An artificial intelligence-based virtual brain cell annotation method

Abstract

The application discloses a virtual brain cell annotation method based on artificial intelligence and belongs to the technical field of single-cell data analysis.The method first loads single-cell data and a protocol configuration file, checks time point effectiveness and mutual exclusion relationship, and generates a time mapping; then maps the cell data to a fixed hyper-variable gene space, constructs a cell sequence, and directly encodes expression values; obtains a time condition vector by embedding time information, accesses each layer of a Transform model through adaptive layer normalization, completes cell pairing in real time within a batch during training, constructs an intermediate cell state in combination with paired samples, completes model training in cooperation with a plurality of loss functions, finally completes result evaluation through multidimensional indexes, and uniformly saves and exports related data.The application can fully utilize time sequence characteristics, reduce resource occupation, has high prediction accuracy and good stability, can effectively realize brain cell time sequence state prediction and virtual annotation, and is suitable for various single-cell development analysis scenes.

Description

Technical Field

[0001] This invention provides a virtual brain cell annotation method based on artificial intelligence, belonging to the field of single-cell data analysis technology. Background Technology

[0002] The development and differentiation of brain cells exhibit significant temporal dynamics. Single-cell sequencing technology can acquire gene expression data of brain cells at different time points, making it a core tool for analyzing brain cell development mechanisms and completing cell annotation. Currently, the analysis and annotation techniques for temporal single-cell data are mainly divided into two categories: traditional statistical analysis methods and conventional deep learning prediction methods.

[0003] Traditional statistical analysis methods rely heavily on algorithms such as clustering, differential gene analysis, and trajectory inference to classify and annotate cell states. These methods can only perform static analysis of cell data at observed time points and cannot accurately extrapolate cell developmental states at unobserved time points, making it difficult to achieve virtual cell annotation across time dimensions. Furthermore, traditional methods fail to capture continuous expression changes during cell development, exhibiting significant limitations in analyzing the rate of cell state evolution and intermediate transitional cell states, resulting in low analytical accuracy when dealing with the complex temporal dynamics of brain cell development.

[0004] With the development of artificial intelligence technology, some deep learning models have been applied to single-cell time-series data prediction. Existing single-cell time-series prediction models based on Transformer and recurrent neural networks have several shortcomings: First, they lack a unified spatial mapping for highly variable gene features in single cells, resulting in inconsistent feature dimensions across different batches and time points, leading to poor model generalization ability. Second, the temporal information fusion method is simplistic, using only time labels as ordinary feature inputs to the model, failing to achieve fine-grained control of multi-layer operational parameters and resulting in low utilization of temporal features. Third, cell sample pairing often employs offline fixed pairing methods, consuming significant storage resources and failing to adapt to dynamic model training processes. Fourth, the loss function design is simplistic, relying solely on a single regression loss to constrain model training, making it difficult to consider multiple dimensions such as cell population distribution, single-cell matching degree, and feature variance, leading to problems like distorted gene expression values, cell type shifts, and population distribution deviating from the true state in prediction results. Fifth, the prediction stage lacks reasonable iterative update strategies and numerical constraints, resulting in insufficient stability of extrapolation and interpolation tasks, and most solutions lack standardized multi-dimensional result evaluation systems, making it impossible to quantify the reliability of virtual cell annotation and prediction results.

[0005] In summary, existing technologies struggle to simultaneously consider the continuity, temporal correlation, and population distribution characteristics of single-cell brain time-series data, and are unable to efficiently complete virtual annotation and state prediction of brain cells at unobserved time points. There is an urgent need for an intelligent solution that is highly targeted, has high prediction accuracy, and is efficient in operation. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of insufficient prediction accuracy of temporal single-cell data, poor temporal feature fusion effect, high resource consumption for cell pairing, single training constraint method, uncontrollable prediction results, and imperfect evaluation system in existing technologies, and to provide a virtual brain cell annotation method based on artificial intelligence. This invention, by constructing a temporal conditional Transformer network, setting a dynamic cell pairing mechanism within batches, adopting a multi-loss function collaborative constraint strategy, and an Euler iterative prediction framework, can achieve accurate prediction and virtual annotation of the temporal state of brain cells under different types of tasks. Simultaneously, it achieves integrated processing of the entire process of data verification, feature encoding, model training, and result evaluation, improving the automation level and overall accuracy of brain single-cell temporal data analysis and annotation.

[0007] To address the aforementioned problems, the present invention proposes a technical solution: a virtual brain cell annotation method based on artificial intelligence, comprising the following steps:

[0008] Step 1: Load single-cell related data and protocol configuration files, verify the validity and mutual exclusion of various time point information in the files, and generate corresponding time mapping relationships based on task type;

[0009] Step 2: Map all cell expression data to a fixed hypervariable gene space, construct individual cell data into a sequence containing gene identifiers, continuous expression values ​​and mask states, and encode the continuous expression values;

[0010] Step 3: Embed different types of time information, combine them to form a time condition vector, and inject the vector into each layer of the Transformer model to adjust the model's operation parameters;

[0011] Step 4: During model training, source time, target time, and corresponding cell samples are selected in batches, and the pairing and combination of source cells and target cells are completed within each batch.

[0012] Step 5: Construct intermediate cell states using paired cells, input them into the model to predict the cell expression state at the target time point, and deduce the cell state progression speed accordingly. Model training is completed through the collaborative constraints of multiple loss functions.

[0013] Step 6: Make predictions for unobserved time points. Determine the data source based on the task type, and iterate step by step from the cell state at the source time point while constraining the expression value range to obtain cell expression data at the target time point.

[0014] Step 7: Evaluate the prediction results using multi-dimensional indicators, and save and export the prediction data, evaluation results, and related configuration information.

[0015] Furthermore, the protocol configuration file records available time points, observed time points, reserved evaluation time points, task types, source time selection strategies, and highly variable gene dimensions. During verification, it is ensured that observed time points and reserved evaluation time points do not overlap, and that all time points exist in the original data.

[0016] Furthermore, the constructed cell sequences also include classification marker characters, and the continuous expression values ​​of cells are directly encoded without discrete binning, thus preserving the detailed features of gene expression.

[0017] Furthermore, the time condition vector is fed into the Transformer model through adaptive layer normalization to adjust the operating parameters of the model's internal attention structure and feedforward structure.

[0018] Furthermore, the cell pairings are generated in real time within a single training batch, without the need to store fixed cell pairing relationships in advance.

[0019] Furthermore, the loss functions used for training include flow matching loss, multi-spatial distribution distance loss, and population mean matching loss, and variance constraint loss can also be added.

[0020] Furthermore, the cell state is updated step by step using the Euler iteration method. In each iteration, the current cell state, time information and model output results are combined to calculate the propulsion speed and complete the state update.

[0021] Furthermore, the task types include interpolation tasks, short-range extrapolation tasks, long-range extrapolation tasks, and benchmarking tasks; interpolation tasks select observed time points adjacent to the target time point as data sources, and extrapolation tasks select historical observed time points as data sources.

[0022] Furthermore, the indicators used for outcome evaluation included spatial distribution distance, population average expression correlation, single-cell matching correlation, cell type consistency, and characteristic gene distribution difference indicators.

[0023] Furthermore, this method for predicting the dynamic development of a single cell includes seven interconnected functional modules: a data loading and protocol parsing module, a fixed hypervariable gene space construction module, a time-conditional Transformer model module, an online optimal transmission pairing module, a target state flow matching training module, an Euler iterative prediction generation module, and an evaluation and result export module. Each module works sequentially according to the data flow order to jointly complete the modeling, prediction, and result evaluation of the dynamic development of a single cell.

[0024] Due to the adoption of the above technical solution, the beneficial effects of the artificial intelligence-based virtual brain cell annotation method of the present invention are as follows:

[0025] 1. Standardized data verification, adaptable to multiple tasks: This invention manages core parameters such as all time points and task types uniformly through protocol configuration files, and enforces verification of the mutual exclusivity and validity of observed time points and evaluation time points, avoiding data conflict issues from the source; at the same time, it divides tasks into four categories: interpolation, near-distance extrapolation, far-distance extrapolation, and benchmark testing, which can be adapted to different time series prediction and annotation scenarios of brain cells, making it more versatile.

[0026] 2. Reasonable feature encoding, preserving original details: All cell data are uniformly mapped to a fixed hypervariable gene space to eliminate dimensional differences between different batches and time points; classification marker characters are added to cell sequences, and continuous expression values ​​are directly encoded without discrete binning, preserving the continuous detailed features of single-cell gene expression to the greatest extent and improving the accuracy of subsequent model learning.

[0027] 3. Deep integration of temporal features and refined model control: Multiple types of temporal information are embedded to generate temporal conditional vectors, which are then injected into all network layers of the Transformer model. Combined with adaptive layer normalization to adjust the running parameters of attention and feedforward structure, the model can be globally controlled by temporal information, fully explore the temporal correlation features of cell development, and solve the problem of insufficient utilization of temporal features in traditional models.

[0028] 4. Online cell pairing with lower resource consumption: The batch real-time online cell pairing method eliminates the need to store a large number of fixed cell pairing relationships in advance, greatly reducing the storage pressure on the device. At the same time, the dynamic pairing method can adapt to the dynamic training logic of the model, improving training efficiency and sample diversity.

[0029] 5. Multi-loss collaborative constraints for better training results: It integrates flow matching loss, multi-spatial distribution distance loss, and population mean matching loss, and can selectively add variance constraint loss. It constrains model training from multiple dimensions such as cell state transition, population distribution, average expression, and feature variance, effectively avoiding problems such as distribution shift, expression distortion, and variance anomaly in prediction results.

[0030] 6. Controllable iterative prediction and strong result stability: The cell state is updated step by step using the Euler iteration method, which accurately calculates the cell state progression speed and conforms to the objective law of gradual brain cell development. The prediction process constrains the range of gene expression values ​​throughout and matches corresponding data sources with different task types, which greatly improves the prediction stability and robustness of interpolation and extrapolation tasks.

[0031] 7. Standardized multi-dimensional assessment with quantifiable results: Establish an assessment index system that includes spatial distribution, expression correlation, cell type consistency, and other dimensions. This system can comprehensively quantify the quality of virtual brain cell annotation and temporal state prediction results, enabling full-process traceability and evaluation, and providing reliable data support for research on brain single-cell development mechanisms. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0033] Figure 1 This is the overall model architecture and prediction flowchart of Novaeve-Dev, a virtual brain cell annotation method based on artificial intelligence, according to the present invention.

[0034] Figure 2 This is a flowchart of the Novaeve-Dev model training process, which is based on an artificial intelligence-based virtual brain cell annotation method according to the present invention.

[0035] Figure 3 This is a flowchart illustrating the prediction of unobserved time points during the inference phase of Novaeve-Dev, a virtual brain cell annotation method based on artificial intelligence according to the present invention. Detailed Implementation

[0036] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1: Brain Cell Annotation under the Interpolation Task

[0037] This embodiment is applied to the study of brain nerve cell development. It performs cell state interpolation prediction and virtual annotation for unknown time points between two adjacent observed time points to supplement cell development time series data and analyze the characteristics of intermediate transitional cells.

[0038] 1. Basic Data and Configuration Settings

[0039] Single-cell sequencing data of brain nerve cells were selected, and two observed time points, T1 and T3, were set. The unknown time point T2 to be predicted is located between T1 and T3.

[0040] Configuration protocol file content: The available time points of the system are T1, T2, and T3; the observed time points are T1 and T3; no reserved evaluation time points are set for this time; the task type is set as interpolation task; the source time selection rule is to select the observed time points adjacent to the target time point; the dimension of high-variable genes is set to 1800.

[0041] The data verification process confirms that there is no overlap between the observed time points and the evaluation time points, and that all time points marked in the protocol exist in the original sequencing data. After the verification is completed, the time mapping relationship between T1 and T2 and between T2 and T3 is generated in conjunction with the interpolation task.

[0042] Model configuration: A time-conditional Transformer model is used, with a training batch size of 256. Four loss functions are enabled during training: flow matching loss, multi-spatial distribution distance loss, population mean matching loss, and variance constraint loss. All seven linked functional modules are used throughout the entire process.

[0043] 2. Specific Implementation Steps

[0044] Step 1: Start the data loading and protocol parsing module, import brain single-cell sequencing data and protocol configuration files, complete the validity and mutual exclusion verification of time points, and generate the corresponding time mapping relationship in combination with the interpolation task.

[0045] Step two involves activating the fixed hypervariable gene space construction module to uniformly map all cell expression data from time points T1 and T3 to an 1800-dimensional fixed hypervariable gene space. Data sequences are constructed for each individual cell, containing gene identifiers, continuous expression values, mask status, and classification markers. Continuous gene expression values ​​are directly encoded without discrete binning, preserving complete expression detail features.

[0046] Step 3: Activate the temporal conditional Transformer model module to embed various temporal information such as absolute time and time intervals, and integrate them to generate a temporal conditional vector. Through adaptive layer normalization, the temporal conditional vector is connected to each layer of the Transformer model, adjusting the operating parameters of the model's internal attention and feedforward structures.

[0047] Step 4: Activate the online optimal transmission pairing module. During model training, source time T1, target time T3 and corresponding cell samples are selected in batches. The pairing of source cells and target cells is completed in real time within a single batch, without pre-storing cell pairing relationships.

[0048] Step 5: Activate the target state flow matching training module. Utilize paired cell samples within a batch to construct intermediate cell development states corresponding to time point T2. Input the intermediate state data into the model to predict cell expression states at time point T2 and calculate the cell state progression rate. Relying on the synergistic constraints of four types of loss functions, the model training is iterated until the model loss value stabilizes, completing the model training.

[0049] Step six: Activate the Euler iterative prediction generation module. In this embodiment, it is an interpolation task, selecting the observed time points T1 and T3 adjacent to the target time point T2 as data sources. Starting from the cell state at the source time point, the cell state is updated step by step using the Euler iterative method. Each iteration combines the current cell state, time information, and model output results to calculate the state progression speed. The range of gene expression values ​​is strictly limited during the iteration process. After the iteration is completed, the cell expression data and virtual annotation results at time point T2 are obtained.

[0050] Step 7: Activate the evaluation and results export module. Five indicators—spatial distribution distance, population average expression correlation, single-cell matching correlation, cell type consistency, and characteristic gene distribution differences—are used to comprehensively evaluate the prediction results. After the evaluation is completed, the prediction data, evaluation reports, protocol configuration files, etc., are saved uniformly, and the data is exported according to the specified format.

[0051] 3. Implementation Results

[0052] This embodiment successfully completed the cell development data at time point T2 between T1 and T3. The predicted cell type consistency reached over 94%, the continuous gene expression characteristics were not lost, and no abnormal expression values ​​were generated. It was able to accurately complete the virtual annotation of brain cells at intermediate time nodes.

[0053] Example 2: Brain Cell Annotation under a Close Extrapolation Task

[0054] This embodiment is applied to the short-term development trend assessment of brain neural stem cells. Based on a single historical observation time point close to the target time, it carries out short-term temporal extrapolation prediction and cell annotation, which is suitable for predicting the cell development status in the short term.

[0055] 1. Basic Data and Configuration Settings

[0056] Single-cell sequencing data of brain neural stem cells were selected, and the observed time point S1 was set. The target time point S2 to be predicted is close to S1 in terms of time interval and location.

[0057] Configuration protocol file content: The available time points of the system are S1 and S2; the observed time point is S1; the reserved evaluation time point is S2; the task type is set as a near-distance extrapolation task; the source time selection rule is to select historical observation time points that are immediately adjacent to the target time; the dimension of highly variable genes is set to 2000.

[0058] The data verification process confirms that the observed time point S1 and the reserved evaluation time point S2 are independent and do not overlap. All configured time points exist in the original sequencing data. After verification, the time mapping relationship from S1 to S2 is generated.

[0059] Model configuration: A time-conditional Transformer model is used, with a training batch size of 128. During training, flow matching loss, multi-spatial distribution distance loss, population mean matching loss, and variance constraint loss are all enabled simultaneously. All seven linked functional modules are utilized throughout the entire process.

[0060] 2. Specific Implementation Steps

[0061] Step 1: Start the data loading and protocol parsing module, import neural stem cell sequencing data and protocol configuration files, complete the verification of time point validity and mutual exclusion, and generate time mapping relationships by combining the close extrapolation task.

[0062] Step 2: Activate the fixed hypervariable gene space construction module to uniformly map all cell expression data at time point S1 to a 2000-dimensional fixed hypervariable gene space. For each individual cell, construct a data sequence containing gene identifiers, continuous expression values, mask status, and classification marker characters. Continuous expression values ​​are directly encoded without discrete binning.

[0063] Step 3: Activate the temporal conditional Transformer model module to embed various types of temporal information and synthesize temporal conditional vectors. Adaptive layer normalization is then used to connect these vectors to the various network layers of the model, adjusting the operating parameters of the attention and feedforward structures.

[0064] Step 4: Activate the online optimal transmission pairing module. During the model training phase, select cell samples corresponding to source time S1 and target time S2 in batches and complete cell pairing in real time within each batch without pre-reserving pairing data.

[0065] Step 5: Activate the target state flow matching training module. Construct intermediate cell development states based on paired cells, input them into the model, predict the cell expression state at time point S2, and calculate the cell state progression speed. Iteratively update the model using four types of loss functions to collaboratively constrain its updates until training is complete.

[0066] Step six: Activate the Euler iterative prediction generation module. In this embodiment, it is a short-range extrapolation task, selecting the historically observed time point S1 as the sole data source. The cell state is updated step by step using the Euler iterative method. In each iteration, the cell state progression rate is calculated, while constraining the gene expression value range, ultimately obtaining the cell expression data and annotation results at time point S2.

[0067] Step 7: Start the evaluation and results export module, use five evaluation indicators to quantitatively analyze the prediction results, and after the evaluation is completed, save and export all contents such as prediction data, evaluation report, and configuration files.

[0068] 3. Implementation Results

[0069] In this near-range extrapolation prediction, the spatial distribution distance of the cell population was less than 0.06, the single-cell matching correlation was higher than 0.90, the predicted cell types matched the actual cell types well, and the prediction could accurately predict the short-term developmental status of brain neural stem cells. The annotation results are reliable.

[0070] Example 3: Brain Cell Annotation in a Long-Distance Extrapolation Task

[0071] This embodiment is applied to the study of the long-term developmental patterns of brain embryonic cells. Based on multiple consecutive historical observation time points, it performs long-distance extrapolation prediction and cell annotation for unknown time points with long time intervals, which is used to analyze the overall process of long-term development, differentiation and maturation of brain cells.

[0072] 1. Basic Data and Configuration Settings

[0073] Single-cell sequencing data from the embryonic brain development stage were selected, and consecutive observed time points M1, M2, M3, and M4 were set. The target time point M5 to be predicted is significantly different from the above historical time points.

[0074] Configuration protocol file content: The available time points of the system are M1, M2, M3, M4, and M5; the observed time points are M1, M2, M3, and M4; the reserved evaluation time point is M5; the task type is set as a long-distance extrapolation task; the source time selection rule is to use all historical observed time points; the dimension of highly variable genes is set to 2200.

[0075] The data verification process confirms that the observed time points do not overlap with the reserved evaluation time points, and all configured time points can be matched with corresponding content in the original sequencing data. After verification, a complete time-series mapping relationship is generated from M1 to M2, M2 to M3, M3 to M4, and M4 to M5.

[0076] Model configuration: A time-conditional Transformer model is used, with a training batch size of 256; all four loss functions are enabled during training. All seven linked functional modules are enabled throughout the entire process.

[0077] 2. Specific Implementation Steps

[0078] Step 1: Start the data loading and protocol parsing module, import multi-timepoint embryonic cell sequencing data and protocol configuration files, complete the verification of timepoint validity and mutual exclusion, and generate a complete time-series mapping relationship by combining the long-distance extrapolation task.

[0079] Step 2: Activate the fixed hypervariable gene space construction module to uniformly map cell expression data from all time points M1 to M4 to a 2200-dimensional fixed hypervariable gene space. Data sequences are constructed on a single-cell basis, containing gene identifiers, continuous expression values, mask states, and classification marker characters. Continuous expression values ​​are directly encoded without discrete binning.

[0080] Step 3: Activate the temporal conditional Transformer model module, integrate various temporal information and perform embedding processing to generate temporal conditional vectors. Adaptive layer normalization is then used to feed these vectors into each layer of the Transformer model, adjusting the operating parameters of the model's internal attention and feedforward structures.

[0081] Step 4: Activate the online optimal transmission pairing module. During the training process, the source time, target time, and corresponding cell samples are selected in batches. Cell pairing is completed in real time only within the current batch, and fixed pairing relationships are not stored in advance.

[0082] Step 5: Activate the target state flow matching training module. Construct multi-level intermediate states of cell development using paired cells, input them into the model, predict the cell expression state at time point M5, and calculate the state progression rate of long-term cell development. The model training is constrained by four types of loss functions, and training ends after iterating until the loss stabilizes.

[0083] Step six: Activate the Euler iterative prediction generation module. In this embodiment, it is a long-distance extrapolation task, selecting all historical observation time points M1, M2, M3, and M4 as data sources. Starting from the cell state at the last historical time point M4, the cell state is updated step by step using the Euler iterative method. Each iteration combines real-time data to calculate the state progression speed, and the gene expression value range is constrained throughout the process. After the iteration is completed, the cell expression data and virtual annotation results at time point M5 are obtained.

[0084] Step 7: Start the evaluation and results export module. Use five evaluation indicators to comprehensively evaluate the long-term time series prediction results. After the evaluation is completed, save and export the prediction data, evaluation results, model parameters, and configuration files.

[0085] 3. Implementation Results

[0086] For long-term extrapolation scenarios, this solution effectively reduces the data distribution offset caused by excessive time intervals, maintains stable cell type consistency above 92%, and minimizes differences in characteristic gene distribution. It can stably complete the virtual annotation and state prediction of the long-term development process of brain embryonic cells, and can provide effective support for research on brain development mechanisms.

[0087] in Figure 1The overall model architecture and prediction process of Novaeve-Dev are demonstrated. The system takes a single-cell expression vector x∈R^d as input, first standardizing the expression matrix, performing log1p transformation, and sorting high-variance genes to obtain fixed HVG gene coordinates and gene order. Then, the system constructs a gene sequence with a CLS marker for each cell. This sequence contains three types of information: gene ID, continuous expression value, and mask state, enabling the model to simultaneously identify gene identity, expression intensity, and input state. In the encoding stage, continuous expression values ​​are mapped to expression embeddings using ContinuousValueEncoder. The flow time t_flow, source time t_src, and target time t_tgt are then processed through sinusoidal time embedding and MLP encoding to form a unified time condition vector C_all. This conditional vector is injected into a multi-layered AdaLNTransformerBlock to adjust the normalization parameters of the attention layer and feedforward layer, enabling the model to perceive the developmental direction, time span, and dynamic changes between the source and target time points. In the output phase, the model predicts the expression endpoints x'_1, x'_2, ..., x'_d at the target time point based on gene sequences. Unlike directly predicting velocity, Novaeve-Dev employs a target state reparameterization strategy: the model first predicts the target expression state, then derives the velocity field vθ(x, t_flow, t_src, t_tgt) from the difference between the current and target states. Finally, the system progressively advances the cell state from the source time point through EulerUpdate, forming a continuous evolutionary trajectory of x_t0→x_t1→...→x_tn, and outputs the cell expression distribution at the target time point.

[0088] Figure 2The Novaeve-Dev model training workflow is demonstrated. The system first reads preprocessed single-cell time-series data, including the expression matrix, hypervariable gene panel, cell metadata, and observed time point information. During training, the system samples the source time t_src and target time t_tgt from the observed time points, and extracts source cell batches and target cell batches respectively. Since single-cell time-series data typically lacks a true one-to-one correspondence at the cell level, the system constructs a training matching relationship between source and target cells through online optimal transfer or Sinkhorn pairing. After obtaining the source-target cell pair, the system samples the stream time τ and constructs the intermediate state x_τ = (1-τ)x_src + τx_tgt. The model takes the intermediate expression state, gene number, expression value, mask state, and time conditions τ, t_src, and t_tgt as input, and predicts the expression endpoint x̂_tgt at the target time point after time-conditional AdaLNTransformer encoding. The system then derives the velocity v̂=(x̂_tgt-x_τ) / (1-τ+ε) from the predicted endpoints and performs conditional flow matching training with the difference between the predicted endpoints and the actual source-target values. The training objective includes conditional flow matching loss, HVG / PCA spatial sinkhorn distribution loss, population mean matching loss, and optional variance constraints. This process enables the model to learn not only the direction of advancement of individual cell states but also the overall distribution characteristics of the cell population at the target time point, thereby improving the consistency and biological plausibility of the predicted distribution at unobserved time points.

[0089] Figure 3This demonstrates the Novaeve-Dev process for predicting unobserved time points during the inference phase. The system first loads the trained model, prediction protocol, observed time points, and the held-out target time point to be predicted. Based on the task type, the system analyzes the source of the target time point: for interpolation tasks, observation times before and after the target time point can be selected as the source; for extrapolation tasks, the most recent observed time point is selected as the source time. Subsequently, the system uses the cell expression matrix of the source time point as the initial state x_0. In the EulerRollout phase, the system iteratively advances the cell state at a fixed number of steps. At step k, the system sets the flow time τ = k / K and inputs the current expression state x_k, source time t_src, target time t_tgt, and flow time τ into the model. The model predicts the target expression endpoint x̂_tgt, then calculates the velocity v̂_k based on the difference between the current state and the target endpoint, and completes an Euler update using x_{k+1} = x_k + v̂_kΔt. After each update, the system applies a non-negativity constraint to the expression values ​​to ensure that the generated results conform to the data characteristics of the expression matrix. After all iterations are completed, the system outputs the predicted expression matrix x_pred(t_tgt) for the target time point, which is the generated single-cell population at the held-out time point. Subsequently, the evaluation module compares the predicted results with the actual cells at the target time point, calculates HVG / PCA distribution distance, cell type kNN consistency, marker gene indices, etc., and exports the predicted artifact, protocol summary, and evaluation results. This workflow demonstrates Novaeve-Dev's core capability to generate cell population distributions for unobserved developmental stages from observed developmental stages.

[0090] The present invention and its embodiments have been described above. This description is not restrictive. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A virtual brain cell annotation method based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Load single-cell related data and protocol configuration files, verify the validity and mutual exclusion of various time point information in the files, and generate corresponding time mapping relationships based on task type; Step 2: Map all cell expression data to a fixed hypervariable gene space, construct individual cell data into a sequence containing gene identifiers, continuous expression values ​​and mask states, and encode the continuous expression values; Step 3: Embed different types of time information, combine them to form a time condition vector, and inject the vector into each layer of the Transformer model to adjust the model's operation parameters; Step 4: During model training, source time, target time, and corresponding cell samples are selected in batches, and the pairing and combination of source cells and target cells are completed within each batch. Step 5: Construct intermediate cell states using paired cells, input them into the model to predict the cell expression state at the target time point, and deduce the cell state progression speed accordingly. Model training is completed through the collaborative constraints of multiple loss functions. Step 6: Make predictions for unobserved time points. Determine the data source based on the task type, and iterate step by step from the cell state at the source time point while constraining the expression value range to obtain cell expression data at the target time point. Step 7: Evaluate the prediction results using multi-dimensional indicators, and save and export the prediction data, evaluation results, and related configuration information.

2. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The protocol configuration file records available time points, observed time points, reserved evaluation time points, task types, source time selection strategies, and highly variable gene dimensions. During verification, it is ensured that observed time points and reserved evaluation time points do not overlap, and that all time points exist in the original data.

3. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The constructed cell sequences also include classification marker characters, and the continuous expression values ​​of cells are directly encoded without discrete binning, thus preserving the detailed features of gene expression.

4. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The time condition vector is fed into the Transformer model through adaptive layer normalization to adjust the operating parameters of the model's internal attention structure and feedforward structure.

5. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The cell pairings are generated in real time within a single training batch, without the need to store fixed cell pairing relationships in advance.

6. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The loss functions used for training include flow matching loss, multi-spatial distribution distance loss, and population mean matching loss. Variance constraint loss can also be added.

7. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The cell state is updated step by step using the Euler iteration method. In each iteration, the current cell state, time information and model output results are combined to calculate the propulsion speed and complete the state update.

8. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The task types include interpolation tasks, short-range extrapolation tasks, long-range extrapolation tasks, and benchmarking tasks; interpolation tasks select observed time points adjacent to the target time point as data sources, while extrapolation tasks select historical observed time points as data sources.

9. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: The indicators used for result evaluation included spatial distribution distance, population average expression correlation, single-cell matching correlation, cell type consistency, and characteristic gene distribution difference indicators.

10. The virtual brain cell annotation method based on artificial intelligence according to claim 1, characterized in that: This method for predicting the dynamic development of a single cell comprises seven interconnected functional modules: a data loading and protocol parsing module, a fixed hypervariable gene space construction module, a time-conditional Transformer model module, an online optimal transport pairing module, a target state flow matching training module, an Euler iterative prediction generation module, and an evaluation and result export module. Each module works sequentially according to the data flow order to jointly complete the modeling, prediction, and result evaluation of the dynamic development of a single cell.

Images