Cytodynamics prediction method for pseudo-sequential manifold alignment and graph attention

Abstract

The invention discloses a cell dynamics prediction method for pseudo-sequential manifold alignment and graph attention, and belongs to the technical field of bioinformatics and artificial intelligence crossing in a new generation of information technology. The method comprises the following steps: firstly, acquiring single-cell transcriptome data, processing to obtain a single-cell transcriptome feature matrix which is orderly arranged based on a pseudo time sequence track, and acquiring batch transcriptome data and a cell subset proportion standard value at a plurality of discrete clinical time points at the same time; respectively extracting a pseudo time sequence feature sequence and a real time sequence feature sequence through a double-flow heterogeneous neural network; the method comprises the following steps: correcting an individual time sequence phase based on an alignment module capable of realizing Microsoft dynamic time warping, constructing a cell interaction graph by utilizing an aligned pseudo time sequence feature sequence, aggregating microenvironment interaction information through a graph attention network model, and predicting the proportion of a target cell subset at a future time point. According to the method, the defect that dynamic prediction is difficult to carry out in the prior art is overcome, and accurate prediction of cell proliferation dynamics, especially peak time of a key window period, is realized.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and artificial intelligence in the next generation of information technology, and specifically relates to a cell dynamics prediction method for pseudo-temporal manifold alignment and graph attention. Background Technology

[0002] With the development of high-throughput sequencing technology, analyzing the cellular composition of complex tissue microenvironments (TMEs) has become crucial for precision medicine. Early cell proportion deconvolution algorithms were mainly based on linear support vector regression (SVR). CIBERSORT, proposed by Newman et al., is a cornerstone of this field, successfully estimating the proportions of 22 immune cell types using SVR. Subsequently, to improve resolution and address platform differences, Newman's team further developed CIBERSORTx, introducing a batch correction mode to achieve cell type-specific expression inference from Bulk-Seq data. With the popularization of single-cell sequencing (scRNA-Seq), using single-cell data as a reference ensemble has become mainstream. The MuSiC algorithm proposed by Wang et al. effectively addresses the biological variation problem between the reference set and the test sample through cross-subject weighted transfer learning. In recent years, deep learning has been widely used due to its nonlinear feature extraction capabilities. Scaden, developed by Menden et al., uses simulated data to train deep neural networks, eliminating dependence on specific marker genes. Chen et al. proposed the TAPE algorithm, employing a transfer learning autoencoder framework, which improves prediction accuracy while preserving high-dimensional gene expression feature structures.

[0003] Since 2022, research in this field has evolved towards spatial resolution and high precision. The BayesPrism developed by Chu et al. integrates scRNA-Seq and Bulk data using Bayesian methods, achieving high-precision prediction of cell subtypes and gene expression in complex tumors such as glioblastoma. CARD, proposed by Ma et al., combines spatial transcriptome information and uses a conditional autoregressive model to analyze the cellular composition of tissue sections. However, most of these studies focus on refining the "spatial" or "static proportions," and have not yet effectively addressed the dynamic mapping problem in the "temporal dimension."

[0004] Existing technologies in China mainly focus on parameter optimization or sample expansion of static regression models. For example, patent CN113707216A introduces the Particle Swarm Optimization (PSO) algorithm to optimize the kernel function and penalty factor of Support Vector Regression (SVR), which improves the accuracy of linear unmixing to some extent. Patent CN117672359A utilizes the adversarial training mechanism of Generative Adversarial Networks (GANs) to alleviate the problem of sparse training samples by generating pseudo-single-cell expression matrices. In addition, patent CN120808894A attempts to address estimation bias caused by the heterogeneity of the tumor microenvironment by constructing an iteratively updated reference matrix.

[0005] While these technical solutions have improved the deconvolution correlation coefficient to some extent for specific datasets, they generally treat cell development as a static snapshot and lack a mechanism to align the "pseudo-temporal" data of single-cell data with the "physical time" of clinical flow cytometry data, resulting in an inability to accurately predict the cell proliferation dynamics during treatment.

[0006] Defects and shortcomings of existing technology:

[0007] 1. Lack of time alignment mechanism: Existing technology treats biological samples at different time points as isolated snapshots, and cannot establish a mapping relationship between single-cell pseudo-time trajectory and the real clinical physical time axis, resulting in the inability to accurately predict the lag or advance effect of cell proliferation.

[0008] 2. Ignoring microenvironment interactions: Existing technologies often assume that cell subpopulations are statistically independent, failing to capture the nonlinear synergistic or antagonistic feedback regulation between cells in the immune microenvironment through signal communication using mathematical models.

[0009] 3. Poor generalization of cross-modal features: Existing technologies lack effective cross-domain manifold alignment methods, which can easily lead to severe distribution drift when integrating high-dimensional sparse sequencing data with low-dimensional continuous streaming data, resulting in a significant decrease in the generalization performance of the model in real clinical cohorts. Summary of the Invention

[0010] To address the aforementioned issues, this invention proposes a cell dynamics prediction method based on pseudo-temporal manifold alignment and graph attention. This method solves the technical challenges faced by existing technologies in inferring dynamic cell processes using discrete static observation data, such as the misalignment between the microscopic developmental pseudo-temporal sequence and the macroscopic clinical physical time dimension, as well as the lack of information on cell microenvironment interactions.

[0011] The technical solution adopted in this invention is as follows:

[0012] In a first aspect, this invention proposes a cell dynamics prediction method based on pseudo-temporal manifold alignment and graph attention, comprising the following steps:

[0013] (1) Obtain single-cell transcriptome data of the target object, and after preprocessing and pseudo-time series analysis, obtain a single-cell transcriptome feature matrix based on the ordered arrangement of pseudo-time series trajectories; at the same time, obtain batch transcriptome data of the same target object at multiple discrete clinical time points and cell subpopulation ratio measured by flow cytometry, which together constitute real time series observation data.

[0014] (2) Construct a heterogeneous neural network model with two parallel channels; input the single-cell transcriptome feature matrix into the variational autoencoder of the first channel to obtain a pseudo-temporal feature sequence that characterizes the continuous development of cells in the latent space; input the batch transcriptome data in the real temporal observation data into the bidirectional long short-term memory network of the second channel to obtain the real temporal feature sequence of cell population evolution with physical time;

[0015] (3) Input the pseudo-time series feature sequence and the real time series feature sequence into the alignment module based on the configurable dynamic time warping, and output the phase-aligned pseudo-time series feature sequence;

[0016] (4) Construct a cell microenvironment interaction graph with cell subpopulations as nodes, use the phase-aligned pseudo-temporal feature sequence as the initial feature of the node, use the graph attention network model to aggregate the features of the neighboring nodes, simulate the intercellular interaction in the microenvironment, and predict the proportion of the target cell subpopulation at future time points.

[0017] (5) The model is trained by combining the alignment loss of the combined feature sequence and the prediction loss of the target cell subpopulation ratio. The trained model is then used to predict the target cell subpopulation ratio at future time points, thus characterizing the proliferation dynamics trend of the target cell subpopulation.

[0018] Furthermore, in step (1), the pseudo-time series analysis is performed using tools such as Monocle, Slingshot, PAGA, DiffusionPseudotime, or scVelo.

[0019] Further, in step (1), the preprocessing of the single-cell transcriptome data includes quality control and screening for hypervariable genes, thereby obtaining a high-quality single-cell expression matrix. Subsequently, the matrix is ​​processed using a pseudo-temporal analysis tool to obtain a one-dimensional scalar value representing the relative position of each cell during continuous development, i.e., the pseudo-temporal trajectory. Finally, the single-cell expression matrix is ​​reorganized based on this pseudo-temporal trajectory, so that the feature vectors of each row of cells are arranged in ascending order according to their pseudo-temporal values, ultimately generating a single-cell transcriptome feature matrix arranged in an ordered manner based on the pseudo-temporal trajectory.

[0020] Furthermore, in the alignment module based on soft dynamic time warping, the alignment loss is calculated based on the soft dynamic time warping algorithm. It is obtained by calculating the distance matrix between all feature points in the pseudo-time series feature sequence and the real time series feature sequence, and searching for a soft alignment path that minimizes the cumulative alignment cost among all possible alignment paths.

[0021] Furthermore, the alignment loss formula is as follows:

[0022] ;

[0023] in, It is alignment loss. It is a pseudo-temporal feature sequence. It is a true time-series feature sequence. It has a smoothing parameter The soft dynamic time warping operator, It is the distance between feature point i in the pseudo-temporal feature sequence and feature point j in the real temporal feature sequence. It is the set of all legally aligned paths. It is a specific alignment path.

[0024] Furthermore, the alignment module learns a nonlinear mapping function from cellular pseudo-time to clinical physical time based on alignment loss; it uses the nonlinear mapping function to automatically stretch, compress, or translate the pseudo-time feature sequence to achieve optimal alignment with the real time feature sequence on the time axis, correcting the temporal phase deviation caused by differences in developmental rates between individuals, and obtaining a phase-aligned pseudo-time feature sequence.

[0025] Furthermore, in step (4), the edges of the cell microenvironment interaction graph are constructed based on a known receptor-ligand interaction database or expression correlations between cell types.

[0026] Furthermore, in step (4), the process of predicting the proportion of the target cell subpopulation at future time points includes:

[0027] The cell state embedding vector containing microenvironment interaction information generated by the aggregation of the graph attention network model is input into a prediction layer, and the cell state embedding vector is mapped to one or more scalar outputs.

[0028] The scalar output is used to characterize the predicted proportion of the target cell subpopulation at one or more specified clinical time points in the future.

[0029] Secondly, this invention proposes a pseudo-temporal manifold alignment and graph attention cell dynamics prediction system to implement the aforementioned pseudo-temporal manifold alignment and graph attention cell dynamics prediction method.

[0030] Thirdly, the present invention proposes a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned pseudo-temporal manifold alignment and graph attention cell dynamics prediction method.

[0031] Fourthly, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting cell dynamics of pseudo-temporal manifold alignment and graph attention.

[0032] The beneficial effects of this invention are:

[0033] (1) This invention introduces a differentiable time alignment module, which realizes accurate mapping between microscopic pseudo-time and macroscopic physical time. This module uses a differentiable dynamic time warping algorithm to learn the nonlinear mapping relationship between microscopic cellular pseudo-time trajectory and macroscopic clinical physical time series. This mechanism can dynamically correct the temporal phase deviation caused by individual differences in immune response rate, thereby achieving accurate prediction of key dynamic parameters such as cell proliferation peak, overcoming the shortcomings of traditional methods in predicting dynamic change windows.

[0034] (2) This invention constructs a cell microenvironment interaction graph with cell subpopulations as nodes and introduces a graph attention network. Based on the phase-aligned pseudo-temporal feature sequence, the association weights of neighboring cells are dynamically calculated through the attention mechanism, enabling the model to effectively capture and utilize the nonlinear synergistic or antagonistic feedback regulation signals between cells in the microenvironment, thereby achieving more accurate trend prediction in complex heterogeneous samples and enhancing the model's explanatory power for the complex immune environment in vivo.

[0035] (3) The present invention adopts a dual-stream heterogeneous feature extraction architecture combined with a pre-training strategy. Through parallel variational autoencoders and bidirectional long short-term memory networks, it can make full use of the rich manifold structure of single-cell data to make up for the sparsity of real time series. This information complementary design makes up for the shortcomings of the sparse time points of clinical data, improves the generalization ability of sparse clinical time series data, and enables the model to maintain robust predictive performance at key time points (such as the peak period of treatment response), providing a more reliable computing tool for clinical dynamic monitoring and treatment evaluation. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a pseudo-temporal manifold alignment and graph attention-based method for predicting cell dynamics.

[0037] Figure 2 This is a schematic diagram of the framework for a pseudo-temporal manifold alignment and graph attention-based cell dynamics prediction method;

[0038] Figure 3 This is a schematic diagram illustrating the application process of the trained model in a specific scenario.

[0039] Figure 4 This is a schematic diagram of the framework of a cell dynamics prediction system based on pseudo-temporal manifold alignment and graph attention. Detailed Implementation

[0040] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0041] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0042] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0043] like Figure 1 As shown, this invention proposes a cell dynamics prediction method based on pseudo-temporal manifold alignment and graph attention, which mainly includes the following steps:

[0044] S1: Obtain a multimodal time-series mapping map containing single-cell transcriptome data, batch transcriptome data, and flow cytometry data, and process it to obtain the input features.

[0045] In this step, single-cell transcriptome data (scRNA-seq) of the target object is first acquired. After preprocessing and pseudo-time series analysis, a single-cell transcriptome feature matrix based on the ordered arrangement of pseudo-time series trajectories is obtained. At the same time, bulk transcriptome data (Bulk RNA-seq) and flow cytometry data of the same target object at multiple discrete clinical time points are acquired. The bulk transcriptome data of the same target object at multiple discrete clinical time points is used as input features, and the cell proportion data measured by flow cytometry is used as training labels to jointly construct real time series observation data.

[0046] In one specific embodiment of the present invention, the acquired single-cell transcriptome data is first subjected to quality control to remove low-quality cells and background noise, and hypervariable genes are screened to obtain a single-cell expression matrix. Subsequently, a pseudo-time series analysis tool (such as Monocle2 or Monocle3) is used to perform trajectory inference on the single-cell expression matrix, calculating a one-dimensional scalar value representing the relative position of each cell during continuous development. This refers to a pseudo-temporal trajectory; this scalar value defines the sequential order of cells along a potential developmental pathway, serving as a reference manifold reflecting the continuous evolution of cell states. Next, the single-cell expression matrix is ​​reorganized using this pseudo-temporal trajectory to construct a single-cell transcriptome feature matrix based on the ordered arrangement of the pseudo-temporal trajectory. In this feature matrix, each cell, in addition to possessing high-dimensional gene expression characteristics, also includes the aforementioned scalar value representing the developmental sequence. The characteristic matrix is ​​denoted as... Where M is the total number of effective cells after trajectory inference, and G is the number of highly variable genes screened. To reflect the time series dependencies, each row of this matrix (i.e., the feature vector of each cell) has been strictly scaled according to its corresponding pseudo-time series scalar. The data was sorted in ascending order.

[0047] For the same target subject, batch transcriptome data are collected at multiple key discrete time points during treatment or disease development (e.g., baseline period Day 0 before treatment, expansion period Day 7 during treatment, and recovery period Day 14 after treatment). Simultaneously, flow cytometry data at the same time points are collected. This technique allows for precise counting of specific cell subpopulations, and the resulting cell proportion data will serve as the gold standard for subsequent model alignment and training.

[0048] Real-world time-series observation data is a multimodal time-series dataset constructed from two types of data (Bulk-Seq data and flow cytometry data) collected in parallel at multiple discrete clinical time points for the same target object. Bulk-Seq data serves as the input feature variable for the model, providing a mixed gene expression profile of all cells at each time point. Flow cytometry data serves as the model's supervisory gold standard (Ground Truth), providing the precise proportions (abundance) of various cell subpopulations at each time point. It is used only to calculate the loss function to calibrate the model and is not input into the neural network. Real-world time-series observation data can be represented as a set of parameters. It contains all observational data collected from the 1st to the Tth clinical time point, among which It is the batch transcriptome feature vector at time point t. This is the corresponding cell subpopulation proportion vector; these data reflect the overall expression state and cell composition of the cell population at a specific physical time point t; t is an index representing the sequence number of the clinical physical time point; T is the total number of clinical time points.

[0049] like Figure 2 The complete architecture of the cell dynamics prediction model is demonstrated. The input features, after processing in S1, are sequentially processed by a heterogeneous neural network model, a flexible dynamic time warping alignment module, and a graph attention network model, ultimately outputting a prediction of the proportion of the target cell subpopulation. The processing steps at each stage are described in detail below.

[0050] S2 receives input features and performs dual-stream heterogeneous feature extraction.

[0051] A heterogeneous neural network architecture with two parallel channels was constructed. The first channel utilizes a variational autoencoder (VAE) to receive a single-cell transcriptome feature matrix based on a pseudo-temporal trajectory-ordered arrangement. The first channel extracts continuous developmental manifold features from the latent space; the second channel utilizes a bidirectional long short-term memory network (Bi-LSTM) to receive only batch transcriptome data sequences from real time-series observations. Encode the data to extract the dynamic trend characteristics of cell population evolution over physical time.

[0052] In one specific embodiment of the present invention, a heterogeneous neural network performs feature learning on microscopic cell development data and macroscopic clinical observation data respectively, extracting complementary representations, including:

[0053] (2.1) Extraction of single-cell developmental manifold features (first channel):

[0054] This channel captures continuous developmental trajectory information from high-dimensional, sparse single-cell data. The single-cell transcriptome feature matrix obtained in step S1, based on pseudo-temporal trajectories and arranged in an ordered manner, is then used. Input variational autoencoder (VAE), encoder of VAE Mapping to latent variables The posterior distribution. By optimizing the objective function. This allows the model to learn by compressing the data into a smooth, continuous latent space. In this objective function, The posterior distribution of the encoder output Compared with the standard normal prior distribution The KL divergence between them (Kullback-Leibler Divergence); To reconstruct the loss, The prior distribution is a standard normal distribution. To balance the weighting coefficients between reconstruction quality and decoupling capability, a feature sequence arranged according to the pseudo-temporal order of cell development is obtained based on this latent space representation, denoted as the pseudo-temporal feature sequence. This sequence constitutes the reference manifold for cell state evolution. Among them, is the i-th feature vector in the sequence, corresponding to the low-dimensional representation of the i-th cell on the pseudo-temporal trajectory after being encoded by the VAE encoder, containing information on the gene expression pattern of the cell state during development; i is an index variable used to identify the position in the sequence.

[0055] (2.2) Extraction of dynamic features of clinical time series (second channel):

[0056] This channel learns the macroscopic dynamic evolution of cell populations from sparse, discrete clinical time-point observation data. It uses the multi-time-point batch transcriptome data sequences obtained in step S1. After standardization, the input is a bidirectional long short-term memory (Bi-LSTM) network. Bi-LSTM can capture the contextual dependencies of the time series from both forward and backward directions, thus encoding the temporal features of each time point t. For ease of subsequent representation, let... Finally, the true temporal feature sequence over the entire observation period is output. It is used to characterize the dynamic trend of a group changing over time.

[0057] The principle behind this invention's dual-stream heterogeneous feature extraction architecture lies in addressing the differences and complementarities in the essential attributes of the two types of data. A variational autoencoder (VAE) is used as the first channel because single-cell transcriptome data has extremely high dimensionality (20,000+ genes) and severe "zero-inflation" noise, making it difficult for traditional linear dimensionality reduction to capture complex nonlinear developmental manifolds. The VAE, through a probabilistic generative model, can effectively fill in missing values ​​and reduce dimensionality by learning the data distribution, thereby extracting robust and continuous cell homeostatic features as a reference. A Bi-LSTM is used as the second channel because clinical real-time observation data (Bulk RNA-seq) has significant time-series dependencies and sparse sampling points. Cell population evolution is a continuous dynamic process, and the Bi-LSTM can simultaneously capture forward evolutionary information and backward dependency information along the physical time axis, maximizing the extraction of global temporal dynamic features. The high-resolution spatial manifold provided by the VAE compensates for the time-point sparsity of clinical data, while the true physical time dynamics provided by the Bi-LSTM correct for rate bias in pseudo-time series; the combination of the two achieves complementarity in spatiotemporal dimensions.

[0058] S3, cross-modal adaptive timing alignment.

[0059] Within the latent feature space, an alignment module with differentiable soft-DTW is introduced. This is achieved by computing pseudo-temporal feature sequences. Compared with the real time series feature sequence The soft alignment loss between cells establishes a nonlinear mapping function from cellular pseudo-time to clinical physical time. This function can be used to characterize the nonlinear mapping relationship between the microscopic gene expression evolution rate and the macroscopic clinical disease progression. It can automatically correct the temporal phase deviation caused by the difference in immune response rate among different individuals and output the pseudo-temporal feature sequence after phase alignment.

[0060] In one specific embodiment of the present invention, a differentiable alignment module is introduced to learn the correspondence between pseudo-timing and real physical time, reducing phase shifts caused by individual differences. The alignment process is as follows:

[0061] (3.1) Based on pseudo-time series feature sequences Real time series feature sequences Calculate the distance matrix D between the two sequences, where the elements are... Characterizing feature points in pseudo-temporal feature sequences Feature points in the real time series feature sequence The distances between them form a distance matrix. .

[0062] This embodiment uses Soft-DTW as the alignment loss. Using the aforementioned distance matrix, it searches for a soft alignment path that minimizes the cumulative alignment cost among all possible alignment paths. The loss function is as follows:

[0063]

[0064] in, It has a smoothing parameter The soft dynamic time warping operator, For smoothing coefficients, Represents the set of all allowed alignment paths. It is a specific alignment path. This loss... Differentiable, supporting end-to-end training.

[0065] (3.2) By minimizing the alignment loss The model learns a soft mapping from the pseudo-time scalar s to clinical time t. This nonlinear mapping function automatically stretches, compresses, or shifts pseudo-temporal feature sequences to achieve optimal alignment with real temporal feature sequences on the time axis. It accurately projects cell states onto the clinical time axis and reduces nonlinear temporal offset, outputting phase-aligned pseudo-temporal features.

[0066] S4, Intercellular Crosstalk and Dynamics Prediction.

[0067] Based on aligned pseudo-temporal features, a graph attention network (GAT) is used to construct a cellular microenvironment interaction graph. Using cell subpopulations as nodes, edges are constructed based on known receptor-ligand interaction databases or expression correlations between cell types. The interaction weights between nodes of different cell subpopulations are calculated, such as identifying nonlinear regulatory signals of effector T cells on regulatory T cells. Features of neighboring cells are aggregated through an attention mechanism to simulate crosstalk effects and nonlinear feedback regulatory mechanisms between cells in the immune microenvironment (especially effector T cells on regulatory T cells). The predicted abundance values ​​of target cell subpopulations at future time points are output, characterizing the proliferation dynamics trend of the target cell subpopulation.

[0068] In one specific embodiment of the present invention, the process of modeling the interaction relationships between cell subpopulations using a graph attention mechanism includes:

[0069] (4.1) Graph Construction: Using all identified cell subpopulations as nodes, and utilizing known receptor-ligand interaction databases (e.g., CellChatDB, CellPhoneDB, or SingleCellSignalR), query whether the cell subpopulation represented by node i highly expresses ligand L, and whether the cell subpopulation represented by node j highly expresses the corresponding receptor R. If the expression levels both exceed a preset expression threshold (expression rate > 10%), then a directed edge is established between node i and node j to represent the potential communication relationship. The final graph is obtained. The node features are directly derived from the phase-aligned pseudo-timing features output in step S3.

[0070] (4.2) Attention Convergence: The constructed cell interaction graph is input into the graph attention network GAT. Based on the adjacency matrix... Determined neighborhood For any node Execution feature aggregation in, , For the corresponding node features, It is a linear transformation matrix; To normalize the attention weights, the formula is used. The calculation yields a node. For nodes The importance of; For attention parameter vectors, It is a non-linear activation function.

[0071] (4.3) Dynamic prediction output: After one or more layers of GAT updates, the final cell state embedding containing rich microenvironment interaction information is obtained. For target cell subpopulations (such as effector T cells), their corresponding states are embedded. A prediction layer is input. This layer maps the target cell subpopulation to one or more scalars, directly outputting the predicted proportion (abundance) of the target cell subpopulation at one or more specified clinical time points in the future. The multiple time points are determined by the network structure design and training task settings of the prediction layer. The output node dimensions of the prediction layer are pre-designed to correspond to specific time offsets or specific clinical detection points (such as T+1 or T+n). These predicted values, arranged chronologically, constitute the predicted proliferation dynamics curve of the cell subpopulation, which can be used to demonstrate the dynamic evolution trend of the cell population during treatment. In this embodiment, the prediction layer consists of a fully connected layer and an activation function.

[0072] (4.4) Model Training

[0073] The training of the model is ultimately optimized through a joint loss function, which includes prediction loss and alignment loss. The prediction loss refers to the error between the predicted ratio and the gold standard ratio of flow cytometry, and the alignment loss has been introduced in step S3.

[0074] like Figure 3 As shown, the trained model can be used to obtain the proportion prediction curve of specific immune cell subsets changing over time based on the patient's single-cell transcriptome data and batch transcriptome data, enabling dynamic monitoring and evaluation. The specific process is as follows:

[0075] First, patient data is received, including single-cell transcriptome data and clinical time-series observation data. The clinical time-series observation data consists of bulk RNA-seq data of the same new target subject at multiple discrete clinical time points (such as before, during, and after treatment). Flow cytometry data at corresponding time points are no longer needed because they serve as the "gold standard" label during training. The model will learn to infer proportions from the bulk data.

[0076] Secondly, the input single-cell transcriptome data undergoes the same preprocessing and pseudo-temporal analysis as during training, resulting in a single-cell transcriptome feature matrix arranged in an ordered manner based on pseudo-temporal trajectories. This matrix is ​​then input into a trained variational autoencoder (VAE) to extract pseudo-temporal feature sequences. Multi-timepoint Bulk RNA-seq data is then input into a trained bidirectional long short-term memory network (Bi-LSTM) to extract true temporal feature sequences. Finally, the two feature sequences are input into a trained alignment module. This module utilizes the nonlinear mapping learned during training to automatically perform phase adjustment on the pseudo-temporal feature sequences, aligning them with the clinical timeline, and outputting the phase-aligned pseudo-temporal features.

[0077] Finally, a cell interaction graph is constructed using all cell subpopulations as nodes. The phase-aligned pseudo-temporal features are used as initial node features and input into a trained Graph Attention Network (GAT). GAT, through its fixed attention mechanism parameters, aggregates information from the neighboring nodes of each cell node, outputting a final cell state embedding containing the microenvironment interaction context. Feature vectors corresponding to the target cell subpopulation are extracted from the final state embedding and input into a trained prediction layer, outputting the predicted proportion (abundance) of the target cells at future time points. These time-ordered predictions constitute the predicted proliferation kinetics curve of the target cell subpopulation.

[0078] To verify the effectiveness of the invention, it was tested on a clinical validation set that included complete longitudinal time points (covering the baseline, expansion, and recovery periods).

[0079] Three widely used cell deconvolution algorithms were selected as baseline models: CIBERSORTx, xCell, and MuSiC. These comparison methods were run time-by-time on the same Bulk input and the same set of time points, and the serialized results were output according to a unified summarization rule.

[0080] This study employs multiple metrics, including root mean square error (RMSE), Pearson correlation coefficient, and dynamic time warping distance (DTW Distance), to evaluate the model's accuracy in capturing peak cell proliferation during the critical treatment window. RMSE quantifies the absolute numerical error between the predicted cell proportion and the "gold standard" measurement in flow cytometry; a lower RMSE indicates higher numerical accuracy. Pearson correlation coefficient assesses the trend consistency between the predicted and actual curves; a higher R-value indicates a stronger ability to capture the direction of change. DTW Distance measures the degree of overall shape and phase matching between two time series curves. DTW Distance is sensitive to peak lag, premature peaks, or shape distortion; a lower DTW value indicates better time alignment.

[0081] Table 1 Experimental Results

[0082]

[0083] The experiments are shown in Table 1. The results show that the method of the present invention outperforms the comparative method in all three indicators across the entire time range, demonstrating its comprehensive advantages in numerical accuracy and trend correlation, and verifying the effectiveness of the differentiable time-series alignment module.

[0084] In particular, the comparison method in Pearson While there may be some trend consistency, the DTW distance is relatively higher, suggesting that there may be phase misalignment or peak position shift during rapidly changing windows (such as the peak stage near Day 7), making it impossible to accurately predict when the rising peak will appear. This invention uses a differentiable alignment module to accurately map the microscopic pseudo-time series with the macroscopic physical time axis, significantly reducing the time series phase deviation and thus improving the matching ability for dynamic changes in key windows.

[0085] This embodiment also provides a cell dynamics prediction system for pseudo-temporal manifold alignment and graph attention to implement the above method, such as... Figure 4 As shown, it includes:

[0086] The data preparation and input module is used to acquire single-cell transcriptome data of the target object. After preprocessing and pseudo-time series analysis, a single-cell transcriptome feature matrix based on the ordered arrangement of pseudo-time series trajectories is obtained. At the same time, batch transcriptome data of the same target object at multiple discrete clinical time points and the proportion of cell subpopulations measured by flow cytometry are acquired to jointly constitute real time series observation data.

[0087] The dual-stream heterogeneous feature extraction module includes a first feature extraction channel and a second feature extraction channel set in parallel. The first feature extraction channel contains a variational autoencoder, which takes a single-cell transcriptome feature matrix as input to extract pseudo-temporal feature sequences. The second feature extraction channel contains a bidirectional long short-term memory network, which takes batch transcriptome data from real time-series observation data as input to extract real time-series feature sequences.

[0088] Differentiable timing alignment module is used to receive pseudo-timing feature sequences and real timing feature sequences, and output phase-aligned pseudo-timing feature sequences;

[0089] The microenvironment interaction modeling module is used to construct a cell microenvironment interaction graph with cell subpopulations as nodes. It uses the phase-aligned pseudo-temporal feature sequence as the initial feature of the node, and uses a graph attention network model to aggregate the features of neighboring nodes to simulate the intercellular interactions in the microenvironment and predict the proportion of target cell subpopulations at future time points.

[0090] The training module is used to train the model by combining the alignment loss of the feature sequences and the prediction loss of the target cell subpopulation ratio.

[0091] The cell dynamics prediction module is used to predict the proportion of target cell subpopulations at future time points using the trained model, and to characterize the proliferation dynamics trend of the target cell subpopulations.

[0092] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0093] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0094] It should also be noted that the pseudo-temporal manifold alignment and graph attention cell dynamics prediction method in the above embodiments can essentially be executed by a computer program. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the method provided in the above embodiments, which includes a memory and a processor;

[0095] The memory is used to store computer programs;

[0096] The processor is configured to implement the pseudo-temporal manifold alignment and graph attention cell dynamics prediction method in the above embodiments when executing the computer program.

[0097] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium.

[0098] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the method provided in the above embodiments. The storage medium stores a computer program that, when executed by a processor, can realize the cell dynamics prediction method for pseudo-temporal manifold alignment and graph attention in the above embodiments.

[0099] It is understood that the computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0100] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention, characterized in that, Includes the following steps: (1) Obtain single-cell transcriptome data of the target object, and after preprocessing and pseudo-temporal analysis, obtain a single-cell transcriptome feature matrix based on the ordered arrangement of pseudo-temporal trajectories; Meanwhile, the batch transcriptome data of the same target object at multiple discrete clinical time points and the proportion of cell subpopulations measured by flow cytometry together constitute real time-series observation data. (2) Construct a heterogeneous neural network model with two parallel channels; input the single-cell transcriptome feature matrix into the variational autoencoder of the first channel to obtain a pseudo-temporal feature sequence that characterizes the continuous development of cells in the latent space; input the batch transcriptome data in the real temporal observation data into the bidirectional long short-term memory network of the second channel to obtain the real temporal feature sequence of cell population evolution with physical time; (3) Input the pseudo-time series feature sequence and the real time series feature sequence into the alignment module based on the configurable dynamic time warping, and output the phase-aligned pseudo-time series feature sequence; (4) Construct a cell microenvironment interaction graph with cell subpopulations as nodes, use the phase-aligned pseudo-temporal feature sequence as the initial feature of the node, use the graph attention network model to aggregate the features of the neighboring nodes, simulate the intercellular interaction in the microenvironment, and predict the proportion of the target cell subpopulation at future time points. (5) The model is trained by combining the alignment loss of the combined feature sequence and the prediction loss of the target cell subpopulation ratio. The trained model is then used to predict the target cell subpopulation ratio at future time points, thus characterizing the proliferation dynamics trend of the target cell subpopulation.

2. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 1, characterized in that, In step (1), the pseudo-time series analysis is performed using tools such as Monocle, Slingshot, PAGA, Diffusion Pseudotime, or scVelo.

3. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 1, characterized in that, In the alignment module based on soft dynamic time warping, the alignment loss is calculated based on the soft dynamic time warping algorithm. It calculates the distance matrix between all feature points in the pseudo-time series feature sequence and the real time series feature sequence, and finds a soft alignment path that minimizes the cumulative alignment cost among all possible alignment paths.

4. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 3, characterized in that, The alignment loss formula is as follows: ； in, It is alignment loss. It is a pseudo-temporal feature sequence. It is a true time-series feature sequence. It has a smoothing parameter The soft dynamic time warping operator, It is the distance between feature point i in the pseudo-temporal feature sequence and feature point j in the real temporal feature sequence. It is the set of all legally aligned paths. It is a specific alignment path.

5. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 1 or 4, characterized in that, The alignment module learns a nonlinear mapping function from cellular pseudo-time to clinical physical time based on alignment loss; it uses the nonlinear mapping function to automatically stretch, compress, or translate the pseudo-time feature sequence to achieve optimal alignment with the real time feature sequence on the time axis, corrects the temporal phase deviation caused by the difference in developmental rate between individuals, and obtains the phase-aligned pseudo-time feature sequence.

6. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 1, characterized in that, In step (4), the edges of the cell microenvironment interaction graph are constructed based on known receptor-ligand interaction databases or expression correlations between cell types.

7. The method for predicting cell dynamics based on pseudo-temporal manifold alignment and graph attention according to claim 1, characterized in that, In step (4), the process of predicting the proportion of the target cell subpopulation at future time points includes: The cell state embedding vector containing microenvironment interaction information generated by the aggregation of the graph attention network model is input into a prediction layer, and the cell state embedding vector is mapped to one or more scalar outputs. The scalar output is used to characterize the predicted proportion of the target cell subpopulation at one or more specified clinical time points in the future.

8. A cell dynamics prediction system based on pseudo-temporal manifold alignment and graph attention, characterized in that, include: The data preparation and input module is used to acquire single-cell transcriptome data of the target object. After preprocessing and pseudo-time series analysis, a single-cell transcriptome feature matrix based on the ordered arrangement of pseudo-time series trajectories is obtained. Meanwhile, the batch transcriptome data of the same target object at multiple discrete clinical time points and the proportion of cell subpopulations measured by flow cytometry together constitute real time-series observation data. The dual-stream heterogeneous feature extraction module includes a first feature extraction channel and a second feature extraction channel configured in parallel. The first feature extraction channel contains a variational autoencoder that uses the single-cell transcriptome feature matrix as input to extract pseudo-temporal feature sequences. The second feature extraction channel includes a bidirectional long short-term memory network, which takes batch transcriptome data from real time-series observation data as input to extract real time-series feature sequences. Differentiable timing alignment module is used to receive pseudo-timing feature sequences and real timing feature sequences, and output phase-aligned pseudo-timing feature sequences; The microenvironment interaction modeling module is used to construct a cell microenvironment interaction graph with cell subpopulations as nodes. It uses the phase-aligned pseudo-temporal feature sequence as the initial feature of the node, and uses a graph attention network model to aggregate the features of neighboring nodes to simulate the intercellular interactions in the microenvironment and predict the proportion of target cell subpopulations at future time points. The training module is used to train the model by combining the alignment loss of the feature sequences and the prediction loss of the target cell subpopulation ratio. The cell dynamics prediction module is used to predict the proportion of target cell subpopulations at future time points using the trained model, and to characterize the proliferation dynamics trend of the target cell subpopulations.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the cell dynamics prediction method for pseudo-temporal manifold alignment and graph attention as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the cell dynamics prediction method for pseudo-temporal manifold alignment and graph attention as described in any one of claims 1 to 7.

Images