implement: - obtaining data relating to the biological sample;

[0106] - the extraction of a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0107] - the classification of the biological sample by applying a decoder of said neural network to the hidden state vector of the final encoder.

[0108] The second piece of equipment 2 typically includes a user interface 23 such as a screen for interaction.

[0109] The first equipment 1 advantageously stores a training database, i.e. a set of data relating to already classified reference biological samples (e.g. associated with labels, i.e. classes, "presence of anomaly" or "absence of anomaly" indicating the presence or absence of an anomaly in the biological sample).

[0110] Advantageously, the training data is associated with labels relating to the individual from whom the biological sample was taken, for example, "individual's age," "individual's height," "individual's weight," and / or labels defining the test conditions, for example, "sample type," "sample collection time." This training data can be used to learn an initial context vector.

[0111] Process

[0112] With reference to Figure 2, a method for classifying a biological sample is proposed. The method aims to associate one or more classes with the biological sample, potentially to diagnose the individual from whom the sample was taken. More simply, the method can determine the type of at least one cell (for example, "basophil" or "platelet cluster") in the biological sample.

[0113] The method advantageously includes a step a0) of supervised learning of an encoder-decoder neural network based on the training set. As explained, the training set advantageously comprises data relating to reference biological samples associated with reference classes. The neural network is thus trained to generate as output one or more classes of a sample, taking as input data relating to that sample. In addition, the neural network includes a self-attentive block that takes as input a predetermined first context vector.

[0114] The first context vector advantageously shares no data with the biological sample data. The first context vector is independent of the biological sample data. The first context vector is not redundant with the biological sample data. Such a first context vector is described in the scientific publication Luna: Linear Unified Nested Attention, 2021, Xuezhe Ma et al.

[0115] Advantageously, the process includes a step a1) for learning the first context vector. The first context vector refers to the input context vector of the first self-attentive block of the neural network encoder. As will be described later, there are other context vectors used within the neural network. The first context vector is the first context vector used by the neural network when classifying a sample. The training of this first context vector is advantageously performed simultaneously with the training of the neural network in step a0). In this embodiment, the first context vector is therefore one of the parameters (i.e., hyperparameters) of the neural network, learned along with the other parameters of the neural network in step a0).

[0116] Alternatively, or in addition, the first context vector is determined from data relating to the individual from whom the biological sample originates. While linked to the individual, the first context vector is independent of the biological sample. Such individual data might include, for example, their age, height, weight, and / or medical conditions. In this case, the first context vector can be considered an input to the encoder.

[0117] The first context vector can be determined through learning and enriched using data relating to the individual from whom the biological sample was taken. In this case, the first context vector is both learned and determined from data relating to the individual from whom the biological sample was taken.

[0118] Advantageously, the training step aO) is implemented once, before the initial commissioning of the neural network, and possibly thereafter occasionally for updates to the neural network. Thus, advantageously, the training step aO) is not implemented for each subsequent second device 2, for installation on each second device. The step aO) is implemented once to obtain a trained neural network that can then be installed on different second devices.

[0119] The process includes a step a) of obtaining data relating to the biological sample.

[0120] Advantageously, the data relating to the biological sample include impedance measurements of the volume of the cells in the biological sample, optical absorption measurements of the cells in the biological sample, diffraction measurements at different angles generated by passing the cells of the biological sample through a laser beam, and / or fluorescence measurements of nucleic acids made fluorescent by one or more fluorochromes previously exposed to the cells of the biological sample. Patent application no. FR3078777 describes such measurements.

[0121] Advantageously, the data relating to the biological sample include immunological measurements (immunohematology). To obtain these measurements, the sample is placed in the presence of reagents containing fluorochrome-conjugated antibodies specific to receptors expressed on the surface of certain cells. Passing the cells through one or more laser beams generates signals proportional to the cell surface markers, thus allowing the cells to be characterized.

[0122] A person skilled in the art will know that they can use raw or pre-processed measurements as data relating to the biological sample. For example, measurements can be pre-processed to reduce their dimensionality.

[0123] Advantageously, the measures are pre-processed so that some of these measures are concatenated.

[0124] Advantageously, the measures are pre-processed so that a logarithmic transformation is applied to them.

[0125] Advantageously, the measurements are preprocessed using a "positional encoding" transformation of the Fourier decomposition type, with the decomposition parameters learned during the aO) training step of the neural network. An example of such a transformation is equation (7) from the following article: "Vision Transformer for Ne RF-Based View Synthesis from a Single Input Image," https: / / arxiv.org / pdf / 2207.05736. This transformation is called a NERF transformation. In this case, the dimensionality of the measurements is increased. This transformation allows for very efficient processing. Indeed, this transformation facilitates the separation of the signal (i.e., measurements in the frequency domain) from the background noise. Moreover, this transformation provides input data that is easily usable by the neural network.

[0126] Preferably, step a) of data acquisition includes acquiring data relating to the biological sample using at least one flow cytometer. The process includes a step b) of extracting a final encoder hidden state vector (VEF) representative of the biological sample data and a final context vector (VCF) relating to the context of the biological sample. Extraction step b) is implemented by applying encoder E of the RED encoder-decoder neural network, taking as input the biological sample data VE1 as schematically represented in Figure 3.

[0127] The final encoder vector is advantageously a high-level semantic representation of the first state vector taken as input by the encoder. This high-level semantic representation is then classified to obtain one or more classes of the biological sample.

[0128] The network encoder comprises at least one self-attentive block that takes as input a context vector, the context vector being relative to the context of the biological sample, and an encoder hidden state vector representing the data relating to the biological sample. The self-attentive block outputs the updated context vector and the updated encoder hidden state vector. The first self-attentive block of encoder E also takes as input the first context vector VC1.

[0129] A self-attentive block includes an attention layer. An attention layer processes a sequence of input data by applying weights, i.e., attention weights, to each of its parts according to its importance for the specific task to be performed. Typically, attention weights are encompassed in three weight matrices: query weights (WQ), key weights (WK), and value weights (Wv). This set of matrices, denoted (WQ, WK, WV), is called an attention head. Figure 4, which schematically represents a self-attentive block in a particular implementation, illustrates the paths Q, V, and K, which mean that a vector entering via that path will have the corresponding matrix WQ, WK, or WV applied to it, respectively.

[0130] It is worth noting that self-attentive block encoders are well-known for generative language models such as LLMs, and surprisingly prove to be very well-suited to the present case of classifying a biological sample. Moreover, while known language model architectures are generally very complex and require high-powered data processing capabilities that would have deterred a professional from using such architectures, the present method, on the contrary, works very well with a much lighter structure, as will be seen later.

[0131] Any self-attentive block network structure known in natural language processing can be used, for example, a transformer-type network described in the paper "Attention is all you need" (https: / / arxiv.org / p 4 ■ (06.03762), LUNA (Linear Unified Nested Attention) described in the paper https: / / arxiv.org / pdf / 2106.01540), etc. In any case, the self-attentive block of the invention is specific in that it updates a context vector (which is not the case in a basic transformer).

[0132] A self-attentive block updates a context vector and a state vector. In other words, a self-attentive block takes a context vector and a state vector as input and outputs the updated context vector VCmaj and the updated state vector VEmaj.

[0133] A state vector is a vector representing the data related to the sample. It is updated by the self-attentive block by taking into account the context vector VC. An updated state vector is a state vector, therefore a vector representing the data related to the sample, including parts of which have been weighted and modified according to an attention weight and the context vector.

[0134] We will use the term "encoder hidden state vector" to refer to a state vector taken as input or generated by a part of the encoder. We will use the term "decoder hidden state vector" to refer to a state vector taken as input or generated by a part of the decoder.

[0135] A context vector is not redundant with the data relating to the biological sample. A context vector includes contextual information. A context vector is updated by taking into account a state vector. Thus, an updated context vector is a context vector enriched with contextual information from a state vector.

[0136] Thus, a self-attentive block allows a state vector to be updated by taking into account contextual information contained within the context vector, and the context vector is updated by taking into account contextual information that would be contained within the state vector. This makes it possible to obtain vectors, in particular to obtain a final encoder hidden state vector, that contain the important and useful information for performing reliable classification.

[0137] Thus, in the present process, two distinct and different vectors are generated by the self-attentive block: a state vector (the updated state vector) and a context vector (the updated context vector). The resulting context vector can be manipulated and used independently of the state vector, notably, for example, for the subsequent generation, by another self-attentive block, of a new, updated, hidden encoder state vector, as will be described later. This differs from known encoder-decoder architectures that include, for example, ISAB (Induced Self-Attention Block) layers, as described in the paper "Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks," by Juho Lee et al., which relate events and contextualize events and generate a state vector but do not also generate a context vector.

[0138] As we will see later, the encoder E can have a succession of hidden state vectors corresponding to successive processing, and, for example, the state vector taken as input by the i-th self-attentive block can be the i-th hidden state vector, and thus said updated hidden state vector is the i+1-th hidden state vector.

[0139] Similarly, the encoder E may have a succession of context vectors corresponding to successive processing, and, for example, the context vector taken as input by the i-th self-attentive block may be the i-th context vector, and thus said updated context vector is the i+1-th context vector.

[0140] Advantageously, as illustrated in Figures 3 and 5, the encoder E includes a first self-attentive block BAE1. The first self-attentive block takes as input the data relating to the biological sample as the first encoder hidden state vector VE1 and also takes as input the first predetermined context vector VC1. Thus, the data relating to the biological sample forms the first encoder hidden state vector.

[0141] Advantageously, as illustrated in Figure 4, at least one self-attentive block of the encoder includes at least one multihead attention layer (MTL). A multihead attention layer comprises several attention heads. In other words, input data to a multihead attention layer is processed in parallel by the layer's heads. This allows for more robust processing because values ​​from some of these heads, particularly at initialization, may be unreliable and destabilize the processing. The plurality of heads helps prevent such destabilization.

[0142] Advantageously, as illustrated in Figure 4, at least one self-attentive block of the encoder comprises at least two multi-head attention layers CMT1, CMT2.

[0143] Advantageously, each self-attentive block of the encoder includes at least two multi-head attention layers.

[0144] The two multihead attention layers comprise a first multihead attention layer (CMT1) and a second multihead attention layer (CMT2). ​​The first multihead attention layer takes as input a state vector and a context vector. The second multihead attention layer, positioned downstream of the first multihead attention layer, takes as input the same state vector taken as input by the first multihead attention layer and a vector generated from a vector generated as output by the first multihead attention layer and the context vector. This results in more stable and reliable processing. Indeed, the state vector(s) are updated based on the context (i.e., contextual information contained in the context vector) of the biological sample.The second multi-head attention layer is positioned downstream of the first multi-head attention layer, but there is advantageously a residual addition and AN2 normalization layer between the first multi-head attention layer and the second multi-head attention layer.

[0145] Advantageously, at least one self-attentive block of the encoder, preferably each self-attentive block of the encoder, comprises exactly two multi-head attention layers as illustrated in Figure 5. The processing is therefore reliable without being too computationally intensive or time-consuming.

[0146] Advantageously, at least one self-attentive block of the encoder, preferably every self-attentive block of the encoder, includes a feedforward neural network (RPA). This network allows for simple output generation. In such a network, information moves in only one direction: forward, from the input nodes, through the hidden layers (if any), and to the output nodes. There are no cycles or loops in such a network.

[0147] Preferably, the forward-propagating network is positioned downstream of a layer, preferably the second or last, of multihead attention. The forward-propagating network preferably takes as input a combination of a vector generated by the last multihead attention layer and the state vector taken as input by the self-attentive block to which the forward-propagating network belongs.

[0148] Advantageously, the combination of the vector generated by the last multi-head attention layer and the state vector taken as input by the self-attentive block is achieved using a residual addition and normalization (Add & Norm) layer. Such a layer adds vectors and then normalizes the result of the addition.

[0149] Preferably, the self-attentive block(s) of the encoder each include two multi-head attention layers CMT1, CMT2 and a forward propagation network RPA as illustrated in Figure 4.

[0150] Preferably, the encoder's self-attentive block(s) are LUNA (Linear Unified Nested Attention) type blocks as described in the scientific publication Luna: Linear Unified Nested Attention, 2021, Xuezhe Ma et al. A LUNA type self-attentive block comprises, as illustrated in Figure 4:

[0151] - a first multi-head attention layer CMT1 taking as input a context vector VC and a hidden encoder state vector VE,

[0152] - a second multi-head attention layer CMT2 taking as input a second primary intermediate state vector VI12 generated at output by the second residual addition and normalization layer AN2 (described below) and the encoder hidden state vector VE, - a first residual addition and normalization layer AN1 (called Add & Norm in English) taking as input a secondary intermediate state vector VI2 generated at output by the second multi-head attention layer CMT2 and the encoder hidden state vector VE,

[0153] - a second residual addition and normalization layer AN2 taking as input a first primary intermediate state vector VI11 generated as output by the first multi-head attention layer CMT1 and the context vector VC, the second residual addition and normalization layer AN2 generating as output an updated context vector VCmaj which corresponds to one of the outputs of the self-attentive block and the second primary intermediate state vector VI12, the second primary intermediate state vector VI12 therefore corresponds in the embodiment of figure 4 to the updated context vector VCmaj,

[0154] - a feed-forward neural network (RPA) taking as input a tertiary intermediate state vector VI3 generated at output by the first residual addition and normalization layer AN1,

[0155] - a third residual addition and normalization layer AN3 taking as input a quaternary intermediate state vector VI4 generated at output by the forward propagating network RPA and the tertiary intermediate state vector VI3, the third residual addition and normalization layer AN3 generating at output the updated encoder hidden state vector VEmaj. In other words, this third residual addition and normalization layer AN3 generates the updated encoder hidden state vector that corresponds to one of the outputs of the self-attentive block.

[0156] Optionally, as illustrated in Figure 4bis, the RPA forward propagation network taking as input a tertiary intermediate state vector VI3 is called the first RPA forward propagation network, and the self-attentive block includes a second RPA forward propagation network RPA2 located downstream of the second residual addition and normalization layer AN2, and a fourth residual addition and normalization layer AN4 located downstream of this second RPA forward propagation network RPA2. In this case, the second RPA forward propagation network RPA2 takes as input a first intermediate context vector VC11, which is identical to the second primary intermediate state vector VI12. The fourth residual addition and normalization layer AN4 takes as input a second intermediate context vector VCI2, generated by the second RPA forward propagation network RPA2, and the first intermediate context vector VCI1.The fourth residual addition and normalization layer AN4 generates the updated context vector VCmaj. Thus, the updated context vector VCmaj is the vector generated by the fourth residual addition and normalization layer AN4. According to this embodiment illustrated in Figure 4bis, unlike the embodiment illustrated in Figure 4, the updated context vector VCmaj is therefore not the second primary intermediate state vector VI12.

[0157] We note that, concerning the first residual addition and normalization layer AN1, said layer takes as input a hidden state vector of encoder VE which has ignored (i.e. bypassed) the second multi-head attention layer CMT2. This phenomenon is also called connection skipping or shortcut and is a known characteristic in residual neural networks (ResNets, as described for example in the paper "Deep Residual Learning for Image Recognition", Kaiming He et al.).

[0158] Advantageously, the E encoder comprises several self-attentive blocks (SABs) as illustrated in Figures 3 and 5. These blocks are advantageously arranged sequentially. In other words, the blocks are consecutive, i.e., they form an open chain.

[0159] Preferably, encoder E comprises a sequence of at least three self-attentive blocks (SABs) as illustrated in Figures 3 and 5. Encoder E may comprise 12 or more self-attentive blocks, or even 48 self-attentive blocks.

[0160] The invention is not limited to a specific number of self-attentive blocks of the encoder E, particularly to a maximum number of self-attentive blocks of the decoder D. Indeed, the input data of the encoder—the first context vector, and then the hidden encoder and context state vectors input to the self-attentive blocks of the encoder—are lightweight. Furthermore, when a self-attentive block comprises two multi-head attention layers, the complexity of the attention operations increases linearly with the input size, thus allowing for lightweight processing. Therefore, the encoder E can include a large plurality of self-attentive blocks (one or more than ten) while not requiring excessive computing resources and generating a highly relevant output vector that enables very reliable classification or, more generally, prediction.The encoder of the invention allows efficient processing, which does not require too many computing resources compared to network encoders of language models which generally include about six self-attentive blocks but of quadratic complexity of the input size.

[0161] Advantageously, the encoder consists solely of one or more self-attentive blocks.

[0162] Preferably, as illustrated in Figures 3 and 5, the encoder comprises a sequence of n>1 consecutive self-attentive blocks such that the i+1th self-attentive block, where i is strictly greater than 1 and strictly less than n, takes as input the i+1th context vector and the i+1th hidden encoder state vector corresponding to the i-th updated context vector (VC) and the i-th updated hidden encoder state vector (VE) generated by the i-th self-attentive block (BAE). In other words, a self-attentive block, if it is not the first in the encoder's block sequence, takes as input, as a state vector and context vector, the state and context vectors generated by the preceding self-attentive block.

[0163] As explained, if we consider the i-th self-attentive block, then its hidden state vector VE (i.e. the state vector that the i-th block takes as input) is the i-th hidden state vector VE and said updated hidden state vector VEmaj that it generates is the i+1-th hidden state vector.

[0164] In other words, the self-attentive BAE blocks of the E encoder form an open chain and each block takes as input vectors generated by only one other block which is the one preceding it (i.e. the one upstream) as illustrated in figures 3 and 5.

[0165] Advantageously, the final encoder hidden state vector (FELV) is the n+1th encoder hidden state vector. Recall that the encoder E allows the extraction (step b)) of this final encoder hidden state vector (FELV), which enables classification. The final encoder hidden state vector is the state vector generated (i.e., an updated state vector) by the last self-attentive block of the encoder.

[0166] Advantageously, the final context vector VCF is the n+1-th context vector.

[0167] In the case of Figures 3 and 5, the encoder comprises a sequence of n=3 consecutive self-attentive blocks such as:

[0168] - The first self-attentive block BAE1 takes as input the first context vector VC1 and the first state vector VE1, and generates a second context vector VC2 and a second state vector VE2 corresponding to the updated first context vector VC1 and the updated first hidden encoder state vector VE1,

[0169] - The second self-attentive block BEA2 takes as input the second context vector VC2 and the second state vector VE2, and generates a third context vector VC3 and a third state vector VE3 corresponding to the updated second context vector VC2 and the updated second hidden encoder state vector VE2,

[0170] - the 3rd self-attentive block BEA3 takes as input the third context vector VC3 and the third state vector VE3, and generates a fourth context vector VC4 (which is the final vector VCF) and a fourth state vector VE4 (which is the final vector VEF) corresponding to the updated third context vector VC3 and the updated third hidden encoder state vector VE3.

[0171] Alternatively, if the encoder comprises a single self-attentive block, the final encoder hidden state vector is the state vector generated by this self-attentive block of the encoder, which is indeed the last (and simultaneously the first) block of the encoder. The process further includes a step c) of classifying the biological sample by applying a decoder D of said neural network to the final encoder hidden state vector VEF. The decoder D therefore takes as input the final encoder hidden state vector VEF, as illustrated in Figure 3. The decoder D is also called the classification decoder.

[0172] As explained previously, biological sample classification refers to assigning one or more classes to the biological sample as a whole and / or assigning one or more classes to components of the biological sample. These components include, for example, cells, particles, debris, and molecules (e.g., proteins). Examples of components are also detailed previously in the sample example list.

[0173] Advantageously, the decoder includes at least one self-attentive block BAD and one classifier block BC as illustrated in Figure 3. Preferably, the self-attentive block(s) BAD of the decoder are positioned upstream of the classifier block.

[0174] Preferably, the first self-attentive block of the decoder takes as input the final context vector VCF and the final encoder hidden state vector VEF and outputs at least one hidden state vector of the decoder VD.

[0175] The decoder can include a single self-attentive block generating at output a single hidden state vector of the VD decoder, which is itself taken as input by the classifier.

[0176] Advantageously, as with encoder E, the invention is not limited to a specific number of self-attentive blocks in decoder D, particularly not to a maximum number of self-attentive blocks in decoder D. Even if decoder D comprises at least three self-attentive blocks, or even more than 12, or even 48, decoder D and the overall neural network will remain lightweight in terms of computing resources while still achieving excellent results in classifying a biological sample. Conversely, even though conventional language model decoder architectures generally include six self-attentive blocks (i.e., at least 12 with the encoder), these architectures are very resource-intensive and therefore energy-intensive.

[0177] The decoder's self-attentive block(s) are preferably similar to the encoder's self-attentive blocks. Preferably, each decoder's self-attentive block(s) comprise two multi-head attention layers and a forward-propagating network as illustrated in Figure 5. If the decoder comprises several self-attentive blocks, they are preferably arranged like those of the encoder, namely in sequence, with each block taking as input a context vector and a state vector generated by the preceding block.

[0178] Advantageously, at least the last self-attentive ADB block of the D decoder does not generate a context vector as its output, as illustrated in Figures 3 and 5. This block only generates a final hidden state vector of the D decoder. Figure 5 shows that the decoder's self-attentive ADB block does not include a layer that enables the generation of a context vector. More precisely, the decoder's self-attentive ADB block does not include the second residual addition and normalization layer AN2.

[0179] According to an unrepresented embodiment, the decoder's self-attentive block(s) are identical to the encoder's self-attentive block(s) and therefore each generate a context vector, including the last one. This context vector will, in practice, be discarded (i.e., not used in subsequent processing).

[0180] Advantageously, as illustrated in Figures 3 and 5, the decoder D includes a classifier block BC that generates a vector of classification scores C for a plurality of possible classes of the biological sample. For example, the sum of the scores could be 1, and for a sample, a score of 0.95 could be obtained for "presence of an abnormality" and a score of 0.05 for "absence of an abnormality." In this case, it would be concluded, based on the scores, that the biological sample presents an abnormality. As a second example, a component of the sample could obtain a score of 0.8 for "lymphocyte," 0.15 for "monocyte," and 0.05 for "neutrophil." In this case, it could be concluded, based on the scores, that this component of the sample can be classified as a "lymphocyte" with a high level of confidence. Thus, advantageously, step c) of classification consists of determining the type of at least one cell in the biological sample.

[0181] The classifier block could, for example, include a convolutional neural network (CNN), any dense network (DNN), a fully connected layer (FC) and / or a softmax layer.

[0182] The classifier advantageously takes as input the hidden state vector of the final decoder. This vector is representative of the data relating to the biological sample and has been processed by the encoder and decoder so that the classification can be reliably and efficiently implemented on this vector.

[0183] Alternatively, there may be no self-attentive block in the D decoder, and the decoder may even consist of the classifier, i.e., the latter directly takes the final vector as input. Advantageously, according to an unshown embodiment, the classifier block of the decoder comprises a dense network, a forward-propagating network, and / or a fully connected layer. This network and / or these layers reduce the dimensionality of the final decoder's hidden state vector.

[0184] Advantageously, the final encoder hidden state vector (FELV) taken as input by the decoder is a vector of dimensions NxD, where N preferably corresponds to the number of cells in the biological sample and D to the number of channels. Thus, for each cell, there are D channels, i.e., pieces of information about the cell. Advantageously, the decoder D, preferably the classifier block, comprises a dense network on dimension D so as to project the final encoder hidden state vector (FELV) onto a matrix of size NxA, where A is advantageously the number of classes. This NxA-sized vector preferably indicates, for each cell, a score associated with each class, representing the probability that the cell belongs to that class.The decoder D, preferably the classifier block, then includes a softmax type layer allowing to obtain a class (or some classes), from the set of A classes, associated with each cell.

[0185] Thus, according to a preferred embodiment, the classifier block comprises a dense network and a softmax layer arranged at the output of the dense network.

[0186] According to a preferred embodiment, the classes include the following classes: basophil (a type of white blood cell), erythroblast (Nucleated Red Blood Cell), lymphocyte (a type of white blood cell), monocyte (a type of white blood cell), blast (which is an immature hematopoietic cell), neutrophil (a type of white blood cell), immature granulocyte, eosinophil (a type of white blood cell), red blood cell (RBC), reticulocytes (a type of red blood cell), platelets, and platelet clump.

[0187] As explained previously, classification advantageously allows us to obtain a cell type for some cells, or even for each cell, in the biological sample. In other words, the process makes it possible to determine the type of each cell in the biological sample, thus allowing us to determine the composition of the sample (for example, a complete blood count).

[0188] In a specific embodiment, the classification process allows for the classification of several biological samples. Advantageously, these biological samples originate from the same individual. In this way, different biological samples can be classified, taking into account data from the various samples, thus enriching the classification process. For example, one might have a first biological sample containing the individual's white blood cells and a second biological sample containing the individual's red blood cells. In practice, as explained previously, an unlysed blood sample can be used to study red blood cells, and a lysed blood sample to study white blood cells.In this case, the first biological sample, which includes white blood cells, corresponds to the lysed blood sample and the second biological sample, which includes the individual's red blood cells, corresponds to the unlysed blood sample.

[0189] Preferably, the biological samples are different from each other, but they could be identical. Conversely, preferably, if the biological samples are identical, the relative data are different. For example, data relating to the first biological sample and data relating to the second biological sample: if the first and second biological samples are identical, their relative data are advantageously different, for example, because different information has been extracted from the biological sample (for example, data relating to red blood cells on the one hand and data relating to white blood cells on the other).

[0190] Thus, step a) of the process includes obtaining data for each biological sample. In the example above, step a) includes obtaining data for the first biological sample and the second biological sample.

[0191] Advantageously, at least one self-attentive block of the encoder takes as input a context vector relating to the contexts of the biological samples and, for each biological sample, an encoder hidden state vector representing the data relating to the biological sample. The self-attentive block outputs said updated context vector and each updated encoder hidden state vector.

[0192] In the example shown in Figure 4ter, the encoder's self-attentive block takes as input a context vector relating to the contexts of the first biological sample a and the second biological sample b, a first encoder hidden state vector VEa representing the data relating to the first biological sample, and a second encoder hidden state vector VEb representing the data relating to the second biological sample. The self-attentive block outputs the updated context vector, the first updated encoder hidden state vector VEamaj, and the second updated encoder hidden state vector VEbmaj.

[0193] Advantageously, the self-attentive block according to this "multiple samples" embodiment comprises the same elements as the self-attentive block described previously, but multiplied by the number of said plurality of biological samples, except for certain elements used to generate the context vector. "TJ lines" can be defined.

[0194] "Sample processing," each line being associated with a sample from the plurality of biological samples and comprising the elements necessary to generate an updated state vector for that sample. A "context processing line" can be defined to generate the updated context vector. For example, in Figure 4ter, there is a first processing line for a first sample comprising the blocks indicated with the letter a, a second processing line for a second sample comprising the blocks indicated with the letter b, and the context processing line in the middle. The processing lines share the same context line so that the context vector is updated based on all samples and the state vectors are updated based on the context of all samples.Thus, this embodiment does not simply replicate the structure of the self-attentive block for the analysis of a single sample. This embodiment cleverly allows for the consideration of data relating to the same individual, but from different samples, in order to classify each sample.

[0195] For this, as illustrated in figure 4ter, advantageously, the self-attentive block includes at least a first primary multi-head attention layer CMT1a, which takes as input the context vector VC and the state vector relative to the first biological sample VEa, and a first secondary multi-head attention layer CMT1b, which takes as input the context vector VC and the state vector relative to the second biological sample VEb.

[0196] Advantageously, the self-attentive block further includes a second primary multihead attention layer CMT2a and a second secondary multihead attention layer CMT2b. The second primary multihead attention layer CMT2a takes as input the state vector relating to the first biological sample VEa and a vector VI12a generated from a vector VI11a generated as output from the first primary multihead attention layer CMT1a, a vector VI11b generated as output from the first secondary multihead attention layer CMT1b and the context vector VC. The second secondary multi-head attention layer CMT2b takes as input the state vector relating to the second biological sample VEb and a vector VI12b generated from a vector VI11a generated as output from the first primary multi-head attention layer CMT1a, a vector VI11b generated as output from the first secondary multi-head attention layer CMT1b and the context vector VC.Therefore, vectors VI12a and VI12b are identical.

[0197] Advantageously, the vectors VI12a and VI12b generated from a vector generated as output from the first primary multihead attention layer, a vector generated as output from the first secondary multihead attention layer, and the context vector are vectors generated by the second residual addition and normalization layer AN2. The second residual addition and normalization layer AN2 therefore preferably takes as input the context vector, the vector generated as output from the first primary multihead attention layer, and the vector generated as output from the first secondary multihead attention layer. According to an embodiment not shown in Figure 4ter, the second residual addition and normalization layer AN2 generates the updated context vector VCmaj.Alternatively, as explained previously and illustrated in Figure 4ter, the self-attentive block can include a second propagating network before RPA2 and a fourth residual addition and normalization layer AN4, which generates the updated context vector VCmaj. In this case, the second residual addition and normalization layer AN2 generates a first intermediate context vector VC11, which is identical to vectors VI12a and VI12b.

[0198] The vector generated by the second residual addition and normalization layer AN2 therefore depends on information from the different state vectors VEa, VEb and the initial context vector VC. The updated context vector VCmaj obtained at the output of the self-attentive block thus expresses the contexts of the different biological samples.

[0199] Furthermore, the second multi-head attention layers CMT2a and CMT2b take as input a context vector (i.e., the updated context vector VCmaj or the first intermediate context vector VCI1 if the block includes an RPA2 network and an AN4 layer) that takes into account the first and second biological samples. The updated state vectors are therefore generated based on contextual information from the first and second biological samples.

[0200] Thus, the self-attentive block takes as many state vectors as there are samples analyzed and generates as many updated state vectors as there are samples analyzed, but only one context vector is taken as input and only one updated context vector is generated as output. The context vector includes contextual information for all the samples.

[0201] The resulting state vectors are more relevant for classifying the biological sample. Similarly, the context vector is more relevant, for example, for predicting an individual's state.

[0202] As described previously, the encoder can comprise a succession of several self-attentive blocks. According to this embodiment, each self-attentive block takes as input a state vector for each sample to be classified and a context vector.

[0203] The encoder thus allows us to obtain several final encoder hidden state vectors (one for each sample) VEFa, VEFb, and a single final context vector VCF as illustrated in Figure 3bis. Figure 3bis corresponds to Figure 3 adapted according to this embodiment with several samples; in Figure 3bis, two samples are shown (sample a and sample b).

[0204] According to this embodiment, step c) comprises the classification of the biological sample by applying at least one decoder of the neural network to at least one final encoder hidden state vector. For example, there may be only one decoder and therefore only one final encoder hidden state vector associated with a biological sample of interest, taken as input by a decoder. The other final encoder hidden state vectors will have been used to obtain a final encoder hidden state vector and a final context vector that are highly relevant for the classification of the biological sample of interest.

[0205] Alternatively, as illustrated in Figure 3bis, the encoder-decoder can comprise several decoders, each decoder being adapted to classify a distinct biological sample. Each decoder takes as input a hidden state vector from its respective final encoder and the final context vector. Each decoder generates a respective classification score vector C. As illustrated in Figure 3bis, there is a decoder Da relative to sample a and a decoder Db relative to sample b. According to this embodiment, it is therefore possible to classify several biological samples very reliably because the state vectors are determined based on contextual information from the other samples in a very clever way. Classifying white blood cells by taking into account data relating to red blood cells, and vice versa, allows for a much better classification of these two groups of blood cells.

[0206] This embodiment is also very useful for working with data from biological samples that have different formats. For example, red blood cell and white blood cell data are typically acquired separately by different sensors. For instance, the number of measurement channels differs for the two groups (4 for white blood cells and 5 for red blood cells), and the domains of these channels also differ. Therefore, it is difficult to create a single dataset for a single blood sample that includes both white and red blood cells. The present method allows for the reliable and efficient classification of both white and red blood cells through the use of both data.

[0207] The process preferably includes a step d) of predicting the state of an individual from whom the biological sample was taken by applying a second decoder of the neural network to the final context vector VCF. Recall that the final context vector is the context vector generated as output by the encoder. This step allows the final context vector to be used to diagnose the individual from whom the biological sample was taken. Indeed, the final context vector carries relevant information about the individual and the sample. This vector can therefore be taken as input by a second decoder D2, called the prediction decoder, trained to predict the individual's state. In terms of structure, what has been described concerning the first decoder D, i.e., the classification decoder, applies to the second decoder D2. The second decoder, i.e., the prediction decoder, can have the same structure as the first decoder, i.e.the classification decoder. The second decoder advantageously includes a predictor block adapted to generate as output a data point (for example a vector) representative of a state of the individual.

[0208] The second decoder can, for example, have a structure similar to that of the first decoder (i.e., the decoder classifying the encoder's final hidden state vector). In this case, the predictor block is similar to the classifier block of the first decoder, and this predictor block is therefore configured to classify the final context vector in such a way as to assign a class to the individual from whom the biological sample originates. This class advantageously corresponds to a state of the individual, for example, a pathology.

[0209] Figure 8 illustrates an encoder-decoder neural network comprising a first decoder D1 (classification decoder) which corresponds to the decoder enabling the classification of the biological sample and the second decoder D2 (prediction decoder) enabling the prediction of a state of the individual ET from which the biological sample comes.

[0210] With reference to Figure 6, a method for predicting the state of an individual from whom the biological sample is taken is proposed. The method comprises steps a') and b') which advantageously correspond to steps a) and b) of the classification method described above. The method further includes a step c') of predicting a state of the individual by applying a decoder, called the prediction decoder, of the neural network to the final context vector.

[0211] In other words, the process of predicting the state of an individual from which a biological sample is taken includes the following steps implemented by data processing means of a second piece of equipment:

[0212] a) obtaining data relating to the biological sample;

[0213] b) extraction of a final encoder hidden state vector representative of the data relating to the biological sample and of a final context vector relative to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relative to the context of the biological sample and a hidden encoder hidden state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder hidden state vector;

[0214] c') prediction of the state of the individual from which said biological sample is taken by applying a prediction decoder of said neural network to the final context vector VCF.

[0215] The term "state" can refer, for example, to a pathology such as a blood disorder. It can also refer to the origin of a disease. Indeed, the pathology may be confirmed, but one might wish to know its source.

[0216] Step a') is preferably identical to step a) of the biological sample classification process described above.

[0217] Step b') is preferably substantially identical to step b) of the biological sample classification process described above.

[0218] Step c') is preferably identical to step d) of the biological sample classification process described above.

[0219] Figure 7 illustrates a RED encoder-decoder neural network including a D2 prediction decoder to implement prediction. As shown, the encoder-decoder neural network can predict the state of an individual ET from which a biological sample is taken.

[0220] Advantageously, the prediction process includes a step d') of classifying the biological sample by applying a classification decoder (similar to the first decoder) of said neural network to the final encoder hidden state vector. This step d') is advantageously identical to step c) of the classification process described previously.

[0221] Advantageously, for this purpose, the neural network includes a classification decoder D1 which corresponds to the decoder enabling the classification of the biological sample and a prediction decoder D2 enabling the prediction of a state of the individual ET from which the biological sample comes as illustrated in figure 8.

[0222] Alternatively, the D2 prediction decoder may be different from the D1 classification decoder. For example, the D2 prediction decoder may include a clustering algorithm, with the final context vector being very well suited for clustering.

[0223] In one embodiment, the prediction process includes a step a0') of supervised learning of the neural network based on a training dataset of reference biological samples associated with reference states. During this step, in one embodiment, only the encoder may be trained, or both the encoder and the classification decoder may be trained. Then, the prediction decoder may be added to the output of the neural network encoder or may replace the classification decoder. The prediction decoder, equipped with the pre-trained encoder, can thus be trained.

[0224] Advantageously, the prediction method includes a step ai') of learning the first context vector. This learning step a1') is preferably identical to step a1) of the biological sample classification method described previously.

[0225] Furthermore, a system is proposed for predicting the state of an individual from whom a biological sample is taken. The system is advantageously similar to the classification system 100 already described. The system includes a second piece of equipment comprising data processing means 21 configured to implement:

[0226] - obtaining data relating to the biological sample;

[0227] - the extraction of a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector;

[0228] - the prediction of the state of the individual from whom said biological sample is taken by applying a prediction decoder of said neural network to the final context vector.

[0229] Product: computer program and readable storage media

[0230] Also proposed is a computer program product comprising code instructions for the execution (in particular on the data processing means 11, 21 of the first equipment 1 and / or the second equipment 2) of a method for classifying a biological sample, as well as computer-readable storage means (a memory 12, 22 of the first equipment 1 and / or the second equipment 2) on which this computer program product is located.

[0231] Also proposed is a computer program product comprising code instructions for the execution (in particular on the data processing means 11, 21 of the first equipment 1 and / or the second equipment 2) of a method for predicting the state of an individual from which a biological sample is taken, as well as computer-readable storage means (a memory 12, 22 of the first equipment 1 and / or the second equipment 2) on which this computer program product is found.

Claims

33 DEMANDS 1. Method for classifying a biological sample comprising the following steps implemented by data processing means (21) of equipment (2): a) obtaining data relating to the biological sample; b) extraction of a final encoder hidden state vector representative of the data relating to the biological sample and of a final context vector relative to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relative to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector; c) classification of the biological sample by applying a decoder of said neural network to the hidden state vector of the final encoder.

2. A method according to claim 1, wherein the encoder comprises a sequence of n>1 consecutive self-attentive blocks such that the i+1-th self-attentive block, 1 <i<n, prend en entrée des i+1-ème vecteur de contexte et i+1-ème vecteur d’état caché d’encodeur correspondant aux i-ème vecteur de contexte mis à jour et i-ème vecteur d’état caché d’encodeur mis à jour générés par le i-ème bloc auto-attentif, le vecteur d’état caché d’encodeur final étant le n+1 -ème vecteur d’état caché d’encodeur.

3. A method according to any one of claims 1 and 2, wherein the encoder comprises a first self-attentive block, the first self-attentive block taking as input the data relating to the biological sample as a first encoder hidden state vector and a first predetermined context vector.

4. Method according to claim 3, wherein the first context vector is learned with said neural network and / or related to data relating to an individual from which said biological sample is taken.

5. A method according to any one of claims 1 to 4, wherein the encoder comprises a sequence of at least three self-attentive blocks.

6. A method according to any one of claims 1 to 5, wherein at least one self-attentive block of the encoder comprises at least one multi-head attention layer and a forward-propagating neural network.34 7. A method according to any one of claims 1 to 6, wherein at least one self-attentive block of the encoder comprises at least two multi-head attention layers.

8. A method according to any one of claims 1 to 7, wherein the decoder comprises a classifier block that outputs a vector of classification scores for a plurality of possible classes of the biological sample and / or components of the biological sample.

9. Method according to claim 8, wherein the decoder comprises at least one self-attentive block upstream of said classifier block.

10. Method according to claim 9, wherein the self-attentive block of the decoder takes as input the final context vector and the final state vector and generates as output a hidden state vector of the decoder.

11. A method according to any one of claims 1 to 10, wherein the biological sample is or comprises at least one of the following liquids: blood, plasma, serum, urine, cerebrospinal fluid, synovial fluid, cell suspension, lymphatic fluid, aspiration, lavage or bronchoalveolar lavage fluid, and particle suspension in biological or industrial media.

12. A method according to any one of claims 1 to 11, wherein the classification step c) consists of determining the type of at least one cell of the biological sample.

13. A method according to any one of claims 1 to 12, wherein the obtaining step a) comprises the acquisition of data relating to the biological sample by means of at least one flow cytometer (3).

14. A method according to any one of claims 1 to 13, wherein the data relating to the biological sample include impedance measurements of the volume of the cells of the biological sample, optical absorption measurements of the cells of the biological sample, diffraction measurements at different angles generated by passing the cells of the biological sample through a laser beam and / or fluorescence measurements of nucleic acids made fluorescent by one or more fluorochromes previously brought into contact with the cells of the biological sample.

15. A method according to any one of claims 1 to 14, comprising a step d) of predicting a state of an individual from which said biological sample is taken by applying a second decoder of said neural network to the final context vector.

16. A method according to any one of claims 1 to 15, comprising a step a0) of supervised learning of the neural network on the basis of a training database relating to reference biological samples associated with reference classes.

17. A method according to any one of claims 1 to 16, wherein the biological sample is a first biological sample, step a) further comprises obtaining data relating to a second biological sample, the final encoder hidden state vector is a first final encoder hidden state vector representative of the data relating to the first biological sample, step b) comprises extracting a second final encoder hidden state vector representative of the data relating to the second biological sample, the final context vector is relative to the context of each biological sample, the encoder takes as input the data relating to each biological sample, the context vector taken as input by the self-attentive block is relative to the context of each biological sample,The encoder hidden state vector taken as input by the self-attentive block is representative of the data relating to the first biological sample, and the self-attentive block further takes as input an encoder hidden state vector representative of the data relating to the second biological sample, and the self-attentive block generates as output said updated context vector and each updated encoder hidden state vector. Step c) includes the classification of each biological sample by applying at least one decoder of said neural network to the final encoder hidden state vectors.

18. A method according to claim 17, wherein the encoder generates a single final context vector and the self-attentive block generates a single updated context vector.

19. A biological sample classification system (100) comprising at least one piece of equipment (2) including data processing means (21), characterized in that said data processing means are configured to implement: - obtaining data relating to the biological sample; - the extraction of a final encoder hidden state vector representative of the data relating to the biological sample and a final context vector relating to the context of the biological sample, by applying an encoder of an encoder-decoder type neural network taking as input the data relating to the biological sample, the network encoder comprising at least one self-attentive block taking as input a context vector relating to the context of the biological sample and a hidden encoder state vector representative of the data relating to the biological sample and generating as output said updated context vector and said updated hidden encoder state vector; - the classification of the biological sample by applying a decoder of said neural network to the hidden state vector of the final encoder.

20. System according to claim 19 further comprising at least one flow cytometer (3) configured to acquire biological sample data from the biological sample.

21. Product computer program comprising code instructions for the execution of a process according to any one of claims 1 to 18 of classification of a biological sample, when said program is executed on a computer.

22. Computer-readable storage means on which a computer program product includes code instructions for performing a process according to any one of claims 1 to 18 for classifying a biological sample.

Images