Artificial intelligence driven virtual cell biology simulation platform and method

Abstract

The application discloses an artificial intelligence driven virtual cell biology simulation platform and method, belongs to the field of computer simulation, and comprises a multi-cell line co-culture simulation module, a molecular circuit simulation module and a multi-omics analysis module. Culture environment states are represented by four mutually orthogonal scalar parameters, i.e. an environment reaction rate multiplier, an environment attenuation multiplier, a basal stress level and an initial inoculation density. The co-culture module calls a large language model to predict population quantity and culture medium health index by steps and feeds back to the next step. The molecular circuit module pre-compiles a circuit directed graph into a reaction instruction set, integrates the chemical state of a cell agent according to the Michaelis enzyme kinetics equation and triggers threshold type phenotype rules. The multi-omics module analyzes reagent description into expression folding change parameter objects through pre-evaluation, and then calls a single cell transcriptome basic model and a large language model to output omics data, so that consistent simulation across levels can be realized without establishing differential equations for each type of culture environment.

Description

Technical Field

[0001] This invention relates to the field of computer simulation technology at the intersection of computational biology and artificial intelligence, specifically to an artificial intelligence-driven virtual cell biology simulation platform and method. Background Technology

[0002] Modern cell biology research faces three independent yet intertwined technical bottlenecks in the experimental design phase.

[0003] The first bottleneck lies in the high cost, long cycle, and lack of pre-screening tools for multi-cell line co-culture experiments. In vitro cell co-culture experiments require specialized technicians to operate in biosafety cabinets, with experimental cycles typically lasting 3 to 7 days. This consumes expensive cell culture consumables and serum, and the results are highly uncertain due to competition for space and nutrients among multiple cell lines, depletion of culture medium components over time, and the specific tolerance differences of each cell line to temperature and pH. Existing population dynamics tools mostly use the Lotka-Volterra equation or its variants to describe interspecies competition. These equations require a pre-defined competition coefficient matrix between each cell line, which cannot be determined without wet experimental data, preventing researchers from effectively pre-screening multiple culture protocols using computers before the experiment.

[0004] The second bottleneck lies in the high barrier to entry for modeling intracellular signal transduction pathways, which is disconnected from cellular physical behavior. Cellular signaling pathway research typically relies on specialized systems biology software, such as MATLAB SimBiology or COPASI. This software requires researchers to manually input numerous kinetic parameters and describe reaction networks using systems biology markup languages, posing a significant barrier to use for experimental researchers without modeling expertise. More importantly, this software solves a system of ordinary differential equations relating molecular concentration over time, with the state variable being the average concentration of the entire cell population, which cannot be coupled with the spatial location, movement, collisions, and division of individual cells. On the other hand, while agent-based multi-cell simulation frameworks such as PhysiCell and PhysiBoSS can characterize cellular physical behavior, their intracellular decision-making layers typically employ Boolean networks or simplified empirical rules, lacking realistic enzyme reaction kinetics. To simultaneously obtain molecular-level kinetics and cellular-level physical behavior, researchers must manually transfer parameters between the two software programs, and the entire model needs to be recalibrated every time a different culture medium formulation is used.

[0005] The third bottleneck lies in the lack of efficient tools for multi-omics prediction in gene perturbation experiments. Gene knockout or knockdown experiments, including small interfering RNA, short hairpin RNA, and techniques involving regularly clustered short palindromic repeats, require 48 to 72 hours of experimental manipulation followed by several weeks of data analysis, resulting in high costs and long cycles. Existing transcriptome prediction tools, such as PerturbNet and GenePT, require pre-structured gene perturbation vectors as input and lack the ability to directly generate simulation parameters from natural language reagent descriptions. Furthermore, these tools output a single-modality expression matrix, failing to provide a comprehensive dashboard integrating multi-dimensional omics indicators such as protein abundance, pathway activity, oxidative stress, and epigenetic status, and offering no actionable experimental validation suggestions. Moreover, these tools are typically deployed as GPU inference services; when the inference service becomes unavailable due to cold starts, capacity expansion, or network issues, the entire analysis process fails, and the user's invested waiting time and quota are completely lost.

[0006] The reason why existing technologies have long failed to break through the above three types of bottlenecks lies in a widely accepted but not valid understanding: those skilled in the art generally believe that molecular-level reaction dynamics and phenotypic behavior at the cellular population level belong to two independent modeling systems. The former's state carrier is the concentration field and the time scale is from milliseconds to seconds, while the latter's state carrier is the individual cell and the time scale is from hours to days. Therefore, they must be modeled and solved separately. At the same time, it is generally believed that the influence of the culture environment on the simulation system is multi-channel and strongly coupled. Nutrient limitation, changes in oxygen tension, and toxic stress each act on different biochemical processes. Therefore, a separate differential equation system or a separate empirical model must be established for each type of environment.

[0007] Research has revealed biases in both of these understandings. First, the difficulty in unifying the molecular and cellular levels stems not from differences in their physical mechanisms, but solely from the lack of uniformity in their state carriers and time steps. Once molecular circuits are pre-compiled into a reaction instruction set decoupled from the graph structure, and chemical states are reassigned from the population average concentration field to each cellular agent, the two levels can share the same time step for synchronous advancement. Second, the entire impact of the culture environment on the simulation system can be compressed into four orthogonal channels acting on reaction rate, decay rate, basal apoptosis probability, and initial cell number, without requiring separate equations for each type of environment. This is because nutrient limitation ultimately manifests biochemically as an overall downregulation of enzymatic reaction rates, changes in oxidation and proteasome activity ultimately manifest biochemically as an overall upregulation of degradation rates, and toxic stress ultimately manifests biochemically as a basal mortality probability independent of concentration states. Mathematically, these three fall precisely at three non-overlapping positions: the multiplicative rate channel, the multiplicative decay channel, and the additive probability channel.

[0008] The aforementioned cognitive biases have led to a fragmented landscape in existing tools: some only calculate molecules without considering cellular behavior, while others only calculate cells without considering real enzyme kinetics, and the entire model must be recalibrated for every change in culture conditions. This invention aims to correct these cognitive biases by providing a virtual cell biology simulation platform and method that unifies the molecular, cellular, and population layers within the same simulation system. Summary of the Invention

[0009] To address the core bottlenecks in existing cell biology simulation technologies, namely the disconnect between molecular-level dynamics and cellular-level physical behavior, and the need to establish separate differential equation systems for each type of culture environment, this invention provides an artificial intelligence-driven virtual cell biology simulation platform and method. By compressing the culture environment state into four mutually orthogonal scalar parameters and simultaneously modulating three functional modules, precompiling molecular circuits into reaction instruction sets decoupled from graph structures, and attaching chemical states to the cell agent, this invention achieves consistent cell biology simulation across molecular, cellular, and population levels from the principle of unified state carrier and time step, without relying on wet experimental calibration of competition coefficient matrices or establishing separate differential equation systems for each type of culture environment.

[0010] The technical solution of this invention is as follows: an artificial intelligence-driven virtual cell biology simulation platform, comprising a multi-cell line co-culture simulation module, a molecular circuit simulation module, and a multi-omics analysis module sharing the same culture environment state. The culture environment state is characterized by four mutually orthogonal scalar parameters: environmental reaction rate multiplier, environmental decay multiplier, basal stress level, and initial seeding density. The multi-cell line co-culture simulation module calls a large language model at a fixed simulation step size to predict the population size and culture medium health index of each cell line. It performs non-negative integer verification and clamping on the population size and injects the verified culture medium health index into the prompt context of the next simulation step. The molecular circuit simulation module pre-compiles the directed graph of the molecular circuit into a reaction instruction set, modulates the reaction rate constant and decay rate constant with the environmental reaction rate multiplier and the environmental decay multiplier, respectively, performs a first-order Euler integral on the chemical state of the cell agent according to the Michaelis enzyme kinetic equation, clamps the concentration to a closed interval between zero and one, triggers the behavior of the cell agent according to threshold-type phenotypic rules, and determines the basal apoptosis probability of the cell agent with the basal stress level. In the pre-evaluation phase, the multi-omics analysis module parses the natural language reagent description into gene expression folding change parameter objects, and then uses these parameter objects to call the single-cell transcriptome basic model inference service and the large language model to output multi-omics dashboard data.

[0011] This invention also provides an AI-driven gene perturbation cellular omics simulation method, comprising: step M1, receiving detailed specifications of cell line type, passage number, seeding density, culture medium type, oxygen tension, carbon dioxide concentration, and at least one experimental reagent; step M2, performing a pre-assessment without deducting running quotas, converting the natural language reagent description into a structured parameter object using a bidirectional deterministic numerical mapping; step M3, invoking a single-cell transcriptome basic model inference service deployed on a graphics processor cloud computing infrastructure, and switching to a synthesis inference mode based on a large language model when the service is unavailable; step M4, calculating the time series of cell number, viability percentage, and reactive oxygen species index using pathway activity scores as parameters; step M5, injecting the large language model to generate multidimensional omics dashboard data and performing structural validation; and step M6, presenting the results with an interactive graphical user interface.

[0012] Beneficial effects:

[0013] First, this invention reduces the number of modeling parameters for the culture environment from one set of kinetic parameters per type of environment to four scalars. The mechanism lies in the fact that the three environmental factors—nutrient limitation, oxidative degradation, and toxic stress—mathematically fall into the multiplicative rate channel, the multiplicative decay channel, and the additive probability channel, respectively. These three are non-overlapping and therefore can be orthogonally decomposed. Assuming the platform supports… Culture environment, molecular circuits Each reaction requires calibration according to the existing method of calibrating each type of environment separately. The dynamic parameters of the group, but according to the present invention only need to be calibrated. Group reference parameters and An environmental scalar. When , In this case, the number of dynamic parameter sets that need to be calibrated is reduced from 36 sets to 9 sets, plus 16 scalars, reducing the parameter calibration workload by approximately 75%. This comparison can be directly verified using the parameter tables described in this manual.

[0014] Second, this invention precompiles the directed graph of molecular circuits into a set of reaction instructions that can be directly addressed by source node indexes. The mechanism is that the circuit topology does not change during a single simulation, thus the results of graph traversal can be reused. Compared to the implementation method of re-traversing the circuit graph at each simulation step, this invention reduces the time complexity of single-step reaction lookup from... Reduced to ,in For the number of nodes, Let O be the number of edges; the total time complexity of a single step is O(n). ,in For the number of cell agents, The number of response commands. When , , In this way, each simulation step can avoid 400 graph traversals, which means avoiding approximately 8,000 visits to nodes and edges.

[0015] Third, this invention changes the chemical state from a population average concentration field to one that is attached to each cellular agent, allowing the first-order Euler integral at the molecular level to share the same time step with Brownian motion, elastic collisions, and mitosis at the cellular level. Its nonlinear synergistic effect lies in the fact that daughter cell agents inherit the complete chemical state of the mother cell agent during mitosis, thereby amplifying the concentration differences at the molecular level at the cellular level, producing clonal heterogeneity that neither simple concentration field models nor simple agent models can generate. Simultaneously, when the concentration of specific nodes in the cytoplasm exceeds a threshold, it triggers detachment and cell cycle arrest, enabling molecular events to reversely alter the cell's physical parameters, forming a bidirectional coupling between the molecular and cellular levels.

[0016] Fourth, this invention establishes a unique narrow interface between the natural language semantic layer and the single-cell statistical layer, using a bidirectional deterministic numerical mapping between the knockdown percentage and expression folding changes as the sole information channel between the two layers. The mechanism lies in the fact that describing gene expression fold changes using natural language reagents is a semantic parsing problem requiring inference from world knowledge and mechanisms, while describing gene expression fold changes to the whole transcriptome response is a distribution extrapolation problem requiring pre-trained representations using large-scale single-cell corpora. The inductive biases required for both are orthogonal, and using the same model will inevitably degrade one end. Setting a narrow interface produces three effects that cannot be achieved by setting either layer alone: ​​first, the two models can be replaced independently without affecting each other; second, when the graphics processor inference service is unavailable, the large language model can generate synthetic inference results with the same structure, increasing the success rate of the simulation process from the availability of the inference service. The cost of upgrading to 1 is only to mark the data source identifier in the response; thirdly, the semantic parsing stage does not require graphics processor resources, thus eliminating the need for runtime quotas and allowing users to obtain risk level assessments before consuming advanced quotas.

[0017] Fifth, this invention uses a single logistic decay function to characterize passage-dependent cell aging. The two parameters, half-life passage number and passage decay width, precisely define the boundaries between the five stages—early passage, mild aging, moderate aging, severe aging, and crisis period—falling precisely at passages 5, 10, 15, and 20, respectively, eliminating the need for separate segmented models for each stage. According to the parameter values ​​described in this specification, the proliferation capacity of mesenchymal stem cells at passage 18 decreased by 38.9% compared to passage 4, and the pathway response amplitude decreased by 34.2%, consistent with the well-known phenomenon in the art that mesenchymal stem cells exhibit significant replicative aging after in vitro expansion to passage 15 and beyond.

[0018] Sixth, the three functional modules of this invention share a unified runtime AI provider configuration management system. The configuration is encrypted using a hardware or software-based secure encryption storage mechanism and stored in a relational database. It is read according to a three-level priority system: database encrypted configuration, environment variables, and built-in default configuration. This allows system administrators to switch large language model providers at runtime without redeploying the code. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the human-computer interaction interface of the multi-cell line co-culture simulation module of the present invention;

[0020] Figure 2 This is a schematic diagram of the human-computer interaction interface of the molecular circuit simulation module of the present invention;

[0021] Figure 3 This is a schematic diagram of the experimental configuration and four-dimensional time monitoring interface of the multi-omics analysis module in Embodiment 3 of the present invention;

[0022] Figure 4 This is a schematic diagram of the protein abundance estimation interface in Embodiment 3 of the present invention;

[0023] Figure 5 This is a schematic diagram of the transcriptome volcano diagram interface in Embodiment 3 of the present invention;

[0024] Figure 6 This is a schematic diagram of the eight-axis path active radar map interface in Embodiment 3 of the present invention;

[0025] Figure 7 This is a schematic diagram of the oxidative stress analysis interface in Embodiment 3 of the present invention;

[0026] Figure 8 This is a schematic diagram of the epigenetic state inference interface in Embodiment 3 of the present invention;

[0027] Figure 9 This is a schematic diagram of the user-defined parameter time series interface in Embodiment 3 of the present invention;

[0028] Figure 10 This is a schematic diagram of the artificial intelligence comprehensive summary interface in Embodiment 3 of the present invention;

[0029] Figure 11 This is a schematic diagram of the cell configuration and custom monitoring parameter input interface in Embodiment 5 of the present invention;

[0030] Figure 12 This is a schematic diagram of the time series interface for the three custom parameters in Embodiment 5 of the present invention;

[0031] Figure 13 This is a schematic diagram of the four-dimensional time monitoring and pre-evaluation interface in Embodiment 5 of the present invention;

[0032] Figure 14 This is a schematic diagram of the transcriptome volcano diagram interface in Embodiment 5 of the present invention;

[0033] Figure 15 This is a schematic diagram of the protein abundance estimation interface in Embodiment 5 of the present invention;

[0034] Figure 16 This is a schematic diagram of the eight-axis path active radar map interface in Embodiment 5 of the present invention;

[0035] Figure 17 This is a schematic diagram of the oxidative stress analysis interface in Embodiment 5 of the present invention;

[0036] Figure 18 This is a schematic diagram of the epigenetic state inference interface in Embodiment 5 of the present invention;

[0037] Figure 19 This is a schematic diagram of the comprehensive summary interface for artificial intelligence in Embodiment 5 of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0039] The virtual cell biology simulation platform of this invention is provided as a service through a unified network interface and includes three mutually cooperating functional modules. For example... Figure 1 As shown, the interface of the multi-cell line co-culture simulation module provides a cell population configuration area and a culture condition parameter area on the left, a microscopic view of the culture dish in the middle, a multi-cell line population-time line graph in the upper right, and a simulation log area at the bottom. Figure 2 As shown, the interface of the molecular circuit simulation module provides a node type toolbar on the left, a molecular circuit canvas divided into four biological spatial partitions at the top center, a physical rule bar at the bottom center, a cell agent canvas at the top right, and an artificial intelligence pathway generation dialog panel at the bottom right. Figure 3 and Figure 11 As shown, the interface of the multi-omics analysis module provides a cell configuration area, a custom monitoring parameter area, and a reagent workshop on the left, a four-dimensional time monitor and a pre-assessment report area on the right, and an omics dashboard consisting of seven tabs at the bottom. The three modules share the following common technical facilities.

[0040] The first is a runtime AI provider configuration management system. The large language model provider configuration includes interface keys, service addresses, model identifiers, temperature parameters, and system prompts specific to each function module. This configuration is stored in a relational database after being symmetrically encrypted using a hardware- or software-based secure encryption mechanism. The system reads configurations according to a three-tier priority: first, it attempts to read the encrypted configuration from the database; if it does not exist, it reads the environment variables; if it still does not exist, it uses the built-in default configuration. Configuration changes retain a timestamped version history. The user interfaces of all three modules provide visual indicators showing the source of the current configuration: database configurations are displayed as a C-mode badge, and environment variable configurations are displayed as a G-mode badge.

[0041] Secondly, there is quota and billing management. Each step of the multi-cell co-culture simulation deducts one credit point, with the service identifier being the co-culture simulation service; each query for the artificial intelligence design of molecular circuits deducts one credit point, with the service identifier being the tissue culture simulation service; each call to the multi-omics simulation deducts one advanced run quota; no credit points or quotas are deducted during the pre-evaluation phase.

[0042] Thirdly, there is user authentication and access control. All interface requests are authenticated through the identity verification framework. If there are insufficient credit points, a 402 status code response is returned, and all functional modules are inaccessible to guest users.

[0043] One of the core mechanisms of this invention is to compress the entire influence of the culture environment on the simulation system into four mutually orthogonal scalar parameters. Let the culture environment state be denoted as... ,but ,in, The environmental reaction rate multiplier is used to apply a uniform multiplicative modulation to the reaction rate constants of all enzyme-catalyzed reactions. Its value ranges from 0.1 to 2.0 and is dimensionless. The environmental attenuation multiplier is used to apply a uniform multiplicative modulation to the attenuation rate constant of all nodes. Its value ranges from 0.5 to 3.0 and is dimensionless. The basal stress level is used to determine the single-step basal apoptosis probability independent of chemical state, and its value ranges from 0 to 1.0, and is dimensionless. The initial inoculation density refers to the number of cell agents created when the simulation starts, ranging from 10 to 400, measured in cells. It should be noted that the orthogonality here refers to the four parameters acting on non-overlapping computational channels, not statistically uncorrelatedness.

[0044] This embodiment provides four standardized culture environment conditions, and the parameter values ​​are shown in Table 1.

[0045] Table 1. Parameter values ​​for four standardized culture environment conditions.

[0046]

[0047] The basis for determining the values ​​of each parameter in Table 1 is as follows. The baseline value is 1.0, corresponding to the signal transduction intensity maintained by conventional complete culture medium with a serum concentration of 10%; the complete culture medium is 1.5 because it contains complete growth factors and amino acid supply, which can maintain kinase activity higher than the baseline; the starvation culture medium is 0.3, corresponding to the phosphorylation level of extracellular signal-regulated kinases reduced to the baseline level under serum starvation treatment; the toxic culture medium is 0.4, reflecting the damage to the signaling pathway. The baseline value is 1.0, corresponding to the protein half-life under normal conditions; 0.8 for complete culture medium corresponds to the inhibition of protein degradation by growth factors; 1.5 for starvation culture medium corresponds to the upregulation of autophagy and proteasome activities; and 2.5 for toxic culture medium corresponds to rapid protein hydrolysis. These values ​​can be corrected online according to the measured semi-quantitative blot grayscale ratio of phosphorylated proteins, with the correction rule being... ,in and These are the measured and predicted grayscale ratios, respectively. The learning rate is set to 0.2. When the user does not specify the culture environment state, the system degenerates to the parameter values ​​used with complete culture medium.

[0048] The transmission relationship of the four parameters to the three modules is as follows. For the molecular circuit simulation module, the first... The effective reaction rate constant of the first reaction and the second reaction The effective decay rate constants for each node are as follows:

[0049] ,in For the first The baseline reaction rate constant for the given reaction, in units of ; For the first The reference decay rate constant for each node, in units of For basal apoptosis of cell agents, within each simulation step, any cell agent with a random number less than the probability given by the following formula enters a near-death state and is removed in subsequent steps when it enters a death state:

[0050] ,in is the basal apoptosis rate constant, taken as 0.1, in units of . ; The simulation step size is set to 0.01, with units of [unit missing]. It should be noted here that... This is an implicit constant with dimensions. Omitting its unit would cause dimensional inconsistencies on both sides of the equation; therefore, this invention explicitly states its unit. Taking a toxic culture medium as an example... The probability of single-step apoptosis is Based on 100 steps per second, the cumulative apoptosis rate per second is: The rate of acute cell death under the action of toxic agents is on the same order of magnitude.

[0051] For the multi-cell line co-culture simulation module and Write the context description of the prompt in natural language. Used to determine the initial population size for each cell line.

[0052] Thus, nutritional restriction was achieved through Channels, oxidation, and proteasome degradation via Channels, toxic stress pathways Each channel exerts its influence independently, and the three channels do not overlap computationally, thus eliminating the need to establish a separate differential equation system for each type of culture environment.

[0053] like Figure 2 As shown, molecular circuits are represented by a directed graph. Representation. Node set Each node in the dataset possesses type, position, initial normalized concentration, and decay rate attributes. Node types include eight categories: ligands, chemicals, mixtures, pathogens, receptors, kinases, phosphatases, and transcription factors. Each type has a different visual shape representation: ligands are tri-circular molecular clusters, receptors are cup-shaped with a transmembrane domain, kinases are bilobed kidney-shaped, phosphatases are pod-shaped, transcription factors are rounded rectangles containing DNA binding, chemicals are hexagonal benzene rings, and pathogens are starburst-shaped. The scalable vector graphics canvas is divided into four biological spatial regions: extracellular, membrane, cytoplasm, and nucleus. Each region is marked with a specific color background. Nodes automatically attach to their corresponding regions based on their position attributes. The vertical coordinate ranges for the four regions are 20-80, 100-180, 200-300, and 320-450, respectively. Edge set Each edge in the algorithm has a source node, a target node, a reaction type, and a reaction rate constant. The reaction types include three types: activation, inhibition, and binding.

[0054] The system has built-in preset signaling pathway networks for 11 cell lines, covering 10 standard signaling pathways. Each preset network contains 6 to 11 nodes and 5 to 9 edges, and the specific correspondence is shown in Table 2.

[0055] Table 2 Preset cell line signaling pathway network

[0056]

[0057] Users can create nodes by dragging and dropping, hold down the Shift key and click to establish a connection edge, and double-click to edit node properties. Node properties include name, type, initial concentration, decay rate and mutation state flag, where the initial concentration value ranges from 0 to 1 and the decay rate value ranges from 0 to 0.5.

[0058] Before the simulation starts, the loop compiler will generate a directed graph. Translated into a reaction instruction set The reaction instruction set is a linear list consisting of four tuples: ,in and The first The source node index and target node index of each reaction are both integers; This is an identifier for the reaction type, with values ​​of activation, inhibition, or binding. The total number of response instructions is equal to the number of edges. The reaction instruction set is organized into a directly addressable structure based on the source node index, making the reaction lookup overhead within each simulation step constant, thereby eliminating the overhead generated by traversing the directed graph at each step. The compiler also extracts the initial concentration and decay rate of each node and marks four types of input nodes—ligands, chemicals, mixtures, and pathogens—as exempt from decay, representing exogenous molecules that persist in the culture medium.

[0059] For each cellular agent, all biochemical reactions were numerically integrated according to the Michaelis enzyme kinetic equation. For the first The normalized concentration of each node is a dimensionless quantity. For the activation reaction, the concentration increment of the target node is: In the formula The normalized concentration of the upstream node. This is the residual capacity term for the target node, used to simulate substrate saturation effects. Here, the residual capacity term is used instead of the one in the classical Michaelis equation. The form is because when the upper limit of the normalized concentration is 1, the two are in The limit is equivalent to the remaining capacity term, which naturally guarantees that the concentration does not exceed the upper bound.

[0060] For the inhibition reaction, the concentration reduction of the target node is: This formula achieves a reduction in enzyme concentration, and is suitable for phosphatase inhibition and negative feedback regulation. When As the concentration decreases, it approaches zero, thus the concentration does not exceed the lower limit.

[0061] For the binding reaction between the ligand and the receptor, the same kinetics as the activation reaction are used to simulate the establishment of receptor occupancy.

[0062] For each non-input node's decay and degradation, the concentration reduction is: ,

[0063] In the Within the simulation step, the first The concentration of each node is updated as follows:

[0064] ,in This is a clamping operator used to prohibit negative concentrations and ensure that the concentration does not exceed the upper bound. , The upper limit is At that time, the absolute value of the single-step concentration increment does not exceed The truncation error of the first-order Euler method is much smaller than the width of the concentration interval by 1, therefore the truncation error is acceptable in this application scenario, and there is no need to use a higher-order integral scheme. Other notations, such as the summation sign, absolute value, and minimax operators, are standard notations in this field and will not be elaborated further.

[0065] Within a single simulation step, for all The time complexity for an agent cell to perform the above integration is: .when , At that time, the number of floating-point operations per step is approximately This can be completed in milliseconds on a regular personal computer, thus supporting real-time rendering.

[0066] Each cell agent maintains the following state variables: unique identifier, cell lineage type, and canvas coordinates. Speed ​​of movement Orientation angle, cell radius Cell cycle stages, chemical states The data includes cell age, infection status, detachment status, and the time of the most recent division. Cell cycle stages are represented by six states: G1, S, G2, M, near-death, and death. Canvas coordinates and cell radius are in pixels, and the simulation area is 600 pixels by 400 pixels.

[0067] It is necessary to clarify the mapping relationship between simulation time and biological time here. For time-scale mapping coefficients, this implementation takes... That is, 1 second of simulation time corresponds to 8 hours of culture time, then the biological time The existence of this mapping coefficient allows parameters such as the minimum division interval discussed later to correspond to real cell biological timescales.

[0068] Within each simulation step, a random Brownian force is applied to each cell agent, with the velocity increment being: ,in The coefficients are cell line-specific motility coefficients, dimensionless, with 1.2 for fast-moving human embryonic kidney 293T cells, 1.0 for medium-moving cervical cancer HeLa cells, and 0.4 for slow-moving cardiomyocytes. Let be the Brownian force intensity constant, and take . ; Let be a dimensionless random vector that follows a two-dimensional standard normal distribution. Then, update the position and apply velocity decay as follows: ,in Let be the velocity retention coefficient, dimensionless, 0.90 for adherent cells and 0.99 for detached cells. The velocity retention coefficient is higher in the detached state because after trypsin digestion, the cells retract into a spherical shape, significantly reducing the adhesion area to the culture medium substrate and thus decreasing motion damping. For any two cell agents, let the center-to-center distance be denoted as . .when If overlap is detected, a symmetrical repulsive displacement is applied along the line connecting the centers of the two cells using the following formula:

[0069] In the formula This is a unilateral repulsive displacement, measured in pixels. This formula ensures that the two cells are exactly tangent after repulsion, and the centroid position remains unchanged. Simultaneously, boundary reflection is performed, constraining the canvas coordinates to... , .

[0070] Mitosis is triggered when the cell agent simultaneously meets the following four conditions: Condition 1, the cell is in the G1 phase; Condition 2, the time since the most recent division exceeds the minimum interval between divisions. Condition 3: The current total population size does not exceed the maximum population limit. Condition four: The random sampling passes the probability test, i.e. ,in, In the formula The basic splitting probability is set to 0.005, which is the dimensionless probability per simulation step. The cell line-specific proliferation coefficient is dimensionless, with 1.2 to 1.5 for rapidly proliferating cell lines and 0.6 to 0.8 for slowly proliferating cell lines. This is the passage aging correction factor defined in Section 8 below. Conversion, Corresponding to a biological timeframe of 24 hours, consistent with the typical doubling time of mammalian adherent cells. The daughter cell agent inherits the complete chemical and infection states of the parent cell agent and is subjected to an initial velocity in the opposite direction to achieve physical separation. This inheritance of infection state enables vertical viral transmission.

[0071] Users or artificial intelligence can define multiple phenotypic rules, each rule being a quadruple: ,in For monitoring node identifiers; The concentration threshold is dimensionless. For comparison conditions; These are the trigger actions. There are six trigger actions: accelerated mitosis, programmed cell death, accelerated migration, reprogramming to an induced pluripotent stem cell state, differentiation into a specialized lineage, and viral infection propagation. Within each simulation step, each cell agent is evaluated individually for each trigger action. Once the condition is met, the corresponding action will be executed.

[0072] threshold The values ​​are determined based on the following criteria: For transcription factor monitoring nodes, A value of 0.70 to 0.80 corresponds to the transcription factor occupying more than two-thirds of the promoter region, at which point the transcription of downstream target genes enters a plateau phase; for toxin and pathogen monitoring nodes, A value of 0.30 to 0.50 corresponds to the initiation threshold of cellular stress response; for differentiation factor monitoring nodes, Take 0.60. When the user does not specify... At that time, the system automatically selects the median value of the above interval according to the type of monitoring node. The online correction can be performed based on the deviation between the simulation results and the measured phenotypic ratio. The correction rule is as follows: ,in and The figures show the proportions of positive cells in simulation and actual measurements, respectively. Set the value to 0.1. This rule is skipped without interrupting the simulation when the monitoring node identifier does not exist in the current loop.

[0073] Cell detachment modeling is as follows: When the normalized concentration of trypsin nodes in the cytoplasm exceeds the detachment threshold of 0.3, the cell agent is marked as detached, its velocity retention coefficient switches from 0.90 to 0.99, and the cell cycle is arrested to simulate the enzymatic passage process.

[0074] like Figure 2 As shown in the physical rules bar below, the currently active phenotypic rules are displayed in real time as labels, and users can enable or disable any rule immediately.

[0075] like Figure 1As shown on the left, the cell population configuration system supports no fewer than 60 preset mammalian cell lines, including cervical cancer HeLa cells, human embryonic kidney 293T cells, human embryonic kidney 293 cells, lung adenocarcinoma A549 cells, breast cancer MCF-7 cells, triple-negative breast cancer MDA-MB-231 cells, acute T-cell leukemia Jurkat cells, Chinese hamster ovary cells, liver cancer HepG2 cells, mouse embryonic fibroblast NIH / 3T3 cells, rat pheochromocytoma PC-12 cells, neuroblastoma SH-SY5Y cells, African green monkey kidney Vero cells, colon adenocarcinoma Caco-2 cells, and glioblastoma U87MG cells. Each cell line has a unique identifier, custom color coding, and biological characteristic descriptions, which are used as simulation context by the large language model. Users can also create custom cell lines and input any biological characteristics.

[0076] The culture condition parameter system includes no fewer than seven independently adjustable parameters: temperature, pH, glucose level, serum percentage, antibiotic presence, culture medium type, and custom reagents. Temperature is in degrees Celsius; pH ranges from 0 to 14; glucose levels are categorized as low, normal, and high; serum percentage ranges from 0 to 100%; culture medium types include Dalberg modified Eagle medium, Roswell Park Memorial Institute 1640 medium, Ham F12 medium, neural basal medium, and a custom minimalist medium; and custom reagents include drug name and concentration.

[0077] Culture medium health index Expressed as a percentage, it is used to simulate the cumulative effect of waste accumulation and nutrient depletion. Within each simulation step, the large language model is adjusted according to the current... Update the population size of each cell line and output the updated values. The system performs security verification on the output of the large language model: the population size is constrained to a non-negative integer range, and backoff processing is performed for non-numerical and negative values. During backoff, the health of the culture medium is estimated using the following formula: ,in For the first The total population size at each step, expressed in individuals; The maximum carrying capacity is set to 1500, with the unit being individuals; The culture medium consumption coefficient is taken as 0.08, which is dimensionless. This represents the incubation time corresponding to the simulation step size, defaulted to 24, in units of... ; For reference duration, 24 is used, and the unit is... . The basis for using 0.08 is: the standard operating procedure for mammalian cell culture requires changing the culture medium every 48 to 72 hours when the cell density is close to the confluence, based on a full-capacity population. Calculations show that the culture medium health decreases by 8 percentage points every 24 hours, reaching approximately 76% after 72 hours, which precisely corresponds to the operational threshold requiring medium replacement every three days. This coefficient can be corrected online using actual measured pH indicator readings of the culture medium, with the correction rule being... , Take 0.3. When When not configured, this formula degenerates into a fixed decay independent of population size, i.e. When the health of the culture medium drops to a low threshold, it triggers accelerated apoptosis; when the population of a certain cell line drops to zero, it triggers an irreversible extinction determination for that cell line, after which the cell line will no longer participate in the population update of subsequent simulation steps.

[0078] The key to this invention lies in the fact that the artificial intelligence-driven simulation engine explicitly injects the following six types of constraints into the prompt context of each simulation step, thereby driving the large language model to simultaneously handle three intercoupled biological problems: multicellular competition, environmental sensitivity, and pharmacological effects.

[0079] Constraint 1: Competition for space and nutrients among multiple cell lines, with the relative competitive advantage of each cell line dynamically adjusted based on population density and culture medium health status.

[0080] Constraint 2: The nonlinear inhibitory effect of culture medium health on cell proliferation rate, and the threshold effect of low culture medium health triggering cell death;

[0081] Constraint 3: Each cell line has a specific tolerance window for temperature and pH, and cell viability is automatically reduced when it deviates from the optimal range;

[0082] Constraint 4: Concentration-dependent pharmacological effects of added reagents, including growth inhibition, apoptosis induction, and specific drug resistance in various cell lines;

[0083] Constraint 5: Random biological variation, simulating the population fluctuations inherent in in vitro culture;

[0084] Constraint 6: When the population size of a certain cell line drops to zero, an irreversible extinction determination is triggered.

[0085] The output of the large language model is rigorously validated structured data, including the number of new populations for each cell line, the updated health of the culture medium, and event logs.

[0086] like Figure 1 As shown in the center, the visualization of the petri dishes is achieved using a dynamic visual rendering engine. The color of the culture medium is interpolated using a radiometric gradient based on the health status of the medium. ,in The values ​​for the pink three-channel display of fresh culture medium. The values ​​for the three yellow channels represent depleted culture medium, all integers from 0 to 255. When... hour, It is light orange, consistent with the actual observation that the culture medium gradually changes from pink to light yellow.

[0087] When the total population size exceeds the visualization limit At that time, the system automatically calculates the scaling factor and renders the cell distribution proportionally: the number of drawing points is taken as... The radius is drawn according to the principle of area conservation. .like Figure 1 As shown in the upper right corner, the system also provides a multi-cell line graph of population over time based on the chart library, supporting interactive tooltips.

[0088] The multi-omics analysis module of this invention adopts a two-stage processing pipeline, with the dividing point set between the natural language semantic layer and the single-cell statistical layer.

[0089] Phase 1: Pre-assessment: The pre-assessment phase receives user input regarding cell line, passage number, seeding density, culture medium type, oxygen tension, and the name, concentration, dosage unit, addition time point, and duration of action of one or more reagents. The system invokes a large language model to perform fuzzy matching and biological analysis on the input reagents, identifies affected target genes, infers the gene knockdown percentage for each reagent, and assesses the risk level of reagent combinations. The rules for inferring the knockdown percentage for each type of reagent are shown in Table 3.

[0090] Table 3. Rules for inferring gene regulation methods and knockout percentages

[0091]

[0092] The system incorporates a target mapping database of no fewer than 18 commonly used compounds as a reference library for fuzzy matching, including irapasine mapped to solute carrier family 7 member 11, RSL3 mapped to glutathione peroxidase 4, thionine sulfoxide imine mapped to glutamate cysteine ​​ligase catalytic subunit, ferritin 1 mapped to lipid peroxidation protection, doxorubicin mapped to DNA insertion, cisplatin mapped to DNA crosslinking, U0126 mapped to mitogen-activated protein kinase kinase, womanpemium mapped to phosphatidylinositol 3 kinase, Bay11-7082 mapped to nuclear factor κB, JQ1 mapped to bromine-containing domain protein 4 and MYC, and Nutlin-3a mapped to mouse bimicrosome 2 and p53 axis, etc.

[0093] The pre-assessment report includes basic cell line biology annotations, reagent analysis results, literature references, and a risk assessment, without deducting any credit points. It also generates structured parameter objects for use in the second phase.

[0094] The entire information pathway between the pre-evaluation module and the basic single-cell transcriptome model consists solely of the following bidirectional deterministic numerical mapping. For knockdown-type regulation, the knockdown percentage is denoted as... Then the gene The expression folding change is as follows: For overexpression-based regulation, the overexpression fold is denoted as... ,but: ,

[0095] This represents a dimensionless relative multiple change, with negative values ​​indicating a downward adjustment and positive values ​​indicating an upward adjustment. The direction is unique across all platforms. Specific values ​​are shown in Table 4.

[0096] Table 4. Examples of the mapping from gene regulation to changes in expression folding

[0097]

[0098] The technical significance of setting up this narrow interface is as follows: First, the models on the semantic parsing side and the statistical inference side can be replaced independently; second, failures on both sides of the interface can be degraded independently; third, the semantic parsing side does not occupy graphics processor resources, thus eliminating the need for quotas.

[0099] Phase Two: Simulation and Interpretation: Phase Two uses structured parameter objects as input and requests the single-cell transcriptome basic model inference service deployed on a GPU cloud computing infrastructure via a secure hypertext transfer protocol. The server endpoint is injected at runtime via environment variables or database configuration, with a default timeout of 90 seconds to tolerate model cold start latency, and a maximum of two retries. The request body includes the cell line name and a list of perturbation parameters, supporting combined perturbations of multiple genes. The response body contains no fewer than 20 differentially expressed genes, each including the gene name, base-2 logarithmic folding change, and corrected p-value, as well as activity scores for 8 core pathways, with values ​​ranging from [value range missing]. to The eight core pathways are, in order: oxidative stress response, DNA damage repair, apoptosis, cell cycle, mitogen-activated protein kinases and extracellular signal-regulated kinases, phosphatidylinositol 3-kinases and protein kinase B, inflammatory signaling, and stemness. Multiple perturbation sets support sequential batch execution with a 500-millisecond delay between batches to avoid service overload. The system provides dedicated health check functions to perform pre-checks on server endpoints.

[0100] When the inference service is unavailable or times out, the system automatically switches to the synthetic inference mode based on the large language model, generates a result with the same structure as the output of the inference service, and marks the data source as false in the response to transparently explain the data source.

[0101] Subsequently, the system injects the differentially expressed gene list, pathway activity score, and original experimental context into the large language model prompts to generate complete seven-dimensional omics dashboard data, and performs 16 structural validations on the output; if the validation fails, it returns partial results and marks the simulation status as partially completed rather than completely failed.

[0102] The number of cells in the control group was calculated using a pure exponential growth model: ,in This represents the initial number of cells, expressed in units of cells. The training period is [duration], and the unit is [unit]. ; The doubling time is expressed in units of 1. ; The growth rate constant is expressed in units of 1000 mcg. Cell lines are classified into three types based on doubling time: rapid type... , Including human embryonic kidney 293T cells, HeLa cells, A549 cells, Jurkat cells, NIH / 3T3 cells, and Chinese hamster ovary cells; intermediate-rate type to , to Including MCF-7 cells, THP-1 cells, and U87-MG cells; slow-growing type to , to It includes human umbilical vein endothelial cells, mesenchymal stem cells, and induced pluripotent stem cells.

[0103] The growth rate constant of the treatment group was corrected for cell cycle pathway scores: ,in This is a dimensionless score for cell cycle pathway activity. (Coefficient) The value is based on the fact that when the cell cycle pathway is completely inhibited, the proliferation rate drops to half of the baseline rather than completely arresting, which is consistent with the fact that cell cycle checkpoint arrest does not immediately terminate the division of cells that have entered S phase. Therefore... The value range is self-consistent and there is no negative growth rate.

[0104] This invention constructs a lethal driving force by combining apoptosis score and oxidative stress response score:

[0105] ,

[0106] ,

[0107] in As a score for apoptosis pathway activity, For the activity score of oxidative stress response pathways, both values ​​are within the range of [missing information]. to Dimensionless Negative values ​​indicate damage to the antioxidant system; The apoptosis lethality weight is set to 1.2, which is dimensionless. The oxidation lethality weight is set to 4.0, which is dimensionless. The oxidative stress tolerance threshold is set to 0.3, which is dimensionless. For reference duration, 72 is used, and the unit is... ; As the baseline mortality rate, take The unit is It should be noted that a negative exponential form is used here instead of a non-linear decreasing form because the linear form will give a negative vitality percentage when the score approaches the boundary of the value range, while the value range of the negative exponential form is naturally smaller. It is self-consistent without the need for additional constraints.

[0108] The basis for setting it to 0.3 is that when the absolute value of the oxidative stress score is less than 0.3, the corresponding mitochondrial health score is higher than 74, which is at the warning level rather than the high stress level. At this time, the cells can compensate through the endogenous antioxidant system without net death, so it is set as the allowable threshold. and The ratio reflects that the slope of oxidative stress-induced death is approximately 3.3 times that of apoptosis-related death, which is consistent with the rapid and irreversible process of oxidative death forms such as ferroptosis after initiation. The above weights can be corrected online according to the measured propidium iodide positivity rate, with the correction rule being... , Take 0.25. When the pathway activity score is missing, Degenerate into containing only baseline terms At this point, the vitality percentage was 97.9% after 72 hours, which is close to the control behavior without treatment.

[0109] The number of cells in the treatment group was: The reactive oxygen species index is calculated in two branches based on the sign of the oxidative stress response score. When That is, when the antioxidant system is damaged: ,when That is, when antioxidant capacity is enhanced: ,in Let be the cumulative rate constant of reactive oxygen species, taken as 0.02, in units of . ; This represents the upper limit of the increase in reactive oxygen species, and is dimensionless. (Two branches are mentioned.) The value is 1 at all points, and the function is continuous. The time points are set to 0, 6, 24, 48, and 72 hours, fully covering the typical time windows of cell biology experiments.

[0110] like Figures 4 to 10 as well as Figures 14 to 19 As shown, the dashboard contains seven tabs that are updated synchronously.

[0111] The transcriptome tag generates a volcano plot with no fewer than 100 gene points. The horizontal axis represents the logarithmic folding change with base 2, and the vertical axis represents the negative value of the logarithmic p-correction value with base 10. Significantly upregulated genes are marked in red, significantly downregulated genes are marked in green, and genes with no significant difference are marked in gray. The significance criterion is that the absolute value of the logarithmic folding change is greater than 1 and the corrected p-value is less than 0.05. The list of the top 5 upregulated and downregulated genes is also displayed.

[0112] The proteome tag estimates protein levels from changes in messenger RNA folding using the following formula:

[0113] ,in The efficiency coefficient for messenger RNA to protein transfer is taken as 0.7, dimensionless. Here, it is taken as... The operator is necessary: ​​if omitted, when That is, less than The estimated value will be negative, while protein abundance cannot be negative. The basis for using 0.7 is that multiple quantitative studies of paired proteomics and transcriptomics have reported rank correlation coefficients of approximately 0.4 to 0.6. Using 0.7 reflects that the buffering effect of translation efficiency and protein degradation makes the variation in protein abundance smaller than the variation in messenger RNA. In the absence of paired protein data, Degradation is assigned a value of 1, indicating proportional transmission. Three confidence levels are assigned: high, medium, and low. Enzymes and structural proteins are assigned high confidence, signal intermediates are assigned medium confidence, and transcription factors are assigned low confidence.

[0114] The pathway tab displays the activation status of eight core pathways in radar chart format, with coordinate values ​​normalized from pathway scores using the following formula: The oxidative stress tab displays a time-series line graph of the reactive oxygen species index, a gradation of oxidative stress severity, a mitochondrial health score, and the status of key antioxidant genes. The mitochondrial health score is calculated using the following formula:

[0115] ,in The reactive oxygen species damage coefficient is taken as 22, which is dimensionless. The tolerance for physiological fluctuations in reactive oxygen species (ROS) is set to 0.6, which is dimensionless. The reason for introducing the ROS term here is that the score obtained solely through a linear mapping of pathway scores is limited to a range of 60 to 100, failing to characterize situations where mitochondrial health scores drop below 40 under severe oxidative damage such as ferroptosis. Therefore, the ROS term is introduced... After setting the threshold-type injury term as the independent variable, the scoring range covers 0 to 100, consistent with actual observations. Oxidative stress severity is categorized as follows: It is divided into four levels: normal, warning, high stress, and critical, with the dividing points being 85, 74, and 50 points respectively.

[0116] Epigenetic tags are inferred from messenger RNA folding changes to determine promoter accessibility and CpG methylation probability using the following formula: ,in For promoter accessibility, The CpG methylation probability is a dimensionless quantity ranging from 0 to 1. The accessibility conversion width is set to 2.0, which is dimensionless. The basis for choosing 2.0 is: the logarithmic folding change with base 2 reaches... When the expression level changes fourfold, the accessibility reaches 0.73 or 0.27, which corresponds precisely to the commonly used reading boundary between open and closed regions in chromatin accessibility sequencing. hour, If the value falls between 0.44 and 0.56, it is considered partially reachable. The advantage of this formula compared to the three-level discrete interpretation is that it outputs a continuous value instead of a step value, avoiding discontinuous jumps at the criterion boundaries.

[0117] The custom parameters tab allows users to specify an unlimited number of custom biological indicators, such as p53 phosphorylation, telomere length, adenosine triphosphate production, mitochondrial membrane potential, autophagy markers, etc. The system generates quantitative prediction time series and artificial intelligence biological interpretations for each indicator across 5 time points.

[0118] The comprehensive abstract tab generates a structured comprehensive report from a large language model, which includes the mechanism of action of the experimental design, a list of 3 to 5 major molecular effects, a description of the predicted cell phenotype, a three-tiered risk assessment (low, medium, and high), and 3 to 5 recommended follow-up validation experiments, each including specific experimental methods, detection targets, and time points.

[0119] The platform incorporates a comprehensive pathway annotation knowledge graph for no fewer than 82 human genes. Each gene's annotation data includes its pathway classification from the Kyoto Genome and Genome Encyclopedia, Reactome functional classification, 14 functional categories, radar axis assignment, expression direction under stress conditions, and a human-readable functional description. The knowledge graph covers five functional gene sets: 13 antioxidant genes, including superoxide dismutase 2, superoxide dismutase 1, catalase, heme oxygenase 1, glutamate-cysteine ​​ligase catalytic subunit, glutamate-cysteine ​​ligase regulatory subunit, glutathione peroxidase 4, glutathione peroxidase 1, thioredoxin, nuclear factor E2-related factor 2, acyl-CoA synthase long chain family member 4, prostaglandin intraperoxide synthase 2, and transferrin receptor; 9 DNA repair genes; 10 apoptosis pathway genes; 8 cell cycle genes; and 8 stemness genes. The knowledge graph is not used as a real-time parameter of the simulation engine, but rather as an analytical basis for the interpretation stage of the large language model, used for gene function clustering, pathway scoring directionality judgment, and report generation.

[0120] For passage-sensitive cell lines, including mesenchymal stem cells, induced pluripotent stem cells, hematopoietic stem cells, BJ fibroblasts, and human umbilical vein endothelial cells, this invention implements passage-dependent aging correction using a single logistic decrement function. The passage number is denoted as... The aging correction factor is then...

[0121] ,in The half-life algebra is taken as 19.6, and the unit is generations; The generation attenuation width is set to 4.0, with the unit being generations; Dimensionless. The values ​​of the two parameters above are based on the following: the boundaries of the five generations fall exactly at generations 5, 10, 15, and 20. The threshold values ​​were 0.97, 0.90, 0.75, and 0.45, respectively. The values ​​of the correction factors and biological performance at each passage stage are shown in Table 5.

[0122] Table 5. Classification of passage stages and passage aging correction factors

[0123]

[0124] The correction factor acts on the structured parameter object according to the following five rules. Rule 1: The proliferation rate is corrected to... Rule 2, the activity score of the stemness and differentiation pathway is revised to... ,in The aging sensitivity index is taken as 0.85, which is dimensionless. This indicates that the sensitivity of the pathway response to passage is lower than that of proliferation capacity. The mechanism is that pathway activation is an acute event, while proliferation capacity is a cumulative capacity; the former is relatively less affected by replicative senescence. Rule 3, the expression folding changes of the telomerase reverse transcriptase gene are revised to... Rule 4, the expression folding changes of the cyclin-dependent kinase inhibitor 2A gene are corrected to... Rule 5, the changes in the expression folding of aging-related secretory phenotypic factors are revised to... .

[0125] For non-passaged sensitive cell lines, such as HeLa cells, human embryonic kidney 293T cells, A549 cells, and other established cancer cell lines, take... The algebraic number does not significantly correct the simulation results.

[0126] Researchers hope to simulate the competitive dynamics of ternary co-culture of tumor cells, immune cells, and normal cells in the tumor microenvironment and evaluate the efficacy of adding paclitaxel.

[0127] The input configuration was as follows: Cell line 1 was HeLa cells, color-coded red, with an initial quantity of 500; Cell line 2 was THP-1 monocytes, color-coded yellow, with an initial quantity of 200; Cell line 3 was human embryonic kidney 293T cells, color-coded blue, with an initial quantity of 300; The culture medium was Roswell Park Memorial Institute 1640 medium, with 10% serum, a temperature of 37 degrees Celsius, a pH of 7.4, and normal glucose levels; The reagent was paclitaxel, with a concentration of 100 nanomoles per liter.

[0128] like Figure 1 As shown, after the user completes the above input in the experimental configuration area on the left, the simulation is started. The microscopic view of the petri dish in the middle and the population-time line graph in the upper right are then updated synchronously in 24-hour increments. Event descriptions are gradually added to the simulation log area below.

[0129] At step 1, 24 hours into the culture, the system construction indicated the current population size and biological characteristics of the three cell lines, with the culture medium health at 98%. The new population size output by the large language model was: 620 HeLa cells, 230 THP-1 cells, and 370 human embryonic kidney 293T cells; the updated culture medium health was 94%; the event log recorded that HeLa cells exhibited the fastest proliferation, with a doubling time of approximately 24 hours, and that 100 nanomoles per liter of paclitaxel began to disrupt the assembly of the mitotic spindle. Early G2 / M phase arrest was observed in HeLa cells, while THP-1 cells grew at a moderate rate, and early lactate accumulation appeared in the culture medium.

[0130] According to the backtracking verification method in Section 6: the total population size increased from 1000 to 1250, taking... , The expected decline in the health of the culture medium is then... The percentage point is on the same order of magnitude as the 4 percentage point output by the large language model, so it does not trigger a rollback.

[0131] At step 3, 72 hours into the culture, the new population output by the large language model was: 450 HeLa cells, 480 THP-1 cells, and 390 human embryonic kidney 293T cells; the health of the updated culture medium was 78%; the event log showed that G2 / M phase arrest of paclitaxel had become the dominant effect, the net HeLa cell count decreased due to mitotic catastrophe, THP-1 cells showed M1 activation due to phagocytosis of dying HeLa cells, human embryonic kidney 293T cells remained tolerant at a concentration of 100 nanomoles per liter, and the culture of HeLa cell lysis products decreased.

[0132] Visual output of petri dish as follows Figure 1 As shown in the middle: According to the color interpolation formula in Section 6, a culture medium health of 78% corresponds to the three-channel values. The culture dish color gradually changed from pink to light yellow; the proportion of HeLa red cells decreased from 58% to 34%, while the proportion of THP-1 yellow cells increased, reflecting the passive expansion of immune cells. Figure 1 As shown in the population-time line graph in the upper right corner, the three curves intersect at 72 hours, with the THP-1 cell curve surpassing the HeLa cell curve, visually demonstrating the reversal of competitive advantage.

[0133] This embodiment illustrates that, since the health of the culture medium is used as a shared state quantity across cell lines to provide feedback to the next step, the large language model can spontaneously generate negative feedback that accelerates the proliferation of the dominant cell line and depletes the culture medium, thereby inhibiting its own proliferation, without needing to pre-define the competition coefficient matrix. This achieves the interspecies competition dynamics that existing Lotka-Volterra-like models require wet experimental calibration to obtain.

[0134] Researchers designed and simulated the self-excited feedback loop of leukemia inhibitory factor-signal transducer and transcription activator factor 3-pluripotency factor required for the maintenance of induced pluripotent stem cells.

[0135] like Figure 2As shown, after selecting a node type in the left-hand node type toolbar, the user drags it to the central canvas. The system automatically attaches it to one of the four biological spatial partitions—extracellular, membrane, cytoplasm, and nucleus—based on its positional attributes. The loaded induced pluripotent stem cell preset network contains 6 nodes and 5 edges: node n1 is a leukemia inhibitory factor, type ligand, located in the extracellular partition; node n2 is a leukemia inhibitory factor receptor, type receptor, located in the membrane partition; node n3 is a signal transduction and transcription activator factor 3, type kinase, located in the cytoplasmic partition; nodes n4, n5, and n6 are octamer-binding transcription factor 4, sex-determining region Y-box protein 2, and Nanog homologous box protein, respectively, all of which are transcription factors, located in the nucleus partition. The five edges are as follows: Edge e1 points from leukemia inhibitory factor to leukemia inhibitory factor receptor, type: binding; Edge e2 points from leukemia inhibitory factor receptor to signal transduction and transcription activator 3, type: activation; Edge e3 points from signal transduction and transcription activator 3 to octamer-binding transcription factor 4, type: activation; Edge e4 points from octamer-binding transcription factor 4 to sex-determining region Y-box protein 2, type: activation; Edge e5 points from octamer-binding transcription factor 4 to Nanog homeobox protein, type: activation.

[0136] Users can also Figure 2 The AI ​​pathway generation dialog panel in the lower right corner describes the target pathway in natural language. For example, inputting the design of the epidermal growth factor receptor downstream signal cascade, the AI ​​pathway generation interface generates a directed graph of molecular circuits containing a complete set of nodes and edges. Node coordinates are automatically assigned according to biological spatial partitioning, and known cross-pathway crosstalk is automatically connected. AI can also automatically infer phenotypic rules based on the topology of the current molecular circuit graph: toxin and pathogen nodes are inferred as apoptosis or infection rules, growth factor activation nodes are inferred as proliferation rules, differentiation factor concentration exceeding the threshold is inferred as differentiation rules, and high concentration of pluripotency factors is inferred as reprogramming rules.

[0137] Under complete culture medium conditions , ,Pick The initial normalized concentration of leukemia inhibitory factor was 1.0, the initial normalized concentration of leukemia inhibitory factor receptor was 0.2, and the baseline reaction rate constant was taken as... The reference decay rate constant is taken as The first step of the Euler integral process is as follows.

[0138] The turnover rate of leukemia inhibitory factor receptors is:

[0139] ,

[0140] ,

[0141] ,

[0142] The update amount of signal transduction and transcription activator 3 is:

[0143] ,

[0144] ,

[0145] After about 100 simulation steps, the activation level of signal transduction and transcription activator 3 reached a steady state of about 0.65, driving the normalized concentration of octamer-binding transcription factor 4, sex-determining region Y-box protein 2, and Nanog homologous box protein to rise to above 0.85, forming a stable pluripotency self-excitation feedback.

[0146] Based on this, two phenotypic rules are set: Rule 1 is that reprogramming is triggered when the normalized concentration of Nanog homeobox protein is greater than 0.75, i.e., pluripotency is maintained; Rule 2 is that differentiation is triggered when the normalized concentration of octamer binding transcription factor 4 is less than 0.30, i.e., low octamer binding transcription factor 4 triggers spontaneous differentiation.

[0147] like Figure 2 As shown in the upper right cell agent canvas, induced pluripotent stem cell agents proliferate rapidly with a cell line-specific proliferation coefficient of 1.2 under complete culture conditions, and the number of purple circular agents on the canvas steadily increases. Agents that meet the condition of a Nanog homeobox protein normalized concentration greater than 0.75 trigger reprogramming actions, visually marked as green fluorescent rings. Figure 2 As shown in the physical rules bar below, the two currently active phenotypic rules are displayed in real time as labels.

[0148] This example illustrates that because the chemical state is attached to each cell agent rather than maintained as a population average concentration field, the daughter cell agents produced by division inherit the complete chemical state of the mother cell agent. This leads to the differentiation of Nanog homologous box protein concentrations from different lineages within the same culture system. Consequently, both reprogrammed and spontaneously differentiated phenotypes can be observed simultaneously on the same canvas. This clonal heterogeneity cannot be produced by a simple ordinary differential equation concentration field model.

[0149] Researchers plan to use small interfering RNA to knock down 11 genes of 7 members of the solute carrier family in HeLa cells at a concentration of 50 nanomoles per liter for 72 hours to predict multidimensional omics effects related to ferroptosis.

[0150] Phase 1: Pre-assessment. For example... Figure 3As shown on the left, the user selects HeLa cells in the cell configuration area, sets the passage number to P5, the seeding density to 10,000 cells / mL, the culture medium to complete Dalberg modified Eagle medium, the oxygen tension to normoxic 21%, and the carbon dioxide concentration to 5%. In the custom monitoring parameter area, the user adds a multi-dimensional omics effect tag to predict ferroptosis-related factors. In the reagent workshop, the user enters the reagent description as Small Interfering RNA Knockdown Solute Vector Family 7 Member 11, dosage 50, unit nanomoles per liter, addition time 0 hours, and duration of action 72 hours.

[0151] The large language model's analysis output is as follows: Cell line annotations indicate that HeLa cells exhibit p53 degradation of the human papillomavirus type 18 E6 / E7 oncoprotein, are resistant to classical apoptosis, and highly sensitive to ferroptosis; reagent analysis results indicate that the target gene is member 11 of the solute vector family 7, with a knockdown percentage of 85%, and the mechanism of action is RNA-induced silencing complex-mediated messenger RNA degradation. The expected main effect is reduced cysteine ​​uptake leading to inhibited glutathione synthesis, which in turn leads to inactivation of glutathione peroxidase 4, ultimately resulting in lipid peroxidation and ferroptosis; the predicted risk level is moderate. Figure 3 As shown in the lower right corner of the pre-assessment report area, the risk level label indicates medium risk.

[0152] According to the bidirectional deterministic numerical mapping in Section 7, 85% knockdown corresponds to expression folding changes. .

[0153] Phase Two: Inference using a single-cell transcriptome basic model and interpretation using a large language model. The inference service returned activity scores for eight pathways: oxidative stress response. Apoptosis Cell cycle DNA damage repair Mitogen-activated protein kinases and extracellular signal-regulated kinases Phosphatidylinositol 3-kinase and protein kinase B Inflammatory signals Dry .

[0154] Calculate the time series using the mathematical model in Section 8. HeLa cell doubling time. Therefore The corrected growth rate constant for the treatment group is: The two lethal driving forces are respectively and Therefore The calculation results are shown in Table 6.

[0155] Table 6. Results of four-dimensional time series calculations in Example 3

[0156]

[0157] like Figure 3 As shown in the four-dimensional time monitor in the upper right corner, the interface simultaneously plots three curves: the control group curve, the treatment group curve, and the vitality percentage curve. The top statistics label shows that the vitality percentage at 72 hours is 18%, which is consistent with the 17.9% calculated in Table 6.

[0158] The output of the 7D dashboard is as follows.

[0159] like Figure 5 As shown in the volcano plot of the transcriptome tag page, heme oxygenase 1, glutathione-specific γ-glutamyl cyclotransferase 1 (a cysteine-containing acidic protease inhibitor), DNA damage-induced transcript 3, activation transcription factor 4, and asparagine synthase are the top 5 upregulated genes, marked in red; solute carrier family 7 member 11, glutathione peroxidase 4, glutamate-cysteine ​​ligase catalytic subunit, solute carrier family 3 member 2, and the proliferation marker Ki-67 are the top 5 downregulated genes, marked in green. The base-2 logarithmic folding change of heme oxygenase 1 is shown in the figure. Acyl-CoA synthase long chain family member 4 is Transferrin receptor is Glutathione peroxidase 4 is The catalytic subunit of glutamate-cysteine ​​ligase is .

[0160] like Figure 4 As shown, the proteomics label is arranged by Estimate protein levels. The estimated value for heme oxygenase-1 is... ,and Figure 4 The horizontal axis reading corresponding to the bar for this gene treatment group was consistent with 3.95; the estimated value for the transferrin receptor was... The estimated value of the catalytic subunit of glutamate-cysteine ​​ligase is... It remains non-negative; the original estimate for glutathione peroxidase 4 was [value missing]. ,through The operator clamps the result to 0, which is reflected here. The necessity of operators. Figure 4 The confidence level list below shows that members of the solute carrier family 7, heme oxygenase 1, and glutathione peroxidase 4 are at high confidence, glutathione-specific γ-glutamyl cyclotransferase 1, a cysteine-containing acidic protease inhibitor, is at medium confidence, and activating transcription factor 4 is at low confidence.

[0161] like Figure 6 As shown, the eight-axis radar chart of the pathway tab is arranged according to... After normalization, the normalized coordinates of the oxidative stress response axis are plotted as follows: The radar polygons show significant inward contraction, with an apoptosis axis of 0.675 and an inflammation signal axis of 0.60, while the radar polygons bulge outward, exhibiting a distinctly asymmetrical shape.

[0162] like Figure 7 As shown, the time series line graph of the reactive oxygen species (ROS) index under the oxidative stress label starts at 1.0 and rises to 2.42 after 72 hours. The upper limit of the ROS increase is... Mitochondrial health score: ,and Figure 7 The mitochondrial health score was consistent with 40; the severity was indicated as critical. The antioxidant gene status list showed superoxide dismutase 2 as... Heme oxygenase 1 is The catalytic subunit of glutamate-cysteine ​​ligase is Catalase is .

[0163] like Figure 8 As shown, the epigenetic label page is arranged by Calculate the CpG methylation probability. The base-2 logarithmic folding change of solute carrier family 7 member 11 is: Therefore ,and Figure 8 The methylation probability readings of this gene were 88% consistent, and the promoter accessibility was determined to be closed; the logarithmic folding change of glutathione peroxidase 4 was... , The logarithmic folding change of prostaglandin intraperoxidase 2 is as follows: , The logarithmic folding change of heme oxygenase 1 is as follows: , The promoter accessibility is determined to be open. Figure 8 The chromatin environment notes below state that knockdown of member 11 of the solute carrier family 7 triggers a large-scale integration stress response, and the chromatin pattern shifts to favor the activation of transcription factor 4-driven amino acid transporters and endoplasmic reticulum stress marker transcription, while antioxidant defense sites exhibit progressive heterochromatization due to metabolic depletion.

[0164] like Figure 9 As shown, the custom parameters tab displays the time series of the ferroptosis potential index specified by the user. The baseline value is 1.00 arbitrary units, and the final value after 72 hours is 12.50 arbitrary units, with the curve showing an exponential increase. The AI ​​biology explanation below states that: member 11 of the solute carrier family 7 knockdown blocks the rate-limiting step of cystine uptake in glutathione synthesis. When glutathione is lacking, glutathione peroxidase 4 cannot neutralize lipid peroxides, leading to an exponential increase in the ferroptosis induction process.

[0165] like Figure 10 As shown in the summary tag output: The mechanism of action is that member 11 of the solute carrier family 7 inhibits and blocks the cysteine / glutamate antitransporter, leading to acute intracellular cysteine ​​starvation and subsequent glutathione depletion; molecular effects include severe depletion of the antioxidant glutathione pool, inactivation of glutathione peroxidase 4 leading to the accumulation of lethal lipid hydroperoxides, induction of an integrated stress response via activation of transcription factor 4 and the cysteine-containing acidic protease inhibitor glutathione-specific γ-glutamyl cyclotransferase 1, and heme oxygenase 1 and transferrin receptor... The study indicated an increase in the unstable iron pool; the predicted phenotype was rapid ferroptosis characterized by lipid peroxidation, mitochondrial shrinkage, and loss of plasma membrane integrity, with an 82% decrease in cell viability at 72 hours; the risk assessment was critical; recommended follow-up experiments included co-treatment with ferroin 1 or liprisstatin 1 to confirm ferroptosis rescue, measurement of the intracellular reduced glutathione to oxidized glutathione ratio at 6 and 12 hours, quantification of lipid peroxidation by flow cytometry using C11-BODIPY staining, and assessment of the unstable iron pool using the calcein methyl ester quenching method.

[0166] This embodiment illustrates that by connecting the semantic parsing in the pre-evaluation stage and the statistical prediction in the inference stage through a narrow interface containing only expression folding variations, the 85% knockdown given by the semantic parsing is losslessly converted. The expression folds and changes, and all subsequent multidimensional omics outputs are derived from this single value through statistical model inference; because the interface is narrow, when the inference service is unavailable, the large language model can generate synthetic inference results according to the same structure without changing any downstream computational logic.

[0167] Researchers treated 18th-passage mesenchymal stem cells with Wnt3a to assess the response of high-passage mesenchymal stem cells to osteogenic induction.

[0168] The system execution process is as follows: First, identify mesenchymal stem cells as passage-sensitive cell lines. Second, calculate the passage aging correction factor according to Section 9: If the value falls within the range of 0.45 to 0.75, it is determined to be in the severe aging stage, which is consistent with the generation range of P16 to P20.

[0169] The third step involves injecting aging correction during the pre-assessment phase: superimposed changes in the expression and folding of telomerase reverse transcriptase genes. Significant downregulation was observed; the expression of cyclin-dependent kinase inhibitor 2A gene was altered by folding changes. The expression of aging-related secretory phenotypic factors such as interleukin-6, interleukin-8, and matrix metalloproteinase-3 was upregulated; the expression of these factors was superimposed with changes in folding. Activation of aging-related secretory phenotypes; mild downregulation of superoxide dismutase 2.

[0170] Step 4, Preliminary Assessment Report Supplementary Warning: Mesenchymal stem cells from passage 18 are in a severely aged stage, with a decreased proliferative capacity compared to passage 4. The magnitude of Wnt3a-induced β-catenin activation was as follows: Correction, a decrease relative to the 4th generation. The activation efficiency of Runt-related transcription factor 2 osteogenic transcription factor is limited; it is recommended to consider a low-passage mesenchymal stem cell control experiment at passages 4 to 6.

[0171] In the fifth step, the scores of the β-catenin and Runt-related transcription factor 2 pathway output by the single-cell transcriptome basic model were automatically corrected according to the above coefficients. The osteogenic pathway response score of the 18th passage mesenchymal stem cells was about 34% lower than that of the 4th passage reference.

[0172] The sixth step involves a seven-dimensional dashboard that compares and displays the differential responses of 18th and 4th passage mesenchymal stem cells to the same Wnt3a stimulation, helping researchers assess the feasibility of the experiment.

[0173] This embodiment illustrates that a single logistic function can simultaneously provide two decay curves: one for proliferation capacity and one for pathway response, with different decay magnitudes. This eliminates the need to build a segmented model for each generation stage or to model proliferation and pathways separately.

[0174] Researchers performed genome-wide knockout of the regularly clustered short palindromic repeats of the TP53 gene in MCF-7 cells and customized monitoring of three indicators: caspase 3 activation level, p21 protein level, and cell cycle G1 / S transition rate.

[0175] like Figure 11 As shown, the user selects MCF-7 cells in the cell configuration area. The interface indicates that this cell line has a relatively low response to estrogen and low invasiveness. The passage number is set to P5, the seeding density is 10,000 cells / mL, the culture medium is complete Dalberg modified Eagle medium, the oxygen tension is normoxic 21%, and the carbon dioxide concentration is 5%. The user adds the above three indicators in the custom monitoring parameter area. The upper right corner of the interface shows that three items have been added. In the reagent workshop, the reagent description is entered as TP53 regular clustered short palindromic repeat sequence whole genome knockout, dose 10, unit micromoles per liter, addition time 0 hours, and duration of action 72 hours.

[0176] According to the mapping rules in Section 7, knocking out genes with regularly clustered, short palindromic repeat sequences corresponds to a 100% knockdown. MCF-7 cells are an established cancer cell line and are non-passage-sensitive. The generation number P5 does not introduce aging correction.

[0177] The pathway score returned by the single-cell transcriptome basic model was: apoptosis Cell cycle Oxidative stress response Dry DNA damage repair .

[0178] Calculate the doubling time of MCF-7 cells using the mathematical model in Section 8. Therefore The corrected growth rate constant is: The levels were higher than in the control group, reflecting accelerated proliferation after the G1 / S barrier was removed due to p53 deficiency. The two lethal drivers were... and Therefore The vitality percentage is .

[0179] like Figure 13 As shown, the top statistical label of the four-dimensional time monitor shows that the viability percentage at 72 hours is 98%, consistent with the calculated 97.9%, and shows that the increase in cell number in the treatment group relative to the control group is positive; the curve of the treatment group has been consistently higher than that of the control group since 24 hours, and the growth rate constant has changed from... Rise to The calculation results are in the same direction. Figure 13 The pre-assessment report area below shows a low risk level and records that MCF-7 cells are estrogen receptor-positive and progesterone receptor-positive breast cancer cell lines. Unlike most cancer models, they have wild-type p53 and are highly responsive to cell cycle arrest induced by estrogen and DNA damage.

[0180] The output of the 7D dashboard is as follows.

[0181] like Figure 14 As shown in the volcano plot of the transcriptome tag page, cyclin D1, proliferation marker Ki-67, proliferating cell nuclear antigen, cyclin-dependent kinase 4, and MYC are the top 5 upregulated genes; TP53, cyclin-dependent kinase inhibitor 1A, mouse double microsomes 2, BCL2-associated X protein, and BCL2-binding component 3 are the top 5 downregulated genes.

[0182] like Figure 15 As shown in the proteomics label, the bars for the treatment groups of p53, p21, mouse double microsomes 2, and BAX proteins were significantly shorter than those for the control group, while the bars for the treatment group of cyclin D1 were about twice as long as those for the control group. The confidence level list showed that p21 and BAX were at high confidence, while p53, cyclin D1, and mouse double microsomes 2 were at medium confidence.

[0183] like Figure 16 As shown, in the eight-axis radar chart of the pathway tab, the normalized coordinates of the cell cycle axis are: Plugs outward; the apoptosis axis is Significantly contracted; stem axis was 0.65; the radar polygon exhibited a morphology of strong proliferation and weak apoptosis, consistent with... Figure 6 The strong oxidative stress and weak proliferation morphology in Example 3 form a stark contrast.

[0184] like Figure 17 As shown, in the oxidative stress tab, the upper limit of the increase in reactive oxygen species is... The reactive oxygen species index rose to [a certain level] after 72 hours. Mitochondrial health score: ,and Figure 17 The mitochondrial health score was consistent with 76; the severity level was indicated as warning. The antioxidant gene status list showed Sestrin2 as... Glutathione peroxidase 1 is Both are direct target genes of p53, and their downregulation is consistent with the direction of TP53 knockout.

[0185] like Figure 18 As shown, in the epigenetic tag page, the promoter accessibility of cyclin-dependent kinase inhibitor 1A and BCL2-associated X protein was interpreted as closed, with CpG methylation probabilities of 78% and 65%, respectively; the promoter accessibility of cyclin D1 and the proliferation marker Ki-67 was interpreted as open. The chromatin environment annotation below states that TP53 deficiency leads to reduced recruitment of histone acetyltransferases to the p21 and BAX promoters, resulting in a denser chromatin at these sites and increased accessibility to genes related to cell cycle progression.

[0186] like Figure 12 As shown, the Custom Parameters tab displays the time series of three user-defined indicators. Caspase 3 activation decreased from 1.00 random units at baseline to 0.35 random units at 72 hours, showing a downward trend. This is explained by the fact that p53 deficiency reduces the transcriptional readiness of the endogenous apoptosis pathway, leading to a decrease in both basal and induced caspase 3 activation levels. p21 protein levels decreased from a relative abundance of 1.00 at baseline to 0.12 relative abundance at 72 hours, showing a significant downward trend. This is explained by the fact that p21, a cyclin-dependent kinase inhibitor 1A, is a direct transcriptional target of p53, and its protein level rapidly decreases after TP53 knockdown. The cell cycle G1 / S transition rate increased from 35.00% at baseline to 74.00% at 72 hours, showing a significant upward trend. This is explained by the fact that the decrease in p21 levels relieved the inhibition of cyclin-dependent kinases 4 / 6 and cyclin-dependent kinase 2, significantly accelerating the transition from G1 to S phase.

[0187] like Figure 19 As shown in the summary tag page, the mechanism of action is that TP53 knockdown in MCF-7 cells removes the cell cycle inhibitor p21 and the major transcriptional activator of pro-apoptotic factors BAX and BCL2 binding component 3; molecular effects include significant downregulation of cyclin-dependent kinase inhibitor 1A leading to loss of G1 / S checkpoint control, reduction of pro-apoptotic BCL-2 family members and increased resistance to cell death, and upregulation of cell cycle D1 and proliferation markers such as Ki-67; the predicted phenotype is a superproliferative phenotype with increased doubling rate and significant resistance to DNA damage-induced apoptosis; the risk assessment is a high risk of genomic instability over time, and cells are likely to develop resistance to standard chemotherapy drugs that induce p53-dependent cell death; recommended follow-up experiments include confirming increased S-phase entry with bromodeoxyuridine incorporation assays, assessing chemotherapy resistance with doxorubicin-treated annexin V and propidium iodide staining, and performing Western blotting detection of p21 and phosphorylated retinoblastoma proteins.

[0188] This embodiment uses the exact same computational pipeline and formula set as Embodiment 3, differing only in the sign and magnitude of the pathway activity score. Specifically, it yields a highly toxic phenotype with an activity percentage of 17.9% and a hyperproliferative phenotype with an activity percentage of 97.9%, respectively. This demonstrates that the lethal driving force negative exponential model of this invention remains self-consistent across the entire range of pathway scores, without requiring switching model forms for different perturbation types.

[0189] Those skilled in the art will understand that the specific parameter values ​​in the above embodiments can be adjusted without departing from the technical solution of the present invention. For example, simulation step size. The value can be between 0.005 and 0.05 seconds; the smaller the value, the smaller the truncation error of the first-order Euler method, but the greater the computational cost; maximum population limit. The value can be between 200 and 2000 depending on rendering performance; time-scale mapping coefficient. Values ​​can be taken from 1 to 24 hours per second, depending on the biological process of interest; the number of standardized culture environment states can be expanded from four to more, simply by providing a set for each new state. The value can be taken directly without modifying any reaction equation, which is the direct benefit of the four-parameter orthogonal modulation mechanism.

[0190] The node types of molecular circuits can be expanded beyond the eight types mentioned above, for example, by adding microRNA nodes and metabolite nodes. This can be achieved by specifying the shape visual representation and the corresponding biological spatial region for each new type. The biological spatial regions can be expanded from four to include more subdivided regions encompassing mitochondria, endoplasmic reticulum, and Golgi apparatus, while the node adsorption rules remain unchanged.

[0191] Single-cell transcriptome baseline models can be replaced with other single-cell baseline models or perturbation response prediction models, as long as they accept perturbation parameter objects of the same structure and return a list of differentially expressed genes and pathway activity scores of the same structure. This is a direct benefit of the narrow interface design. Similarly, large language model providers can switch between them at runtime through a configuration management system without redeploying code.

[0192] The present invention can also be implemented in the form of a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described platform or method.

[0193] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-driven virtual cell biology simulation platform, characterized in that, The platform includes a multi-cell co-culture simulation module, a molecular circuit simulation module, and a multi-omics analysis module that share the same culture environment state. The culture environment state is characterized by four mutually orthogonal scalar parameters: environmental response rate multiplier, environmental decay multiplier, basal stress level, and initial seeding density. The multi-cell co-culture simulation module is used to call a large language model to predict the population size and culture medium health index of each cell line according to a fixed simulation step size, perform non-negative integer verification and clamping on the population size output by the large language model, and inject the verified culture medium health index into the prompt context of the next simulation step size. The molecular circuit simulation module is used to precompile the directed graph of the molecular circuit into a reaction instruction set, modulate the reaction rate constant and decay rate constant in the reaction instruction set with the environmental reaction rate multiplier and the environmental decay multiplier, respectively, perform a first-order Euler integral on the chemical state of the cell agent according to the Michaelis enzyme kinetic equation and clamp the concentration to a closed interval of zero to one, and then trigger the proliferation, apoptosis, migration, reprogramming, differentiation or infection behavior of the cell agent according to the threshold phenotype rule, and determine the basal apoptosis probability of the cell agent with the basal stress level; The multi-omics analysis module is used to parse natural language reagent descriptions into gene expression folding change parameter objects during the pre-evaluation stage, and then use the gene expression folding change parameter objects to call the single-cell transcriptome basic model inference service and the large language model to output multi-omics dashboard data.

2. The artificial intelligence-driven virtual cell biology simulation platform according to claim 1, characterized in that, The culture environment is configured as one of four standardized states: complete culture medium, serum-reduced culture medium, starved culture medium, and toxic culture medium. Each of the four standardized states is associated with a set of values ​​for the environmental response rate multiplier, the environmental decay multiplier, the basal stress level, and the initial inoculation density. The environmental response rate multiplier is set to 1.5, 0.7, 0.3, and 0.4 in the order of complete culture medium, serum-reduced culture medium, starved culture medium, and toxic culture medium. The environmental decay multiplier is set to 0.8, 1.0, 1.5, and 2.5 in the same order. The basal stress level is set to 0, 0, 0.1, and 0.6 in the same order. The platform multiplies the environmental reaction rate multiplier by the baseline reaction rate constant of each reaction in the reaction instruction set, multiplies the environmental decay multiplier by the baseline decay rate constant of each node, multiplies the basal stress level by the simulation step size and the basal apoptosis rate constant to obtain the single-step basal apoptosis probability, and determines the initial number of cell agents by the initial seeding density, thereby applying unified modulation to three types of environmental factors: nutrient limitation, oxidative degradation and toxic stress with four scalars.

3. The artificial intelligence-driven virtual cell biology simulation platform according to claim 2, characterized in that, The molecular circuit simulation module includes a circuit compiler and an integrator; The circuit compiler is used to translate each edge of the directed graph of the molecular circuit into a reaction instruction containing a source node index, a target node index, a reaction type, and a reaction rate constant before the simulation starts, and to organize all the reaction instructions into a set of reaction instructions that can be directly addressed by the source node index, so that the reaction lookup cost within each simulation step is on the order of constant. The integrator is used to calculate the concentration increment for activation reactions in the reaction instruction set by multiplying the normalized concentration of the upstream node, the remaining capacity term of the target node, and the modulated reaction rate constant; to calculate the concentration decrement for inhibition reactions by multiplying the normalized concentration of the upstream node, the current normalized concentration of the target node, and the modulated reaction rate constant; to use the same kinetic form for ligand-receptor binding reactions as for activation reactions; to calculate the decay amount for each non-input node by multiplying the modulated decay rate constant and the current normalized concentration of that node; and to exempt four types of input nodes—ligands, chemicals, mixtures, and pathogens—from decay.

4. The artificial intelligence-driven virtual cell biology simulation platform according to claim 3, characterized in that, The cell agent maintains a unique identifier, cell line type, canvas coordinates, movement speed, orientation angle, cell radius, cell cycle stage, chemical state, age, infection status, detachment status, and time of the most recent division; The molecular circuit simulation module sequentially performs the first-order Euler integral, applies a random Brownian force proportional to the cell line specific motion coefficient, applies an elastic repulsive displacement along the line connecting the centers of the two cells according to the circular overlap amount, and reflects and constrains the canvas coordinates within the simulation area for each cell agent within each simulation step. The molecular circuit simulation module triggers mitosis when the cell agent simultaneously meets four conditions: it is in the G1 phase, the time since the last division exceeds the minimum division interval, the current total population size does not exceed the maximum population limit, and random sampling passes a probability test with the product of the basic division probability and the cell line-specific proliferation coefficient as the threshold. The resulting daughter cell agent inherits the chemical state and infection state of the mother cell agent and is given an initial velocity in the opposite direction.

5. The artificial intelligence-driven virtual cell biology simulation platform according to claim 4, characterized in that, Each threshold-type phenotypic rule specifies a monitoring node identifier, a concentration threshold within a closed interval of zero to one, a comparison condition greater than or less than, and a trigger action. The trigger action is selected from accelerated mitosis, programmed cell death, accelerated migration, reprogramming to an induced pluripotent stem cell state, differentiation into a specialized lineage, and viral infection and transmission. When the normalized concentration of trypsin nodes in the chemical state of the cell agent exceeds the desorption threshold, the cell agent is marked as the desorption state, its velocity retention coefficient switches from the adherent state value to the desorption state value, and the cell cycle is blocked. The cell agent, marked as infected by the viral infection and transmission action, transmits the infection status to the daughter cell agent during mitosis.

6. The artificial intelligence-driven virtual cell biology simulation platform according to claim 1, characterized in that, The multi-cell co-culture simulation module explicitly injects the following constraints in the context of each simulation step: the competition for space and nutrients among cell lines is dynamically adjusted with population density and the health index of the culture medium; the nonlinear inhibition of the proliferation rate by the health index of the culture medium and the apoptosis threshold effect when it is close to zero; the specific tolerance window of each cell line to temperature and pH; the concentration-dependent pharmacological effects of added reagents and cell line-specific drug resistance; and random biological variation. When the output of the large language model contains a non-numerical or negative value, the multi-cell co-culture simulation module replaces the output of the large language model with a backtracking recursive formula that decreases the culture medium health index by the product of the ratio of the current total population to the maximum carrying population, the culture medium consumption coefficient, and the simulation step size. When the population of a certain cell line drops to zero, an irreversible extinction determination is triggered for that cell line.

7. The artificial intelligence-driven virtual cell biology simulation platform according to claim 1, characterized in that, The multi-omics analysis module uses a two-way deterministic numerical mapping that divides the knockdown percentage by 100 and takes a negative value, and directly takes a positive value for the overexpression folding, to convert the gene regulation corresponding to the natural language reagent description into the expression folding change in the gene expression folding change parameter object. The pre-evaluation stage does not deduct the running quota. When the single-cell transcriptome basic model inference service is unavailable or times out, the multi-omics analysis module switches to the synthesis inference mode based on the large language model, outputs a list of differentially expressed genes and pathway activity scores with the same structure as the single-cell transcriptome basic model inference service, and marks the data source identifier in the response; The multi-omics analysis module performs structural verification on the multi-omics dashboard data output by the large language model. When structural verification fails, it returns partial results and marks the simulation status as partially completed.

8. The artificial intelligence-driven virtual cell biology simulation platform according to claim 1, characterized in that, The platform also includes a runtime AI provider configuration management system, which is used to store the large language model provider configuration, which includes interface keys, service addresses, model identifiers, temperature parameters and specific system prompts for functional modules, in a relational database after being encrypted with a hardware or software-based secure encryption storage mechanism. The large language model provider configuration is read according to a three-level priority of database encrypted configuration, environment variables and built-in default configuration, so as to switch the large language model provider at runtime without redeploying the code. The multi-cell co-culture simulation module, the molecular circuit simulation module, and the multi-omics analysis module are each bound to independent service identifiers and quota deduction rules, and their respective user interfaces provide visual indicators that indicate whether the current configuration source is database configuration or environment variable configuration.

9. An AI-driven gene perturbation cytokine simulation method, employing the AI-driven virtual cell biology simulation platform described in any one of claims 1-9, characterized in that the method comprises: Step M1: Receive the cell line type, passage number, seeding density, culture medium type, oxygen tension, carbon dioxide concentration, and the name, dosage, addition time point, and duration of action of at least one experimental reagent. Step M2, perform a pre-assessment without deducting running quota: call the large language model to perform target gene identification on the experimental reagents, infer expression folding changes by a two-way deterministic numerical mapping that takes a negative value after dividing the knockdown percentage by one hundred and a positive value directly after taking the overexpression fold, generate a structured parameter object containing cell line names and perturbation lists, and output the risk level; Step M3: Using the structured parameter object as input, call the single-cell transcriptome basic model inference service deployed on the graphics processor cloud computing infrastructure to obtain a list of differentially expressed genes and pathway activity scores; When the single-cell transcriptome basic model inference service is unavailable or times out, switch to the synthesis inference mode based on the large language model, output the results with the same structure and mark the data source identifier; Step M4: Correct the exponential growth rate constant with the cell cycle score in the pathway activity score, construct the lethal driving force with the weighted sum of the portions of the apoptosis score and oxidative stress response score in the pathway activity score that exceed their respective thresholds, calculate the viability percentage according to the negative exponent of the lethal driving force, and calculate the cell number and reactive oxygen species index at multiple time points. Step M5: Inject the differentially expressed gene list, the pathway activity score, and the parameters received in step M1 into the large language model prompt, generate multi-omics dashboard data, and perform structure verification. If the structure verification fails, return partial results and mark the simulation status as partially completed. Step M6 presents the multi-omics dashboard data in an interactive graphical user interface. The multi-omics dashboard data includes seven dimensions: transcriptomics, proteomics, pathway activity, oxidative stress, epigenetic status, custom parameters, and comprehensive summary.

10. The AI-driven gene perturbation cytomic simulation method according to claim 9, characterized in that, Step M2 further includes: when the cell line type belongs to a preset set of passage-sensitive cell lines, calculating a passage aging correction factor using the passage number as the independent variable, the half-life passage number between the early passage and the crisis passage as the center, and the passage decay width as the scale using a logistic decrement function, and correcting the structured parameter object as follows: multiplying the passage aging correction factor by the exponential growth rate constant, multiplying the aging sensitivity of the passage aging correction factor by the pathway activity score of the stemness and differentiation pathway, adding the logarithm of the passage aging correction factor to the expression folding change of the telomerase reverse transcriptase gene by the base-2 logarithm of the passage aging correction factor, adding the reciprocal of the passage aging correction factor to the expression folding change of the cyclin-dependent kinase inhibitor 2A gene by the base-2 logarithm of the passage aging correction factor, and adding twice the difference between the passage aging correction factor and one to the expression folding change of aging-related secretory phenotype factors. When the cell line type does not belong to the passage-sensitive cell line set, the passage aging correction factor is set to one.

Images