Adaptive mechanism fusion method for live cell density prediction oriented to multi-stage cell culture

Abstract

This invention discloses an adaptive mechanism fusion method for predicting live cell density in multi-stage cell culture. Specifically, it involves: acquiring the process variable sequence of sample cells during multi-stage culture; inputting this sequence into a temporal feature extraction network to obtain a state vector; then passing this vector through three multilayer perceptron branch networks to obtain a mechanism parameter vector, gating weights, and data branch prediction values; substituting the mechanism parameter vector into a cell growth-death dynamics mechanism model to obtain a mechanism prediction sequence; using the third multilayer perceptron branch network to obtain a data prediction sequence; optimizing the temporal feature extraction network and the three multilayer perceptron branch networks to obtain a converged prediction model; and finally, obtaining the live cell density prediction result based on the prediction model and the third multilayer perceptron branch network. The beneficial effects of this invention are: it can significantly improve the prediction accuracy and stability at stage inflection points during multi-stage cell culture.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to an adaptive mechanism fusion method for predicting live cell density for multi-stage cell culture. Background Technology

[0002] Viable cell density (VCD) is one of the core process parameters for evaluating cell expansion and product quality, especially in cell therapy and regenerative medicine applications such as human umbilical cord mesenchymal stem cells (hUC-MSCs). Accurate prediction of VCD is crucial for optimizing culture strategies, implementing process control, and ensuring the quality of the final product. Meanwhile, hUC-MSCs exhibit multiple phases during in vitro expansion, including a slow proliferation phase, an exponential proliferation phase, a plateau phase, and a decline phase. The cell's response to nutrients and metabolic inhibitors changes significantly with each phase, resulting in significant time-varying and nonlinear growth kinetics, posing challenges to modeling and prediction.

[0003] Existing technologies mainly include two types of paths: (1) Mechanism-based kinetic models. Typical growth models, such as the Monod equation, describe cell growth under ideal conditions by setting a functional relationship between specific growth rate and substrate concentration. However, such models usually assume that key parameters (such as maximum specific growth rate, half-saturation constant, etc.) remain constant throughout the process, making it difficult to reflect the time-varying physiological state of cells in multi-stage culture. When the culture process enters a stage transition or metabolic pressure changes significantly, fixed-parameter models often exhibit systematic biases, resulting in insufficient cross-stage prediction accuracy.

[0004] (2) Data-driven deep learning models. Models represented by recurrent neural networks (RNNs) and their variant, long short-term memory networks (LSTMs), can learn complex nonlinear relationships from historical time-series data and, theoretically, can capture the dynamic changes in the cell culture process. However, in applications such as hUC-MSCs where it is difficult to obtain massive amounts of samples, data-driven models are easily limited by small sample sizes and suffer from insufficient generalization. In particular, they are difficult to stably characterize the key inflection point features of the stage transition period, resulting in unstable predictions of future VCDs.

[0005] To alleviate the aforementioned contradictions, a hybrid modeling approach combining mechanism and data fusion has emerged in recent years, attempting to combine mechanistic equations with learning models to balance interpretability and fitting ability. However, existing fusion schemes mostly use fixed-structure mechanistic terms with general learners for correction, and still suffer from the following common limitations: (i) mechanistic parameters are difficult to adaptively adjust with cell state, resulting in insufficient cross-stage timeliness; (ii) there is a lack of a synergistic mechanism between mechanistic constraints and data fitting that can dynamically balance with process state, making it difficult to automatically balance their weights at different stages; (iii) in small sample scenarios, the fusion model's identification and utilization of stage inflection points are still insufficient, and prediction errors are easily amplified at stage transitions. These problems make it difficult for existing schemes to achieve VCD prediction results that are both accurate and robust in multi-stage, highly time-varying, and sample-limited cell culture prediction tasks.

[0006] Therefore, there is an urgent need for a live cell density prediction technology that can be applied to multi-stage cell culture processes and still has good generalization ability under small sample conditions. This technology should introduce a state-adaptive mechanism for key mechanistic parameters while maintaining the interpretability of the mechanism, and construct a data-mechanism trade-off strategy that dynamically coordinates with the process state to improve the prediction accuracy and overall robustness at stage transitions. Summary of the Invention

[0007] To overcome the problems of existing fixed-parameter mechanism models in cell culture being unable to accurately predict across stages, and pure data-driven models having poor generalization under small sample sizes, especially with large errors at stage inflection points, this invention provides an adaptive mechanism fusion method for predicting live cell density in multi-stage cell culture.

[0008] To achieve the above-mentioned objectives, this invention provides an adaptive mechanism-integrated live cell density prediction method for multi-stage cell culture, which includes the following steps: Step S1: Obtain the process variables of cells in the multi-stage culture process and form a time-based process variable sequence. Then, perform cleaning, alignment, resampling and normalization to obtain the preprocessed process variable sequence. Step S2: Input the preprocessed process variable sequence into the time-series feature extraction network to obtain the state vector c representing the current cultivation state; Step S3: Based on the state vector c, the data is processed through three structurally independent multilayer perceptron branch networks that do not share weights, respectively, to obtain: a mechanism parameter vector p for the mechanism model, and gate weights for allocating the relative weights of the data branch and the mechanism branch. λ t and data branch prediction values ; Step S4: Substitute the mechanism parameter vector p into the preset cell growth-death kinetic mechanism model, and calculate the mechanism branch prediction value for the time corresponding to the process variable sequence. And the predicted values ​​of the mechanism branches at each time point Predicting sequences based on time aggregation mechanism X bio ; Step S5: Index the third multilayer perceptron branch network at each time point. t k Data branch prediction value Data prediction sequences are compiled chronologically to form a time index that corresponds to the process variable sequence. X data ; Step S6: Construct a structure with a real live cell density The data error term, obtained from the difference between the predicted values ​​and the data branch values, is based on the true live cell density. Mechanism branch prediction value The difference in the mechanism error term and the relationship between the gating weights The regularization term is a component of the total loss function, wherein the mechanistic error term is determined according to the gating weights. Time-varying weighting is performed; based on the total loss function, end-to-end joint optimization is performed on the temporal feature extraction network and the three multilayer perceptron branch networks in step S3 to obtain a converged prediction model; Step S7: The cell culture process samples to be predicted are preprocessed according to Step S1 to form a time-varying process variable sequence, and the state vector c is obtained by Step S2; the prediction model converged in Step S6 is used, and the data branch prediction value at the corresponding time point is output by the linear output layer of the third multilayer perceptron branch network in Step S3. This yields the predicted live cell density at the target time.

[0009] The process variables in step S1 are multidimensional state variables characterizing cell growth and the culture environment, including: a) Target variable: Live cell density ; b) Mechanism-related variables: glucose concentration S 1. Glutamine concentration S 2. Lactic acid concentration P 1. Ammonia concentration P 2; c) Auxiliary process variables: dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate, and glutamine feeding rate.

[0010] In step S2, the temporal feature extraction network includes a multi-layer long short-term memory network for encoding the preprocessed process variable sequence to obtain the hidden state sequence. ; in the hidden state sequence Positional encodings are superimposed and input into a multi-head self-attention layer to obtain an attention representation. The hidden state sequence With the attention representation An enhanced representation is obtained through residual connection, Dropout processing, and layer normalization, and the last time step of the enhanced representation is taken as the state vector c. Auxiliary process variables (dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate, and glutamine feeding rate) are used as input features in this step to extract the state vector c.

[0011] Step S3 specifically involves: The mechanism parameter vector p The first multilayer perceptron branch network Generates a four-dimensional vector. p k ,

[0012] Furthermore, the output layer employs element-wise Softplus activation to constrain the output to be positive, with each component corresponding sequentially to the maximum specific growth rate. μ max ( t k-1 Metabolite inhibition constant K P1 ( t k-1 Metabolite inhibition constant K P2 ( t k-1 ) and the rate of death μ d ( t k-1 ); The gating weight λ t The second multilayer perceptron branch network The generation process employs Sigmoid activation in its output layer to obtain a one-dimensional scalar located at (0,1). λ t ; The data branch prediction value The third multilayer perceptron branch network Directly output the corresponding time. t kData branch prediction value Its output layer is a linear mapping.

[0013] In step S4: The cell growth-death dynamics mechanism model satisfies:

[0014]

[0015] in, S 1, S 2 represents the glucose concentration and glutamine concentration, respectively. P 1, P 2 represents the concentrations of lactic acid and ammonia, respectively. K S1 , K S2 For fixed parameters, , , and Taken from the mechanism parameter vector p and according to sampling interval Discretization yields:

[0016] The above Mechanistic loss terms used during training as soft constraints based on biological first principles.

[0017] Step S5 specifically involves: Indexing the third multilayer perceptron branch network at each time point t k Data branch prediction value By compiling the data chronologically, a data prediction sequence with a time index consistent with the process variable sequence is obtained. , in, Let be a scalar at the corresponding time, and satisfy: .

[0018] The total loss function in step S6 is:

[0019] Among them, data loss for:

[0020] Used to measure data-driven predicted values Compared with the true value Differences; biological mechanism loss for:

[0021] Used to measure mechanism branch prediction values Compared with the true value Differences; Dynamic gating weights for:

[0022] The time-varying weights are generated by a gating network based on the state variable c and activated by Sigmoid; Regular terms for:

[0023] Used to inhibit Collapse towards zero; where N is the batch size and ε is a positive constant. is the regularization coefficient.

[0024] The beneficial effects of this invention are: it can significantly improve the prediction accuracy of live cell density and the stability at the inflection point of the stage in the multi-stage cell culture process, while maintaining good generalization and interpretability under small sample conditions, thus meeting the actual needs of process optimization and online monitoring. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture of the model in Embodiment 2 of the present invention.

[0026] Figure 2 The gating weight in Embodiment 2 of the present invention A schematic diagram of the time-varying characteristics (coaxial with the live cell density curve).

[0027] Figure 3 The stage transition heat map in Embodiment 2 of the present invention ( (The mechanism error changes inversely during the critical period).

[0028] Figure 4 This is a schematic diagram showing the comparison of prediction accuracy in the experimental examples of this invention.

[0029] Figure 5 This is a statistical graph showing the correlation between gating weights and mechanistic errors in the experimental examples of this invention.

[0030] Figure 6 This is a schematic diagram of time series prediction after migration in an experimental example of the present invention.

[0031] Figure 7 This is a schematic diagram of the post-migration prediction-observation scatter plot in the experimental example of this invention. Detailed Implementation

[0032] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0033] Example 1 This invention provides an adaptive mechanism fusion method for predicting live cell density in multi-stage cell culture, comprising the following steps: Step S1: Obtain the process variables of cells in the multi-stage culture process and form a time-based process variable sequence. Then, perform cleaning, alignment, resampling and normalization to obtain the preprocessed process variable sequence. Step S2: Input the preprocessed process variable sequence into the time-series feature extraction network to obtain the state vector c representing the current cultivation state; Step S3: Based on the state vector c, the data is processed through three structurally independent multilayer perceptron branch networks that do not share weights, respectively, to obtain: a mechanism parameter vector p for the mechanism model, and gate weights for allocating the relative weights of the data branch and the mechanism branch. λ t and data branch prediction values ; Step S4: Substitute the mechanism parameter vector p into the preset cell growth-death kinetic mechanism model, and calculate the mechanism branch prediction value for the time corresponding to the process variable sequence. And the predicted values ​​of the mechanism branches at each time point Predicting sequences based on time aggregation mechanism X bio The three branches mentioned above receive the input of the state vector c in parallel and do not share weights with each other. Step S5: Index the third multilayer perceptron branch network at each time point. t k Data branch prediction value Data prediction sequences are compiled chronologically to form a time index that corresponds to the process variable sequence. X data ; Step S6: Construct a structure with a real live cell density The data error term, obtained from the difference between the predicted values ​​and the data branch values, is based on the true live cell density. Mechanism branch prediction value The difference in the mechanism error term and the relationship between the gating weights The regularization term is a component of the total loss function, wherein the mechanistic error term is determined according to the gating weights. Time-varying weighting is performed; based on the total loss function, end-to-end joint optimization is performed on the temporal feature extraction network and the three multilayer perceptron branch networks in step S3 to obtain a converged prediction model; Step S7: The cell culture process samples to be predicted are preprocessed according to Step S1 to form a time-varying process variable sequence, and the state vector c is obtained by Step S2; the prediction model converged in Step S6 is used, and the data branch prediction value at the corresponding time point is output by the linear output layer of the third multilayer perceptron branch network in Step S3. This yields the predicted live cell density at the target time.

[0034] Wherein, the process variables in step S1 are multidimensional state quantities characterizing cell growth and the culture environment, including: a) Target variable: Live cell density ; b) Mechanism-related variables: glucose concentration S 1. Glutamine concentration S 2. Lactic acid concentration P 1. Ammonia concentration P 2; used for the mechanistic model calculation in step S4; c) Auxiliary process variables: dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate and glutamine feeding rate; these serve as input features for step S2 and are involved in the extraction of state vector c.

[0035] Furthermore, during the training phase, the sequence of process variables includes time indices. t k The corresponding true live cell density .

[0036] In step S2, the temporal feature extraction network includes a multi-layer long short-term memory network for encoding the preprocessed process variable sequence to obtain the hidden state sequence. ; in the hidden state sequence Positional encodings are superimposed and input into a multi-head self-attention layer to obtain an attention representation. The hidden state sequence With the attention representation An enhanced representation is obtained through residual connection, Dropout processing, and layer normalization, and the last time step of the enhanced representation is taken as the state vector c. Auxiliary process variables (dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate, and glutamine feeding rate) are used as input features in this step to extract the state vector c.

[0037] Step S3 specifically involves: The mechanism parameter vector p The first multilayer perceptron branch network Generates a four-dimensional vector. p k ,

[0038] Furthermore, the output layer employs element-wise Softplus activation to constrain the output to be positive, with each component corresponding sequentially to the maximum specific growth rate. μ max ( t k-1 Metabolite inhibition constant K P1 ( t k-1 Metabolite inhibition constant K P2 ( t k-1 ) and the rate of death μ d ( t k-1 ); The gating weight λ t The second multilayer perceptron branch network The generation process employs Sigmoid activation in its output layer to obtain a one-dimensional scalar located at (0,1). λ t ; The data branch prediction value The third multilayer perceptron branch network Directly output the corresponding time. t k Data branch prediction value Its output layer is a linear mapping (without activation); The three branches mentioned above receive the state vector in parallel. c The inputs are complementary and share weights.

[0039] In step S4: The cell growth-death dynamics mechanism model satisfies:

[0040]

[0041] in, S 1, S 2 represents the glucose concentration and glutamine concentration, respectively. P 1, P 2 represents the concentrations of lactic acid and ammonia, respectively. KS1 , K S2 For fixed parameters, , , and Taken from the mechanism parameter vector p and according to sampling interval Discretization yields:

[0042] The above Mechanistic loss terms used during training as soft constraints based on biological first principles.

[0043] Step S5 specifically involves: Indexing the third multilayer perceptron branch network at each time point t k Data branch prediction value By compiling the data chronologically, a data prediction sequence with a time index consistent with the process variable sequence is obtained. , in, Let be a scalar at the corresponding time, and satisfy: .

[0044] The total loss function in step S6 is:

[0045] Among them, data loss for:

[0046] Used to measure data-driven predicted values Compared with the true value Differences; biological mechanism loss for:

[0047] Used to measure mechanism branch prediction values Compared with the true value Differences; Dynamic gating weights for:

[0048] The time-varying weights are generated by a gating network based on the state variable c and activated by Sigmoid; Regular terms for:

[0049] Used to inhibit Collapse towards zero; where N is the batch size and ε is a positive constant. is the regularization coefficient.

[0050] Example 2 This invention provides an adaptive mechanism fusion method for predicting live cell density in multi-stage cell culture. For ease of understanding, Figure 1 The method framework of this embodiment is given. This embodiment takes a multidimensional process variable sequence as input, obtains a state vector c through a time-series feature extraction network, and generates a dynamic mechanism parameter vector p and a gating weight by three multilayer perceptron branch networks. λ t The data branch predictions are used; dynamic parameters are substituted into the extended growth-death dynamics mechanism model to obtain the mechanism branch predictions, and end-to-end co-training is performed (where the data branch predictions are obtained from the data branch predictions). λ t By setting logarithmic regularization and balancing the loss term, the data branch and the mechanistic branch can adaptively coordinate at different culture stages; in the online prediction stage, the data branch directly outputs the predicted live cell density. This method significantly improves cross-stage robustness and prediction accuracy in small sample scenarios while ensuring mechanistic interpretability, especially effectively reducing errors in the stage transition interval. The prediction method is as follows: I. Obtain process variables of cell samples during multi-stage culture and form a time-varying process variable sequence. Perform cleaning, alignment, resampling, and normalization to obtain a preprocessed process variable sequence. The process variables are multidimensional state variables characterizing cell growth and the culture environment, including: a) Target variable: Live cell density ; b) Mechanism-related variables: glucose concentration S 1. Glutamine concentration S 2. Lactic acid concentration P 1. Ammonia concentration P 2; used for the mechanistic model calculation in step S4; c) Auxiliary process variables: dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate and glutamine feeding rate; these serve as input features for step S2 and are involved in the extraction of state vector c.

[0051] Furthermore, during the training phase, the sequence of process variables includes time indices. t k The corresponding true live cell density .

[0052] 2. The preprocessed process variable sequence is input into a temporal feature extraction network to obtain a state vector c representing the current cultivation state. The temporal feature extraction network includes a multi-layer long short-term memory network for encoding the preprocessed process variable sequence to obtain a hidden state sequence. ; in the hidden state sequence Positional encodings are superimposed and input into a multi-head self-attention layer to obtain an attention representation. The hidden state sequence With the attention representation An enhanced representation is obtained through residual connection, Dropout processing, and layer normalization, and the last time step of the enhanced representation is taken as the state vector c. Auxiliary process variables (such as dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate, and glutamine feeding rate) are used as input features in this step to extract the state vector c.

[0053] III. Based on the aforementioned state vector c, the following results are obtained through three structurally independent multilayer perceptron branch networks that do not share weights: 1. Mechanism parameter vector p used in mechanism models Mechanism parameter vector p : By the first multilayer perceptron branch network Generates a four-dimensional vector. p k ,

[0054] Furthermore, the output layer employs element-wise Softplus activation to constrain the output to be positive, with each component corresponding sequentially to the maximum specific growth rate. μ max ( t k-1 Metabolite inhibition constant K P1 ( t k-1 Metabolite inhibition constant K P2 ( t k-1 ) and the rate of death μ d ( t k-1 ); 2. Gating weights used to assign relative weights between data branches and mechanism branches. λ t ; Gating weights λ t : By the second layer perceptron branch network The generation process employs Sigmoid activation in its output layer to obtain a one-dimensional scalar located at (0,1). λ t ; 3. And the predicted value of the data branch .

[0055] Data branch prediction value : By the third multilayer perceptron branch network Directly output the corresponding time. t k Data branch prediction value Its output layer is a linear mapping (without activation); The three branches mentioned above receive the state vector in parallel. c The inputs are complementary and share weights.

[0056] Fourth, substitute the mechanistic parameter vector p into a preset cell growth-death kinetic mechanism model, and calculate the mechanistic branch prediction value for the time corresponding to the process variable sequence. And the predicted values ​​of the mechanism branches at each time point Predicting sequences based on time aggregation mechanism X bio The three branches mentioned above receive the input of the state vector c in parallel and do not share weights; wherein, the cell growth-death dynamics mechanism model satisfies:

[0057]

[0058] in, S 1, S 2 represents the glucose concentration and glutamine concentration, respectively. P 1, P 2 represents the concentrations of lactic acid and ammonia, respectively. K S1 , K S2 For fixed parameters, , , and Taken from the mechanism parameter vector p and according to sampling interval Discretization yields:

[0059] The above Mechanistic loss terms used during training as soft constraints based on biological first principles.

[0060] 5. Index the third multilayer perceptron branch network at each time point.t k Data branch prediction value By compiling the data chronologically, a data prediction sequence with a time index consistent with the process variable sequence is obtained. ,in, Let be a scalar at the corresponding time, and satisfy: .

[0061] VI. Constructing a system based on the density of real living cells The data error term, obtained from the difference between the predicted values ​​and the data branch values, is based on the true live cell density. Mechanism branch prediction value The difference in the mechanism error term and the relationship between the gating weights The regularization term is a component of the total loss function, wherein the mechanistic error term is determined according to the gating weights. Perform time-varying weighted average. Figure 2 It is to gating weights When plotted coaxially with the live cell density curve, the transition points between stages can be observed. The occurrence of adaptive adjustments indicates that the model will adjust the mechanism and relative dependence of data at key stages, which also reflects the dynamic collaborative strategy of this embodiment. Furthermore, Figure 3 To use heatmap comparison The inverse relationship between the mechanism error and the critical period visually demonstrates the adjustment pattern of increasing error and decreasing mechanism weight, indicating that the gating strategy is effective.

[0062] The total loss function is used to measure the data-driven predicted value. Compared with the true value The differences are as follows:

[0063] Among them, data loss Used to measure mechanism branch prediction values Compared with the true value The differences are as follows:

[0064] biological mechanism loss for:

[0065] Dynamic gating weights The time-varying weights are generated by a gating network based on the state variable c and activated by a Sigmoid function, specifically as follows:

[0066] Regular terms for:

[0067] Used to inhibit Collapse towards zero; where N is the batch size and ε is a positive constant. is the regularization coefficient.

[0068] Based on the total loss function, the temporal feature extraction network and the three multilayer perceptron branch networks in step S3 are jointly optimized end-to-end to obtain a converged prediction model.

[0069] 7. The cell culture process samples to be predicted are preprocessed according to step one to form a time-varying sequence of process variables, and the state vector c is obtained from step two. The prediction model converged in step six is ​​used, and the data branch prediction value at the corresponding time point is output by the linear output layer of the third multilayer perceptron branch network in step three. This yields the predicted live cell density at the target time.

[0070] Experimental Example To verify the prediction accuracy, robustness at stage inflection points, effectiveness of the gating mechanism, and transferability of the adaptive mechanism-integrated live cell density prediction method for multi-stage cell culture under multi-stage scenarios, the following comparative and transfer experiments were designed. All experiments followed the data preprocessing and time alignment process (cleaning, alignment, resampling, and normalization) described in Example 2, and used the same time window to construct a time-varying sequence of process variables as model input. The target values ​​for both training and validation phases were real live cell densities. .

[0071] 1. Experimental subjects and baseline comparison The tested method (Example 2 of this invention): The temporal feature extraction network outputs a state vector c, and three structurally independent MLP branches that do not share weights output a dynamic mechanism parameter vector p, a gate weight λt, and a data branch prediction value, respectively. The mechanistic branch is obtained by combining the extended Monod model with discrete recursion. Training period Perform end-to-end collaborative training.

[0072] Baseline comparison: LSTM, Improved Attention LSTM (implementation details and training / validation split are consistent with this invention).

[0073] 2. Experimental Setup Data processing: outlier handling, interpolation and equal-interval resampling, normalization / standardization, consistent with Example 2, to ensure comparable magnitude of loss terms.

[0074] Evaluation metrics: MAE, RMSE,R 2 ; Observe the curve fit separately at the inflection point segment.

[0075] Gating validity test: statistics The correlation and significance with mechanistic error (or its normalized value).

[0076] Migration setup: A migration strategy of freezing the lower layers and fine-tuning the upper layers was adopted on small sample data of hUC-MSC to verify the generalization under cross-system small sample.

[0077] 3. Overall Results and Comparison like Figure 2 As shown, four representative batches are selected to compare the actual X(t) with the predictions of each model. It can be observed that: At the inflection point of the stage, the curve of the present invention fits the real curve better, and the peak / valley position and amplitude error are significantly lower than the baseline. Smaller MAE and RMSE across all time periods R 2 Higher (see the accompanying table for a statistical summary).

[0078] This indicates that, compared to purely data-driven models, this method has stronger trend capture and inflection point robustness in long-term, multi-stage cultivation processes.

[0079] 4. Effectiveness of the gating mechanism To examine whether the dynamic collaborative strategy of prioritizing mechanisms over data is truly effective, two types of complementary evidence are presented: Time-varying behavior (e.g.) Figure 2 (as shown) As can be seen from the coaxial display with the VCD, It will make adaptive adjustments when the phase transitions or dynamic modes change, reflecting its response to critical periods.

[0080] Statistical related (e.g.) Figure 5 As shown): Correlation analysis of the four batches shows It is significantly negatively correlated with mechanism error (most p-values ​​are significant), that is, when mechanism error increases... Automatically decreases and increases, thus achieving the desired effect during the training period. L bio It is a time-varying trade-off rather than a static summation.

[0081] also, Figure 3 The phase transition heatmap visually presents the above phenomena: The mechanism error changes inversely at key time grids, further reinforcing the correspondence between the gating learned strategy and the process stages.

[0082] 5. Transfer and small sample scenarios On small sample data of hUC-MSC, a migration strategy of freezing the bottom layer and fine-tuning the top layer was adopted: Figure 6 The data shows that the predicted trend at the time series level is in good agreement with the actual trend. Figure 7 The predicted-observed scatter plots for the entire sample are close. The diagonal lines indicate a high overall fitting accuracy.

[0083] The results show that this method can maintain stable performance under small sample conditions across cell lines, thanks to the shared structure of mechanism-related variables and the adaptability of the gating strategy.

[0084] In summary, compared to LSTM / attention LSTM, this invention is more stable at stage inflection points and has lower overall error. Figure 4 ); Gating weights Negative correlation with mechanism error and its stage-wise adaptive ( Figure 2 , Figure 4 This demonstrates that the dynamic collaborative strategy can adaptively allocate weights between the mechanism and the data based on the process state; it also exhibits good generalization in cross-system small sample scenarios. Figure 6 , Figure 7 Therefore, this invention combines mechanistic interpretability with data-driven flexibility, and can significantly improve the accuracy and robustness of live cell density prediction in multi-stage cell culture processes.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-adaptive mechanism fusion viable cell density prediction method for multi-stage cell culture, characterized in that, The prediction method includes the following steps: Step S1: Obtain the process variables of cells in the multi-stage culture process and form a time-based process variable sequence. Then, perform cleaning, alignment, resampling and normalization to obtain the preprocessed process variable sequence. Step S2: Input the preprocessed process variable sequence into the time-series feature extraction network to obtain the state vector c representing the current cultivation state; Step S3: Based on the state vector c, the data is processed through three structurally independent multilayer perceptron branch networks that do not share weights, respectively, to obtain: a mechanism parameter vector p for the mechanism model, and gate weights for allocating the relative weights of the data branch and the mechanism branch. λ t and data branch prediction values , Wherein, the mechanism parameter vector p The first multilayer perceptron branch network Generates a four-dimensional vector. p k , Furthermore, the output layer employs element-wise Softplus activation to constrain the output to be positive, with each component corresponding sequentially to the maximum specific growth rate. μ max ( t k-1 Metabolite inhibition constant K P1 ( t k-1 Metabolite inhibition constant K P2 ( t k-1 ) and the rate of death μ d ( t k-1 ); Step S4: Substitute the mechanism parameter vector p into the preset cell growth-death kinetic mechanism model, and calculate the mechanism branch prediction value for the time corresponding to the process variable sequence. And the predicted values ​​of the mechanism branches at each time point Predicting sequences based on time aggregation mechanism X bio , The cell growth-death kinetics model satisfies the following: Where X represents the live cell density, μg represents the specific cell growth rate, and μd represents the specific cell death rate; the μg is determined by the maximum specific growth rate μ * max Glucose concentration S1, glutamine concentration S2, lactate concentration P1, ammonia concentration P2, glucose half-saturation constant K S1 Glutamine half-saturation constant K S2 Lactic acid inhibition constant K * P1 and ammonia inhibition constant K * P2 Calculated together; μ * max 、K * P1、K * P2 is the notation of the components of the mechanism parameter vector pk in the mechanism model, in, , , and Taken from the mechanism parameter vector p and according to sampling interval Discretization yields: The above As a soft constraint based on the first biological principle for the training period mechanism loss term; Step S5: a third multi-layer perceptron branch network is applied to each time index t k of the data branch prediction values According to the time sequence, the data prediction sequence with the same time index as the process variable sequence is formed X data ; Step S6: Construct a structure with a real live cell density The data error term, obtained from the difference between the predicted values ​​and the data branch values, is based on the true live cell density. Mechanism branch prediction value The difference in the mechanism error term and the relationship between the gating weights The regularization term is the total loss function of the components; based on the total loss function, the temporal feature extraction network and the three multilayer perceptron branch networks in step S3 are jointly optimized end-to-end to obtain a converged prediction model; Step S7: The cell culture process samples to be predicted are preprocessed according to Step S1 to form a time-varying process variable sequence, and the state vector c is obtained by Step S2; the prediction model converged in Step S6 is used, and the data branch prediction value at the corresponding time point is output by the linear output layer of the third multilayer perceptron branch network in Step S3. This yields the predicted live cell density at the target time.

2. The prediction method of claim 1, wherein, The process variables in step S1 are multidimensional state variables characterizing cell growth and the culture environment, including: a) Target variable: viable cell density ; b) Mechanism-related variables: glucose concentration S 1. glutamine concentration S 2. lactate concentration P 1. ammonia concentration P 2. c) Auxiliary process variables: dead cell density, number of lysed cells, culture temperature, culture medium pH, glucose feeding rate, and glutamine feeding rate.

3. The prediction method of claim 1, wherein, In step S2, the temporal feature extraction network includes a multi-layer long short-term memory network for encoding the preprocessed process variable sequence to obtain the hidden state sequence. ( ) ; in the hidden state sequence ( ) Positional encodings are superimposed and input into a multi-head self-attention layer to obtain an attention representation. The hidden state sequence ( ) With the attention representation An enhanced representation is obtained through residual connection, Dropout processing, and layer normalization, and the last time step of the enhanced representation is taken as the state vector c.

4. The prediction method of claim 1, wherein, The gating weights in the step S3 λ t , by a second multi-layer perceptron branch network generated, whose output layer employs a sigmoid activation to obtain a one-dimensional scalar lying in (0, 1) λ t .

5. The prediction method according to claim 1, characterized in that, The data branch prediction value in the step S3 By a third multi-layer perception branch network Directly output the corresponding moment t k The data branch prediction value The output layer is a linear mapping.

Images