Method for generating an intermediate cellular state from anchoring states of a cellular state evolution

Abstract

The present invention relates to a method, preferably a computer-implemented method, and to an apparatus for generating at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state. The present invention also relates to a computer program for generating at least one intermediate cellular state and to a non-transitory computer readable data medium storing the computer program. Moreover, the present invention relates to a method and to an apparatus for training a model for generating at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state. The present invention also relates to a computer program for training a model for generating at least one intermediate cellular state and to a non-transitory computer readable data medium storing the computer program. Furthermore, the present invention relates to a method for diagnosing a disease and forecasting patients' future clinical outcome making use of the method for generating at least one intermediate cellular state. Moreover, the present invention relates to an in-vitro method for transforming the identity of a cell.

Description

[0001]Munich, 20 March 2024Our Ref.: JM 5225-02WO DREApplicant:Julius-Maximilians-Universität WürzburgSerial Number: Subsequent Application based on EP24165826.9Julius-Maximilians-Universität Würzburg Sanderring 2, 97070 Würzburg, Germany Method for generating an intermediate cellular state from anchoring states of a cellular state evolution TECHNICAL FIELD The present invention relates to a method, preferably a computer-implemented method, and to an apparatus for generating at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state. The pre- sent invention also relates to a computer program for generating at least one intermediatecellular state and to a non-transitory computer readable data medium storing the computerprogram. Moreover, the present invention relates to a method and to an apparatus for train-ing a model for generating at least one intermediate cellular state of a cellular state evolu-tion from an initial cellular state to a corresponding advanced cellular state. The presentinvention also relates to a computer program for training a model for generating at leastone intermediate cellular state and to a non-transitory computer readable data mediumstoring the computer program. Furthermore, the present invention relates to a method fordiagnosing a disease and forecasting patients’ future clinical outcome making use of themethod for generating at least one intermediate cellular state. Moreover, the present in-vention relates to an in-vitro method for transforming the identity of a cell.* *2022500025859607*058967* TECHNICAL BACKGROUND Diseases evolve and manifest due to accumulation of genetic and epigenetic dysregula- tions over our lifetime. The evolution of diseases is a continuous process and patients usu- ally undergo sequences of certain biological events from a “dysregulated state”, which is often undiscovered and not yet accompanied by symptoms, up to the complete manifesta- tion of the disease. These evolutions can be described as “disease-state” trajectories. Many methods leverage the power of single-cell RNA-sequencing for reconstructing cellstate trajectories. Such approaches generally require sufficient sampling of cell states cov-ering the entire trajectory. Since patients typically undergo treatment only after symptoms have already manifested, clinical samples covering intermediate disease states are gener-ally unavailable, limiting the ability to understand the underlying causal paths of diseaseevolution.Over the last decade, single cell genomics techniques have been developed which allowdescribing the regulatory information of several biological systems. Such technologies cap- ture high dimensional data of a cell in a certain state, time point and spatial location.Computational tools have been established to analyse and describe such snapshot singlecell genomics data to define cell types and marker genes associated with underlying cellu- lar states. However, to truly understand the underlying dynamics of a biological system, adensely sampled time-series strategy is needed, which can stitch together the discrete dis-continuous snapshots of cellular states to construct the underlying continuum process.1 Optimal-Transport approach (Waddington-OT)In 2019, the Waddington-OT method was published by (Schiebinger et al., 2019, Cell). In this work, single-cell RNA sequencing was performed for 315,000 cells in a cellular repro- gramming experiment, collected at half-day intervals over a period of 18 days. To infer how these probability distributions evolve over time, the Waddington-OT method uses the mathematical approach of optimal transport (OT) by utilizing scRNA-seq data col- lected across a time course. Even though the biological process in the experiment is densely sampled over time with half-day intervals, the Waddington-OT method cannot in- terpolate between cellular states that were sampled with longer time gaps.2 RNA velocity approachRNA velocity is a computational method that uses single-cell RNA sequencing data to pre- dict the future state of a cell. It is based on the observation that the ratio of unspliced to spliced RNA molecules within a cell provides information about the rate of transcription and splicing, and hence can be used to infer the direction and speed of a cell's differentiation trajectory. Although RNA velocity can be a useful tool for analysing short-term biological processes (up to a week), it is limited to scenarios of scRNA-seq data without temporal gaps of more than a few hours between cell states. RNA velocity can only predict dynamics on the time- scale of RNA turnover dynamics corresponding to several hours. Larger gaps in the sam- pling of cell states make it impossible for the method to reconstruct intermediate states. Therefore, if any state is missing, the method may not be able to recover that information. This limitation can be particularly challenging in real-world scenarios where some cell types or states are rare or difficult to obtain.Additionally, different implementations of RNA velocity have been shown to produce con-tradicting directionalities for the underlying biological process development (Gorin G et al., 2022, PLoS Comput Biol).3 Variational autoencoders (VAEs) approachAnother tool is the scGen method (Lotfollahi et al, 2019, Nat Methods), which uses varia-tional autoencoders and latent space vector arithmetic to predict perturbation outcomes. Although it was shown that this approach could be useful in modelling simple perturbation and infection responses (one effect at a time), it is limited when predicting complex pertur- bations or constructing complex intermediate states. This is because scGen, and similar approaches, rely on a simple assumption that arithmetic functions can resolve complex intermediate states in the latent space. Consequently, with the presently available tools, it is only possible to track the evolution of a (already diagnosed / manifested) disease by collecting and analysing samples over many years of the progression of the disease. However, for providing sufficient therapeutic options, it is important to interfere already dur- ing the evolution of the disease and with the presently available tools, a mechanistic un- derstanding of the underlying causes of such evolving biological systems is missing. More specifically the “intermediate states” are missing which have deep and intimate links to the underlying causal paths of diseases. SUMMARY OF THE INVENTIONIt is an object of the present invention to provide possibilities for analysing and preferablyinterfering with the evolution of a disease, particularly preferably before the disease has manifested (e.g. before first symptoms arise).According to the present invention, a method, preferably a computer-implemented method,for generating at least one intermediate cellular state of a cellular state evolution from aninitial cellular state to a corresponding advanced cellular state is proposed. The methodcomprises the steps of:- providing transcriptome data representing the initial cellular state and the corre-sponding advanced cellular state,- providing a model trained by a machine learning for generating the at least one in-termediate cellular state on the basis of the initial cellular state and the corresponding advanced cellular state, and- generating the at least one intermediate cellular state using the provided trainedmodel on the basis of the initial cellular state and the corresponding advanced cel-lular state.The invention is based on the recognition that with the known tools mentioned above, it isonly possible to analyse and describe a single snapshot in the evolution of a disease. Inaddition, current single cell analytical approaches are constrained by the observed data and are not capable of reconstructing additional, parallel cellular states or molecular events. Thus, the present tools do not provide (enough) information to gain access to the full con- tinuous process of the evolution of a disease. However, it is particularly the knowledge ofthe “intermediate states”, which provides crucial information, since the disease has not fullymanifested yet and the evolution may be slowed, redirected to a less severe direction, or in some cases even be stopped, e.g. by therapeutic treatment. Thus, with information onsuch “intermediate states”, it is possible to screen subjects before a disease has mani-fested, detect such “intermediate states” in a rather early stage of disease evolution, e.g.before the first symptoms arise, and to interfere with the evolution as described above.Consequently, there is a great need for finding mechanisms to analyse and to interfere with the evolution of a disease, particularly before the disease has manifested (e.g. before first symptoms arise). It was found that with the method according to the invention, the evolution of diseases can be modelled. As described in the example section, the obtained models were verified, wherein known intermediate states were hidden from the trained model, which however was able to reconstruct these states. Thus, with the method according to the invention, it is possible reconstruct the intermediate disease states, which was a major open challenge in the state of the art. It is a particular advantage of the method according to the invention that valuable mecha- nistic insights into the development and progression of diseases at high temporal resolution can be gained. Preferably, the transcriptome provided in the method is constructed by a DNA microarray,is constructed by a RNA sequencing or is constructed by a single-cell transcriptomics. Par-ticularly preferably, the transcriptome provided in the method is constructed by single cell RNA sequencing.Preferably, the transcriptome data representing the initial cellular state and the correspond-ing advanced cellular state are clinical data obtained experimentally from a patient.It is particularly preferred that the initial and advanced cellular state refer to- different phases of the cells in cell division,- different phases of the cells in cellular differentiation,- different phases of the cells in the transformation of cellular identity, particularly incellular trans-differentiation,- different phases of the cells in the transformation of a cell to a cancer cell, or- different phases of the cells in the transformation of a cancer cell to a metastaticcancer cell. Preferably, the term “different phases of the cells in cell division” refers to the G0, G1, S, G2 and M phase, wherein the term “different phases” describes that the initial cellular state is selected from the G0, G1, S, G2 and M phase and the advanced cellular state is selected from the G0, G1, S, G2 and M phase but is not the same phase as the initial cellular state. Preferably, the term “different phases of the cells in cell division” refers to the interphase, prophase, prometaphase, metaphase, anaphase, telophase and the cytokinesis phase, wherein the term “different phases” describes that the initial cellular state is selected from the interphase, prophase, prometaphase, metaphase, anaphase, telophase and the cyto- kinesisphase, and the advanced cellular state is selected from the interphase, prophase, prometaphase, metaphase, anaphase, telophase and the cytokinesis phase but is not the same phase as the initial cellular state. In case the initial and advanced cellular state refer to different phases of the cells in cell division, the provided transcriptome data representing the initial cellular state and the cor-responding advanced cellular state may be from cells in one phase (initial cellular state,e.g. G1 phase or interphase) and from cells in another phase (advanced cellular state, e.g. M phase or telophase). In this case, the at least one intermediate cellular state may be the S phase. Preferably, the term “cellular differentiation” describes the process in which a stem cell or progenitor cell changes to a cell of less potency, e.g. from totipotent to pluripotent, or mul-tipotent to unipotent. Preferably, the term “cellular differentiation” describes the process inwhich a stem cell or progenitor cell changes to a differentiated, i.e. mature, cell. During the process of cellular differentiation, several phases are passed. A skilled person knows howto discriminate between the different phases, for example by determining the activity ofparticular genes. Typically, specific genes are more / less active (measured e.g. on mRNA or protein level) in different phases. Determining the activity of such genes may be applied to determine the phase of cellular differentiation. Additionally or alternatively, a cell may also be assessed for its phase of cellular differentiation by morphological or phenotypical features, e.g. determined by microscopy.Preferably, the term “cellular identity” describes the type of a cell, preferably wherein thetype of a cell may be selected from the group consisting of Brunner's gland cell, Insulated goblet cell, Foveolar cell, Chief cell, Parietal cell, Pancreatic acinar cell, Paneth cell, TypeII pneumocyte, Club cell, Type I pneumocyte, Gall bladder epithelial cell, Centroacinar cell,Intestinal brush border cell (with microvilli), K cell, L cell, I cell, G cell, Enterochromaffin cell, Enterochromaffin-like cell, N cell, S cell, D cell, or M cell, Thyroid epithelial cell, Parafollic-ular cell, Parathyroid chief cell, Oxyphil cell, Alpha cell, Beta cell, Delta cell, Epsilon cell,PP cell (gamma cell), Salivary gland mucous cell, Salivary gland serous cell, Von Ebner's gland cell, Mammary gland cell, Lacrimal gland cell, Ceruminous gland cell, Eccrine sweat gland dark cell, Eccrine sweat gland clear cell, Apocrine sweat gland cell, Gland of Moll cell in eyelid, Sebaceous gland cell, Bowman's gland cell, Corticotropes, Gonadotropes, Lac- totropes, Melanotropes, Somatotropes, Thyrotropes, Magnocellular neurosecretory cells,, Parvocellular neurosecretory cells, Chromaffin cells (adrenal gland), Keratinocyte, Epider- mal basal cell (stem cell), Melanocyte, Trichocyte, Medullary hair shaft cell, Cortical hair shaft cell, Cuticular hair shaft cell, Huxley's layer hair root sheath cell, Henle's layer hair root sheath cell, Outer root sheath hair cell, Surface epithelial cell, basal cell (stem cell), Striated duct cell, Lactiferous duct cell, Ameloblast, Odontoblast, Cementoblast, Auditory inner hair cells of organ of Corti, Auditory outer hair cells of organ of Corti, Basal cells of olfactory epithelium, Cold-sensitive primary sensory neurons, Heat-sensitive primary sen- sory neurons, Merkel cells of epidermis, Olfactory receptor neurons, Pain-sensitive primary sensory neurons, Photoreceptor rod cells, Photoreceptor blue-sensitive cone cells of eye, Photoreceptor green-sensitive cone cells of eye, Photoreceptor red-sensitive cone cells of eye, Proprioceptive primary sensory neurons, Touch-sensitive primary sensory neurons, Chemoreceptor glomus cells of carotid body cell, Outer hair cells of vestibular system of ear, Inner hair cells of vestibular system of ear, Taste receptor cells of taste bud, Choliner-gic neurons, Adrenergic neural cells, Peptidergic neural cells, Inner pillar cells of organ ofCorti, Outer pillar cells of the organ of Corti, Inner phalangeal cells of organ of Corti, Outer phalangeal cells of organ of Corti, Border cells of organ of Corti, Hensen's cells of organ of Corti, Vestibular apparatus supporting cells, Taste bud supporting cells, Olfactory epithe- lium supporting cells, Olfactory ensheathing cells, Schwann cells, Satellite glial cells, En- teric glial cells, Basket cells, Cartwheel cells, Stellate cells, Golgi cells, Granule cells, Lu- garo cells, Unipolar brush cells, Martinotti cells, Chandelier cells, Cajal–Retzius cells, Dou- ble-bouquet cells, Neurogliaform cells, Retina horizontal cells, Starburst amacrine cells, Renshaw cells, Spindle neurons, Fork neurons, Place cells, Grid cells, Speed cells, Head direction cells, Betz cells, Boundary cells, Bushy cells, Purkinje cells, Medium spiny neu- rons, Astrocytes, Oligodendrocytes, Tanycytes, Pituicytes, Anterior lens epithelial cell, Crystallin-containing lens fiber cell, White fat cell, Brown fat cell, Liver lipocyte, Cells of the Zona glomerulosa, Cells of the Zona fasciculata, Cells of the Zona reticularis, Theca Interna cell, Corpus luteum cell, Granulosa lutein cells, Theca lutein cells, Leydig cell, Seminal vesicle cell, Prostate gland cell, Bulbourethral gland cell, Bartholin's gland cell, Gland of Littre cell, Uterus endometrium cell, Juxtaglomerular cell, Macula densa cell, Peripolar cell, Mesangial cell, Parietal epithelial cell, Podocyte, Proximal tubule brush border cell, Loop of Henle thin segment cell, Kidney distal tubule cell, Principal cell, Intercalated cell, Transi- tional epithelium, Duct cell, Efferent ducts cell, Epididymal principal cell, Epididymal basal cell, Endothelial cells, Planum semilunar epithelial cell of vestibular system of ear, Organ of Corti interdental epithelial cell, Loose connective tissue fibroblasts, Corneal fibroblasts, Tendon fibroblasts, Bone marrow reticular tissue fibroblasts, Other nonepithelial fibro- blasts, Hepatic stellate cell (Ito cell), Nucleus pulposus cell, Hyaline cartilage chondrocyte, Fibrocartilage chondrocyte, Elastic cartilage chondrocyte, Osteoblast / osteocyte, Osteopro- genitor cell, Hyalocyte, Stellate cell, Pancreatic stellate cell, Red skeletal muscle cell (slow twitch), White skeletal muscle cell (fast twitch), Intermediate skeletal muscle cell, Nuclear bag cell, Nuclear chain cell, Myosatellite cell (stem cell), Cardiac muscle cell, SA node cell, Purkinje fiber cell, Smooth muscle cell, Myoepithelial cell, Erythrocyte, Megakaryocyte, Platelets if considered distinct cells, currently there's debate on the subject., Monocyte, Connective tissue macrophage, Epidermal Langerhans cell, Osteoclast, Dendritic cell, Mi- croglial cell, Neutrophil granulocyte, Eosinophil granulocyte and precursors, Basophil gran- ulocyte and precursors, Mast cell, Helper T cell, Regulatory T cell, Cytotoxic T cell, Natural killer T cell, B cell, Plasma cell, Natural killer cell, Hematopoietic stem cells and committed progenitors for the blood and immune system, Oogonium / Oocyte, Spermatid, Spermato- cyte, Spermatogonium cell, Spermatozoon, Granulosa cell, Sertoli cell, Epithelial reticular cell and Interstitial kidney cells. In case the initial and advanced cellular state refer to different phases of the cells in cellular differentiation, the provided transcriptome data representing the initial cellular state and the corresponding advanced cellular state may be from cells of particular potency (initial cellu- lar state, e.g. multipotency) and from cells of less potency than in the initial cellular state (advanced cellular state, e.g. oligopotency or unipotency). In this case, the at least one intermediate cellular state may be a cell of less or equal potency than in the initial cellular state but of higher or equal potency than in the advanced cellular state. Preferably, the provided transcriptome data representing the initial cellular state and the corresponding advanced cellular state may be from cells of different potency during hem- atopoiesis. Preferably, the term “cellular transdifferentiation” describes the process in which one ma- ture somatic cell is transformed into another mature somatic cell without undergoing anintermediate pluripotent state or progenitor cell type. For example, alpha-cells in the pan-creas may transdifferentiate to beta-cells in the pancreas, as reported in humans. Further, the cellular transdifferentiation may occur spontaneously or may be induced. A skilled per-son knows how to discriminate cell types, for example by determining the activity of partic-ular genes and / or by assessing morphological or phenotypical features, e.g. determined by microscopy. In case the initial and advanced cellular state refer to different phases of the cells in the transformation of cellular identity, the provided transcriptome data representing the initial cellular state and the corresponding advanced cellular state may be from particular mature somatic cells (initial cellular state, e.g. epithelial cells or fibroblasts) and from particularother mature somatic cells (advanced cellular state, e.g. mesenchymal cells or cardiomyo-cytes). In this case, the at least one intermediate cellular state may be a cell in the process of transformation from epithelial identity to mesenchymal identity or from fibroblasts to car- diomyocytes. Preferably, the term “transformation of a cell to a cancer cell” refers to carcinogenesis. Preferably, the term describes a process, in which a cell undergoes changes at the cellular,genetic, and / or epigenetic level and / or shows abnormal cell division. A skilled personknows how to identify a cancer cell, for example by determining the activity of particulargenes and / or by assessing morphological or phenotypical features, e.g. determined by mi- croscopy. In case the initial and advanced cellular state refer to different phases of the cells in the transformation of a cell to a cancer cell, the provided transcriptome data representing the initial cellular state and the corresponding advanced cellular state may be from healthy cells (initial cellular state, e.g. plasma cells) and from cancer cells (advanced cellular state, e.g. myeloma cells). In this case, the at least one intermediate cellular state may be a cell of higher malignancy than the initial cellular state but lower malignancy than the advancedcellular state, for example MGUS (monoclonal gammopathy of undetermined significance)cells or smoldering myeloma cells. Preferably, the term “transformation of a cancer cell to a metastatic cancer cell” describes a process, in which a cancer cell (or a cancerous tumour) gains the ability to spread from the primary cancer site to a secondary cancer site. A skilled person knows how to identify a metastatic cancer cell, for example by determining the cell type of the cancer cell and assessing whether this cell type is typically found at the cancer site, where it was located. In case the initial and advanced cellular state refer to different phases of the cells in the transformation of a cancer cell to a metastatic cancer cell, the provided transcriptome data representing the initial cellular state and the corresponding advanced cellular state may befrom non-metastatic cancer cells (initial cellular state), and from metastatic cancer cells(advanced cellular state). In this case, the at least one intermediate cellular state may be a cell in the process of transformation from a cancer cell to a metastatic cancer cell.Additionally or alternatively, based on the at least one intermediate cellular state, a causaltrajectory for of the cellular state evolution from the initial cellular state to the corresponding advanced cellular state is generated. The model trained by machine learning may be a classification model such as a neural network that is trained for generating at least one intermediate cellular state of a cellularstate evolution from an initial cellular state to a corresponding advanced cellular state. Theneural network may be trained using transcriptome training data. In particular, the model trained by machine learning may have a deep reinforcement learn- ing architecture where a neural network is trained for generating at least one intermediate cellular state of a cellular state evolution trajectory from an initial cellular state to a corre- sponding advanced cellular state. The neural network may be trained using single cell tran- scriptomic data. The deep reinforcement learning architecture preferably implements a pol-icy network, which is parametrized by a multilayer perceptron (MLP) and which may beoptimized via gradient ascent to maximize the reward objective. In addition to the standard reward objective, the policy network may be configured to optimize a maximum entropy objective to enable the model to alternate between exploration and exploitation phases during reconstruction.Optionally, the model is trained by a deep reinforcement learning technique. For example,the trained model may be trained, e.g., by deep reinforcement learning, as described belowusing the method or the apparatus for training a model by machine learning for generatingat least one intermediate cellular state. The present invention also relates to a computer program configured for generating at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state, the computer program including instructions for ex-ecuting the steps of the method for generating at least one intermediate cellular state de-scribed before, when run on a computer. The present invention also relates to a non-tran-sitory computer readable data medium storing the computer program for generating at leastone intermediate cellular state. The present invention also relates to an apparatus for generating at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a correspondingadvanced cellular state. The apparatus comprises a transcriptome data providing unit, atrained model providing unit and an intermediate cellular state generating unit. The tran-scriptome data providing unit is configured for providing transcriptome data representingthe initial cellular state and the corresponding advanced cellular state. The trained modelproviding unit is configured for providing a model trained by a machine learning for gener-ating the at least one intermediate cellular state on the basis of the initial cellular state andthe corresponding advanced cellular state. The intermediate cellular state generating unitis configured for generating the at least one intermediate cellular state using the providedtrained model on the basis of the initial cellular state and the corresponding advanced cel-lular state. The apparatus can be used for carrying out the method for generating at leastone intermediate cellular state described before.According to the present invention, also a method for training a model by machine learningis proposed. The method comprises the steps of:i) providing a plurality of transcriptome training data, each representing an initial cellu- lar state and a corresponding advanced cellular state of a cellular state evolution,ii) providing a model that can be trained by machine learning, andiii) training the model by machine learning using the transcriptome training data suchthat the trained model is configured to receive transcriptome data representing aninitial cellular state and a corresponding advanced cellular state of a cellular state evolution and for generating at least one intermediate cellular state of the cellular state evolution from the initial cellular state to the corresponding advanced cellular state based on the received transcriptome data, preferably via a generative model.Preferably, the transcriptome training data, each representing the initial cellular state andthe corresponding advanced cellular state, respectively, are clinical data obtained experi-mentally from one or more patients. In the training method according to the invention, theto-be-trained model may generate actions in gene expression space and may learn sto-chastic policies to reconstruct trajectories connecting two distant anchoring cellular states,i.e., an initial cellular state to the corresponding advanced cellular state of a cellular stateevolution. It is preferred in the training method according to the invention that step iii) comprises at least one of the steps:iii.1) generating transcriptome data of further cellular states by sampling from a multivar-iate Gaussian distribution to learn combinations of action parameters comprising a control mean and a standard deviation of the generated transcriptome data of tran- scriptome data of further cellular states, e.g., using a Gen-unit,iii.2) classifying the generated transcriptome data, e.g., from the Gen-unit, of further cel-lular states based on the learned combinations of action parameters to identify those combinations of action parameters that fulfil a predefined reward criterion, e.g., using a Do unit,iii.3) calculating a distance of the generated transcriptome data of further cellular statesto the transcriptome data of advanced cellular states of the plurality of transcriptome training data, e.g., using a distance calculating unit,iii.4) rewarding the generated transcriptome data of further cellular states with referenceto a predefined reward criterion, e.g., using a rewarding unit, andiii.5) selecting or emitting a combination of action parameters to generate transcriptomedata of further cellular states, which have the highest probability of an increasing reward from one generated transcriptome data to the next transcriptome data within the sequence, as the transcriptome data of a plurality of sequential intermediate cellular states, e.g., usingan action parameters selecting unit. For example, the action parameters selecting unit maycomprise a policy network. In particular, step iii.5) may be implemented in that the policynetwork emits a combination of action parameters to generate transcriptome data of furthercellular states, which have the highest probability of an increasing reward from one gener- ated transcriptome data to the next transcriptome data within the sequence, as the tran- scriptome data of a plurality of sequential intermediate cellular states, e.g., using an action parameters from the trained policy network. It is further preferred in the method according to the invention that the advanced cellular state in step i) is divided into a plurality of groups of different severity of the advanced cellular state, preferably wherein the transcriptome data of a plurality of sequential intermediate cellular states selected and generated in step iii.5) is assigned to one of the groups of different severity of the advanced cellular state. The groups of different severity may be assigned in any way, for example the different severities are ranked on a numerical scale. It is further preferred in the method according to the invention that step iii) further comprisesiii.6) repeating the steps iii.1) to iii.5) to generate a plurality of transcriptome data of aplurality of sequential intermediate cellular states. According to the present invention, also an apparatus for training a model by machinelearning is proposed. The apparatus comprises a transcriptome training data providing unit,a model-to-be-trained providing unit and a model training unit. The transcriptome trainingdata providing unit being configured for providing a plurality of transcriptome training data,each representing an initial cellular state and a corresponding advanced cellular state of acellular state evolution. The model-to-be-trained providing unit being configured for provid-ing a model that can be trained by machine learning. The model training unit being config- ured for training the model by machine learning using the transcriptome training data such that the trained model is configured to receive transcriptome data representing an initial cellular state and a corresponding advanced cellular state of a cellular state evolution and for generating at least one intermediate cellular state of the cellular state evolution from the initial cellular state to the corresponding advanced cellular state based on the receivedtranscriptome data. With the training apparatus it is possible to carry out the method fortraining a model by machine learning for generating at least one intermediate cellular statedescribed herein.Preferably, the model training unit comprises at least one of the following:- a Gen-unit that is configured for generating transcriptome data of further cellularstates by sampling from a multivariate Gaussian distribution to learn combinations of action parameters comprising a control mean and a standard deviation of the gen- erated transcriptome data of transcriptome data of further cellular states,- a Do-unit that is configured for applying the actions parameters to the Gen-unit togenerate novel transcriptome data of further cellular states.- a distance calculating unit that is configured for calculating a distance of the gener-ated transcriptome data of further cellular states to the transcriptome data of ad- vanced cellular states of the plurality of transcriptome training data,- a rewarding unit that is configured for rewarding the generated transcriptome data offurther cellular states with reference to a predefined reward criterion, and- an action parameters selecting unit that is configured for selecting a combination ofaction parameters to generate transcriptome data of further cellular states, which have the highest probability of an increasing reward from one generated transcrip- tome data to the next transcriptome data within the sequence, as the transcriptome data of a plurality of sequential intermediate cellular states.In particular, the action parameters selecting unit may comprise a policy network that istrained for selecting a combination of action parameters to generate transcriptome data of further cellular states, which have the highest probability of an increasing reward from one generated transcriptome data to the next transcriptome data within the sequence, as the transcriptome data of a plurality of sequential intermediate cellular states.A possible exemplary implementation of the training apparatus will be explained in the fol-lowing:For example, it is possible that a given cellular state may be represented as ^ ∈ ℝ^,, withcell index ^ = ^, ... , ^ and gene index ^ = ^, ... ,^. The initial cellular state may be denotedas ^^ ∈ ℝ^, and the corresponding advanced cellular state of interest as ^^ ∈ ℝ^,^ . It ispossible to generate a sequence of intermediate cellular states ^^^^ ∈ ℝ^, representing acausal path connecting two anchoring cellular states (^^ , ^^).To train a model in order to be able to find the missing cellular state ^^^^, the trainingapparatus may comprise the following components: an action generator and controller(AGC) Unit, an observation space (O), a reward function (R), and a policy network (^^).The environment may be designed to be represented as an infinite-horizon Markov deci-sion process (MDP). The MDP may be defined by a tuple (^, ^,, ^), where ^ represents acomplete description of the environment state and ^ represents a continuous action spacein the environment. The transition probability ^ represents the probability density of thenext state ^^+^ ∈ ^ given the current state of the environment ^^ ∈ ^ and the action ^^ ∈^.A trajectory ^ is defined as a sequence of states and actions in the environment through ^steps: ^^ = (^^, ^^, ^^, ^^,…, ^^, ^^)The transition from one cellular state to another is governed by a stochastic state transition function: ^^+^ ∼ ^(⋅∣ ^^, ^^)The actions are generated by the model according to the policy network ^^ which is para-metrized by : ^^ ∼ ^^(⋅∣ ^^)The environment may release reward ^^ after each transition which is dependent on thecurrent state of the environment, the actions taken, and the environment’s next state given the reward function ^: ^^ = ^ (^, ^, ^^+^)Preferably, it is possible to find the parameters ^ of the optimum policy ^^ (^^ ∣ ^^) bymaximizing the expected total reward (^) summed over the trajectory, where the horizon parameter ^ represents the number of steps on each trajectory:The reward objective may be extended to an infinite-horizon discounted return (^) over atrajectory, with a discounting factor ^ ∈ (^, ^): where ^ < 1 induces a discount effect on the overall return.Actions Generator and Controller (AGC) unit:The AGC may be part of the model training unknit and may be composed of two sub-units,the Gen-Unit mentioned with reference to step iii.1 of the training method according to theinvention and the Do-Unit mentioned with reference to step iii.2 of the training method ac-cording to the invention. I. Gen-Unit:The Gen-Unit that preferably is part of the model training unit may be configured to gener-ate cellular states in the gene expression space.The generation process may start with concatenating ^^ and ^^ into a matrix ^ ∈ ℝ^, :^ = [^^ ^^]The M matrix may be pushed to a standard data pre-processing unit to perform log-nor-malization (using size factor 104molecules for each cell) and to scale the gene expressionvalues for each gene (z-score transformation). The AGC unit may transpose the pre-pro-cessed ^ and calculates the gene-gene covariance matrix Σ ∈ ℝ^,:Σ = ^^^ (MT) To extract the orthogonal axes of gene expression variability characteristic of the generegulatory networks governing ^^ and ^^, the AGC may perform singular value decompo-sition (SVD) on ^, Σ= U Λ ^^where a) U ∈ ℝ ., / represents the left-singular vectors of Σ with f decomposed factors,b) ^ ∈ ℝ .,. represents the singular-value diagonal matrix, andc) ^^ ∈ ℝ / ,. represents the right-singular vectors of Σ.The Gen-Unit may be configured to generate new cellular states ^^^^ by sampling from amultivariate Gaussian distribution ^ ∈ ℝ xgen,f,^ ~ ^ (^^^^, ^^^^)where the two parameters ^^^^ and ^^^^ control mean and standard deviation of thegenerated cellular state ^^^^ and each dimension implements an action on one of thefactors. The chosen dimension ^^^^ controls the number of generated cells. Since U describes the direction of ^ matrix’s underlying maximum action, ^ may be intro-duced as a scaling parameter for ^, to control the action magnitude. Finally, ^^ representsthe mean gene expression of the two cellular states ^, which is “max” normalized in ^^^^.II. Do-unit:The Do-unit that preferably is part of the model training unit may optionally be introducedfor controlling the generated cellular states in the gene expression space via ^^ operationon the (^^^^, ^^^^ and ^ ) Gen-Unit parameters, where ^^^, ^^^ or ^^^ represent possible outcomes of the Do-unit, i.e., Gaussian noise,overexpression, or knockdown, respectively, for the newly generated cellular state ^^^^.In addition, ^^^^^^ may be sued to perturb the singular values by interventions throughtuning the scaling parameter ^. It is desirable to find the controller parameters and ) of optimum policy. Ulti-mately, the optimum policy may guide the model-to-be-trained to provide a sequence ofactions and to reconstruct ^^^^ in ^ steps for a given ^^.In the following, the observation space is introduced and it is outlined how the environment is observed. The Observation Space:The observation space preferably is a cellular state manifold representing all possible con-figurations of a cell. Certain regions of interest within this space can be visited and recon- structed. Such regions of the observation space are termed Reconstruction Field (RF). To enable gaining a sense of directionality and distances in the RF, one or more of a state observation (^), a cellular states distance (CSD), learning bound scores (LBS), and a cel-lular state loss (CSL) may be used.I. Single Observation (^):A given cellular state ^ may be summarized into a single observation ^ according to thefollowing equation, ^= ^^^^^^^ (^^^ (^))where the median gene activity (MGA) of a given cellular state ^ is calculated and rescaledvia SoftMax function to j-dimensional output ranging from [0,1] and is summing to 1. II. Cellular States Distance (CSD):First, a Cellular States Distance (CSD) as mentioned in the iii.3 of the training methodbetween the initial cellular state observation ^^ and the target cellular observation ^^ maybe calculatedThe ^^^ is composed of two terms:a) Cosine similarity term: 1 - Scos (Oi || Of)b) Kullback-Leibler divergence term: DKL(Oi || Of)The parameter ^ defines the importance of the ^^^ terms to the ^^^.III. Learning bound scores (LBS)The LBS preferably define the boundaries of the cellular state reconstruction field, com-prising one or more of the following cellular bound scores:^^^^^ = ^^^ × ^,^^^^^ = ^^^ × ^,^^^^^ = ^^^ × ^,^^^^^ = ^where ^ controls the minimum reward bound ^^^^^, ^ controls the maximum rewardbound ^^^^^, ^ controls the maximum trajectory bound ^^^^^ and ^ controls the min-imum trajectory bound ^^^^^. IV. Cellular State loss (CSL):The cellular state loss (CSL) may be employed to enable the model to learn the distancesbetween generated and target cellular state observations: The CSL may be composed of one or more of the following terms:a) Cosine similarity term: 1 - Scos (Oi || Of)b) Kullback-Leibler divergence term: DKL(Oi || Of)c) Entropy regularization term^^^^ and ^^ represent a generated observation and a target state observation, respec-tively. The ^ and ^ parameters define the weights of the ^^^ and entropy regulation terms.Reward Function:The reward function of the environment as mentioned in step iii.4 of the training methodaccording to the present invention may depend on the current cellular state observationsand the ^^^ incurred by the newly generated cellular state,^ (^, ^, ^^+1) = ^ + (^ − ^^^) ^where ^ is defining the reward signal and ^ is the reward coefficient. The model may beconditioned get positive or negative rewards given the ^^^: If the ^^^ is between ^^^^^ ^^^ ^^^^^, such behavior is reinforced, and a positive ^^is received. In contrast, if the reward bounds are exceeded, a negative ^^ is received.Furthermore, a termination condition ^ with a done signal may be introduced to the envi-ronment, which depends on ^^^^^ and ^^^^^ as well as the number of steps, for a givencellular state trajectory ^^. ^ starts at ^ = (^ − 1) and is decreased by one at each step: Policy Network: The policy ^^ preferably is parametrized by a neural network and optimized via gradientascent, where the ^ parameter controls the learning rate: The gradient may be used to increase the probability of paths with positive ^^ values anddecrease the probability of paths with negative ^^ values. Monte Carlo sampling of ^ tra-jectories may be performed and the policy ^^ may be run. Then, a multilayer perceptron(MLP) may be fit to estimate the return and to improve the policy iteratively: In addition to optimizing for the standard maximum reward objective, the Soft-Actor-Criticmaximum entropy objective may be considered: The entropy term ^ controls the stochastic behavior of model policies ^^. The maximumentropy objective enables the model to alternate between exploration and exploitationphases during training: The model-to-be-trained may get rewards proportional to the entropy term ^ in the policy^^, and the ^ parameter determines the relative importance of the ^ given the reward ateach time step. Higher α corresponds to more exploration and lower α corresponds to moreexploitation. Optimal Causal Paths (OCP): To define the causal paths given the predicted actions, the optimum causal paths (OCP)concept may be employed. The model to-be-trained may generate actions from the trainedpolicy network ^^. The predicted action parameters ^^^^, ^^^^ and ^ values may be pushedto the Gen-Unit. The ^^^ may be calculated for each path ^^ and a matrix with ^ sampledrealizations and cellular state losses ^k,n may be constructed accordingly:^^,^ = ^^^ (^, ^^)The trained model may generate ^ realizations representing generated paths with ^ stepswhere each path connects the same two anchoring cellular states (^ , ^^). Ultimately, todefine the optimal paths ^^^^, the quantiles ^^ maybe calculated for the ^^, matrix,^^^^(^^^^, ^^) = ^^ (^^,^, ^)where ^ <= ^ <= ^ to stratify paths according to the ^^^ distribution:The present invention also relates to a computer program configured for training a model by machine learning to generate at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state, the com- puter program including instructions for executing the steps of the method for training amodel by machine learning, when run on a computer. The present invention also relates toa non-transitory computer readable data medium storing the computer program for traininga model by machine learning. The present invention further relates to a method for diagnosing a disease, comprisinga) providing a sample of a subject, wherein the sample comprises cells associated withthe disease to be diagnosed,b) obtaining the transcriptome data of one or more cells from the sample obtained instep a), andc) comparing the transcriptome data obtained in step b) with the reconstructed tran-scriptome data of the plurality of sequential intermediate cellular states as generated in the method for generating at least one intermediate cellular state according to thepresent invention, wherein the disease is diagnosed if the transcriptome data obtained in step b) corresponds to one of the transcriptome data of the plurality of sequential intermediate cellular states.Preferably, the term “disease” refers to any disease. Preferably, the term “disease” refersto a disease selected from the group consisting of autoimmune diseases, cancerous dis-eases, genetic diseases, infectious diseases, inflammatory diseases, neurodegenerativediseases. Preferably, the term “cells associated with the disease to be diagnosed” describes cells of a cell type, which is typically associated with the disease to be diagnosed. It is a particular advantage of the present invention that a plurality of sequential intermediate cellular states is generated, which are associated with a state in the evolution of a disease.Thus, in case of a subject, which does not (yet) suffer from symptoms of a disease, thetranscriptome data corresponds to the transcriptome data of an intermediate cellular state as described herein, a diagnosis of the disease may be made. The diagnosis may therefore be made before (even long before) first symptoms arise, for example in case the sample of the subject is provided in step a) during a routine check-up or during an examination due to the presence of one or more risk factors for a particular disease. For almost all diseases known today, the chances for preventing or lessening the disease or for slowing down its evolution and / or progression are dramatically increased in case a diagnosis is made as early as possible. With the method according to the invention, it is possible to diagnose a disease (long) before first symptoms arise and thus long time beforea subject usually asks for medical examination.It is further advantageous that for the diagnosis of the disease, only a sample of a subjectneeds to be provided. After providing a sample, no action or interaction of the subject is required for making the diagnosis. The term “corresponds” as used in the method according to the invention describes an overlap or a complete match. Preferably, the term “corresponds” describes an identity of the transcriptomes of at least 95 %, preferably at least 96 %, preferably at least 97 %, preferably at least 98 %, preferably at least 99 %, preferably at least 99.5 %, preferably 100 %. Preferably, the term “identity of the transcriptomes” describes that at least 95 %, preferably at least 96 %, preferably at least 97 %, preferably at least 98 %, preferably at least 99 %, preferably at least 99.5 %, preferably 100 % of the detected transcripts of one transcriptome are present in the other transcriptome. Preferably, the respective quantity of these tran- scripts, which are present in both transcriptomes, does not differ more than 10 %, prefera- bly not more than 7.5 %, preferably not more than 5 %, preferably not more than 2.5 %, preferably wherein the higher quantity of a transcript is set to 100 % for the comparison. Preferably, the term “transcriptome” refers to all RNA in a sample, particularly preferably only to the mRNA in a sample. It is further preferred in the method according to the invention that in step c) the transcrip- tome data obtained in step b) is further compared with the transcriptome data of a plurality of cells in an advanced cellular state as provided in step i) of the method according to the invention, and / or with the transcriptome data of a plurality of cells in an initial cellular state as provided in step i) of the method according to the invention, wherein the disease is diagnosed if the transcriptome data obtained in step b) corresponds to one of the transcriptome data of the plurality of sequential intermediate cellular states or to one of the transcriptome data of the advanced cellular state, and / or wherein the disease is not diagnosed if the transcriptome data obtained in step b) corre- sponds to one of the transcriptome data of the initial cellular state. Thus, in case the transcriptome of a sample corresponds to the initial cellular state, it is preferably assumed that the disease does not (yet) evolve in the subject of which the sam- ple was provided. In contrast, in case the transcriptome of a sample corresponds to the advanced cellular state, it is preferably assumed that the disease has already evolved in the subject of which the sample was provided. It is preferred in the method for diagnosing a disease according to the invention that the disease is cancer, preferably metastasizing cancer. Preferably, the term “metastasizing cancer” refers to cancer, wherein one or more cancercell(s) has / have gained the ability to spread from the primary cancer site to a secondarycancer site. A skilled person knows how to identify metastasizing cancer, for example bydetermining the cell type of a cancer cell and assessing whether this cell type is typicallyfound at the cancer site, where it was located. What was said with regard to the method for providing at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advanced cellular state applies accordingly to the method for diagnosing a disease. The present invention further relates to a method for predicting disease progression, com- prisingA) providing a sample of a subject, wherein the sample comprises cells associated withthe disease to be diagnosed,B) obtaining the transcriptome data of one or more cells from the sample obtained instep A), andC) comparing the transcriptome data obtained in step B) with the reconstructed tran-scriptome data of the plurality of sequential intermediate cellular states as generated in the method for generating at least one intermediate cellular state according to the present invention, to identify one of the transcriptome data of the plurality of sequential intermediate cellular state as generated in the method according to the invention, to which the transcriptome data obtained in step B) corresponds wherein the prediction of the disease is based on the group of the severity of the disease, to which the transcriptome data identified in step C) is assigned. In addition to the advantages described above, it is a particular advantage of the present invention that groups of different severity of the disease may be formed. In step C) of the method for predicting disease progression according to the present invention, it is thus possible to assign the transcriptome in a sample to a particular group, which is assigned to a particular disease severity. The groups of different severity may be assigned in any way, for example the different severities are ranked on a numerical scale. Preferably, the term “predicting disease progression” refers to the severity of the disease. The term “corresponds” as used in the method according to the invention describes anoverlap or a complete match. Preferably, the term “corresponds” describes an identity ofthe transcriptomes of at least 95 %, preferably at least 96 %, preferably at least 97 %, preferably at least 98 %, preferably at least 99 %, preferably at least 99.5 %, preferably 100 %. Preferably, the term “identity of the transcriptomes” describes that at least 95 %, preferably at least 96 %, preferably at least 97 %, preferably at least 98 %, preferably at least 99 %, preferably at least 99.5 %, preferably 100 % of the detected transcripts of one transcriptome are present in the other transcriptome. Preferably, the respective quantity of these tran- scripts, which are present in both transcriptomes, does not differ more than 10 %, prefera- bly not more than 7.5 %, preferably not more than 5 %, preferably not more than 2.5 %, preferably wherein the higher quantity of a transcript is set to 100 % for the comparison. Preferably, the term “transcriptome” refers to all RNA in a sample, particularly preferably only to the mRNA in a sample. What was said with regard to the method for providing at least one intermediate cellular state of a cellular state evolution from an initial cellular state to a corresponding advancedcellular state or with regard to the method for diagnosing a disease applies accordingly tothe method for predicting disease progression. It is preferred in the method for predicting disease progression according to the invention that the disease is cancer, preferably metastasizing cancer. Preferably, the term “metastasizing cancer” refers to cancer, wherein one or more cancer cell(s) has / have gained the ability to spread from the primary cancer site to a secondary cancer site. A skilled person knows how to determine metastasizing cancer, for example by determining the cell type of a cancer cell and assessing whether this cell type is typically found at the cancer site, where it was located.The present invention further relates to an in-vitro method for transforming the identity of acell, the method comprising1) generating at least one intermediate cellular state of a cellular state evolution froman initial cellular state to a corresponding advanced cellular state, with a method according to the invention, wherein the initial and advanced cellular state refer to different phases of the cells in the transformation of cellular identity, particularly in cellular transdifferentiation,2) analysing the at least one intermediate cellular state generated in step 1) to identifya cellular signalling pathway involved in the transformation of the cellular identity,3) modulating, preferably activating or inactivating, the cellular signalling pathway iden-tified in step 2) to promote the transformation of the identity of a cell. Preferably, for identifying a cellular signalling pathway involved in the transformation of thecellular identity, the generated intermediate cellular state is compared with the initial cellularstate. Statistically significant changes in e.g. the protein level or the phosphorylation of a protein between the generated intermediate cellular state and the initial cellular state are preferably considered as being involved in the transformation of the cellular identity. Thus, a cellular signalling pathway including the e.g. synthesis or phosphorylation of said proteinis considered as a cellular signalling pathway involved in the transformation of the cellularidentity. Preferably, the method for transforming the identity of a cell is a non-therapeutic method. The method for transforming the identity of a cell is an in-vitro method. BRIEF DESCRIPTION OF THE DRAWINGSFig.1A: shows an overview of a model environment architecture where blue dots (1)represent the initial cellular state, orange dots (2) represent the generated cellular statesafter the model has taken actions, and the red dots (3) represent the corresponding ad-vanced cellular states;Fig.1B: shows a schematic representation of an observation space of healthy and dis-ease cellular states (right), where each dot represents a real data point collected from ex-perimental and clinical data. By zooming into the data of the two states in the small box, itis focused on the reconstruction field of interest given the initial and corresponding ad-vanced states of interest. The model, through trial and error, tries to generate cellular statesand trajectories in the reconstruction area, approximating the true cellular state manifold;Figs. 1C,D: show a schematic representation of the learning bound scores. By providinginitial and corresponding advanced states, it is possible able to define the reconstructionarea of interest. It is possible also define the positive reward area, where the model isencouraged to take actions and generate cellular states within that area. Conversely, thenegative reward area is defined to discourage the model from generating cellular states inthat region;Fig.1E: shows an action generator and a controller unit, where the covariance matrixof the initial and corresponding advanced cellular states is decomposed using the SVDalgorithm. Subsequently, the singular value and vector matrices are fed into a Gen-Unit togenerate novel cellular states. The generator unit is controlled by the Do-Unit, which themodel uses to manage the generation parameters of the Gen-Unit;Fig.1F: shows a schematic representation of a single-cell RNA-seq data prepro-cessing unit and model environment architecture. The left side represents the data prepro-cessing unit where the initial and corresponding advanced single-cell RNA-seq data areintroduced and perform normalization and scaling steps. The preprocessed data is thenintroduced to the model environment. The model environment is composed of an AGC unitwhere the model can generate novel cellular states, interacting with the action unit (Do-Unit) that governs the generation process (Gen-Unit). The novel generated state is intro-duced to the CSL unit to calculate estimated rewards, which are then pushed to the policy network, aiming to learn optimal policies to find the right sequence of actions to maximizeoverall rewards;Fig.1G: shows a schematic representation of the causal paths that regulate underlyingdisease state trajectories over our lifetime;Fig.1H: shows a schematic representation of the Markov Decision Process where strepresents the model environment state, Ot represents the observation space state, and atrepresents the action taken at time point t. ^^ represents the policy network. P denotes thetransition probability from the current states st to the next state st+1 after the model takes acertain action;Fig.1I: shows a schematic representation depicting the relationship between trajec-tory steps (H) (x-axis) and the reward (R) function (y-axis). The gray color gradient repre-sents the policy (^^) optimality scale. The grey and white broken lines indicate low and highrewards, respectively;Fig. 1J: shows a schematic representation of a two-state challenge, where data fortwo states of an evolving biological process data (at t1 and t4) are introduced to the modelto reconstruct the hidden intermediate states. In the real data, there is ground truth for theintermediate states at time points t2 and t3, which are used to estimate the reconstructionaccuracy of the model;Fig.2A: shows a uniform manifold approximation and projection for dimension reduc-tion (UMAP) representation of the cell cycle cellular states: G1 (green dots), S phase (yel-low dots), and G2 / M phase (red dots);Fig. 2B: shows a UMAP representation of the two-state challenge, indicating the G1and G2 / M phases provided to the model as initial and advanced state, respectively;Fig.2C: shows a UMAP representation of the co-embedding for cell cycle real andgenerated cellular states;Fig. 2D: shows a heat map depicting correlation patterns between real and generateddata in each Leiden cluster;Fig. 2E: shows a density plot of UMAP representation illustrating co-localization pat-terns of cellular states within each cell cycle phase and the generated states overlaying thecellular state manifold;Fig.2F: shows a UMAP representation of the epithelial-mesenchymal transition pro-cess single-cell RNA-seq data across experimental time points;Fig.2G: shows a UMAP representation for the two-state challenge, presenting themodel data from Day 0 and Day 3 cellular states;Fig.2H: shows a density plot of UMAP representation showing co-localization patternsof cellular states from Day 0, Day 3, and generated states overlaying the cellular statemanifold;Fig.2I: shows a UMAP representation displaying co-embedding of real and generateddata. Colors indicate real and generated data within each Leiden cluster;Fig.2J: shows a UMAP representation displaying EMT gene signature scores duringthe transition process;Fig. 2K: shows a violin plot depicting EMT scores of real and generated cellular statesacross the Leiden clusters;Fig. 3A: shows a UMAP representation of murine hematopoietic stem and progenitorsingle-cell RNA-seq data with cell type labels based on marker gene expression. Colorsindicate Leiden clusters;Fig. 3B: shows a UMAP representation for the two-state challenge, presenting neutro-phil progenitors and erythroid progenitors to the model;Fig. 3C: shows a UMAP representation displaying co-embedding of real and generateddata with the Leiden clusters;Fig. 3D: shows a heat map illustrating correlation patterns between real and generateddata within each Leiden cluster;Fig. 3E: shows a barplot indicating Pearson’s correlation coefficients between top 150differentially expressed (DE) genes for real and generated data, and between highly varia-ble genes (HVGs) within Leiden clusters;Fig. 3F: shows a UMAP representation of key hematopoietic stem cell (HSC) markergene expression (normalized). The top panel represents real data gene expression, whilethe lower panel represents generated cellular states data;Fig. 3G: shows a UMAP representation of intermediate cellular states with a color codehighlighting estimated pseudotime values;Fig. 3H: shows a heat map displaying gene expression of key marker genes and tran-scription factors regulating HSCs and MMPs states across latent pseudotime of the modelgenerated cellular states;Fig. 4A: shows a diagram depicting the multistep evolutionary stages of MM tumors;Fig. 4B: shows a UMAP representation of the B-cell compartment in the Vκ*MYCTransgenic MM Mouse Model, highlighting various stages of MM disease activity in Leidenclusters (MM, Intermediate, and Active MM);Fig. 4C: shows a violin plot displaying the normalized gene expression pattern of theScd1 malignancy marker across different Leiden clusters, reflecting varying levels of MMdisease activity (higher expression values indicate a more malignant state);Fig. 4D: shows a UMAP representation for the two-state challenge, presenting earlyMM (Cluster 2) and active MM (Cluster 1) to the model;Fig. 4E: shows a density plot of UMAP representation demonstrating co-localizationpatterns of real and generated states overlaying the cellular states manifold;Fig. 4F: shows a heat map illustrating correlation patterns between gene expressionof real and generated data in each Leiden cluster;Fig. 4G: shows a bar plot indicating Pearson correlation coefficients between the top150 differentially expressed (DE) genes for real and generated data, and between highlyvariable genes (HVGs) within Leiden clusters;Fig. 4H: shows a UMAP representation of plasma cell sub clusters for Patient-27522at Primary, Remission, Relapse-1, and Relapse-2 disease stages. Different colors repre-sent distinct sub clusters within each time point;Fig. 4I: shows a UMAP representation for the two-state challenge, presenting the Pri-mary and Relapse-2 states to the model;Fig. 4J: shows a density plot of UMAP representation displaying co-localization pat-terns of cellular states from Primary, Relapse-1, and Relapse-2, along with generatedstates overlaying the cellular state manifold;Fig. 4K: shows a UMAP representation of co-embedding for real and generated MMcellular states. The highlighted dotted region represents the model predicted unseen / hid-den cellular states, which are not present in the real data;Fig. 5A: shows a UMAP representation of single-nucleus RNA-seq fibroblast and car-diomyocyte populations extracted from the Fibrotic Zone (FZ) of infarcted heart patients.Colors indicate different cell identities;Fig. 5B: UMAP representations illustrating the model novel reconstructed transdiffer-entiation path between fibroblasts and cardiomyocytes;Fig. 5C: shows a UMAP representation displaying the key three transcriptional eventsregulating the transdifferentiation path and direct reprogramming of fibroblast states to car-diomyocyte states; andFig. 5D: shows a heat map depicting transcription factors, ligands, and receptors reg-ulating the three key transcriptional events governing the transdifferentiation path.DETAILED DESCRIPTION OF EMBODIMENTSFigs. 1A, B and F show a deep reinforcement learning architecture of a model that is con-figured to reconstruct intermediate transcriptomic states on cellular differentiation trajecto-ries connecting two anchoring states profiled by scRNA-seq.The model environment is composed of a model which can observe cellular states, takeactions to generate new cellular states, and observe the consequences of taking such ac-tions. The model environment emits a reward signal given the cellular state distance be-tween the generated and advanced states. If the generated state is closer to the advancedstate, the model receives positive reward to reinforce taking such actions consequently infuture steps. Otherwise, if the generated state is far away from the advanced state, themodel receives a negative reward to decrease the probability of taking such actions in fu-ture steps. In this way, the model may experience the environment over hundreds of thou-sands of steps, in order to learn how to generate the right sequence of actions for recon-structing causal cell state trajectories consistent with the permissible cell state manifold asdepicted exemplary in Fig.1G.The input to model may be a merged single-cell gene expression matrix of the two anchor-ing cellular states (^ , ^^), i.e., an initial cellular state and a corresponding advanced cellularstate. The merged matrix may be pushed to the data processing unit and prepared for themodel environment (Fig.1F). To design reasoning models on a cell state manifold, severalcomponents may be incorporated into the model architecture: (1) The model environmentmay be designed as a Markov decision process (MDP) (Fig. 1H), where it is implicitly as-sumed that a cell fate only depends on the current cellular state, regardless of the entirecell state history. (2) The model observation space is defined as the manifold of all possiblecell states which can be visited by the model (Fig. 1B). (3) The reconstruction field (RF) isconstrained by distances to the anchoring cell states in the observation space. It guidesreconstruction of the relevant region within the cell state manifold by the model (Fig. 1C,D). (4) The Actions Generator and Controller (AGC) unit enables the model to reconstructcell states and to navigate in observation space (Fig. 1B). The AGC unit is composed oftwo subunits, termed Gen-Unit and Do-Unit (Fig.1E). To harness the power of linear mod-els for interpretable cell state reconstruction, the AGC unit may calculate a covariance ma-trix of the merged expression matrix of the two anchoring cell states. This covariance matrix is decomposed into principal components (PC) by singular valuedecomposition (SVD). These PCs capture the gene-gene covariance structure and are uti-lized to guide the generation process within the Gen-Unit. The Gen-Unit can generate newcellular states by sampling from a multivariate Gaussian distribution, modeling changes ofthe individual PCs, where three control parameters (µgen, ^gen ^^^ ζ) are governed by theDo-Unit. These actions can generate novel cellular states in gene expression space whichrespect the gene-gene covariance structure inherent in the anchoring cell states, aiming toconfine the observation space to the actual cell state manifold. Hence, the Do-unit canenable the model to control the Gen-Unit during a learning process. (5) The reward functionmay be configured to be dependent on the cellular state loss (CSL), which quantifies thesimilarity of the generated cellular state to the advanced cellular state based on Kullback-Leibler divergence, cosine similarity and entropy regularization terms, and to satisfy thelearning bound score (LBS) objective function (Methods, Figs.1C, D). The reward functioncan reinforce the model to reconstruct cellular events in the reconstruction field of interest.(6) The model policy ^^ can be parametrized by a neural network which is optimized viagradient ascent to maximize the reward objective (Fig. 1I). In addition to the standard re-ward objective, a maximum entropy objective can be configured to enable the model toalternate between exploration and exploitation phases during reconstruction. The modelcan generate stochastic policies by learning a probability distribution over actions given thecollected observations generated by the AGC unit.As demonstrated inter alia for the examples explained below, the model can be used forreconstructing unseen intermediate cell states in scenarios of varying complexity. In eachcase, it is started with a known continuous cell state trajectory profiled by scRNA-seq asground truths, but only an initial and a corresponding advanced anchoring cell state is fed into the model to reconstruct the unseen intermediate states. Thereby, the model’s capacity to reconstruct hidden states can be shown by directly comparing them to the unseenground truths states (Fig. 1J). The challenges for the model span the reconstruction oflinear trajectories, the reconstruction of multipotent states, as well as complex disease sce-narios. The model has shown robustness of intermediate state reconstruction at the levelof cell embeddings and across genes levels.Further aspects and advantages of the invention result from the subsequent description ofpreferred examples. EXAMPLES Example 1: Reconstruction of the cell cycleMurine hematopoietic progenitor cells going through the Interphase stage of the cell cyclewere analysed (Fig.2A).The trained model has been employed by presenting G1 and G2 single cell RNAseq datato reconstruct the S phase (Fig. 2B).During the training phase, the model went through an exploration phase followed by anexploitation phase, and ultimately led to reconstruction of the cellular states in the recon-struction field (RF) of interest. After training, the model was able to generate novel cellularstates, successfully reconstructing the intermediate S phase, as well as G1 and G2 phases(Fig. 2C).These results show that with the method according to the invention, unseen cellular statescan be reconstructed. Example 2: Reconstruction of shifts in cellular identityThe trained model can also be used to reconstruct an entire biological process involvingshifts in cell identities.Exemplary, it was chosen to investigate the epithelial-mesenchymal transition (EMT) pro-cess, which involves a transformation of a cell identity from epithelial to mesenchymal cel-lular states.EMT is a complex process regulated by several key signalling pathways, including Wntsignalling pathways and transforming growth factor beta (TGF-β), which are crucial for can-cer progression in metastatic conditions.To study the temporal dynamics of the EMT process, single-cell RNA-seq data from Lungcancer cell-line (A549) treated with EMT-inducing factor (TGFB1) over a period of sevendays were analysed (Fig.2E).The model was used to introduce the single-cell RNA-seq data of only Day 0 and Day 3 toreconstruct the intermediate states (Fig. 2F). After training the model, the model was ableto reconstruct the entire manifold (H= 10 steps, S2G and H) including 8 hours, 1 day and7 days cellular states (Fig.2G).Moreover, the model successfully reconstructed the cellular states events at both the 2Dembedding level of the real and generated sub-clusters (Fig.2H) as well as key EMT makergenes. To assess the quality of the generated data in reconstructing the EMT process, the EMTsignature score was calculated. It could be observed that the generated cellular statesscores distribution was similar to the real EMT scores distribution, which suggests that themodel was able to reconstruct the entire EMT process accurately (Fig.2J).These findings demonstrate that with the method according to the invention, complex bio-logical processes involving shifts in cell identity can be reconstructed, which provides po-tential applications in disease modelling. Example 3: Reconstructing intermediate states in hematopoiesisTo test the ability of the model to accurately reconstruct complex intermediate states, themodel was subjected to a more intricate biological process, such as hematopoiesis. Thiscomplex process takes place in the bone marrow and involves the gradual differentiation of hematopoietic stem cells (HSCs) into various types of blood cells, including red blood cells, platelets, and various types of white blood cells such as lymphocytes, neutrophils,and monocytes. The process is characterized by a continuous hierarchy of cellular tran-scriptional states, as HSCs progress through differentiation to multipotent progenitors(MPPs) and eventually commit to a single lineage.To investigate the hematopoiesis process, single-cell RNA-seq data from Kit+ hematopoi-etic progenitors at the early stages of the hematopoietic hierarchy were utilized (Fig. 3A).The HSCs population and various haematopoietic progenitor cellular states were identifiedand the model was challenged by presenting two progenitor states, namely the erythroidand neutrophil progenitors, and tasked the model with reconstructing the stem cell andearlier progenitors’ populations (Fig.2B).The model was capable of accurately reconstructing the entire early haematopoietic hier-archy including the stem and progenitors’ states.The accuracy and quality of the intermediate stem and progenitors’ states reconstructedby the model was assessed by performing Leiden clustering and analysing co-embeddingand gene expression correlation between the real and generated data. Nine clusters andselected the intermediate cluster subset (1,0; 2,3 and 5 cluster) for further analyses weredefined.It was possible to observe co-embedding alignment between real and generated data aswell as highly correlation coefficient between real and generated data at the genes highdimensional space. The model was able to accurately reconstruct differentially expressedgenes (n=150) and highly variable genes (n=3000), with up to 0.96 and 0.85 accuracy,respectively.Moreover, the model was able to reconstruct co-expression patterns of HSCs markers(Cd34, Cd48, Hlf, Procr and Slamf1, Megakaryocyte progenitors (Mpl), erythroid progeni-tors’ transcriptional factors (Gata1 and Gata2) and other progenitors marker genes (Fig.2F). Pseudotime inference for both real and generated data manifold was performed andpseudotime ordered the cells and key marker genes and transcriptional factors of the hem-atopoiesis process were examined (Figs.3G, 3H). For example, it was possible to observeupregulation of (Elane) expression together with (Gfi1 and Mpo) maker genes. Then, theupregulation of the co-expression pattern of the HSCs markers was observed (CD48,SOX4, CD27 and CD34). In addition, the exclusive expression pattern of Gata2 and Gata1was observed, which ultimately ended with the expression of CAR1, CAR2 and GFI1B.These results show that the model was able to accurately identify the correct sequence oftranscriptional events underlying the hematopoiesis process. The generative capabilities ofthe model allow for the generation of multiple realizations of connecting causal paths.Example 4: Reconstructing Multiple Myeloma Intermediate StatesThe investigation with the model was expanded to further scenarios, such as cancer evo-lution, focusing on multiple myeloma (MM) disease evolution. MM is a bone marrow malig-nancy characterized by a multi-step process driven by genetic mechanisms and mutations,transforming the normal plasma cells through stages of increasing malignancy, includingmonoclonal gammopathy of undetermined significance (MGUS), smoldering myeloma(SMM), and ultimately to a high-risk malignant state with myeloma and plasma cell leuke-mia predominance. To assess the model's ability to reconstruct intermediate MM states, single-cell RNAseqdata from a Vκ*MYC transgenic mouse model were analysed. These mice exhibit a pro-gressive accumulation of clonal plasma cells in the bone marrow, mirroring the evolutionarytrajectory of human MM (Fig. 4A). By sub-setting the B-cell compartment and performingclustering, four major clusters representing different disease states were identified, rangingfrom early to intermediate and active disease stages (Fig. 4B). The malignancy markerScd1 displayed a pattern of high expression that correlated with disease activity, with clus-ter 2 exhibiting low expression and cluster 1 showing the highest expression (Fig. 4C).To train the model, the data from early MM (Cluster 2) and late MM (Cluster 1) were utilizedas input.The model accurately reconstructed the entire spectrum of MM disease states, includingthe intermediate states (Fig. 4E). Furthermore, alignment in the co-embedding of real andgenerated data (Fig. 4E) and high correlation coefficients in the high-dimensional geneexpression space between the real and generated data was observed (Figs. 4F and 4G).These results demonstrate the capacity of the model in predicting the intermediate trans-formation processes during the tumorigenesis evolutionary steps. Example 5: Reconstructing the disease progressionThe malignant plasma cell compartment of a MM patient (MM-27522) who underwent re-mission (RM), relapse-1 (RL1), and relapse-2 (RL2) after the initial diagnosis (Primary) over a period of more than three years were analysed (Fig.4H). The primary state and RL2 dataand presented them to the model were extracted (Fig. 4J). Following the training phase,the model successfully reconstructed the entire disease progression with high accuracy(Fig. 4I). Notably, the model demonstrated the ability to predict unseen or hidden cellularstates that were not previously observed in the data (Fig.4K).These results show that the model is able to accurately reconstruct the progression of anindividual patients' disease, paving the way for novel personalized medicine and targetedtherapy approaches.Example 6: Predicting a novel path for fibroblasts-to-cardiomyocytes transdifferentiationMyocardial infarction (MI) results in functional loss of cardiac muscle cells and impaired overall heart function. Following the occurrence of an infarction event, remodelling pro-cesses are initiated, characterized by an extensive inflammatory response. This responseleads to the mobilization of innate and adaptive immune cells, notably macrophages, whilethe damaged tissue undergoes replacement by populations of fibroblasts. To address thisissue, inducing fibroblasts-to-cardiomyocytes transdifferentiation has emerged as a poten-tial approach to convert fibroblasts into functional cardiomyocytes, which has been testedin human and mouse. However, the underlying mechanism governing this process hasremained elusive due to the lack of a systematic approach for predicting such transdiffer-entiating cellular states and the underlying molecular events.Fibroblast and cardiomyocyte populations single-cell RNA-seq data from the fibrotic zoneof MI patients’ damaged heart tissue were utilized (Fig. 5A). Through training the model(S5K and S5M) on that two states data, it was possible to predict a novel fibroblasts-to-cardiomyocytes transdifferentiation path (Fig.5B). This path is characterized by three keytranscriptional events (Fig. 5C). Pseudotime inference and differential expression (DE)analysis was conducted, followed by pseudotime ordering of the DE genes based on theinferred pseudotime component.The findings reveal the presence of three distinct modules that govern the predicted trans-differentiation path. The modules are composed of transcription factors (TFs), receptors(S5L), and ligands (Fig.5D).In the first event, upregulation of transcription factors (TFs) such as SOX5, RUNX1,NR4A2, and HEYL was observed (Event 1). This was followed by another event character-ized by the upregulation of TEAD4, HES5, ONECUT3, ZIC1, POU2F3, BARX2, PAX7,SHOX, and ELF3 (Event 2). To ultimately commit to the cardiac fate lineage, the transdif-ferentiation path showed upregulation of FHL2, ESRRG, and NR4A1 (Event 3), along withthe previously known GATA4, MEF2C, and TBX5 (GMT) reprogramming factors.The results shed light on the molecular events driving fibro-to-cardiac transdifferentiation,providing insights into potential targets for therapeutic interventions. Based on these resultsenhancing cardiac regeneration and improving heart function in MI patients may beachieved. It shall be understood that a preferred embodiment of the invention can also be any com- bination of the dependent claims or above embodiments with the respective independent claim. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single unit or device may fulfil the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Procedures like providing transcriptome data, providing a model trained by a machine learning and generating the at least one intermediate cellular state, etc. performed by one or several units or devices can be performed by any other number of units or devices. These procedures can be implemented as program code means of a computer program and / or as dedicated hardware. A computer program may be stored / distributed on a suitable medium, such as an opticalstorage medium or a solid-state medium, supplied together with or as part of other hard-ware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any units described herein may be processing units that are part of a classical computing system. Processing units may include a general-purpose processor and may also include a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Any memory may be a physical system memory, which may be volatile, non-volatile, or some combination of the two. The term “memory” may include any computer-readable storage media such as a non-volatile mass storage. If the computing system is distributed, the processing and / or memory capability may be distrib- uted as well. The computing system may include multiple structures as “executable com- ponents”. The term “executable component” is a structure well understood in the field of computing as being a structure that can be software, hardware, or a combination thereof. For instance, when implemented in software, one of ordinary skill in the art would under- stand that the structure of an executable component may include software objects, rou- tines, methods, and so forth, that may be executed on the computing system. This mayinclude both an executable component in the heap of a computing system, or on computer- readable storage media. The structure of the executable component may exist on a com-puter-readable medium such that, when interpreted by one or more processors of a com-puting system, e.g., by a processor thread, the computing system is caused to perform a function. Such structure may be computer readable directly by the processors, for instance, as is the case if the executable component were binary, or it may be structured to be inter- pretable and / or compiled, for instance, whether in a single stage or in multiple stages, so as to generate such binary that is directly interpretable by the processors. In other in- stances, structures may be hard coded or hard wired logic gates, that are implemented exclusively or near-exclusively in hardware, such as within a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Accordingly, the term “executable component” is a term for a structure that is well under- stood by those of ordinary skill in the art of computing, whether implemented in software, hardware, or a combination. Any embodiments herein are described with reference to acts that are performed by one or more processing units of the computing system. If such acts are implemented in software, one or more processors direct the operation of the computing system in response to having executed computer-executable instructions that constitute an executable component. Computing system may also contain communication channels that allow the computing system to communicate with other computing systems over, for exam-ple, network. A “network” is defined as one or more data links that enable the transport ofelectronic data between computing systems and / or modules and / or other electronic de- vices. When information is transferred or provided over a network or another communica- tions connection, for ex-ample, either hardwired, wireless, or a combination of hardwired or wireless, to a computing system, the computing system properly views the connection as a transmission medium. Transmission media can include a network and / or data links which can be used to carry desired program code means in the form of computer-execut- able instructions or data structures and which can be accessed by a general-purpose or special-purpose computing system or combinations. While not all computing systems re- quire a user interface, in some embodiments, the computing system includes a user inter- face system for use in interfacing with a user. User interfaces act as input or output mech- anism to users for instance via displays. Those skilled in the art will appreciate that at least parts of the invention may be practiced in network computing environments with many types of computing system configurations, including, personal computers, desktop computers, laptop computers, message proces- sors, hand-held devices, multi-processor systems, microprocessor-based or programma- ble consumer electronics, network PCs, minicomputers, main-frame computers, mobile tel- ephones, PDAs, pagers, routers, switches, data-centres, wearables, such as glasses, and the like. The invention may also be practiced in distributed system environments where local and remote computing system, which are linked, for example, either by hardwired data links, wireless data links, or by a combination of hardwired and wireless data links, through a network, both perform tasks. In a distributed system environment, program mod- ules may be located in both local and remote memory storage devices. Those skilled in the art will also appreciate that at least parts of the invention may be prac- ticed in a cloud computing environment. Cloud computing environments may be distributed, although this is not required. When distributed, cloud computing environments may be dis- tributed internationally within an organization and / or have components possessed across multiple organizations. In this description and the following claims, “cloud computing” is defined as a model for enabling on-demand network access to a shared pool of configura- ble computing resources, e.g., networks, servers, storage, applications, and services. The definition of “cloud computing” is not limited to any of the other numerous advantages that can be obtained from such a model when deployed. The computing systems of the figures include various components or functional blocks that may implement the various embodi- ments disclosed herein as explained. The various components or functional blocks may be implemented on a local computing system or may be implemented on a distributed compu- ting system that includes elements resident in the cloud or that implement aspects of cloud computing. The various components or functional blocks may be implemented as software, hardware, or a combination of software and hardware. The computing systems shown in the figures may include more or less than the components illustrated in the figures and some of the components may be combined as circumstances warrant. Any reference signs in the claims should not be construed as limiting the scope.

Claims

Claims1. A computer-implemented method for generating at least one intermediate cellularstate of a cellular state evolution from an initial cellular state to a corresponding ad- vanced cellular state, the method comprising the steps of: -providing transcriptome data representing the initial cellular state and the cor-responding advanced cellular state,- providing a model trained by a machine learning, preferably deep reinforce-ment learning, for generating the at least one intermediate cellular state onthe basis of the initial cellular state and the corresponding advanced cellularstate, wherein training said model comprises the steps i) providing a plurality of transcriptome training data, each representingan initial cellular state and a corresponding advanced cellular state of a cellular state evolution, ii) providing a model that can be trained by machine learning, preferablydeep reinforcement learning, and iii) training the model by machine learning using the transcriptome trainingdata such that the trained model is configured to receive transcriptome data representing an initial cellular state and a corresponding advanced cellular state of a cellular state evolution and for generating at least one intermediate cellular state of the cellular state evolution from the initial cellular state to the corresponding advanced cellular state based on the received transcriptome data, and -generating the at least one intermediate cellular state using the providedtrained model on the basis of the initial cellular state and the corresponding advanced cellular state,wherein the initial and advanced cellular state refer to -different phases of the cells in cell division,- different phases of the cells in cellular differentiation,- different phases of the cells in the transformation of cellular identity,particularly in cellular transdifferentiation, -different phases of the cells in the transformation of a cell to a cancercell, or -different phases of the cells in cell the transformation of a cancer cell toa metastatic cancer cell wherein step iii) of training the model comprises at least one of the following steps:iii.1) generating transcriptome data of further cellular states by sampling from amultivariate Gaussian distribution to learn combinations of action parameters comprising a control mean and a standard deviation of the generated tran- scriptome data of transcriptome data of further cellular states,iii.2) classifying the generated transcriptome data of further cellular states basedon the learned combinations of action parameters to identify those combina- tions of action parameters that fulfil a predefined reward criterion,iii.3) calculating a distance of the generated transcriptome data of further cellularstates to the transcriptome data of advanced cellular states of the plurality of transcriptome training data,iii.4) rewarding the generated transcriptome data of further cellular states with ref-erence to a predefined reward criterion, andiii.5) selecting a combination of action parameters to generate transcriptome dataof further cellular states, which have the highest probability of an increasing reward from one generated transcriptome data to the next transcriptome datawithin the sequence, as the transcriptome data of a plurality of sequential in- termediate cellular states.

2. The method according claim 1, wherein the transcriptome data representing the ini-tial cellular state and the corresponding advanced cellular state are clinical data ob-tained experimentally from a patient.

3. The method according to any of the preceding claims, wherein based on the at leastone intermediate cellular state, a causal trajectory for of the cellular state evolutionfrom the initial cellular state to the corresponding advanced cellular state is gener-ated.

4. The method according to any of the preceding claims, wherein the model is trainedby a deep reinforcement learning technique.

5. The method according any of the preceding claims, wherein the transcriptome train-ing data in step iii) of training the model, each representing the initial cellular state and the corresponding advanced cellular state, respectively, are clinical data ob-tained experimentally from one or more patients.

6. The method according to any of the preceding claims, wherein the advanced cellularstate in step i) of training the model is divided into a plurality of groups of different severity of the advanced cellular state, preferably wherein the transcriptome data of a plurality of sequential intermediate cellular states selected and generated in step iii.5) of training the model is assigned to one of the groups of different severity of the advanced cellular state.

7. The method according to any of the preceding claims, wherein step iii) of training themodel further comprises iii.6) repeating the steps iii.1) to iii.5) to generate a plurality of transcriptome dataof a plurality of sequential intermediate cellular states.

8. A method for diagnosing a disease, comprisinga) providing a sample of a subject, wherein the sample comprises cells associ-ated with the disease to be diagnosed, b) obtaining the transcriptome data of one or more cells from the sample ob-tained in step a), and c) comparing the transcriptome data obtained in step b) with the reconstructedtranscriptome data of the plurality of sequential intermediate cellular states as generated in the method according to any of the preceding claims, wherein the disease is diagnosed if the transcriptome data obtained in step b) cor- responds to one of the transcriptome data of the plurality of sequential intermediate cellular states.

9. The method according to claim 8, wherein in step c) the transcriptome data obtainedin step b) is further compared with the transcriptome data of a plurality of cells in an advanced cellular state as provided in the method according to any of claims 1 to 7, and / or with the transcriptome data of a plurality of cells in an initial cellular state as provided in the method according to any of claims 1 to 7, wherein the disease is diagnosed if the transcriptome data obtained in step b) cor- responds to one of the transcriptome data of the plurality of sequential intermediate cellular states or to one of the transcriptome data of the advanced cellular state, and / or wherein the disease is not diagnosed if the transcriptome data obtained in step b) corresponds to one of the transcriptome data of the initial cellular state.

10. A method for predicting disease progression, comprisingA) providing a sample of a subject, wherein the sample comprises cells associ-ated with the disease to be diagnosed, B) obtaining the transcriptome data of one or more cells from the sample ob-tained in step a), andC) comparing the transcriptome data obtained in step B) with the reconstructedtranscriptome data of the at least one intermediate cellular state as generatedin the method according to claims 1 to 7, to identify one of the transcriptome data of the at least one intermediate cellu- lar state as generated in the method according to claims 1 to 7, to which the transcriptome data obtained in step B) corresponds wherein the prediction of the disease is based on the group of the severity of the disease, to which the transcriptome data identified in step C) is assigned.

11. The method according to any of claims 8 to 10, wherein the disease is cancer, pref-erably wherein the disease is metastasizing cancer.

12. An in-vitro method for transforming the identity of a cell, the method comprising1) generating at least one intermediate cellular state of a cellular state evolutionfrom an initial cellular state to a corresponding advanced cellular state, with a method according to any of claims 1 to 7, wherein the initial and advanced cellular state refer to different phases of the cells in the transformation of cellular identity, particularly in cellular transdiffer- entiation, 2) analysing the at least one intermediate cellular state generated in step 1) toidentify a cellular signalling pathway involved in the transformation of the cel- lular identity, 3) modulating, preferably activating or inactivating, the cellular signalling path-way identified in step 2) to promote the transformation of the identity of a cell.

Images