A method for cell fate trajectory inference and prediction

Abstract

The application discloses a cell fate trajectory inference and prediction method, and belongs to the technical field of computational biology. The method comprises the following steps: acquiring single-cell transcriptome data and mapping to a high-dimensional shared semantic space to construct a continuous transcriptome dynamics vector field; constructing a fate evolution engine based on a neural stochastic differential equation, wherein the neural stochastic differential equation comprises a deterministic drift term and a random diffusion term parameterized by a neural network; inputting the vector field into the fate evolution engine, performing numerical integration along a continuous physical time axis to obtain a continuous fate trajectory of cell evolution along the physical time; at a lineage bifurcation decision point, introducing a random fluctuation mechanism through the random diffusion term and performing Monte Carlo path integral simulation to calculate the trend probability of cell differentiation to different branches. The application realizes long-period and high-precision continuous deduction of cell fate, can predict non-classical conversion trajectories such as cell reprogramming in a forward-looking manner, and has white-box causal explainability and self-evolution closed-loop capability.

Description

Technical Field

[0001] This invention relates to the fields of computational biology and bioinformatics, and in particular to a method for inferring and predicting cell fate trajectories. Background Technology

[0002] In the fields of life sciences and regenerative medicine, accurately predicting the fate of cells during development, reprogramming, transdifferentiation, and disease remodeling is a crucial foundation for discovering therapeutic targets, conducting stem cell organ repair, and developing innovative drugs. However, cell differentiation and evolution is a complex dynamic process controlled by a highly nonlinear gene regulatory network. How to construct a cell fate trajectory inference method that can span long periods, accurately decouple lineage bifurcation probabilities, and proactively predict unknown non-classical transition paths is a pressing technical challenge in computational biology.

[0003] Currently, the prior art most closely related to this invention can be mainly divided into the following three categories:

[0004] Category 1: Static pseudo-time trajectory inference system based on the topological similarity of high-dimensional spatial graphs.

[0005] Typical technical sources:

[0006] Monocle framework: Trapnell, C., et al. "The dynamics and regulators of cellfate decisions are revealed by pseudotime ordering of single cells." NatureBiotechnology (2014).

[0007] PAGA algorithm: Wolf, FA, et al. "PAGA: graph-abstraction addresses clustering-free trajectory inference and multiplet-testing correction." Genome Biology (2019).

[0008] This type of method projects the gene expression matrix of discrete single cells into a low-dimensional feature space, calculates the geometric distance and topological similarity between cell communities, constructs a minimum spanning tree or graph abstraction network, and assigns a "pseudo-time" scalar value to each cell along the topological curve to qualitatively reproduce the developmental sequence. Its drawbacks are: pseudo-time is not equal to real physical time, it cannot calculate the true rate of cell state migration and evolutionary velocity field, and it completely loses its prospective extrapolation prediction ability for non-classical paths (such as artificial reprogramming and transdifferentiation) that do not exist in natural datasets.

[0009] The second category: Short-cycle RNA velocity calculation tools based on cell shearing kinetics rate equations.

[0010] Typical technical sources:

[0011] Velocyto system: La Manno, G., et al. "RNA velocity of single cells." Nature (2018).

[0012] scVelo framework: Bergen, V., et al. "Generalizing RNA velocity to transient cell states through dynamical modeling." Nature Biotechnology (2020).

[0013] This type of method extrapolates the gene expression change vector of each cell within a very short future time step by establishing a transcriptional shearing kinetic rate constant equation and using the ratio of unspliced ​​to spliced ​​mRNA. Its drawbacks are: the mathematical foundation assumes that the gene transcription and degradation rates are constant throughout the entire life cycle, failing to capture complex nonlinear feedback regulation over long periods; the effective extrapolation span is only a few hours to a few days, leading to cascading amplification of errors in long-term extrapolations; and the use of deterministic ordinary differential equations ignores random molecular fluctuations at bifurcation points, making trajectory overlap and signal smoothing prone to occur at major lineage bifurcation points.

[0014] The third category: pre-trained models based on large-parameter attention mechanisms.

[0015] Typical technical sources:

[0016] scGPT large model: Cui, H., et al. "scGPT: towards a generative foundation model for single-cell biology." Nature Methods (2024).

[0017] This type of method learns a static probability distribution of high-dimensional semantics through a self-supervised mask prediction task by stacking Transformer parameters and massive tabular data, relying on attention weight matrices for perturbation prediction. Its drawbacks include: the architecture does not embed the cellular's intrinsic physical potential topology or gene causal regulatory topology; the output fate transfer paths are prone to producing "trajectory illusions" that violate the basic logic of life sciences; and it has poor robustness to highly heterogeneous clinical noise samples, resulting in low generalization and reproducibility across datasets.

[0018] In summary, existing technologies suffer from several shortcomings, including a lack of a continuous physical timeline and causal direction, extremely short extrapolation periods with smoothed-out bifurcation decision point signals, and a lack of embedded biophysical mechanisms leading to a black-box trajectory illusion. To address these deficiencies, this invention proposes a method for inferring and predicting cell fate trajectories based on neural stochastic differential equations and causal regulatory constraints. Summary of the Invention

[0019] This invention provides a method for inferring and predicting cell fate trajectories. It constructs a continuous transcriptome dynamics vector field model based on neural stochastic differential equations, and organically integrates microscopic causal regulation topology with probabilistic potential energy topography to achieve high-precision, long-span, and predictable cell fate trajectory deduction.

[0020] To achieve the above objectives, the present invention adopts the following technical solution: a method for inferring and predicting cell fate trajectories, comprising the following steps:

[0021] Single-cell transcriptome data is acquired, and the single-cell transcriptome data is mapped to a high-dimensional shared semantic space to obtain the hidden state feature vector of each cell in the high-dimensional shared semantic space, and a continuous transcriptome dynamic vector field is constructed.

[0022] A fate evolution engine based on neural stochastic differential equations is constructed. The neural stochastic differential equations include a deterministic drift term parameterized by a first neural network and a stochastic diffusion term parameterized by a second neural network. The deterministic drift term is used to characterize the backbone dynamic velocity field of cell differentiation, and the stochastic diffusion term is used to characterize the intrinsic transcriptional fluctuations and noise of the cell.

[0023] The continuous transcriptome dynamics vector field is input into the fate evolution engine, and numerical integration is performed along the continuous physical time axis to obtain the continuous fate trajectory of the cell as it evolves with physical time in the high-dimensional shared semantic space.

[0024] In the numerical integration process, at the lineage bifurcation decision point, a random fluctuation mechanism is introduced through the random diffusion term, and multiple Monte Carlo path integral simulations are performed to calculate the probability of the cell differentiating into different lineage branches.

[0025] As an improvement, the method further includes: introducing a scalar potential energy function in the high-dimensional shared semantic space, and adding the negative gradient flow constraint of the scalar potential energy function as a regularization penalty term to the training loss function of the fate evolution engine, so as to force the deterministic drift term to satisfy the negative gradient flow constraint, so that the fate evolution path of the cell spontaneously flows to the low potential energy state.

[0026] Furthermore, a known transcription factor-target gene causal regulatory topology network is obtained, which is then transformed into a directed weight mask matrix. This directed weight mask matrix is ​​then applied to the network layer of the first neural network to restrict the causal transmission direction of gene expression changes.

[0027] As an improvement, the external intervention command is encoded as an external control vector, and the external control vector is added as an input to the neural stochastic differential equation to apply a directional perturbation to the cell fate trajectory. When the external control vector is a non-zero vector, the fate evolution engine shuts down the natural developmental gradient and performs numerical integration along the artificial intervention velocity field to extrapolate and predict the non-classical cell fate transition trajectory that does not exist naturally in the prospective domain.

[0028] As an improvement, the step of "mapping the single-cell transcriptome data to a high-dimensional shared semantic space to obtain the latent state feature vector of each cell in the high-dimensional shared semantic space, and constructing a continuous transcriptome dynamic vector field" specifically includes:

[0029] The standardized single-cell transcriptome expression matrix is ​​input into the pre-trained encoder, and each cell is mapped to a hidden state feature vector in the high-dimensional shared semantic space, which has a dimension of 128 to 512.

[0030] The relative abundance ratio of uncut transcripts to cut transcripts in each cell is extracted to construct a continuous-scale dynamic rate matrix. This dynamic rate matrix is ​​then used as the initial fluid directional derivative at each cell node in the high-dimensional shared semantic space to construct the continuous transcriptome dynamic vector field.

[0031] As an improvement, the specific form of the neural stochastic differential equation is as follows:

[0032] dX(t) = [f_φ(X(t), u(t))]dt + [g_ψ(X(t), u(t))]dW(t)

[0033] Where X(t) represents the hidden state feature vector of the cell at time t, f_φ(·) is the deterministic drift term parameterized by the first neural network, u(t) is the external control vector, g_ψ(·) is the random diffusion term parameterized by the second neural network, and dW(t) is the standard Brownian motion increment.

[0034] As an improvement, the training hyperparameters of the fate evolution engine are as follows:

[0035] The total number of parameters of the first neural network and the second neural network is 50M to 1B;

[0036] The optimizer used is the AdamW optimizer, with an initial learning rate of 2×10. -4 ;

[0037] The numerical integration was solved using the Euler-Maruyama solver, with a simulation calibration step size Δt = 0.01 to 0.1 time units.

[0038] As an improvement, the non-classical cell fate transition trajectory includes: cell reprogramming trajectory, reverse development trajectory, or cross-lineage transdifferentiation trajectory; the external intervention instructions include: artificial transcription factor combination tags or drug physicochemical stimulation tags.

[0039] As an improvement, the method further includes:

[0040] The combination of key regulatory factors in the predicted output is used as the prediction result of the dry experiment to guide the targeted physical verification of the wet experiment.

[0041] Acquire the failure case data that was not hit in the wet experiment verification, and feed the failure case data back to the training end of the fate evolution engine as a negative sample stream;

[0042] The negative sample stream triggers adaptive incremental fine-tuning of the fate evolution engine to automatically prune anomalous divergent branches in the continuous transcriptome dynamics vector field.

[0043] As an improvement, the method also includes a top-level automated exploration scheduling step:

[0044] The task of inferring and predicting cell fate trajectories is broken down into multiple sub-tasks;

[0045] The scheduling data analysis agent performs cell quality control and filtering, the scheduling code generation agent automatically performs the numerical integration solution within a safe sandbox, and the scheduling biological expert agent performs causal attribution review on the deduced lineage branches, generating a white-box auditable trajectory inference report.

[0046] As an improvement, the initial data source for the continuous transcriptome dynamics vector field is splicing dynamics data based on the ratio of unspliced / spliced ​​mRNA abundance, or data based on the ratio of newly synthesized / historically present transcripts based on metabolic markers; the pre-trained encoder is a tabular multi-task pre-trained encoder, a graph autoencoder, or a cross-modal contrastive representation learning encoder.

[0047] The present invention has the following beneficial technical effects:

[0048] This invention completely abandons the traditional static pseudo-time modeling method by introducing a continuously changing velocity field model parameterized by neural stochastic differential equations. It can perform long-term forward-looking fate prediction and lineage origin tracing along a real continuous time axis without accumulating drift error in developmental tasks spanning several weeks, thus achieving an essential leap from static topological fitting to continuous dynamic deduction.

[0049] This invention explicitly introduces a stochastic diffusion term and Brownian motion constraint parameterized by a neural network into the core engine of neural stochastic differential equations. Through large-scale Monte Carlo path integral simulation at key lineage bifurcation decision points, it can accurately capture the critical threshold of cell fate bifurcation and give the probability of branch direction. The accuracy of bifurcation node topology recognition is significantly improved, effectively avoiding the mechanical smoothing defect of branch signals in traditional algorithms.

[0050] This invention deeply integrates high-dimensional dynamic vector fields with a causal inference engine. By simply changing the external control vector, it can prospectively predict the entire trajectory of cell reprogramming or transdifferentiation that does not exist naturally, without changing the base parameters. This provides a high-confidence dry experimental prediction scheme for induction formulation screening in regenerative medicine.

[0051] This invention employs a compact, high-density mechanism embedding architecture, introducing the negative gradient flow constraint of the Wattington potential energy topography map and the weight mask constraint of the directed gene regulatory network in the latent space. The dual biophysical boundaries block invalid cell fate transitions, ensuring the white-box causal interpretability of the predicted trajectory, and significantly improving robustness to multi-center heterogeneous noise data.

[0052] This invention establishes a self-evolving closed loop of model prediction, wet experiment targeted verification, conversion of failed cases into negative sample streams, and adaptive pruning of dynamic vector fields. This enables the system to continuously evolve with the continuous inflow of physical feedback data, overcoming the limitation of the performance cap of traditional static calculation models. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Example 1

[0054] Probabilistic decoupling of the fate of decision points in multiple lineages of hematopoietic stem cells.

[0055] This embodiment demonstrates the specific application of the method of the present invention in overcoming the challenges of trajectory smoothing and divergence at complex branch nodes.

[0056] Step 1: Data Input and Preprocessing. Single-cell transcriptome sequencing data of the hematopoietic system were acquired from multiple clinical centers, and the data were standardized, including quality control and normalization.

[0057] Step 2: High-dimensional shared semantic space mapping. The standardized single-cell transcriptome expression matrix is ​​input into the pre-trained encoder, mapping each cell to a hidden state feature vector in a high-dimensional shared semantic space, which is set to 256 dimensions. Simultaneously, the relative abundance ratio of uncut and cut transcripts within each cell is extracted to construct a continuous-scale dynamic rate matrix, serving as the initial fluid directional derivative at each cell node.

[0058] Step 3: Construct a neural stochastic differential equation fate evolution engine. Establish a neural stochastic differential equation of the following form:

[0059] dX(t) = [f_φ(X(t), u(t))]dt + [g_ψ(X(t), u(t))]dW(t)

[0060] The deterministic drift term f_φ is parameterized by the first multilayer perceptron neural network, representing the backbone dynamic velocity field of cell differentiation; the random diffusion term g_ψ is parameterized by the second multilayer perceptron neural network, representing the intrinsic transcriptional fluctuations and noise intensity within the cell. During training, the AdamW optimizer is used, with an initial learning rate set to 2×10⁻⁶. -4 The number of parameters was controlled at 500M, the Euler-Maruyama solver was used, and the simulation calibration step size Δt was set to 0.05 time units.

[0061] Step 4: Probabilistic decoupling of the lineage bifurcation decision point. When the trajectory advances to the bifurcation decision point where pluripotent stem cells differentiate into myeloid progenitors and megakaryocyte-erythroid progenitors, the system automatically activates the random diffusion term. Through 10,000 Monte Carlo path integral simulations, the random molecular fluctuations of cells at the bifurcation point are explicitly simulated. The simulation results accurately decouple the probability of this batch of cells differentiating into myeloid progenitors to be 62.3% and into megakaryocyte-erythroid progenitors to be 37.7%. Compared with the gold standard of classical genetic lineage tracing, the lineage branch prediction topology achieves a decision point topology identification accuracy of over 88.4%. Example 2

[0062] Prospective extrapolation prediction of non-classical pathways for cross-lineage artificial reprogramming of fibroblasts.

[0063] This embodiment demonstrates a specific application of the method of the present invention in predicting artificially intervened cell fate transition trajectories beyond the natural developmental tree.

[0064] Step 1: Load the pre-trained pedestal model. Load the fate evolution engine pedestal model trained on healthy human skin fibroblast samples. This model has been trained using negative gradient flow constraints from the Wattington potential topography map and directed weight mask constraints from the gene regulatory network.

[0065] Step 2: Apply external control perturbation. Researchers input counterfactual intervention commands to force the simulation to knock in the Yamanaka four-factor (OSKM) combination. The system encodes this command as an external control vector u(t), which is then used as a non-zero vector input to the neural stochastic differential equation.

[0066] Step 3: Non-classical trajectory extrapolation solution. The system disables the natural developmental gradient, allowing the deterministic drift term to be numerically integrated along the artificially intervened velocity field. A directed weight mask of the causal regulatory topology network filters out invalid transitions that violate biochemical mechanisms in real time. The model extrapolates to future developmental steps, predicting a non-classical transition trajectory from fibroblasts directly across lineages to induced pluripotent stem cells, including core intermediate states and rate-limiting differentiation bottlenecks during the transition.

[0067] Step 4: Evaluation of Prediction Performance. In the absence of natural single-cell training data for this reprogramming process, the list of differentially expressed genes generated by prediction was compared head-to-head with the results of subsequent physical validation using real single-cell perturbation sequencing. The area under the prediction consistency curve reached 0.85. Example 3

[0068] The dynamic vector field self-evolutionary pruning closed loop.

[0069] This embodiment demonstrates the specific process by which the method of the present invention achieves model self-evolution through wet experimental physical feedback.

[0070] Step 1: Dry Experiment Prediction Output. The system uses the combination of key regulatory factors from the prospective extrapolation prediction output in Example 2 as the dry experiment prediction result.

[0071] Step 2: Targeted wet experimental validation. The predicted combination of key regulatory factors is sent to a wet laboratory for low-throughput physical validation using targeted Perturb-seq or organoid screening.

[0072] Step 3: Negative Sample Stream Backflow. The failure case data that did not hit in the wet experiment validation is transformed into a heterogeneous negative sample stream and backflowed to the training end of the fate evolution engine.

[0073] Step 4: Adaptive Incremental Fine-tuning. Through inverse gradient adjustment, the system automatically identifies and prunes invalid and erroneous branches in the dynamic vector field caused by algorithm divergence, completing the dynamic self-growth evolution of the model. Example 4

[0074] Multi-agent automated exploration scheduling.

[0075] This embodiment illustrates the specific execution flow of the top-level automated exploration scheduling step of the present invention.

[0076] Step 1: Task Decomposition. The top-level scheduling agent decomposes the cell fate trajectory inference and prediction task into multiple sub-tasks, including cell quality control and filtering, feature extraction, dynamic vector field construction, numerical integration solution, and phylogenetic branch causal attribution review.

[0077] Step 2: Agent Scheduling. The data analysis agent is scheduled to perform cell quality control and filtering steps. The code generation agent is scheduled to automatically perform Euler-Maruyama numerical integration within a secure Docker sandbox. The biological expert agent is scheduled to conduct causal attribution review on the inferred phylogenetic branches and generate a white-boxed, auditable trajectory inference report that includes gradient visualization and SHAP feature contribution mapping.

[0078] Step 3: End-to-end full-process execution. In the authoritative BAISBench bioinformatics agent benchmark test, this method independently completed all core steps, achieving an end-to-end fully automated task pass rate of 67.5%, and the overall accuracy of cell annotation approached the level of human bioinformatics experts.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0081] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for inferring and predicting cell fate trajectories, characterized in that, Includes the following steps: Single-cell transcriptome data is acquired, and the single-cell transcriptome data is mapped to a high-dimensional shared semantic space to obtain the hidden state feature vector of each cell in the high-dimensional shared semantic space, and a continuous transcriptome dynamic vector field is constructed. A fate evolution engine based on neural stochastic differential equations is constructed. The neural stochastic differential equations include a deterministic drift term parameterized by a first neural network and a stochastic diffusion term parameterized by a second neural network. The deterministic drift term is used to characterize the backbone dynamic velocity field of cell differentiation, and the stochastic diffusion term is used to characterize the intrinsic transcriptional fluctuations and noise of the cell. The continuous transcriptome dynamics vector field is input into the fate evolution engine, and numerical integration is performed along the continuous physical time axis to obtain the continuous fate trajectory of the cell as it evolves with physical time in the high-dimensional shared semantic space. In the numerical integration process, at the lineage bifurcation decision point, a random fluctuation mechanism is introduced through the random diffusion term, and multiple Monte Carlo path integral simulations are performed to calculate the probability of the cell differentiating into different lineage branches.

2. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: The method further includes: introducing a scalar potential energy function in the high-dimensional shared semantic space, and adding the negative gradient flow constraint of the scalar potential energy function as a regularization penalty term to the training loss function of the fate evolution engine, so as to force the deterministic drift term to satisfy the negative gradient flow constraint, so that the fate evolution path of the cell spontaneously flows to the low potential energy state. Furthermore, a known transcription factor-target gene causal regulatory topology network is obtained, which is then transformed into a directed weight mask matrix. This directed weight mask matrix is ​​then applied to the network layer of the first neural network to restrict the causal transmission direction of gene expression changes.

3. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: External intervention instructions are encoded into external control vectors, and these external control vectors are added as input terms to the neural stochastic differential equations to apply directional perturbations to the fate trajectories of cells. When the external control vector is a non-zero vector, the fate evolution engine shuts down the natural developmental gradient and performs numerical integration along the artificially intervened velocity field to extrapolate and predict the entire trajectory of non-classical cell fate transitions that do not exist naturally in the look-ahead domain.

4. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: The step of "mapping the single-cell transcriptome data to a high-dimensional shared semantic space, obtaining the latent state feature vector of each cell in the high-dimensional shared semantic space, and constructing a continuous transcriptome dynamic vector field" specifically includes: The standardized single-cell transcriptome expression matrix is ​​input into the pre-trained encoder, and each cell is mapped to a hidden state feature vector in the high-dimensional shared semantic space, which has a dimension of 128 to 512. The relative abundance ratio of uncut transcripts to cut transcripts in each cell is extracted to construct a continuous-scale dynamic rate matrix. This dynamic rate matrix is ​​then used as the initial fluid directional derivative at each cell node in the high-dimensional shared semantic space to construct the continuous transcriptome dynamic vector field.

5. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: The specific form of the neural stochastic differential equation is as follows: dX(t) = [f_φ(X(t), u(t))]dt + [g_ψ(X(t), u(t))]dW(t) Where X(t) represents the hidden state feature vector of the cell at time t, f_φ(·) is the deterministic drift term parameterized by the first neural network, u(t) is the external control vector, g_ψ(·) is the random diffusion term parameterized by the second neural network, and dW(t) is the standard Brownian motion increment.

6. The method for inferring and predicting cell fate trajectories according to claim 5, characterized in that: The training hyperparameters of the fate evolution engine are as follows: The total number of parameters of the first neural network and the second neural network is 50M to 1B; The optimizer used is the AdamW optimizer, with an initial learning rate of 2×10. -4 ; The numerical integration was solved using the Euler-Maruyama solver, with a simulation calibration step size Δt = 0.01 to 0.1 time units.

7. The method for inferring and predicting cell fate trajectories according to claim 3, characterized in that: The non-classical cell fate transition trajectory includes: cell reprogramming trajectory, reverse development trajectory, or cross-lineage transdifferentiation trajectory; the external intervention instructions include: artificial transcription factor combination tags or drug physicochemical stimulation tags.

8. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: The method further includes: The combination of key regulatory factors in the predicted output is used as the prediction result of the dry experiment to guide the targeted physical verification of the wet experiment. Acquire the failure case data that was not hit in the wet experiment verification, and feed the failure case data back to the training end of the fate evolution engine as a negative sample stream; The negative sample stream triggers adaptive incremental fine-tuning of the fate evolution engine to automatically prune anomalous divergent branches in the continuous transcriptome dynamics vector field.

9. The method for inferring and predicting cell fate trajectories according to claim 1, characterized in that: The method also includes a top-level automated exploration scheduling step: The task of inferring and predicting cell fate trajectories is broken down into multiple sub-tasks; The scheduling data analysis agent performs cell quality control and filtering, the scheduling code generation agent automatically performs the numerical integration solution within a safe sandbox, and the scheduling biological expert agent performs causal attribution review on the deduced lineage branches, generating a white-box auditable trajectory inference report.

10. A method for inferring and predicting cell fate trajectories according to claims 1-9, characterized in that: The initial data source for the continuous transcriptome dynamics vector field is splicing dynamics data based on the ratio of unspliced / spliced ​​mRNA abundance, or data based on the ratio of newly synthesized / historically present transcripts based on metabolic markers; the pre-trained encoder is a tabular multi-task pre-trained encoder, a graph autoencoder, or a cross-modal contrastive representation learning encoder.