Estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model
The system estimates RNA velocity for genes with insufficient expression using a machine learning model, addressing calculation limitations and enhancing applications in drug discovery and cell dynamics analysis.
Patent Information
- Application Number
- PCT/JP2025/017233
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-12
- Publication Date
- 2026-11-19
AI Technical Summary
Conventional methods fail to calculate RNA velocity for genes with insufficient RNA expression or contradictory results, limiting their applicability in drug discovery research.
An estimation system and learning system that utilize machine learning to generate an estimation model, allowing the estimation of RNA velocity for genes that cannot be calculated by conventional methods, using gene expression level information as input and outputting RNA velocity information.
Enables the estimation of RNA velocity for genes previously uncalculable by conventional methods, facilitating deeper insights into cell dynamics and supporting applications like trajectory estimation, marker gene analysis, and cell fate prediction.
Smart Images

Figure JP2025017233_19112026_PF_FP_ABST
Abstract
Description
Estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model
[0001] This invention relates to an estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model for estimating the RNA velocity of genes.
[0002] Methods based on gene RNA velocity are known for predicting the future state of individual cells (see, for example, Non-Patent Documents 1 and 2). An algorithm has been proposed to calculate gene RNA velocity from the expression level of RNA from genes in cells, and software that calculates RNA velocity using this algorithm is in use (see, for example, Non-Patent Document 3).
[0003] La Manno, G. , Soldatov, R. , Zeisel, A. , Braun, E. , Hochgerner, H. , Petukhov, V. , Lidschreiber, K. , Kastriti, M. E. , Lonnerberg, P. , Furlan, A. , et al. (2018). RNA velocity of single cells. Nature 560, 494-498.10.1038 / s41586-018-0414-6. Chen, Z. , King, W. C. , Hwang, A. , Gerstein, M. , and Zhang, J. (2022). DeepVelo: Single-cell transcriptomic deep velocity field learning with neural ordinary differential equations. Sci Adv 8, eabq3745.10.1126 / sciadv.abq3745. Bergen, V. , Lange, M. , Peidli, S. , Wolf, F. A. , and Theis, F. J. (2020). Generalizing RNA velocity to transient cell states through dynamic modeling. Nat. Biotechnol. 38, 1408-1414. 10. 1038 / s41587-020-0591-3.
[0004] The inventors of this invention believe that RNA velocity can be useful information for drug discovery research, such as for the search for drug targets. However, if the amount of RNA expression from the gene to be measured is insufficient for the calculation of RNA velocity using the above algorithm, or if the amount of RNA expression from the gene to be measured results in a contradiction in the calculation using the above algorithm, then the RNA velocity of the gene cannot be calculated using the above algorithm. Therefore, the inventors of this invention believe that, for example, in drug discovery research, target genes are expected to extend beyond those for which RNA velocity has been calculated to include those for which it could not be calculated, and have therefore diligently investigated means for estimating the RNA velocity of genes that could not be calculated by conventional methods.
[0005] One embodiment of the present invention has been made in view of the above, and aims to provide an estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model that enable the estimation of RNA velocity of genes that cannot be calculated by conventional methods.
[0006] To achieve the above objective, an estimation system according to one embodiment of the present invention comprises: an acquisition means for acquiring estimation expression level information indicating the amount of RNA expression of a gene to be estimated in a cell; and an estimation means for estimating the RNA velocity of a target gene from the estimation expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[0007] In the estimation system according to one embodiment of the present invention, an estimation model generated by machine learning is used to estimate the RNA velocity of the target gene. By using the estimation model, if estimation expression level information can be prepared, the RNA velocity of the target gene can be estimated. Therefore, according to the estimation system according to one embodiment of the present invention, it is possible to estimate the RNA velocity of a gene that cannot be calculated by conventional methods.
[0008] The gene being estimated may be a different gene from the gene related to the information used in machine learning. With this configuration, it is possible to estimate the RNA velocity of a gene different from the gene related to the information used in machine learning.
[0009] The target gene for estimation may be one for which RNA velocity cannot be calculated using the RNA velocity calculation tools scVelo, UniTVelo, and velocyto. This configuration allows for the estimation of RNA velocity for genes for which RNA velocity cannot be calculated using these tools.
[0010] The expression level information may be a gene expression vector containing values indicating the RNA expression levels for multiple pre-configured cells, and the velocity information may be an RNA velocity vector containing values for multiple cells. With this configuration, the RNA velocity of a gene can be estimated appropriately and reliably.
[0011] One embodiment of this invention can be described as an estimation system invention as described above, or as an estimation method and estimation program invention as described below. These are substantially the same invention, differing only in category, and produce similar functions and effects.
[0012] That is, an estimation method according to one embodiment of the present invention is an estimation method which is an operation method of an estimation system, and includes an acquisition step of acquiring estimation expression level information indicating the amount of RNA expression of a gene to be estimated in a cell, and an estimation step of estimating the RNA velocity of a gene to be estimated from the estimation expression level information acquired in the acquisition step, using an estimation model which is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[0013] An estimation program according to one embodiment of the present invention causes a computer to function as an acquisition means for acquiring estimation expression level information indicating the amount of RNA expression of a gene to be estimated in a cell, and an estimation means for estimating the RNA velocity of a target gene from the estimation expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[0014] To achieve the above objective, a learning system according to one embodiment of the present invention comprises: a learning acquisition means for acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and a learning means for generating an estimation model that performs machine learning using the learning expression level information and learning velocity information acquired by the learning acquisition means, inputting expression level information indicating the amount of RNA expression of a gene in a cell and outputting velocity information indicating the RNA velocity of the gene.
[0015] In a learning system according to one embodiment of the present invention, machine learning is performed to generate an estimation model. By using the generated estimation model, if estimation expression level information can be prepared, the RNA velocity of the target gene can be estimated. Therefore, the learning system according to one embodiment of the present invention makes it possible to estimate the RNA velocity of genes that cannot be calculated by conventional methods.
[0016] The expression level information may be a gene expression vector containing values indicating the RNA expression levels for multiple pre-configured cells, and the velocity information may be an RNA velocity vector containing values for multiple cells. This configuration allows for the generation of an accurate and reliable estimation model.
[0017] One embodiment of the present invention can be described as a learning system invention as described above, but it can also be described as an invention of a learning method and a learning program as follows. These are substantially the same invention, differing only in category, and produce similar functions and effects.
[0018] That is, a learning method according to one embodiment of the present invention is a learning method which is a method for operating a learning system, and includes a learning acquisition step of acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene, and a learning step of performing machine learning using the learning expression level information and learning velocity information acquired in the learning acquisition step to generate an estimation model which takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[0019] A learning program according to one embodiment of the present invention causes a computer to function as a learning acquisition means that acquires learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene, and a learning means that uses the learning expression level information and learning velocity information acquired by the learning acquisition means to perform machine learning and generate an estimation model that takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[0020] The estimation model itself, generated by the learning system according to one embodiment of the present invention, is also an invention with a novel configuration. Specifically, the estimation model according to one embodiment of the present invention is generated by the learning system described above, and the computer is made to function by taking expression level information, which indicates the amount of RNA expression of a gene in a cell, as input and outputting velocity information, which indicates the RNA velocity of the gene.
[0021] According to one embodiment of the present invention, it is possible to estimate the RNA velocity of a gene, which cannot be calculated by conventional methods.
[0022] This figure shows the configuration of the estimation system and learning system according to an embodiment of the present invention. This figure shows an overview of the estimation of gene RNA velocity and the generation of an estimation model according to the embodiment. This flowchart shows the estimation method, which is the process performed by the estimation system according to an embodiment of the present invention. This flowchart shows the learning method, which is the process performed by the learning system according to an embodiment of the present invention. This table shows the dataset used in the example. This is an example of a graph that visualizes the cell-specific values of gene RNA velocity by embedding them in two dimensions. This is an example of a scatter plot of RNA velocity values calculated by the tool and RNA velocity values estimated by the estimation model. This is an example of a graph showing the correlation coefficient between RNA velocity calculated by the tool for each gene and RNA velocity estimated by the estimation model, and an example of a graph showing the distribution of cosine similarity between RNA velocity estimated by the estimation model and RNA velocity calculated by the tool. This is an example of a graph that visualizes the estimated RNA velocity and calculated gene expression vector for a specific gene by embedding them in two dimensions. This is an example of a graph that visualizes the cell-specific RNA velocity values of a gene by embedding them in two dimensions. This is an example of a scatter plot of RNA velocity values calculated by the tool and RNA velocity values estimated by the estimation model. This is an example of a graph showing the correlation coefficient between RNA velocity calculated by the tool for each gene and RNA velocity estimated by the estimation model, and an example of a graph showing the distribution of cosine similarity between RNA velocity estimated by the estimation model and RNA velocity calculated by the tool. This is an example of a graph that visualizes the estimated RNA velocity and calculated gene expression vector for a specific gene by embedding them in two dimensions. This is an example of a graph that visualizes the RNA velocity values of genes for each cell by embedding them in two dimensions.This is an example of a scatter plot showing the RNA velocity values calculated by the tool and the RNA velocity values estimated by the estimation model. This is an example of a graph showing the correlation coefficient between the RNA velocity calculated by the tool for each gene and the RNA velocity estimated by the estimation model, and an example of a graph showing the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity calculated by the tool. This is an example of a graph visualizing the estimated RNA velocity and calculated gene expression vector for a specific gene by embedding them in two dimensions. This is an example of a graph visualizing the cell-specific RNA velocity values of a gene by embedding them in two dimensions. This is an example of a scatter plot showing the RNA velocity values calculated by the tool and the RNA velocity values estimated by the estimation model. Examples of graphs showing the correlation coefficient between RNA velocity calculated by a gene-specific tool and RNA velocity estimated by a estimation model, and examples of graphs showing the distribution of cosine similarity between RNA velocity estimated by a estimation model and RNA velocity calculated by a tool. Examples of graphs visualizing the estimated RNA velocity and calculated gene expression vector for a specific gene by embedding them in two dimensions. Examples of graphs visualizing the cell-specific values of RNA velocity for a gene by embedding them in two dimensions. Examples of graphs visualizing the cell-specific values of RNA velocity for a gene by embedding them in two dimensions. Examples of graphs visualizing the estimated RNA velocity and calculated gene expression vector for a specific gene by embedding them in two dimensions. This is an example of a graph showing the correlation coefficient between the training data and the validation data for each cross-validation group. This is an example of a graph showing the correlation coefficient between the training data and the validation data for each cross-validation group. This is an example of a graph showing the correlation coefficient between the training data and the validation data for each cross-validation group.This is an example of a graph that visualizes the cell-specific RNA velocity values of a gene by embedding them in two dimensions. This is an example of a graph showing the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity calculated by the tool. This is an example of a graph that visualizes the cell-specific RNA velocity values of a gene by embedding them in two dimensions. This is an example of a graph showing the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity calculated by the tool. This is an example of a graph that visualizes the cell-specific RNA velocity values of a gene by embedding them in two dimensions. This is an example of a graph showing the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity calculated by the tool. This is an example of a graph that visualizes the cell-specific RNA velocity values of genes by embedding them in two dimensions. This is an example of a graph showing the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity calculated by the tool. This is a diagram showing the configuration of the estimation program and learning program according to an embodiment of the present invention, together with the recording medium.
[0023] Hereinafter, embodiments of the estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant explanations are omitted.
[0024] Figure 1(a) shows the estimation system 10 according to this embodiment. Figure 1(b) shows the learning system 20 according to this embodiment. The estimation system 10 is a system (device) that estimates (predicts, infers) the RNA velocity of a gene. The estimation system 10 performs estimation using a pre-prepared estimation model. The learning system 20 is a system (device) that generates the estimation model. Figure 2 shows an overview of the estimation of the RNA velocity of a gene and the generation of the estimation model according to this embodiment.
[0025] Gene RNA velocity is the direction and speed of the transition of the time-series state of gene expression in a cell. Gene RNA velocity is a concept that captures the transcriptional dynamics of a gene based on the distribution of RNA expression levels in a cell. Gene RNA velocity is shown, for example, in Non-Patent Literature 1. For example, velocity information indicating gene RNA velocity is an RNA velocity vector (a vector with the dimension of the number of predetermined cells) that has values for multiple predetermined cells. The predetermined number of cells are, for example, cells from which the RNA expression level of the gene is measured.
[0026] As described above, the RNA velocity of a gene can be calculated using conventionally proposed algorithms. However, as also mentioned above, there are RNA velocities that cannot be calculated using conventionally proposed algorithms. For example, RNA velocity cannot be calculated for genes that do not undergo splicing, and for genes for which sufficient reads of immature mRNA that has not undergone splicing cannot be obtained. The estimation system 10 and learning system 20 according to this embodiment are for estimating the RNA velocity of genes that cannot be calculated by conventional methods.
[0027] The RNA velocity of genes estimated by this embodiment can be used for downstream analyses such as trajectory estimation analysis of cellular dynamic processes or marker gene analysis, similar to conventional RNA velocity. For example, using the RNA velocity of genes, the direction and magnitude of state changes in the space of gene expression can be estimated as vectors for the short-term future of individual cells, and the dynamic transitions of each cell in the embedded space can be drawn as trajectories. Furthermore, RNA velocity can be used for differential kinetics analysis to extract genes that undergo dynamic transcriptional changes in the differentiation lineage, or for in silico perturbation and cell fate prediction based on the reconstruction of continuous vector fields by machine learning. This allows for deeper insights into cell dynamics. In addition, the RNA velocity of genes estimated by this embodiment may be used for any other purpose.
[0028] The estimation system 10 and the learning system 20 are, specifically, computers including hardware such as a CPU (Central Processing Unit) and memory. The functions of the estimation system 10 and the learning system 20, as described later, are performed by the operation of these components by programs, etc. The estimation system 10 and the learning system 20 may be implemented in a single computer, or they may be implemented in a computer system configured by multiple computers connected to each other by a network. The estimation system 10 and the learning system 20 may have communication functions for acquiring information necessary for the processing of the estimation system 10 and the learning system 20 as described below. In Figure 1, the estimation system 10 and the learning system 20 are shown as separate systems (devices), but they may be implemented in the same system (device).
[0029] Next, the functions of the estimation system 10 and the learning system 20 according to this embodiment will be described. As shown in Figure 1, the estimation system 10 is configured to include an acquisition unit 11 and an estimation unit 12.
[0030] The acquisition unit 11 is an acquisition means for acquiring estimated expression level information that indicates the amount of RNA expression in cells of the gene targeted for RNA velocity estimation. The gene targeted for estimation may be a different gene from the gene related to the information used in machine learning. The gene targeted for estimation may be a gene for which RNA velocity cannot be calculated using the RNA velocity calculation tools scVelo, UniTVelo, and velocyto. The estimated expression level information may be a gene expression vector having values that indicate the amount of RNA expression for each of a set number of cells.
[0031] The genes for which RNA velocity is to be estimated in the estimation system 10 are predetermined. For example, the genes for which RNA velocity cannot be calculated using conventionally proposed algorithms or tools.
[0032] The tools in question are scVelo, UniTVelo, and velocyto. scVelo is a standard tool widely used in Python-based velocity analysis. scVelo is described in Non-Patent Literature 3. UniTVelo is a tool that models gene expression dynamics using a radial basis function. UniTVelo is described in Gao, M., Qiao, C., and Huang, Y. (2022). UniTVelo: temporarily unified RNA velocity reinforces single-cell trajectory inference. Nat Commun 13, 6586.10.1038 / s41467-022-34188-7. (Non-Patent Literature 4). Velocyto comes in two versions: a Python-based version (velocyto.py) and an R-based version (velocyto.py). Velocyto is described in Non-Patent Document 1.
[0033] In estimation system 10, RNA velocity is estimated using an estimation model generated by machine learning. The gene targeted for RNA velocity estimation in estimation system 10 may be a different gene from the gene related to the information used in machine learning (learning gene). Furthermore, the gene targeted for RNA velocity estimation in estimation system 10 does not necessarily have to be one of the above, but any gene from which the information necessary for RNA velocity estimation can be obtained.
[0034] The estimated expression level information acquired by the acquisition unit 11 is used to estimate the RNA velocity of a gene. The estimated expression level information indicates the amount of RNA expression in a cell for the gene whose RNA velocity is to be estimated. For example, the estimated expression level information is a gene expression vector (a vector with the dimension of the number of pre-set cells) that has values indicating the amount of RNA expression for each of a pre-set number of cells. The pre-set number of cells are, for example, cells in which RNA measurement is performed. Furthermore, the pre-set number of cells are the same as the number of cells pre-set when generating the estimation model in the learning system 20.
[0035] Figure 2 schematically shows the gene expression vector 30 (Current exp. of dropped genes, Test set), which is the estimated expression level information. In the gene expression vector 30 shown in Figure 2, the intensity of the color indicates the value of each element (the same applies to vectors shown by intensity of color hereafter). In Figure 2, the gene expression vectors 30 of two genes are placed side by side. In the example shown in Figure 2, the gene expression vector 30 is a five-dimensional vector. That is, the gene expression vector 30 has values that indicate the amount of RNA expression in five cells. The values that indicate the amount of RNA expression in each of the multiple cells, which are elements of the gene expression vector 30, are index values that indicate the degree of expression in the cell at a specific time. The values that indicate the amount of RNA expression in each of the multiple cells included in the gene expression vector 30 can be obtained, for example, by measuring the RNA expression in each of the multiple cells as shown below and performing existing standardization (normalization) (for example, conventional z-score normalization).
[0036] The gene expression vector 30 is generated based on actual preparation of the plurality of cells described above and measurement of RNA expression through experiments on the cells. The dimension of the gene expression vector 30 is the dimension of the plurality of cells for which RNA expression measurement has been performed. The specific timing described above is, for example, the timing at which RNA expression measurement is performed. The gene expression vector 30 is generated, for example, as follows.
[0037] A plurality of cells to be subjected to RNA expression measurement are prepared. For each cell and each gene, the RNA expression of the gene contained in the cell is measured. As the expression measurement, for example, the read count of spliced mRNA that has undergone splicing (i.e., mature mRNA after the entire mRNA processing process has proceeded) and the read count of unspliced mRNA that has not undergone splicing (i.e., immature mRNA (precursor mRNA) before the entire or part of the mRNA processing process has proceeded) are each counted. The measurement may be performed by a conventional technique such as single-cell RNA sequencing analysis.
[0038] Through the measurement, for example, as shown in Figure 2, for each gene, a vector 40 of spliced mRNA read counts (Spliced count) and a vector 41 of unspliced mRNA read counts (Unspliced count) for each cell are obtained. In Figure 2, the read count vectors 40 and 41 for six genes each are arranged side by side. In the example shown in Figure 2, the read count vectors 40 and 41 are 5-dimensional vectors. That is, the read count vectors 40 and 41 have read counts for five cells.
[0039] From the matrix composed of the aforementioned read count vectors 40 and 41, a gene expression vector 42 (Current expr.) for each gene is calculated using existing tools such as scVelo (Raw RNA velocity calculation). The dimension of the gene expression vector 42 is the same as the dimension of the aforementioned read count vectors 40 and 41, which is the dimension corresponding to the number of cells. The calculated gene expression vector 42 is defined as the gene expression vector 30 which is the expression level information for estimation. Normally, the gene expression vector 30 as the expression level information for estimation is a part of the plurality of calculated gene expression vectors 42.
[0040] Using existing tools such as scVelo, RNA velocity vectors 43 and 44 for each gene can be calculated from the matrix composed of the aforementioned read count vectors 40 and 41, together with the gene expression vector 42 for each gene. The RNA velocity vectors 43 and 44 are velocity information indicating the RNA velocity of the gene. The RNA velocity vectors 43 and 44 are vectors having a value for each of a plurality of cells. The dimension of the RNA velocity vectors 43 and 44 is the same as the dimension of the aforementioned read count vectors 40 and 41 and the gene expression vector 42, which is the dimension corresponding to the number of cells. In FIG. 2, the RNA velocity vectors 43 and 44 of four genes are arranged side by side respectively. In the example shown in FIG. 2, the RNA velocity vectors 43 and 44 are 5-dimensional vectors.
[0041] As described above, for some genes, the RNA velocity vectors 43 and 44 cannot be calculated by existing tools. In the present embodiment, genes for which the RNA velocity vectors 43 and 44 (RNA velocity) can be calculated by existing tools are referred to as Passed genes, and genes for which the RNA velocity vectors 43 and 44 (RNA velocity) cannot be calculated are referred to as Dropped genes.
[0042] RNA velocity vector 43 (RNA velocity (*Raw RNA velocity)) is calculated without removing the noise contained in the above read number vectors 40 and 41 (No denoising). RNA velocity vector 44 (RNA velocity (*denoised future-denoised current)) is calculated after removing the noise contained in the above read number vectors 40 and 41 (Denoising option (if necessary)). RNA velocity vector 44 is calculated as follows.
[0043] Using existing tools such as scVelo, a vector 45 (Future expr. (Current expr. + Raw RNA velocity)) is calculated by adding the RNA velocity to the gene expression vector 42 (Current expr.) for each gene, from a matrix consisting of the above-mentioned vectors 40 and 41 of read counts. This vector 45 indicates the future RNA expression level (projected expression value) relative to the time of measurement. Note that the above vector 45 cannot be calculated for dropped genes.
[0044] The RNA velocity vector 44 is calculated by removing noise from a matrix consisting of gene expression vectors 42 for each gene and a matrix consisting of the above vectors 45 using an existing noise reduction tool. An example of an existing noise reduction tool is MAGIC. MAGIC reconstructs a complementary count matrix based on data diffusion between similar cells. Since the matrix consisting of gene expression vectors 42 for each gene (data matrix of current expression values) and the matrix consisting of the above vectors 45 (data matrix of future expression values) have the same number of cells, a square Markov matrix with a size of cell count × cell count calculated from the expression values of all genes, including both passed and dropped genes, can be applied to the predicted future expression values of the passed genes. MAGIC was developed by van Dijk, D., Sharma, R., Nainys, J., Yim, K., Kathail, P., Carr, A. J., and Burdziak, C. This is shown in Moon, K. R., Chaffer, C. L., Pattabiraman, D., et al. (2018). Recovering Gene Interactions from Single-Cell Data Using Data Diffusion. Cell 174, 716-729 e727.10.1016 / j. cell. 2018.05.061. (Non-patent document 5).
[0045] For example, among the genes for which a gene expression vector 42 has been calculated, the Dropped gene is designated as the gene for which RNA velocity is estimated. In this case, the gene expression vector 42 of the Dropped gene becomes the gene expression vector 30, which is the expression level information for estimation.
[0046] The acquisition unit 11 may, for example, receive and acquire a gene expression vector 30, which is estimated expression level information transmitted from another system. Alternatively, the acquisition unit 11 may accept an input operation for the gene expression vector 30 from the user to the estimation system 10 and acquire the gene expression vector 30. Alternatively, the acquisition unit 11 may acquire the read number vectors 40 and 41 mentioned above and calculate and acquire the gene expression vector 30 as described above. Furthermore, the acquisition unit 11 may acquire the gene expression vector 30 by methods other than those described above.
[0047] The estimated expression level information acquired by the acquisition unit 11 is information indicating the amount of RNA expression in cells of the gene targeted for RNA velocity estimation, and may be other than the gene expression vector 30, as long as it is information that can be used for RNA velocity estimation. Furthermore, the estimated expression level information may be generated by methods other than those described above. The acquisition unit 11 outputs the acquired estimated expression level information to the estimation unit 12.
[0048] The estimation unit 12 is an estimation means that estimates the RNA velocity of a target gene from the estimated expression level information acquired by the acquisition unit 11, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input, and outputs velocity information indicating the RNA velocity of the gene. The expression level information is a gene expression vector having values indicating the amount of RNA expression for each of a set number of cells, and the velocity information may be an RNA velocity vector having values for each of the multiple cells. The estimation unit 12 estimates the RNA velocity of a target gene as follows, for example.
[0049] The estimation unit 12 inputs and stores the estimation model generated by the learning system 20, and uses it to estimate RNA velocity. An example of an estimation model is described below. The estimation model is a trained model generated by machine learning. The estimation model takes expression level information, which indicates the amount of RNA expression of a gene in a cell, as input and outputs velocity information, which indicates the RNA velocity of that gene. That is, in the estimation model, expression level information is used as an explanatory variable, and velocity information is used as the target variable.
[0050] Figure 2 schematically shows an example of an estimation model 50. For example, the estimation model 50 takes the gene expression vector 30 of one gene as input and outputs the RNA velocity vector 60 (Predicted velocity valued for dropped genes) of that gene as the estimation result. In Figure 2, the RNA velocity vectors 60 of two genes are shown side by side. In this embodiment, the estimation model 50 is called NARVI (Neural-network Assisted RNA Velocity Imputation).
[0051] The estimation model 50 is composed of, for example, a neural network. The neural network may be a fully connected neural network. The neural network may be multi-layered. That is, the estimation model 50 may be generated by deep learning. The form of the estimation model 50 may be the same as that of models generated by conventional machine learning.
[0052] The input layer of the estimation model 50 is provided with neurons for inputting expression level information. For example, the input layer of the estimation model 50 is provided with neurons, the number of which is equal to the number of dimensions of the gene expression vector 30 (that is, the number of cells), and to which the values of the elements of the gene expression vector 30 are input. The output layer of the estimation model 50 is provided with neurons for outputting velocity information. For example, the output layer of the estimation model 50 is provided with neurons, the number of which is equal to the number of dimensions of the RNA velocity vector 60 (that is, the number of cells), and which output the values of the elements of the RNA velocity vector 60.
[0053] Specifically, the estimation model 50 may have the following configuration. The estimation model 50 is composed of a plurality of blocks each consisting of a fully connected layer, an activation function, and a dropout layer. There is a fully connected layer that transforms from the input layer to the first hidden layer, followed by an activation function application, and then a dropout layer. h (1) is the output of the first layer, the transformation from the input layer to the first layer is as follows. h (1) =Dropout(σ(W (1) x + b (1) ), p) Here, Dropout() is a function that performs dropout, x is an input vector, W (1) is a weight matrix, b (1) is a bias vector, σ is an activation function, and p is a dropout rate. This series of structures is defined as one block, and this block is repeated for each hidden layer. The transformation from the (l)-th layer to the (l+1)-th layer is as follows. h (l+1) =Dropout(σ(W (l+1) h (l) + b (l+1) ), p) Finally, the network is completed with a fully connected layer that returns to the input size. y = W (L) h (L-1) + b (L) Here, L is the index of the final layer (output layer), and y is the output vector.
[0054] The estimation model 50 is designed to cause a computer to function by taking expression level information, which indicates the amount of RNA expression of a gene in a cell, as input and outputting velocity information, which indicates the RNA velocity of the gene. The estimation model 50 is intended to be used as a program module that is part of artificial intelligence software. The estimation model 50 is used, for example, in a computer equipped with hardware such as a CPU and memory, and the computer's CPU operates according to instructions from the estimation model 50 stored in memory. For example, the computer's CPU operates according to the instructions, inputting information to the estimation model 50, performing calculations corresponding to the estimation model 50, and outputting results from the estimation model 50. Specifically, the computer's CPU operates according to the instructions, inputting information to the input layer of a neural network, performing calculations based on learned weighting coefficients in the neural network, and outputting results from the output layer of the neural network. Note that the estimation model 50 may be composed of something other than a neural network.
[0055] The estimation unit 12 receives estimated expression level information from the acquisition unit 11. The estimation unit 12 inputs the estimated expression level information into a pre-stored estimation model 50 and performs calculations according to the estimation model 50, obtaining the velocity information of the gene to be estimated as the estimation result as output from the estimation model 50. For example, the estimation unit 12 inputs the gene expression vector 30 into the estimation model 50 and performs calculations to obtain the RNA velocity vector 60 of the gene to be estimated.
[0056] The estimation unit 12 outputs velocity information, which is the estimation result. For example, the estimation unit 12 transmits information indicating the estimation result to the user's terminal. Alternatively, the estimation unit 12 may output information indicating the estimation result in a format that the user can recognize. For example, the estimation unit 12 may display information indicating the estimation result on a display device provided by the estimation system 10. The output of information indicating the estimation result from the estimation unit 12 may be performed by methods other than those described above, and to output destinations other than those described above. The above describes the functions of the estimation system 10 according to this embodiment.
[0057] As shown in Figure 2, the learning system 20 is configured to include a learning acquisition unit 21 and a learning unit 22.
[0058] The learning acquisition unit 21 is a learning acquisition means that acquires learning expression level information, which indicates the amount of RNA expression of a learning gene in a cell, and learning velocity information, which indicates the RNA velocity of the learning gene. The learning expression level information may be a gene expression vector having values indicating the amount of RNA expression for each of a set number of cells. The learning velocity information may be an RNA velocity vector having values for each of the set number of cells.
[0059] The genes used for learning in the learning system 20 are pre-configured. For example, these genes are genes (Passed genes) that can have their RNA velocity calculated using the aforementioned tool that calculates RNA velocity using a conventionally proposed algorithm.
[0060] The learning expression level information and learning velocity information acquired by the learning acquisition unit 21 are used to generate the estimation model used in the estimation system 10. The learning expression level information is of the same type as the estimation expression level information. The learning expression level information is information that shows the amount of RNA expression of the learning gene in a cell. For example, the learning expression level information is a gene expression vector that has values that show the amount of RNA expression for each of a set of cells that have been set in advance. The set of cells that have been set in advance are, for example, cells in which RNA is measured. Also, the set of cells that have been set in advance are the same as the set of cells that have been set in advance when estimating RNA velocity in the estimation system 10.
[0061] Figure 2 schematically shows the gene expression vector 70 (Current exp. of dropped genes, Training set), which is learning expression level information. In Figure 2, the gene expression vectors 70 of four genes are arranged horizontally. In the example shown in Figure 2, the gene expression vector 70 is a five-dimensional vector. That is, the gene expression vector 70 has values that indicate the RNA expression levels in five cells. The values that indicate the RNA expression levels for each of the multiple cells that make up the gene expression vector 70 are index values of the expression level in cells at a specific time.
[0062] The learning velocity information is of the same type as the velocity information output from the estimation model in the estimation system 10. The learning velocity information is information that indicates the RNA velocity of the learning gene. For example, the learning velocity information is an RNA velocity vector with pre-set values for multiple cells. The dimension of the RNA velocity vector is the same as the dimension of the gene expression vector 70 described above, which is the dimension of the number of cells.
[0063] Figure 2 schematically shows the RNA velocity vector 71 (Velocity of passed genes), which is the velocity information used for learning. In Figure 2, the RNA velocity vectors 71 of four genes are arranged side by side. In the example shown in Figure 2, the RNA velocity vector 71 is a five-dimensional vector.
[0064] The gene expression vector 70 and RNA velocity vector 71 are generated based on the measurement of RNA expression in multiple actual cells. The dimensions of the gene expression vector 70 and RNA velocity vector 71 are the dimensions of the multiple cells in which the RNA expression was measured. The specific timing mentioned above is, for example, the timing at which the RNA expression was measured. The gene expression vector 70 and RNA velocity vector 71 may be generated, for example, in conjunction with the generation of the gene expression vector 30, which is the estimated expression level information described above.
[0065] For example, as described above, gene expression vector 42 and RNA velocity vectors 43 and 44 are calculated. Of the genes for which this calculation was performed, the Passed genes are designated as learning genes. In this case, the gene expression vector 42 of the Passed gene becomes the gene expression vector 70, which is the learning expression level information. Also, the RNA velocity vector 43 or RNA velocity vector 44 of the Passed gene becomes the RNA velocity vector 71, which is the learning velocity information.
[0066] The learning acquisition unit 21 acquires a combination of a gene expression vector 70, which is learning expression level information, and an RNA velocity vector 71, which is learning velocity information, for each gene to be learned. This combination is used for learning the generation of the estimation model. The learning acquisition unit 21 also receives a sufficient number of the above combinations to generate the estimation model.
[0067] The learning acquisition unit 21 receives and acquires, for example, a combination of a gene expression vector 70, which is learning expression level information, and an RNA velocity vector 71, which is learning velocity information, transmitted from another system. Alternatively, the learning acquisition unit 21 accepts an input operation of the combination to the estimation system 10 from the user and acquires the combination. Alternatively, the acquisition unit 11 may acquire the above-mentioned read number vectors 40 and 41 and calculate and acquire the combination as described above. Furthermore, the acquisition unit 11 may acquire the combination by a method other than those described above.
[0068] The combination of learning expression level information and learning velocity information acquired by the learning acquisition unit 21 is a combination of information indicating the RNA expression level of the learning gene in the cell and information indicating the RNA velocity of the learning gene, and may be any combination other than the gene expression vector 70 and RNA velocity vector 71, as long as it is information that can be used to generate the estimation model. Furthermore, the combination of learning expression level information and learning velocity information may be generated by methods other than those described above. The learning acquisition unit 21 outputs the acquired combination of learning expression level information and learning velocity information to the learning unit 22.
[0069] The learning unit 22 is a learning means that uses learning expression level information and learning velocity information acquired by the learning acquisition unit 21 to perform machine learning and generate an estimation model that takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene as input. The expression level information is a gene expression vector having values indicating the amount of RNA expression for each of several pre-set cells, and the velocity information may be an RNA velocity vector having values for each of several cells. The learning unit 22 generates an estimation model as follows, for example.
[0070] The learning unit 22 receives a combination of training expression level information and training velocity information from the training acquisition unit 21. The learning unit 22 uses the input combination to train an estimation model using machine learning and generates an estimation model. The machine learning itself using the combination of training expression level information and training velocity information can be performed in the same way as conventional machine learning.
[0071] For example, when the learning unit 22 inputs a gene expression vector 70, which is learning expression level information, to the estimation model 50, it trains the estimation model 50 to output an RNA velocity vector 71, which is learning velocity information, that is paired with the gene expression vector 70. In other words, the learning unit 22 learns the relationship between the learning expression level information (expression pattern in cells) and the learning velocity information (RNA velocity) to generate the estimation model 50.
[0072] Since the estimation model 50 is based on the above-mentioned relationships between the genes used to generate the estimation model, in order to properly estimate RNA velocity, the genes that are the target of RNA velocity estimation also need to have the above-mentioned relationships. That is, the estimation model 50 depends on the similarity of the expression patterns between the genes used to generate the estimation model (e.g., Passed genes) and the genes that are the target of RNA velocity estimation (e.g., Dropped genes).
[0073] Specifically, the learning unit 22 inputs the gene expression vector 70 into the estimation model 50 during training and obtains the output 72 at that point. The learning unit 22 compares the obtained output 72 with the RNA velocity vector 71 and updates the parameters of the estimation model 50 by backpropagation to optimize the loss function (Loss). The learning unit 22 repeatedly performs the above training using multiple combinations of the gene expression vector 70 and the RNA velocity vector 71 to generate the estimation model 50.
[0074] The learning unit 22 may tune the hyperparameters of the estimation model 50. The tuning of the hyperparameters of the estimation model 50 can be performed, for example, by optimization using a conventional tool (e.g., Optuna).
[0075] The estimation model 50 generated as described above aims to reconstruct (complement) the RNA velocity of the target gene (e.g., a dropped gene) that exhibits an expression distribution similar to that of the passed gene, from the perspective of interpolation and predictable regions in machine learning.
[0076] The training of the estimation model 50 described above is repeated, similar to conventional machine learning training, for example, until a predetermined number of times, or until the generation of the estimation model 50 converges based on predetermined conditions. Once the repeated training of the estimation model 50 is complete, the learning unit 22 outputs the generated estimation model 50 to the estimation system 10. In the estimation system 10, the estimation unit 12 inputs and stores the estimation model 50 and uses it for the estimation of RNA velocity described above. The above describes the functions of the learning system 20 according to this embodiment.
[0077] Next, the processes performed by the estimation system 10 and the learning system 20 according to this embodiment (the operation methods performed by the estimation system 10 and the learning system 20) will be explained using the flowcharts in Figures 3 and 4.
[0078] First, the estimation method, which is a process performed by the estimation system 10 according to this embodiment, will be explained using the flowchart in Figure 3. In this process, the acquisition unit 11 acquires estimation expression level information, which indicates the amount of RNA expression in cells of the gene to be estimated for RNA velocity (S01, acquisition step). Next, the estimation unit 12 uses an estimation model to estimate the RNA velocity of the gene to be estimated from the estimation expression level information acquired by the acquisition unit 11 (S02, estimation step). Subsequently, the estimation unit 12 outputs information indicating the estimation result (S03). The above is the estimation method according to this embodiment.
[0079] Next, the learning method, which is a process performed by the learning system 20 according to this embodiment, will be explained using the flowchart in Figure 4. In this process, the learning acquisition unit 21 acquires learning expression level information, which indicates the amount of RNA expression of the learning gene in the cell, and learning velocity information, which indicates the RNA velocity of the learning gene (S11, learning acquisition step). Subsequently, the learning unit 22 performs machine learning using the learning expression level information and learning velocity information acquired by the learning acquisition unit 21 to generate an estimation model (S12, learning step). Subsequently, the learning unit 22 outputs the generated estimation model to the estimation system 10 (S13). The estimation model is stored in the estimation unit 12 of the estimation system 10 and used for the estimation of RNA velocity as described above. The above is the learning method according to this embodiment.
[0080] In this embodiment, an estimation model generated by machine learning is used to estimate the RNA velocity of the target gene. By using the estimation model, if estimation expression level information can be prepared, the RNA velocity of the target gene can be estimated. Therefore, according to this embodiment, it is possible to estimate the RNA velocity of genes that cannot be calculated by conventional methods.
[0081] Furthermore, as in this embodiment, the gene to be estimated may be a different gene from the gene related to the information used in machine learning. With this configuration, it is possible to estimate the RNA velocity of a gene different from the gene related to the information used in machine learning.
[0082] Furthermore, as in this embodiment, the gene to be estimated may be a gene for which RNA velocity cannot be calculated using the tools for calculating RNA velocity, such as scVelo, UniTVelo, and velocyto. With this configuration, it is possible to estimate the RNA velocity of genes for which RNA velocity cannot be calculated using these tools. However, the gene to be estimated does not necessarily have to be one of the above; it can be any gene for which expression level information for estimation can be prepared.
[0083] Furthermore, in this embodiment, machine learning is performed to generate an estimation model. By using the generated estimation model, if estimation expression level information can be prepared, the RNA velocity of the target gene can be estimated. Therefore, according to this embodiment, it is possible to estimate the RNA velocity of genes that cannot be calculated by conventional methods.
[0084] Furthermore, as in this embodiment, the expression level information may be a gene expression vector having values indicating the RNA expression levels of multiple pre-set cells, and the velocity information may be an RNA velocity vector having values for multiple cells. With this configuration, the RNA velocity of a gene can be estimated appropriately and reliably. Furthermore, with this configuration, an estimation model can be generated appropriately and reliably. However, the expression level information and velocity information do not necessarily have to be as described above.
[0085] Traditionally, there has been no established method for estimating the RNA velocity of dropped genes, and downstream analyses have naturally been limited to a small number of passed genes. By approximating the RNA velocity of dropped genes to the RNA velocity (true value) of passed genes, it becomes possible to perform a wider range of downstream analyses.
[0086] Next, an example of this embodiment will be described, including an example of evaluating the performance of this embodiment. Five types of datasets were used to evaluate the performance of this embodiment. The table in Figure 5 shows the five types of datasets used. All five types of datasets are described in the literature. The five datasets are: a dataset of mouse hematopoietic cell differentiation (Weinreb, Weinreb, C., Rodriguez-Fraticelli, A., Camargo, F.D., and Klein, A.M. (2020). Lineage tracing on transcriptional landscapes links state to fate during differentiation. Science 367.10.1126 / science.aaw3381. (Non-patent Literature 6)), a dataset of mouse lung epithelium stimulated by bleomycin (Lung, Strunz, M., Simon, L.M., Ansari, M., Kathiriya, J.J., Angelidis, I., Mayr, C.H., Tsidiridis, G., Lange, M., Mattner, L.F., Yee, M., et al. (2020). Alveolar regeneration through a Krt8+ transitional stem cell state that persists) in human lung fibrosis. Nat Commun 11, 3559.10.1038 / s41467-020-17358-3. (Non-patent Literature 7)), mouse erythrocyte differentiation dataset (Erythroid, Pijuan-Sala, B., Griffiths, J.A., Guibentif, C., Hiscock, T.W., Jawaid, W., Calero-Nieto, F.J., Mulas, C., Ibarra-Soria, X., Tyser, R.C.V., Ho, D.L.L., et al. (2019). A single-cell molecular map of mouse gastrulation and early organogenesis. Nature 566,490-495.10.1038 / s41586-019-0933-9. (Non-Patent Literature 8)), a dataset of mouse pancreatic endocrine cell differentiation (Pancreas, Bastidas-Ponce, A., Tritschler, S., Dony, L., Scheibner, K., Tarquis-Medina, M., Salinno, C., Schirge, S., Burtscher, I., Bottcher, A., Theis, F.J., et al. (2019). Comprehensive single cell mRNA profiling reveals a detailed roadmap for pancreatic endocrinogenesis. Development 146.10.1242 / dev.173849. (Non-Patent Literature 9)), and a dataset of human bone marrow (Bone marrow, Setty, M., Kiseliovas, V., Levine, J., Gayoso, A., Mazutis, L., and Pe'er, D. (2019). Characterization of cell fate probabilities in single-cell data with Palantir. Nat. Biotechnol. 37, 451-460.10.1038 / s41587-019-0068-4. (Non-patent document 10). Note that the references listed in parentheses above are those that describe the dataset.
[0087] The table in Figure 5 shows, for each of the five datasets, the cited original paper, the method by which the cells in the dataset were obtained, the organism from which the cells in the dataset were obtained, the number of cells in the dataset, the total number of detected genes, the number of genes for which RNA velocity can be calculated using scVelo, and the total number of dropped genes.
[0088] The performance of this embodiment was evaluated using the cross-validation method. Furthermore, since it was not possible to obtain the correct RNA velocity for dropped genes, passed genes were used as validation data. Specifically, data from passed genes, excluding some, were used to generate an estimation model, and the RNA velocity of those passed genes was estimated using the generated estimation model. The performance of this embodiment (estimation model) was evaluated by examining the correlation between the estimated RNA velocity (estimated value, predicted value) and the RNA velocity (true value) calculated using existing tools.
[0089] An example of this embodiment using a mouse hematopoietic cell differentiation dataset obtained via inDrop (Weinreb in the table in Figure 5) will be described. This dataset is available through Pyro-Velocity. In this example, 916 genes were extracted as passed genes by scVelo, and their RNA velocities were calculated. scVelo is a standard tool widely used in Python-based velocity analysis and was used to calculate the RNA velocity (raw RNA velocity) of the passed genes.
[0090] Figure 6(a) shows a graph visualizing the cell-specific RNA velocity values of Passed genes calculated by scVelo, embedded in two dimensions along with RNA expression levels. In this graph, each point corresponds to one cell, and the trajectory of the cell's dynamic processes is shown as a streamline (the same applies to subsequent graphs where cell-specific RNA velocity values are embedded in two dimensions along with RNA expression levels). The position (coordinates) of the points is calculated based on the RNA expression level of the gene in the cell. In the graph of Figure 6(a), the transition from undifferentiated cells to various functional cells, including monocytes and neutrophil cells, can be observed.
[0091] Next, the data for passed genes was split into training data and validation data, and the performance of the estimation model was evaluated using cross-validation. Three-fold cross-validation was performed. The training data consists of the gene data used to generate the estimation model (machine learning). The validation data consists of the gene data used to estimate RNA velocity by the estimation model, but not used to generate the estimation model (machine learning).
[0092] Figure 7 shows a graph comparing the RNA velocity calculated by scVelo for each gene in the validation data (true value, True on the horizontal axis of the graph) with the RNA velocity estimated by the estimation model (estimated value, Predicted on the vertical axis of the graph). This graph is a scatter plot of the values for each element (i.e., each cell) of the RNA velocity vector (the same applies to the graphs of true and estimated values that follow). Figure 7 shows examples of six passed genes (6430548M08Rik, Myo1d, Sp140, Acpp, Abca13, and Slc17a9). These six passed genes were the genes that showed the best accuracy in cross-validation. For the RNA velocity estimation of the genes in the validation data with the best accuracy, a high correlation coefficient of 0.95 or higher was observed between the true value and the estimated value.
[0093] Next, RNA velocity was estimated for both the training data and the validation data using the estimation model, and the correlation coefficient was calculated between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model. Figure 8(a) shows a graph of the calculated correlation coefficients. In this graph, the correlation coefficients are plotted separately for the training data and the validation data. Each point corresponds to one gene (the same applies to subsequent graphs plotting correlation coefficients). For the genes in the validation data, the median and mean correlation coefficients were 0.722 and 0.782, respectively. For each of the three groups (models) of the 3-fold cross-validation, the average of the above correlation coefficients for the training data and the validation data was calculated. The calculated values are shown in the graph in Figure 25(a). The validation data showed an average correlation coefficient of 0.700 across the three groups.
[0094] Next, the RNA velocity of the Dropped gene was estimated using a prediction model. Figure 6(b) shows a graph visualizing the cell-specific RNA velocity values of the Dropped gene estimated by the prediction model, embedded in two dimensions along with the RNA expression level. Figure 6(c) shows a graph similarly visualizing the RNA velocity of both the Passed and Dropped genes estimated by the prediction model.
[0095] Furthermore, Figure 8(b) shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by scVelo. Cosine similarity is calculated between identical cells using the RNA velocity vector of cells on the embedded dimension (reduced dimension) (the same applies to the cosine similarity graphs below). Cosine similarity is an index that shows the similarity of the dynamics of each cell on the embedded dimension. In Figure 8(b), the graph using the RNA velocity of the Dropped gene as the RNA velocity estimated by the estimation model (Dropped) and the graph using the RNA velocity of both the Passed and Dropped genes (All) are shown.
[0096] Based on these findings, we investigated whether similar cell dynamics could be observed when using RNA velocity calculated by scVelo of the Passed gene in the post-embedding dimension. In the case of the Dropped gene, although a reverse flow was observed near megacacytic cells and mast cells compared to the Passed gene case, the cell dynamics from undifferentiated cells to functional cells, and the starting point of velocity flow, which are consistent with biological findings, were observed to be similar in both cases (see Figures 6(a), 6(b), and 8(b)). The vector streamline of the combined gene group of Passed and Dropped genes eliminated the partial reverse flow and reproduced a flow closer to that of the Passed gene case (see Figures 6(a), 6(c), and 8(b)).
[0097] Next, we show examples of RNA velocity estimated by the estimation model for representative marker genes included in the Dropped genes. Figure 9 shows a graph embedded in two dimensions with the estimated RNA velocity (Recovered velocity) and calculated gene expression vectors (Expression) for the representative marker genes Cd74 (Cluster of differentiation 74) and Lcn2 (Lipocalin-2). Cd74 and Lcn2 are genes that were identified as Dropped genes by both tools. Cd74 is a marker for monocytes and monocyte-derived dendritic cells and is known to play an important role in antigen presentation. Consistent with this finding, the estimated RNA velocity of Cd74 emphasized expression induction in line with differentiation from progenitor cells to monocytes. Lcn2, also known as Neutrophil gelatinase-associated lipocalin (Ngal), is a highly expressed marker in Neutrophil that plays a protective role against bacterial infection and modulates oxidative stress. The predicted RNA velocity of Lcn2 was in good agreement with the trend of increased expression associated with Neutrophil maturation.
[0098] We changed the tool (base tool) for calculating RNA velocity (Raw RNA velocity) from scVelo to UnitVelo and performed the same validation on the mouse hematopoietic cell differentiation dataset (Weinreb in the table in Figure 5). As described above, UnitVelo is a tool that models the dynamics of gene expression using a radial basis function. Unlike the bottom-up approach that derives unspliced mRNA or spliced mRNA from transcription rate defined as a step function, UnitVelo performs parameter estimation and modeling in a top-down manner, starting from the time evolution of spliced mRNA. Using UnitVelo, we obtained RNA velocity for 1348 passed genes in this dataset.
[0099] Figure 10(a) shows a graph visualizing the cell-specific RNA velocity values of the Passed gene calculated by UniTVelo, embedded in two dimensions along with the RNA expression level. The cell dynamics mediated by the Passed gene in UniTVelo showed a flow from undifferentiated to monocyte and basolateral, although there were minor differences in the flow around lymphoids and pDCs compared with scVelo in the 2D-embedded dimension (see Figures 6(a) and 10(a)).
[0100] Figure 11 shows a graph comparing the RNA velocity (true value, True on the horizontal axis of the graph) calculated by UniTVelo for each gene in the validation data with the RNA velocity (estimated value, Predicted on the vertical axis of the graph) estimated by the estimation model. Figure 11 shows examples of six passed genes (Ms4a6c, Lrp12, s100a10, Dab2, Ctsh, and Lilrb3). These six passed genes were the genes with the best accuracy in cross-validation. The gene with the highest estimation accuracy showed a high correlation coefficient of approximately 0.999.
[0101] Figure 12(a) shows a graph of the correlation coefficient between the RNA velocity (true value) calculated by UniTVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model. In this graph, the correlation coefficients are plotted separately for training data and validation data. Figure 25(b) shows a graph of the average values of the above correlation coefficients for the training data and validation data for each of the three groups in the 3-fold cross-validation. In the group that performed best in the cross-validation, the mean and median correlation coefficients with respect to the validation data were 0.639 and 0.871, respectively (see Figure 12(a)), and the average correlation coefficient for the validation data for the three groups was 0.624 (see Figure 25(b)).
[0102] Figure 10(b) shows a graph visualizing the cell-specific RNA velocity values of 489 Dropped genes estimated by the estimation model, embedded in two dimensions along with their RNA expression levels. Figure 10(c) shows a graph similarly visualizing the RNA velocity of both Passed and Dropped genes estimated by the estimation model.
[0103] Furthermore, Figure 12(b) shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by UniTVelo. In Figure 12(b), the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0104] A comparison of the dimensional streamlines of the Dropped and Passed genes after embedding revealed that the flow from undifferentiated to differentiated cells such as monocytes and basolaterals was similar for both (see Figures 10(a) and 10(b)). Furthermore, the similarity of cell-specific vectors in the dimensional after embedding was also high (see Figure 12(b)).
[0105] Figure 13 shows a two-dimensional embedded graph of the estimated RNA velocity (recovered velocity) and calculated gene expression vectors (expression) for representative marker genes, cd74 and Lcn2. The estimated RNA velocities for Cd74 and Lcn2 highlight transcriptional induction in the processes from undifferentiated to monocyte maturation and from undifferentiated to neutrophil, respectively.
[0106] To verify the applicability to a wider range of datasets, the dataset was changed to a dataset of bleomycin-stimulated mouse lung epithelium (Lung in the table in Figure 5), and the same verification was performed. This dataset is available through CellRank. As a tool for calculating RNA velocity (Raw RNA velocity), scVelo with standard parameters was used first. Using scVelo calculations, RNA velocity for 490 passed genes was obtained for this dataset.
[0107] Figure 14(a) shows a graph visualizing the cell-specific RNA velocity values of the Passed gene calculated by scVelo, embedded together with RNA expression levels in two dimensions. As shown in the graph, the cell differentiation flow based on the calculated RNA velocity was reproduced.
[0108] Figure 15 shows a graph comparing the RNA velocity (true value, represented by True on the horizontal axis of the graph) calculated by scVelo for each gene in the validation data with the RNA velocity (estimated value, represented by Predicted on the vertical axis of the graph) estimated by the estimation model. Figure 15 shows examples of six passed genes (Ccdc170, Dnah3, Rsph10b, Cdhr3, Gsta3, and Ccdc141). The gene group with the best accuracy showed a correlation coefficient of 0.950 or higher.
[0109] Figure 16(a) shows a graph of the correlation coefficient between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model. In this graph, the correlation coefficients are plotted separately for the training data and the validation data. Figure 25(c) shows a graph of the average values of the above correlation coefficients for the training data and the validation data for each of the three groups of 3-fold cross-validation. The median correlation coefficients for each gene targeted for estimation with respect to the 327 passed genes and the 163 genes in the validation data were 0.768 and 0.791, respectively (see Figure 16(a)), and no tendency towards overfitting was observed in any of the cross-validations (see Figure 25(c)).
[0110] Figure 14(b) shows a graph visualizing the cell-specific RNA velocity values of 1132 Dropped genes estimated by the estimation model, embedded in two dimensions along with their RNA expression levels. Figure 14(c) shows a graph similarly visualizing the RNA velocity of both Passed and Dropped genes estimated by the estimation model.
[0111] Furthermore, Figure 16(b) shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by scVelo. In Figure 16(b), the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0112] The flow from Goblet cells to Club to Ciliated cells and then to Ciliated cells, the flow from Club activated cells to Club cells in different states and MHC-2+ Club cells, and the flow from ABC to AT1 were all observed in common across all post-embedding dimensions (see Figures 14(a), 14(b), and 14(c)).
[0113] Next, we show examples of RNA velocity estimated by the estimation model for representative marker genes included in the Dropped genes. Figure 17 shows a graph embedded in two dimensions with the estimated RNA velocity (Recovered velocity) and calculated gene expression vectors (Expression) for the representative marker genes Tuba1a (Tubulin Alpha 1a) and Timp2 (Tissue inhibitor of metalloproteinases 2). Tuba1a and Timp2 are genes that were identified as Dropped genes by both tools.
[0114] Tuba1a is a marker of ciliated cells in the airways and is a factor necessary for ciliary formation and maturation. The predicted RNA velocity of Tuba1a suggested transcriptional induction of this gene along the club-to-ciliated to ciliated cluster, which was consistent with the specific increased expression of this gene in ciliated cells. Timp2 is known to be upexpressed in the airway epithelium in animal models of bleomycin-induced lung injury for pulmonary fibrosis studies, and it is thought to contribute to fibrosis by altering the activity of matrix metalloproteinases (MMPs) and regulating the amount of extracellular matrix. In this dataset, Timp2 expression was high in alveolar type cells such as ABC, which increase with bleomycin stimulation, and consistent with this observation, the predicted RNA velocity of Timp2 was high in AT2 activated, ABC, and AT1 cells.
[0115] The tool used to calculate RNA velocity (Raw RNA velocity) was changed from scVelo to UnitVelo, and the same validation was performed on the bleomycin-stimulated mouse lung epithelium dataset (Lung in the table in Figure 5). Using UnitVelo, RNA velocity was obtained for 1078 passed genes in this dataset.
[0116] Figure 18(a) shows a graph visualizing the cell-specific RNA velocity values of Passed genes calculated by UnitVelo, embedded in two dimensions along with RNA expression levels. As shown in this graph, changes in the cell state can be visualized using the calculated RNA velocity.
[0117] Figure 19 shows a graph comparing the RNA velocity calculated by UnitVelo for each gene in the validation data (true value, True on the horizontal axis of the graph) with the RNA velocity estimated by the estimation model (estimated value, Predicted on the vertical axis of the graph). Figure 19 shows examples of six passed genes (Calml4, Capsl, 1700026L06Rik, Chchd6, Nek5, and Rsph10b). Similar to the validation with UnitVelo on the mouse hematopoietic cell differentiation dataset (Weinreb in the table in Figure 5), a very strong correlation was shown between the true value and the estimated value for the gene with the highest estimation accuracy.
[0118] Figure 20(a) shows a graph of the correlation coefficient between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model. In this graph, the correlation coefficients are plotted separately for the training data and the validation data. Figure 25(d) shows a graph of the average values of the above correlation coefficients for the training data and the validation data for each of the three groups in the 3-fold cross-validation. In the group that performed best in the cross-validation, the mean and median correlation coefficients with respect to the validation data were 0.766 and 0.988, respectively (see Figure 20(a)).
[0119] Figure 18(b) shows a graph visualizing the cell-specific RNA velocity values of 691 Dropped genes estimated by the estimation model, embedded in two dimensions along with their RNA expression levels. Figure 18(c) shows a graph similarly visualizing the RNA velocity of both Passed and Dropped genes estimated by the estimation model.
[0120] Furthermore, Figure 20(b) shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by scVelo. In Figure 20(b), the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0121] From these findings, we confirmed the differences in flow between the RNA velocity of dropped genes and the RNA velocity of passed genes in the post-embedding dimension. The flow from the club-to-ciliated precursor cell to ciliated, and the flow from AT2 activated to AT1 and ABC, were similar for both (see Figures 18(a), 18(b), and 18(c)).
[0122] Figure 21 shows a graph embedded in two dimensions with the estimated RNA velocity (Recovered velocity) and calculated gene expression vectors (Expression) of representative marker genes, Tuba1a and Timp2. As expected, the estimated RNA velocity of Tuba1a was elevated in the ciliated cell lineage, and the estimated RNA velocity of Timp2 was high in clusters associated with bleomycin stimulation, such as AT2 activated cells, which is consistent with the specific high expression behavior of this gene in AT1 and ABC.
[0123] To investigate the impact of different tools used to calculate RNA velocity (Raw RNA velocity) on RNA velocity estimation in this embodiment, this embodiment was implemented using a mouse erythrocyte differentiation dataset (Erythroid in the table in Figure 5). This dataset is available through scVelo. In this dataset, several gene groups with multiple rate kinetics contradict the processes of conventional models, and it is known that a simple application of scVelo can lead to distorted genealogical inference.
[0124] Using scVelo calculations, RNA velocity values for 248 passed genes were obtained for this dataset. Figure 22(a) shows a graph visualizing the cell-specific RNA velocity values of passed genes calculated by scVelo, embedded in two dimensions along with RNA expression levels. As expected from the above, in this validation, a portion of the graph above showed the trajectory of differentiation from mature erythroids to more immature ones.
[0125] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the passed genes. RNA velocity was estimated for 1287 dropped genes using the generated model. Figure 22(b) shows a graph visualizing the cell-specific RNA velocity values of dropped genes estimated by the model, embedded in two dimensions along with the RNA expression levels. Figure 22(c) shows a graph similarly visualizing the RNA velocity of both passed and dropped genes estimated by the model.
[0126] Figure 26(a) shows a graph of the average correlation coefficients between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0127] The flow induced by the Dropped gene alone showed a similar trend to that of the Passed gene, namely, a retrograde flow from mature erythroids towards more progenitor cells, which was reproduced in the post-embedding dimension (see Figures 22(a) and 22(b)).
[0128] On the other hand, it is already known that applying UnitVelo to a mouse erythrocyte differentiation dataset (Erythroid in the table in Figure 5) corrects the trajectory toward the expected maturation of erythroid cells.
[0129] Using UnitVelo, we obtained RNA velocity values for 1015 passed genes in this dataset. Figure 23(a) shows a graph visualizing the cell-specific RNA velocity values of passed genes calculated by UnitVelo, embedded in two dimensions along with RNA expression levels. From this graph, it was confirmed that the trajectory was corrected towards the expected maturation of erythroid cells.
[0130] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the passed genes. For the dropped genes, the RNA velocity was estimated using the generated model. Figure 23(b) shows a graph visualizing the cell-specific RNA velocity values of the dropped genes estimated by the model, embedded in two dimensions along with the RNA expression levels. Figure 23(c) shows a graph similarly visualizing the RNA velocity of both the passed and dropped genes estimated by the model.
[0131] Figure 26(b) shows graphs of the average correlation coefficients between the RNA velocity (true value) calculated by UnitVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0132] The flow due to dropped genes alone is similar to that of passed genes. These observations suggest that, regardless of whether the calculation of RNA velocity (raw RNA velocity) by the tool is based on biological findings, the flow due to RNA velocity of dropped genes estimated based on the obtained information of passed genes is similar to that of passed genes.
[0133] Next, UnitVelo, which showed a linage consistent with biological findings in the dimensional space after embedding, was used as a tool to calculate RNA velocity (Raw RNA velocity) for six marker genes included in the Dropped gene (Klf1 (Kruppel-like factor 1), Prdx2, Hba-a2, Alas2, Hba-x, and Hbb-bt (Hemoglobin beta adult t chain)) using a prediction model. Figure 24 shows a graph embedded in two dimensions with the estimated RNA velocity (Recovered velocity) and calculated gene expression vectors (Expression) for the six marker genes.
[0134] All six marker genes are necessary for red blood cell maturation, and their transcription is suggested to be induced in the lineage from progenitor to most mature red blood cell. Interestingly, the transcription factors Klf1 and Hbb-bt had particularly low expression levels, but their transcription induction in this lineage was appropriately predicted. For further RNA velocity-dependent downstream analyses of these genes, the estimation method of this embodiment may offer a possibility.
[0135] Next, this embodiment was carried out using a dataset of mouse pancreatic endocrine cell differentiation (Pancreas in the table in Figure 5). This dataset is available through scVelo. RNA velocity of the Passed gene was obtained for this dataset by calculation using scVelo. Figure 28(a) shows a graph in which the cell-specific RNA velocity values of the Passed gene calculated by scVelo are embedded in two dimensions along with the RNA expression level and visualized.
[0136] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the Passed gene. For the Dropped gene, the RNA velocity was estimated using the generated model. Figure 28(b) shows a graph visualizing the cell-specific RNA velocity values of the Dropped gene estimated by the model, embedded in two dimensions along with the RNA expression level. Figure 28(c) shows a graph similarly visualizing the RNA velocity of both the Passed and Dropped genes estimated by the model.
[0137] Furthermore, Figure 29 shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by scVelo. In Figure 29, the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0138] Figure 26(c) shows a graph of the average correlation coefficients between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0139] Next, RNA velocity of the Passed gene was obtained for this dataset using UniTVelo calculations. Figure 30(a) shows a graph visualizing the cell-specific RNA velocity values of the Passed gene calculated by UniTVelo, embedded in two dimensions along with the RNA expression level.
[0140] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the passed genes. For the dropped genes, the RNA velocity was estimated using the generated model. Figure 30(b) shows a graph visualizing the cell-specific RNA velocity values of the dropped genes estimated by the model, embedded in two dimensions along with the RNA expression levels. Figure 30(c) shows a graph similarly visualizing the RNA velocity of both the passed and dropped genes estimated by the model.
[0141] Furthermore, Figure 31 shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by UniTVelo. In Figure 31, the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0142] Figure 26(d) shows graphs of the average correlation coefficients between the RNA velocity (true value) calculated by UnitVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0143] Next, this embodiment was carried out using a human bone marrow dataset (Bone marrow in the table in Figure 5). This dataset is available through scVelo. RNA velocity of the Passed gene was obtained for this dataset by calculation using scVelo. Figure 32(a) shows a graph in which the cell-specific RNA velocity values of the Passed gene calculated by scVelo are embedded in two dimensions along with the RNA expression level and visualized.
[0144] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the Passed genes. For the Dropped genes, the RNA velocity was estimated using the generated model. Figure 32(b) shows a graph visualizing the cell-specific RNA velocity values of the Dropped genes estimated by the model, embedded in two dimensions along with the RNA expression levels. Figure 32(c) shows a graph similarly visualizing the RNA velocity of both the Passed and Dropped genes estimated by the model.
[0145] Furthermore, Figure 33 shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by scVelo. In Figure 33, the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0146] Figure 27(a) shows a graph of the average correlation coefficients between the RNA velocity (true value) calculated by scVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0147] Next, the RNA velocity of the Passed gene was obtained for this dataset using UniTVelo calculations. Figure 34(a) shows a graph visualizing the cell-specific RNA velocity values of the Passed gene calculated by UniTVelo, embedded in two dimensions along with the RNA expression level.
[0148] Next, a model for estimation was generated from the calculated gene expression vectors and RNA velocity of the passed genes. For the dropped genes, the RNA velocity was estimated using the generated model. Figure 34(b) shows a graph visualizing the cell-specific RNA velocity values of the dropped genes estimated by the model, embedded in two dimensions along with the RNA expression levels. Figure 34(c) shows a graph similarly visualizing the RNA velocity of both the passed and dropped genes estimated by the model.
[0149] Furthermore, Figure 35 shows the distribution of cosine similarity between the RNA velocity estimated by the estimation model and the RNA velocity of the Passed gene calculated by UniTVelo. In Figure 35, the graph shows the case where the RNA velocity of the Dropped gene is used as the RNA velocity estimated by the estimation model (Dropped), and the graph shows the case where the RNA velocity of both the Passed gene and the Dropped gene is used (All).
[0150] Figure 27(b) shows a graph of the average correlation coefficients between the RNA velocity (true value) calculated by UnitVelo for each gene and the RNA velocity (estimated value) estimated by the estimation model for each of the three groups in the 3-fold cross-validation, for both the training data and the validation data.
[0151] For the mouse pancreatic endocrine cell differentiation dataset (Pancreas in the table in Figure 5), scVelo reproduced the cyclical trajectory of ductal cells along the cell cycle, whereas UniTVelo did not show these trajectories. On the other hand, for the human bone marrow dataset (Bone marrow in the table in Figure 5), similar to the results for the mouse erythrocyte differentiation dataset (Erythroid in the table in Figure 5), scVelo showed a flow that reversed from definitively differentiated cells to more undifferentiated cells, whereas UniTVelo obtained a velocity flow based on biological knowledge.
[0152] In this embodiment, the relationship between the RNA expression level of a passed gene and its RNA velocity is learned. Therefore, the RNA velocity estimated for a dropped gene reproduced a flow that is, in principle, similar to that of a passed gene. This fact highlights the importance of determining whether the RNA velocity calculation obtained by the base tool is consistent with biological findings when performing RNA velocity estimation and downstream analysis using these values according to this embodiment.
[0153] The above-described embodiment is merely one example of an application of this embodiment. By utilizing this embodiment, it becomes possible to gain new insights into cell trajectories and to enhance the number of genes that can be analyzed in differential kinetics analysis. This embodiment is a particularly useful tool in fields such as development, immunology, and cancer, where deeper insights into cell dynamics at the single-cell level are required. Applying this embodiment to datasets contributes to the discovery of new regulatory elements and pathways involved in cell differentiation and state transitions, helping to deepen our understanding of cell dynamics and differentiation.
[0154] The details of the above example are described below. The above dataset is downloaded in h5ad format. Each data downloaded in h5ad format was loaded using a scVelo function and used as anndata. The calculation of RNA velocity using scVelo was performed following the tutorial. Filtering and normalization functions were executed with min_shared_counts=20 and highly variable genes=2000, and n_pcs=30 and n_neighbors=30 were specified for moment calculation. The mode for RNA velocity calculation in scVelo was set to stochastic.
[0155] RNA velocity calculation using UniTVelo was also performed following the tutorial. Similar to scVelo, 2000 highly variable genes were extracted. Many of the other specified parameters followed the tutorial, with R2_ADJUST=True, LEARNING_RATE=0.01, FIT_OPTION=1 (unified-time mode), and MAX_ITER=10000.
[0156] For Passed genes for which RNA velocity could be calculated, elements marked True in anndata.var.["velocity_genes"] were extracted. Elements marked False were defined as Dropped genes for which RNA velocity could not be calculated. Furthermore, for Passed genes, the RNA velocity was added to the count value (current expression value), a process used in velocityto and other tools, to obtain the projected future expression value. In the future expression value, any values that became negative after adding RNA velocity were clipped to 0. These processes were performed for application to the count value in MAGIC. As a result, the current and future expression values of Passed genes, as well as the current expression value of Dropped genes, were obtained.
[0157] MAGIC was first applied to the data matrix of current expression values for all genes, including both passed and dropped genes, to calculate an exponentiated Markov affinity matrix for the completion of the count matrix. The calculated matrix was then applied to the data matrices of current and future expression values, respectively, to reconstruct the completion of the count matrix. RNA velocity values were defined as the difference between future and current expression values after MAGIC preprocessing. Furthermore, dropped genes to be estimated were those whose mean current expression value was in the first quartile or higher. That is, data on current expression values and RNA velocity for genes whose mean current expression value was in the bottom 25% or lower were deleted. Finally, z-score normalization was performed on the data matrices of current expression values for both passed and dropped genes. Similarly, z-score normalization was performed on the RNA velocity data matrices for passed genes.
[0158] Nested cross-validation was used for data splitting in the cross-validation. Three-fold cross-validation was performed as the outer-fold. Pearson product-moment correlation coefficients were adopted as the evaluation metric.
[0159] The hyperparameters, loss function, and optimization algorithm of the estimation model were optimized using inner-fold analysis. Specifically, the outer-fold training data was further divided into inner-folds using 3-fold cross-validation, and the minimum loss value of the validation data in each fold was calculated. The Optuna package was used to search within the defined range of values.
[0160] The Optuna search involved minimizing the mean value of all inner-fold data as the objective function, performing 100 trials using the Tree-Structured Parzen Estimator (TPE) algorithm, and pruning based on the median of each trial. Using the best parameter set obtained from the Optuna search, an estimation model was trained on the outer-fold training data, and estimation was performed using the model at the epoch that minimized the loss on the validation data for each outer-fold. The mean of the estimation results for all outer-fold data was used as the final estimate for the validation data. The estimate was scaled back to its original size using the statistic (mean / variance) used during z-score normalization of the RNA velocity data matrix, and the correlation coefficient with the true RNA velocity before normalization was calculated.
[0161] The RNA velocity vectors were calculated at the cell level, each having a length equal to the number of genes, and stored in anndata.layers["velocity"]. The RNA velocities of both passed and dropped genes were then plotted as streamlines on a two-dimensional embedding using the scVelo pl. velocity_embedding_stream function. The base embeddings were reused from those already calculated for the same datasets. Specifically, for the mouse hematopoietic cell differentiation dataset (Weinreb in the table in Figure 5), SPRING was used as the embedding. For the mouse lung epithelial bleomycin-stimulated dataset (Lung in the table in Figure 5), the mouse erythrocyte differentiation dataset (Erythroid in the table in Figure 5), and the mouse pancreatic endocrine cell differentiation dataset (Pancreas in the table in Figure 5), UMAP was used as the embedding. For the human bone marrow dataset (Bone marrow in the table in Figure 5), t-SNE was used as the embedding.
[0162] Next, we will describe the estimation program and the learning program for executing the processing performed by the series of estimation systems 10 and learning systems 20 described above. As shown in Figure 36(a), the estimation program 100 is stored in a program storage area 111 formed on a computer-readable recording medium 110 that is inserted into and accessed by a computer, or is provided by a computer. The recording medium 110 may be a non-temporary recording medium.
[0163] The estimation program 100 comprises an acquisition module 101 and an estimation module 102. The functions realized by executing the acquisition module 101 and the estimation module 102 are the same as the functions of the acquisition unit 11 and the estimation unit 12 of the estimation system 10 described above.
[0164] As shown in Figure 36(b), the learning program 200 is stored in a program storage area 211 formed on a computer-readable recording medium 210 that is inserted into and accessed by the computer, or is provided by the computer. The recording medium 210 may be a non-temporary recording medium. If the estimation program 100 and the learning program 200 are executed on the same computer, the recording medium 210 may be the same as the recording medium 110.
[0165] The learning program 200 comprises a learning acquisition module 201 and a learning module 202. The functions realized by executing the learning acquisition module 201 and the learning module 202 are the same as those of the learning acquisition unit 21 and the learning unit 22 of the learning system 20 described above.
[0166] Furthermore, the estimation program 100 and the learning program 200 may be configured such that part or all of them are transmitted via a transmission medium such as a communication line, received by other devices, and recorded (including installation). Also, each module of the estimation program 100 and the learning program 200 may be installed on any of multiple computers, not just one. In that case, the series of processes described above will be performed by a computer system consisting of these multiple computers.
[0167] The estimation system, estimation method, estimation program, learning system, learning method, learning program, and estimation model of this disclosure have the following configurations: [1] An estimation system comprising: an acquisition means for acquiring estimation expression level information indicating the amount of RNA expression in a cell of a gene to be estimated for RNA velocity; and an estimation means for estimating the RNA velocity of the gene to be estimated from the estimation expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression in a cell of a gene as input and outputs velocity information indicating the RNA velocity of the gene. [2] The estimation system according to [1], wherein the gene to be estimated is a different gene from the gene related to the information used in the machine learning. [3] The estimation system according to [1] or [2], wherein the gene to be estimated is a gene for which RNA velocity cannot be calculated using the tools for calculating RNA velocity, scVelo, UniTVelo, and velocyto. [4] The estimation system according to any one of [1] to [3], wherein the expression level information is a gene expression vector having values indicating the RNA expression levels for each of a plurality of cells set in advance, and the velocity information is an RNA velocity vector having the values for each of the plurality of cells. [5] An estimation method which is a method for operating the estimation system, comprising: an acquisition step of acquiring estimation expression level information indicating the RNA expression level of a gene to be estimated in cells; and an estimation step of estimating the RNA velocity of the gene to be estimated from the estimation expression level information acquired in the acquisition step, using an estimation model that is generated by machine learning and takes expression level information indicating the RNA expression level of a gene in cells as input and outputs velocity information indicating the RNA velocity of the gene.[6] An estimation program that causes a computer to function as: an acquisition means for acquiring estimation expression level information indicating the amount of RNA expression in a cell of a gene to be estimated for RNA velocity; and an estimation means for estimating the RNA velocity of the target gene from the estimation expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression in a cell of a gene as input and outputs velocity information indicating the RNA velocity of the gene. [7] A learning system comprising: a learning acquisition means for acquiring learning expression level information indicating the amount of RNA expression in a cell of a learning gene, and learning velocity information indicating the RNA velocity of the learning gene; and a learning means for generating an estimation model that uses the learning expression level information and learning velocity information acquired by the learning acquisition means to perform machine learning and take expression level information indicating the amount of RNA expression in a cell of a gene as input and output velocity information indicating the RNA velocity of the gene. [8] The learning system according to [7], wherein the expression level information is a gene expression vector having values indicating the RNA expression levels for each of a plurality of cells set in advance, and the velocity information is an RNA velocity vector having values for each of the plurality of cells. [9] A learning method which is a method of operating the learning system, comprising: a learning acquisition step of acquiring learning expression level information indicating the RNA expression level of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and a learning step of performing machine learning using the learning expression level information and learning velocity information acquired in the learning acquisition step to generate an estimation model that takes expression level information indicating the RNA expression level of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene as input.
[10] A learning program that causes a computer to function as: learning acquisition means for acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and learning means for performing machine learning using the learning expression level information and learning velocity information acquired by the learning acquisition means to generate an estimation model that takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
[11] An estimation model generated by the learning system described in [7], which causes a computer to function to take expression level information indicating the amount of RNA expression of a gene in a cell as input and output velocity information indicating the RNA velocity of the gene.
[0168] 10... Estimation system, 11... Acquisition unit, 12... Estimation unit, 20... Learning system, 21... Acquisition unit for learning, 22... Learning unit, 100... Estimation program, 101... Acquisition module, 102... Estimation module, 110... Recording medium, 111... Program storage area, 200... Learning program, 201... Acquisition module for learning, 202... Learning module, 210... Recording medium, 211... Program storage area.
Claims
1. An estimation system comprising: an acquisition means for acquiring estimated expression level information indicating the amount of RNA expression of a gene to be estimated in a cell; and an estimation means for estimating the RNA velocity of the target gene from the estimated expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
2. The estimation system according to claim 1, wherein the gene to be estimated is a different gene from the gene related to the information used in the machine learning.
3. The estimation system according to claim 1 or 2, wherein the gene to be estimated is a gene for which RNA velocity cannot be calculated using the tools for calculating RNA velocity, scVelo, UniTVelo, and velocyto.
4. The estimation system according to any one of claims 1 to 3, wherein the expression level information is a gene expression vector having values indicating the expression levels of RNA for each of a plurality of pre-set cells, and the velocity information is an RNA velocity vector having the values for each of the plurality of cells.
5. An estimation method which is a method for operating an estimation system, comprising: an acquisition step of acquiring estimation expression level information indicating the amount of RNA expression of a gene to be estimated for RNA velocity in a cell; and an estimation step of estimating the RNA velocity of the gene to be estimated from the estimation expression level information acquired in the acquisition step, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
6. An estimation program that causes a computer to function as: an acquisition means for acquiring estimated expression level information indicating the amount of RNA expression of a gene to be estimated in a cell; and an estimation means for estimating the RNA velocity of the target gene from the estimated expression level information acquired by the acquisition means, using an estimation model that is generated by machine learning and takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene.
7. A learning system comprising: a learning acquisition means for acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and a learning means for generating an estimation model that performs machine learning using the learning expression level information and learning velocity information acquired by the learning acquisition means, inputting expression level information indicating the amount of RNA expression of a gene in a cell and outputting velocity information indicating the RNA velocity of the gene.
8. The learning system according to claim 7, wherein the expression level information is a gene expression vector having values indicating the RNA expression levels for each of a plurality of pre-set cells, and the velocity information is an RNA velocity vector having values for each of the plurality of cells.
9. A learning method which is a method for operating a learning system, comprising: a learning acquisition step of acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and a learning step of performing machine learning using the learning expression level information and learning velocity information acquired in the learning acquisition step to generate an estimation model that takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene as output.
10. A learning program that causes a computer to function as: a learning acquisition means for acquiring learning expression level information indicating the amount of RNA expression of a learning gene in a cell, and learning velocity information indicating the RNA velocity of the learning gene; and a learning means for performing machine learning using the learning expression level information and learning velocity information acquired by the learning acquisition means to generate an estimation model that takes expression level information indicating the amount of RNA expression of a gene in a cell as input and outputs velocity information indicating the RNA velocity of the gene as output.
11. An estimation model generated by the learning system described in claim 7, which causes a computer to function by taking expression level information indicating the amount of RNA expression of a gene in a cell as input and outputting velocity information indicating the RNA velocity of the gene.
