Artificial intelligence-based prediction method for serum-free medium components

Abstract

This application relates to the field of culture medium prediction technology and discloses an artificial intelligence-based method for predicting serum-free culture medium components. The method includes: constructing a phenotypic condition vector based on cell type identifiers and cell phenotypic specification parameters; constructing a phenotypic-recipe pairing dataset using historical culture batch data; training a conditional variational autoencoder using the phenotypic condition vector as conditional constraints; sampling latent variables from a standard normal distribution and decoding them through a decoder to obtain a set of candidate culture medium recipes; and using a forward validator to screen the candidate recipes with the smallest Euclidean distance to obtain a recommended culture medium recipe. This application solves the problem that existing artificial intelligence-based serum-free culture medium component prediction methods use cell number as the optimization objective and cannot reverse-engineer culture medium recipes based on clinical phenotypic quality specifications.

Description

Technical Field

[0001] This application relates to the field of culture medium prediction technology, and in particular to an artificial intelligence-based method for predicting the components of serum-free culture media. Background Technology

[0002] The rapid development of the cell and gene therapy (CGT) industry has placed stringent demands on the in vitro culture of therapeutic cells. The expansion and culture of therapeutic cells such as CAR-T cells, NK cells, and mesenchymal stem cells rely on serum-free culture media. The types and concentrations of components in the culture medium, such as amino acids, vitamins, inorganic salts, and cytokines, directly affect the in vitro expansion behavior of cells. In existing technologies, artificial intelligence-based methods for predicting the components of serum-free culture media have been preliminarily applied. A typical implementation uses the concentration of each component in the culture medium as input features and the cell density, expansion fold, or cell viability at the end of culture as output labels. A positive prediction model is trained using gradient boosting decision trees, random forests, or active learning frameworks. Based on this model, the component concentrations are iteratively optimized to screen for the culture medium formulation that maximizes the number of cells.

[0003] However, the aforementioned existing technologies have fundamental shortcomings. Current methods use cell number as an optimization target, which is fundamentally misaligned with the core quality requirements of CAR-T cell clinical products—the clinical efficacy of CAR-T cell products depends on the phenotypic quality of the final product, specifically reflected in quantifiable phenotypic indicators such as the proportion of stem memory cells, the expression level of exhaustion markers, and mitochondrial metabolic status, rather than simply the number of cells expanded. A positive predictive framework that uses cell number as an optimization target cannot translate the manufacturer's pre-set phenotypic quality specifications into specific culture medium component formulations. This could lead to the selected formulations, while meeting the expansion quantity requirements, potentially driving cells to differentiate into end-effect phenotypes, impairing the product's in vivo persistence and anti-tumor activity. Summary of the Invention

[0004] This application provides an artificial intelligence-based method for predicting serum-free culture medium components, which solves the problems of existing artificial intelligence-based methods for predicting serum-free culture medium components that use cell number as the optimization target and cannot reverse generate culture medium formulations based on clinical phenotype quality specifications, as well as the problem that a single phenotype target-driven approach cannot adapt to the different component requirements of CAR-T cell multi-stage culture.

[0005] In a first aspect, this application provides an artificial intelligence-based method for predicting the components of serum-free culture media, the method comprising: Step S1: Based on the cell type identifier and cell phenotype specification parameters, construct the phenotype target vector and the cell type embedding vector, and concatenate the phenotype target vector and the cell type embedding vector to obtain the phenotype condition vector. Step S2: Divide the historical culture batch data into culture stages, and construct a phenotype-formula pairing dataset using the measured phenotypic vectors at the end of each stage as labels. Step S3: Using the phenotypic condition vector as a conditional constraint, the phenotypic-formula pairing dataset is input into a conditional variational autoencoder for training. The encoder of the conditional variational autoencoder takes the concatenation of the formulation vector and the phenotypic condition vector in the phenotypic-formula pairing dataset as input and outputs the mean and log-variance of the latent variables. The latent variables are obtained through reparameterized sampling. The decoder of the conditional variational autoencoder takes the concatenation of the latent variables and the phenotypic condition vector as input and optimizes the loss function by using the sum of reconstruction loss and KL divergence. During inference, the phenotypic condition vector corresponding to the phenotypic target vector is used as a condition. Multiple sets of latent variables are sampled from the standard normal distribution and decoded by the decoder to obtain a set of candidate culture medium formulations. Step S4: Input the set of candidate culture medium formulations into the positive validator, and filter by the minimum Euclidean distance between the predicted phenotypic vector of each candidate formulation and the target phenotypic vector to obtain the recommended culture medium formulation.

[0006] The technical solution provided in this application fundamentally changes the existing positive prediction framework that uses culture medium component concentration as the sole input by constructing cell type identifiers and cell phenotype specification parameters into a phenotype target vector and a cell type embedding vector, and then concatenating them into a phenotype conditional vector. This introduces the manufacturer's pre-defined clinical phenotype quality specifications into the starting point of formulation generation, ensuring that all subsequent calculations are performed under the constraint of the phenotype target. Based on this, historical culture batch data is segmented according to culture stages, and a phenotype-formulation pairing dataset is constructed using the measured phenotype vectors at the end of each stage as labels. This establishes a supervised mapping relationship from quantifiable phenotype indicators to culture medium component concentrations, providing structured pairing samples for the training of the conditional variational autoencoder. This enables the model to establish a learnable causal correspondence between the phenotype space and the formulation space, rather than relying on the statistical average of a single batch result. The conditional variational autoencoder employs a dual-end conditional constraint structure: the encoder takes the concatenation of the formulation vector and the phenotypic condition vector as input, while the decoder takes the concatenation of the latent variable and the phenotypic condition vector as input. This structure ensures that both sampling and decoding of the latent space are explicitly guided by the phenotypic target vector. During inference, multiple sets of latent variables are sampled from the standard normal distribution and decoded by the decoder to obtain a set of candidate culture medium formulations. This achieves the reverse generation from the phenotypic target to the formulation distribution, which differs from the forward prediction path from formulation to cell number in existing technologies.

[0007] The forward validator takes the concatenation of candidate culture medium formulations and cell type embedding vectors in the candidate formulation set as input and the Euclidean distance between the predicted phenotypic vector and the phenotypic target vector as the screening criterion. It forms a closed-loop structure with the conditional variational autoencoder, creating a reverse generation and forward verification mechanism. This establishes a quantitative constraint relationship between generation diversity and formulation reliability. The component safety concentration range filtering further incorporates formulation feasibility into the screening criteria, ensuring that the recommended culture medium formulation meets phenotypic matching requirements while complying with safe operating procedures for cell culture. These technical features work synergistically in the specific application scenario of serum-free culture medium component prediction: the phenotypic conditional vector transforms clinical quality specifications into numerical constraints that the model can handle; the reverse generation structure of the conditional variational autoencoder applies this constraint throughout the sampling process of the formulation space; and the closed-loop verification of the forward validator transforms phenotypic matching errors into quantifiable screening criteria, forming a complete prediction chain driven by phenotypic quality targets and outputting culture medium component formulations. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of an embodiment of the serum-free culture medium component prediction method based on artificial intelligence in this application. Figure 2 This is a schematic diagram illustrating the loss convergence during the training process of the conditional variational autoencoder in an embodiment of this application. Figure 3 This is a schematic diagram showing the distribution of cytokine concentrations in the recommended culture medium formulations for the three culture stages in this application embodiment. Detailed Implementation

[0010] This application provides an artificial intelligence-based method for predicting serum-free culture medium components. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0011] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the serum-free culture medium component prediction method based on artificial intelligence in this application includes: Step S1: Based on the cell type identifier and cell phenotype specification parameters, construct the phenotype target vector and the cell type embedding vector, and concatenate the phenotype target vector and the cell type embedding vector to obtain the phenotype condition vector. Specifically, cell phenotype specification parameters are a set of quantitative indicators pre-set by the manufacturer according to regulatory-approved product quality standards. These are seven phenotypic percentage values ​​detected by flow cytometry, including the CCR7 positivity rate, CD62L positivity rate, CD45RA positivity rate, PD-1 single positivity rate, PD-1 and TIM-3 double positivity rate, CD4 to CD8 ratio, and JC-1 staining positivity rate. All values ​​are continuous numerical parameters that can be directly detected in the laboratory. Arranging these seven parameters into a one-dimensional vector yields the phenotypic target vector. The cell type embedding vector maps discrete cell type identifiers into a 32-dimensional continuous real-number vector using a pre-trained embedding layer. Setting the embedding dimension to 32 is an engineering choice that balances expressive power and parameter quantity. The phenotypic target vector and the cell type embedding vector are concatenated end-to-end according to their dimensions to obtain the phenotypic condition vector. This vector carries both phenotypic quality targets and cell type information, serving as the sole conditional input for the subsequent generative network.

[0012] Step S2: Divide the historical culture batch data into culture stages, and construct a phenotype-formula pairing dataset using the measured phenotypic vectors at the end of each stage as labels. Specifically, historical culture batch data is segmented into three stages: activation, amplification, and harvest preparation. Each segment ends with a label representing the actual phenotypic data detected by flow cytometry, paired with the corresponding culture medium component record to form a phenotypic-formula pairing dataset. The core significance of this dataset lies in establishing a supervised mapping relationship from culture medium components to cell phenotypes, providing training samples for the inverse generation network in step S3. The concentrations of each component must be Z-score normalized to eliminate numerical scale differences between components of different dimensions, ensuring that the gradient contribution of each component to the loss function is consistent during network training.

[0013] Step S3: Using the phenotypic condition vector as a conditional constraint, the phenotypic-formula pairing dataset is input into the conditional variational autoencoder for training. The encoder of the conditional variational autoencoder takes the concatenation of the formulation vector and the phenotypic condition vector in the phenotypic-formula pairing dataset as input and outputs the mean and log-variance of the latent variables. The latent variables are obtained through reparameterized sampling. The decoder of the conditional variational autoencoder takes the concatenation of the latent variables and the phenotypic condition vector as input and optimizes the loss function with the sum of the reconstruction loss and the KL divergence. During inference, the phenotypic condition vector corresponding to the phenotypic target vector is used as a condition. Multiple sets of latent variables are sampled from the standard normal distribution and decoded by the decoder to obtain the candidate culture medium formulation set. Specifically, the conditional variational autoencoder is the core structure that distinguishes this scheme from existing forward prediction methods. Existing technologies predict cell numbers using the formulation as input. In this scheme, the encoder takes the concatenation of the formulation vector and the phenotypic condition vector as input, learning the latent distribution of the formulation under given phenotypic target conditions. The decoder reconstructs the formulation using the concatenation of latent variables and the phenotypic condition vector as input. During inference, the phenotypic condition vector is fixed, and latent variables are sampled from a standard normal distribution. The decoder output is the candidate formulation that satisfies the phenotypic target conditions, achieving reverse generation from the phenotypic target to the culture medium formulation. The reconstruction loss constrains the consistency between the decoder output and the original formulation, and the KL divergence constrains the latent space to follow a standard normal distribution. The weighted sum of these two constraints ensures a balance between generation diversity and reconstruction accuracy.

[0014] Step S4: Input the set of candidate culture medium formulations into the positive validator, and filter them by the minimum Euclidean distance between the predicted phenotypic vector and the phenotypic target vector of each candidate formulation to obtain the recommended culture medium formulation.

[0015] Specifically, the forward validator, acting as a validation network independent of the conditional variational autoencoder, takes the concatenation of candidate formulations and cell type embedding vectors as input to predict the actual phenotypic vectors generated by the formulation. It then calculates the Euclidean distance between the predicted phenotypic vector and the target phenotypic vector as the phenotypic matching error value. Prior to this, a component safety concentration range filter must be performed to eliminate candidate formulations with any component concentration exceeding a preset safety range. This safety concentration range is set based on published cytotoxicity experimental data and GMP specifications. Finally, the candidate formulation with the smallest phenotypic matching error value is selected from the filtered feasible candidate formulations, denormalized to the true concentration range, and the recommended culture medium formulation is obtained.

[0016] In one specific embodiment, step S1 includes: The cell type identifier is input into the pre-trained embedding layer, and the cell type identifier is processed by embedding mapping to obtain the cell type embedding vector. Seven cell phenotypic parameters detected by flow cytometry, namely CCR7 positive rate, CD62L positive rate, CD45RA positive rate, PD-1 single positive rate, PD-1 and TIM-3 double positive rate, CD4 to CD8 ratio and JC-1 staining positive rate, were vectorized and encoded to obtain the phenotypic target vector. The phenotypic target vector and the cell type embedding vector are concatenated according to their dimensions to obtain the phenotypic condition vector.

[0017] Specifically, the pre-trained embedding layer uses cell type identifiers as discrete category inputs. The cell type identifier is an element from a predefined category set, which includes four types: CD3-positive T cells, NK cells, mesenchymal stem cells, and iPSC-derived T cells, encoded as integers from 0 to 3. The embedding layer maps these integer codes to a 32-dimensional continuous real-valued vector, i.e., the cell type embedding vector. The 32-dimensional dimension is chosen because: too low a dimension leads to overlapping expression spaces for different cell types, while too high a dimension introduces redundant parameters, causing training instability. 32 dimensions achieve an engineering balance between the discriminative power of the four cell types and the number of parameters. The embedding layer is trained jointly with a conditional variational autoencoder, and after training, its weights are fixed for use in the inference phase.

[0018] The seven cell phenotypic parameters are all proportional values ​​directly detected by flow cytometry, ranging from 0 to 1 as continuous real numbers. The specific meanings of each parameter are as follows: the CCR7 positivity rate, CD62L positivity rate, and CD45RA positivity rate reflect the stemness and memory phenotype of T cells; the PD-1 single positivity rate reflects the early exhaustion level; the PD-1 and TIM-3 double positivity rate reflects the terminal exhaustion level; the CD4 to CD8 ratio is normalized to a range of 0 to 1 by dividing the number of CD4 cells by the sum of the number of CD4 and CD8 cells; and the JC-1 staining positivity rate reflects the integrity of mitochondrial membrane potential, i.e., the level of cellular oxidative phosphorylation metabolism. Arranging these seven values ​​in a fixed order into a seven-dimensional vector yields the phenotypic target vector. Concatenating this vector with the 32-dimensional cell type embedding vector along its dimensions results in a 39-dimensional phenotypic conditional vector, which remains consistent in the input layers of the subsequent encoder and decoder.

[0019] In one specific embodiment, step S2 includes: Historical culture batch data were segmented into activation stage, amplification stage, and harvest preparation stage to obtain culture medium component records and flow cytometry detection data at the end of each stage. Based on flow cytometry data at the end of each stage, seven indicators—CCR7 positive rate, CD62L positive rate, CD45RA positive rate, PD-1 single positive rate, PD-1 and TIM-3 double positive rate, CD4 to CD8 ratio, and JC-1 staining positive rate—were vectorized and encoded to obtain the measured phenotypic vectors at the end of each stage. Based on the cell viability threshold and Grubbs test, abnormal batch removal was performed on the measured phenotypic vectors at the end of each stage to obtain the filtered effective stage sample set. The culture medium component records of each stage in the effective stage sample set are organized and processed according to the pairing relationship with the corresponding measured phenotypic vectors at the end of each stage to obtain the phenotypic-formula pairing dataset.

[0020] Specifically, historical culture batch data are divided into three stages based on time nodes: the activation stage (Days 0-3), the expansion stage (Days 3-10), and the harvest preparation stage (Days 10-14). These time nodes are determined according to the biological patterns of T cell activation completion, proliferation peak, and harvest window in the standard operating procedures for CAR-T cell manufacturing. The flow cytometry data at the end of each stage includes the same seven indicators as the phenotypic target vector in step S1, arranged in the same fixed order as a seven-dimensional vector, resulting in the measured phenotypic vector at the end of each stage, ensuring a one-to-one correspondence between the dimension and meaning of the label vector and the target vector.

[0021] Abnormal batch removal was performed in two steps: First, batches with a cell viability threshold below 70% at the end of each stage were removed entirely. 70% is the minimum acceptable viability standard in GMP regulations for cell therapy products. Second, for the batches filtered in the first step, Grubbs' test was independently performed on each dimension of the measured phenotypic vector at the end of each stage, with a significance level set at 0.05. Batches whose test statistic for any dimension exceeded the corresponding critical value were identified as outliers and removed, resulting in the filtered effective stage sample set. The culture medium component records included the molar concentrations of 28 components, including amino acids, vitamins, inorganic salts, and cytokines. The concentration of each component was standardized using Z-scores based on the mean and standard deviation of that component across all effective batches to eliminate dimensional differences between components. The standardized 28-dimensional formulation vector was paired one-to-one with the corresponding seven-dimensional measured phenotypic vectors at the end of each stage according to batch number, forming a phenotypic-formula paired dataset. Each sample consisted of both the standardized formulation vector and the measured phenotypic vector.

[0022] In one specific embodiment, in step S3, the phenotypic-recipe pairing dataset is input into the conditional variational autoencoder for training. The encoder of the conditional variational autoencoder takes the concatenation of the recipe vector and the phenotypic condition vector in the phenotypic-recipe pairing dataset as input and outputs the mean and log-variance of the latent variables. The latent variables are obtained through reparameterization sampling, including: The standardized recipe vector and phenotypic condition vector in the phenotypic-recipe pairing dataset are concatenated by dimension to obtain the encoder input vector; The encoder input vector is input into an encoder consisting of a three-layer fully connected network with ReLU activation function in each layer. Feature compression is performed on the encoder input vector to obtain the latent mean vector and the latent log-variance vector. Based on the latent mean vector and the latent log-variance vector, the latent mean vector and the latent log-variance vector are reparameterized and sampled to obtain the latent variables.

[0023] Specifically, the standardized recipe vector has a dimension of 28, and the phenotypic condition vector has a dimension of 39. These are concatenated end-to-end according to their dimensions to obtain a 67-dimensional encoder input vector. The encoder consists of a three-layer fully connected network with 256, 128, and 64 nodes respectively. A ReLU activation function is used to introduce non-linearity between adjacent layers, compressing the 67-dimensional encoder input vector layer by layer to a 64-dimensional hidden layer representation. Following the 64-dimensional hidden layer representation, two independent linear mapping layers are connected, each outputting a 16-dimensional vector, which serves as the latent mean vector and the latent log-variance vector, respectively. The latent space dimension is set to 16 because: too low a dimension results in insufficient expressive power, failing to cover the diversity of recipe distributions corresponding to different phenotypic targets; too high a dimension leads to a sparse latent space and an increased probability of sampling points falling into low-density regions. 16 dimensions strike a balance between generating diversity and sampling effectiveness.

[0024] The specific execution process of reparameterized sampling is as follows: First, the exponentiation of each dimension of the latent log-variance vector is taken to obtain the variance vector. Then, the square root of the variance vector is taken to obtain the standard deviation vector. Next, a 16-dimensional noise vector is independently sampled from the standard normal distribution. Finally, the noise vector and the standard deviation vector are multiplied element-wise along each dimension, and then added element-wise along each dimension to the latent mean vector to obtain the 16-dimensional latent variable. The reparameterization operation transforms the random sampling process into a deterministic transformation of the noise vector, allowing the gradient to propagate back from the latent variable to the latent mean vector and the latent log-variance vector, thereby synchronously optimizing the encoder parameters through gradient descent during training.

[0025] In one specific embodiment, the decoder of the conditional variational autoencoder in step S3 takes the concatenation of latent variables and phenotypic conditional vectors as input and optimizes it using the sum of reconstruction loss and KL divergence as the loss function, including: The latent variables and phenotypic conditional vectors are concatenated along their dimensions to obtain the decoder input vector. The decoder input vector is input into a decoder consisting of a three-layer fully connected network, with ReLU activation function in each layer and linear activation function in the output layer. The decoder input vector is then processed by recipe space mapping to obtain the reconstructed recipe vector. Based on the reconstruction loss between the reconstructed recipe vector and the original standardized recipe vector in the phenotype-recipe pairing dataset, and the KL divergence between the latent mean vector and the latent log-variance vector, the reconstruction loss and KL divergence are weighted and summed to obtain the training loss value. The training loss value is then used as the target to optimize the parameters of the conditional variational autoencoder, resulting in the trained conditional variational autoencoder.

[0026] Specifically, the latent variable dimension is 16, and the phenotypic conditional vector dimension is 39. These are concatenated end-to-end according to their dimensions to obtain a 55-dimensional decoder input vector. The decoder consists of a three-layer fully connected network with 64, 128, and 256 nodes respectively. In the opposite direction to the encoder's compression, the 55-dimensional decoder input vector is progressively expanded to a 256-dimensional hidden layer representation. Nonlinearity is introduced between hidden layers using the ReLU activation function. The output layer has 28 nodes, consistent with the dimension of the normalized formulation vector. The output layer uses a linear activation function instead of ReLU because the normalized formulation vector, after Z-score processing, covers the negative value range for each dimension. Linear activation does not impose truncation constraints on the output value range, ensuring that the decoder output covers the complete formulation concentration range, resulting in a 28-dimensional reconstructed formulation vector.

[0027] The reconstruction loss is calculated as follows: the reconstructed recipe vector is subtracted element-wise from the corresponding original standardized recipe vector in the phenotypic-recipe pairing dataset along each dimension, and the squared result is calculated. The mean value over all 28 dimensions is then obtained to obtain the reconstruction loss value in the form of mean squared error. This calculation method applies an equal-weight penalty to the reconstruction bias of each component dimension. The KL divergence is calculated based on the latent mean vector and the latent log-variance vector. The KL divergence between the standard normal distribution and the encoder output distribution is calculated dimension-wise for each of the 16 dimensions of the latent space. The calculation method is to sum the log-variance, mean square, and variance of each dimension according to a fixed algebraic relationship and then take the mean value over all 16 dimensions to obtain the KL divergence value. This constraint makes the distribution of the latent space converge towards the standard normal distribution, ensuring that the latent variables sampled from the standard normal distribution during inference fall within the effective response range of the decoder. The training loss is obtained by weighted summation of the reconstruction loss and the KL divergence. The weighting coefficients are set to 1 for the reconstruction loss and 0.5 for the KL divergence. The reason for setting the KL divergence weight to less than 1 is that if the KL divergence weight is too large in the early stages of training, the encoder tends to compress all inputs to near the mean of the same latent distribution, causing the decoder to be unable to distinguish different recipes. Appropriately reducing the KL divergence weight allows the encoder to retain sufficient recipe discrimination information in the early training. Parameter optimization is performed using the Adam optimizer, with a learning rate of 0.001, momentum parameters of 0.9 and 0.999, a batch size of 32, and 300 training epochs. The reconstruction loss is calculated on the validation set every 10 epochs, and the encoder and decoder weights corresponding to the minimum reconstruction loss on the validation set are saved as the conditional variational autoencoder after training.

[0028] Figure 2 This is a schematic diagram illustrating the loss convergence during the training process of the conditional variational autoencoder in an embodiment of this application. Figure 2As shown, the horizontal axis represents the number of training epochs, and the vertical axis represents the training loss value obtained by weighted summation of reconstruction loss and KL divergence. The solid line represents the training loss, and the dashed line represents the validation loss. Both curves show a monotonically decreasing trend with the increase of training epochs and tend to stabilize around the 180th epoch. The validation loss reaches its lowest point at the 180th epoch. The corresponding saved encoder and decoder weights are the conditional variational autoencoder after training. The difference between the training loss and the validation loss always remains within a reasonable range, indicating that the conditional variational autoencoder has not overfitted on the training data.

[0029] In one specific embodiment, during inference in step S3, the phenotypic condition vector corresponding to the phenotypic target vector is used as a condition. Multiple sets of latent variables are sampled from the standard normal distribution and decoded by a decoder to obtain a set of candidate culture medium formulations, including: The phenotypic target vector and the cell type embedding vector are concatenated according to their dimensions to obtain the phenotypic condition vector for inference. Based on the phenotypic conditional vector for inference, multiple sets of latent variables are independently sampled from the standard normal distribution. Each set of latent variables and the phenotypic conditional vector for inference are concatenated by dimension and then input into the decoder of the trained conditional variational autoencoder. The concatenated vectors are decoded to obtain multiple sets of reconstructed formula vectors. Multiple sets of reconstructed formulation vectors are denormalized to restore each set of reconstructed formulation vectors to the true concentration domain formulations, thus obtaining a set of candidate culture medium formulations.

[0030] Specifically, the phenotypic condition vector construction method in the inference phase is completely consistent with that in the training phase: the seven-dimensional phenotypic target vector and the 32-dimensional cell type embedding vector are concatenated end-to-end according to their dimensions to obtain a 39-dimensional phenotypic condition vector for inference. The dimension of this vector is the same as that of the phenotypic condition vectors input to the encoder and decoder in the training phase, ensuring that the decoder of the trained conditional variational autoencoder receives conditional input with a consistent structure during inference. 100 sets of 16-dimensional noise vectors are independently sampled from the standard normal distribution as latent variables. The reason for setting the number of sampling sets to 100 is that: if the number of sets is too small, the scope of the candidate formulation set covering the latent space will be limited, and the deviation between the candidate scheme with the smallest phenotypic matching error and the true optimal formulation will be large; if the number of sets is too large, the computational cost will increase significantly, which is not practically feasible in laboratory batch verification scenarios. 100 sets achieve a balance between candidate space coverage and computational efficiency. Each group of 16 latent variables and 39-dimensional inference phenotypic conditional vectors are concatenated end-to-end according to their dimensions to obtain a 55-dimensional decoder input vector. This vector is then input into the decoder of the trained conditional variational autoencoder and mapped through a three-layer fully connected network. Each group outputs a 28-dimensional reconstructed formula vector, resulting in a total of 100 reconstructed formula vectors.

[0031] When performing Z-score standardization on the concentration of each component, the mean and standard deviation of each component in all valid batches were recorded, totaling 28 pairs of values. During the inference phase, the j-th dimension value of each reconstructed formulation vector was multiplied by the standard deviation of the j-th component, and the mean of the j-th component was added. The above reduction calculation was performed dimension by dimension to restore the 28-dimensional standardized values ​​to the true molar concentration or cytokine activity unit value of the corresponding component. After performing the above de-standardization operation on all 100 reconstructed formulation vectors, 100 culture medium formulations represented by the true concentration domain were obtained, forming a candidate culture medium formulation set. Each formulation contains the true concentration values ​​of 28 components, including amino acids, vitamins, inorganic salts, and cytokines.

[0032] Figure 3 This diagram illustrates the distribution of cytokine concentrations in the recommended culture medium formulations for the three culture stages in this application embodiment. Figure 3 As shown in the figure, the horizontal axis represents the types of cytokines, and the vertical axis represents the concentration of cytokines. The oblique-lined bars, gray-filled bars, and cross-filled bars correspond to the recommended concentration values ​​for the activation phase, the expansion phase, and the harvest preparation phase, respectively. As can be seen from the figure, the IL-2 concentration decreases from 1200 IU / mL in the activation phase to 80 IU / mL in the harvest preparation phase, while the concentrations of IL-7 and IL-15 increase in stages. The above concentration change pattern is consistent with the guiding direction of the phenotypic sub-targets in each stage, that is, the activation phase drives rapid proliferation with high concentrations of IL-2, while the expansion phase and the harvest preparation phase gradually increase IL-7 and IL-15 to guide the retention of the stem memory phenotype.

[0033] In one specific embodiment, step S4, inputting the candidate culture medium formulation set into the positive validator, includes: The candidate culture medium formulation set and the cell type embedding vector are concatenated according to their dimensions to obtain the validator input vector. The validator input vector is input into the forward validator, and phenotypic prediction processing is performed on the validator input vector to obtain the predicted phenotypic vector corresponding to each candidate formulation.

[0034] Specifically, each candidate culture medium formulation in the candidate formulation set is denormalized into a 28-dimensional true concentration domain vector. During the inference stage, it needs to be renormalized into a 28-dimensional normalized formulation vector. The mean and standard deviation used for normalization are consistent with the component statistics recorded in step S2, ensuring that the formulation vector input to the positive validator is on the same numerical scale as the formulation vector used when training the positive validator. The 28-dimensional normalized formulation vector and the 32-dimensional cell type embedding vector are concatenated end-to-end according to their dimensions to obtain a 60-dimensional validator input vector. The positive validator consists of a four-layer fully connected network. The input layer receives the 60-dimensional validator input vector, and the number of nodes in the three hidden layers are 256, 128, and 64 respectively. The activation function of each hidden layer is ReLU. The number of nodes in the output layer is 7, consistent with the dimension of the phenotypic target vector. The activation function of the output layer is Sigmoid, which constrains the output value to the interval between 0 and 1, corresponding to the proportional value range of the seven phenotypic indicators.

[0035] The forward validator uses the phenotype-recipe pairing dataset constructed in step S2 as training data, takes the concatenation of standardized recipe vectors and cell type embedding vectors as input, and uses the measured phenotype vectors at the end of each stage as labels. Supervised training is performed to minimize the mean squared error between the predicted phenotype vectors and the measured phenotype vectors. The training hyperparameters are consistent with those of the conditional variational autoencoder, with a learning rate set to 0.001 and a batch size set to 32. Training continues until the mean absolute error of each dimension of the seven phenotypic indicators on the validation set is less than 0.03, meaning the prediction bias of each phenotypic indicator does not exceed 3 percentage points. This threshold is set based on the batch-to-batch repeatability error range of flow cytometry detection itself, ensuring that the prediction accuracy of the forward validator is not lower than the experimental detection accuracy. After training, the weights of the forward validator are fixed, and forward inference is performed one by one on each of the 100 candidate recipes in the candidate culture medium recipe set. Each candidate recipe outputs a 7-dimensional predicted phenotype vector through the forward validator, resulting in a total of 100 sets of predicted phenotype vectors.

[0036] In one specific embodiment, the forward validator consists of a four-layer fully connected network. The input layer receives the validator input vector, the activation function of each hidden layer is ReLU, and the activation function of the output layer is Sigmoid. The validator input vector is mapped to obtain the predicted phenotypic vector corresponding to each candidate formulation.

[0037] Based on the preset safe concentration range of each component, the concentration of each component in each candidate formulation in the candidate culture medium formulation set is subjected to feasibility filtering. Candidate formulations with any component concentration exceeding the safe concentration range are eliminated, resulting in a set of feasible candidate formulations.

[0038] Based on the predicted phenotypic vector and the phenotypic target vector, the Euclidean distance between the predicted phenotypic vector and the phenotypic target vector is calculated for each candidate formulation in the feasible candidate formulation set to obtain the phenotypic matching error value of each candidate formulation. The candidate formulation with the smallest phenotypic matching error value is then de-standardized to obtain the recommended culture medium formulation.

[0039] Specifically, the safe concentration ranges for each component were set based on published cytotoxicity experimental data and GMP specifications. Taking amino acid components as an example, the safe concentration range for L-glutamine was set at 1.0 to 8.0 mmol / L. Exceeding the upper limit would cause cytotoxicity due to ammonia accumulation, while below the lower limit would be insufficient to maintain the nitrogen source required for T cell proliferation. For cytokine components, the safe concentration range for IL-2 was set at 10 to 2000 IU / mL, IL-7 at 1 to 100 ng / mL, and IL-15 at 1 to 100 ng / mL. One by one, the 100 candidate formulations in the candidate culture medium formulation set were examined. Candidate formulations whose concentrations of any of the 28 components exceeded the upper limit or fell below the lower limit of their corresponding safe concentration range were completely eliminated. The remaining candidate formulations constituted the feasible candidate formulation set.

[0040] The Euclidean distance is calculated as follows: For each candidate formulation in the feasible candidate formulation set, the 7-dimensional predicted phenotypic vector output by the positive validator is taken, and the difference between it and the 7-dimensional phenotypic target vector is calculated element-wise along each dimension, and the square is then calculated. The sum of the squared differences of the 7 dimensions is then taken as the square root to obtain the phenotypic matching error value of the candidate formulation. The smaller this value, the closer the cell phenotype generated by the candidate formulation under the prediction of the positive validator is to the phenotypic target input by the manufacturer. The phenotypic matching error values ​​of all candidate formulations in the feasible candidate formulation set are compared, and the candidate formulation with the smallest phenotypic matching error value is selected. Its 28-dimensional standardized formulation vector is then de-standardized according to the mean and standard deviation of each component recorded in step S2. The standardized value of each dimension is multiplied by the standard deviation of the corresponding component and then added to the mean of the corresponding component to restore the true molar concentration or cytokine activity unit value of the 28 components, thus obtaining the recommended culture medium formulation. This formulation is directly output in the form of the true concentration of each component for the manufacturer to use in actual culture operations.

[0041] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting serum-free culture medium components based on artificial intelligence, characterized in that, The method includes: Step S1: Based on the cell type identifier and cell phenotype specification parameters, construct the phenotype target vector and the cell type embedding vector, and concatenate the phenotype target vector and the cell type embedding vector to obtain the phenotype condition vector. Step S2: Divide the historical culture batch data into culture stages, and construct a phenotype-formula pairing dataset using the measured phenotypic vectors at the end of each stage as labels. Step S3: Using the phenotypic condition vector as a conditional constraint, the phenotypic-formula pairing dataset is input into a conditional variational autoencoder for training. The encoder of the conditional variational autoencoder takes the concatenation of the formulation vector and the phenotypic condition vector in the phenotypic-formula pairing dataset as input and outputs the mean and log-variance of the latent variables. The latent variables are obtained through reparameterized sampling. The decoder of the conditional variational autoencoder takes the concatenation of the latent variables and the phenotypic condition vector as input and optimizes the loss function by using the sum of reconstruction loss and KL divergence. During inference, the phenotypic condition vector corresponding to the phenotypic target vector is used as a condition. Multiple sets of latent variables are sampled from the standard normal distribution and decoded by the decoder to obtain a set of candidate culture medium formulations. Step S4: Input the set of candidate culture medium formulations into the positive validator, and filter by the minimum Euclidean distance between the predicted phenotypic vector of each candidate formulation and the target phenotypic vector to obtain the recommended culture medium formulation.

2. The artificial intelligence-based serum-free medium component prediction method of claim 1, wherein, Step S1 includes: The cell type identifier is input into the pre-trained embedding layer, and the cell type identifier is subjected to embedding mapping processing to obtain the cell type embedding vector. Seven cell phenotypic parameters detected by flow cytometry, namely CCR7 positive rate, CD62L positive rate, CD45RA positive rate, PD-1 single positive rate, PD-1 and TIM-3 double positive rate, CD4 to CD8 ratio and JC-1 staining positive rate, were vectorized and encoded to obtain the phenotypic target vector. The phenotypic target vector and the cell type embedding vector are concatenated by dimension to obtain the phenotypic condition vector. 3.The AI-based serum-free medium component prediction method of claim 1, wherein, Step S2 includes: Historical culture batch data were segmented into activation stage, amplification stage, and harvest preparation stage to obtain culture medium component records and flow cytometry detection data at the end of each stage. Based on the flow cytometry data at the end of the aforementioned stages, seven indicators—CCR7 positive rate, CD62L positive rate, CD45RA positive rate, PD-1 single positive rate, PD-1 and TIM-3 double positive rate, CD4 to CD8 ratio, and JC-1 staining positive rate—were vectorized and encoded to obtain the measured phenotypic vectors at the end of each stage. Based on the cell viability threshold and Grubbs test, abnormal batch removal processing was performed on the measured phenotypic vectors at the end of each stage to obtain the filtered effective stage sample set. The culture medium component records of each stage in the effective stage sample set are paired with the corresponding measured phenotypic vectors at the end of each stage to obtain a phenotypic-formula paired dataset. 4.The AI-based serum-free medium component prediction method of claim 1, wherein, In step S3, the phenotypic-formula pairing dataset is input into a conditional variational autoencoder for training. The encoder of the conditional variational autoencoder takes the concatenation of the formula vector and the phenotypic condition vector in the phenotypic-formula pairing dataset as input and outputs the mean and log-variance of the latent variables. The latent variables are obtained through reparameterized sampling, including: The standardized recipe vector in the phenotype-recipe pairing dataset is concatenated with the phenotype condition vector according to dimension to obtain the encoder input vector; The encoder input vector is input into an encoder consisting of a three-layer fully connected network, with ReLU activation function in each layer. Feature compression processing is performed on the encoder input vector to obtain the latent mean vector and the latent log-variance vector. Based on the latent mean vector and the latent log-variance vector, the latent mean vector and the latent log-variance vector are reparameterized and sampled to obtain latent variables.

5. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 4, characterized in that, The decoder of the conditional variational autoencoder in step S3 takes the concatenation of the latent variables and the phenotypic conditional vector as input and optimizes it using the sum of the reconstruction loss and the KL divergence as the loss function, including: The latent variables and the phenotypic condition vectors are concatenated by dimension to obtain the decoder input vector. The decoder input vector is input into a decoder consisting of a three-layer fully connected network, with ReLU activation function in each layer and linear activation function in the output layer. The decoder input vector is then processed by recipe space mapping to obtain a reconstructed recipe vector. Based on the reconstruction loss between the reconstructed recipe vector and the original standardized recipe vector in the phenotype-recipe pairing dataset, and the KL divergence between the latent mean vector and the latent log-variance vector, the reconstruction loss and the KL divergence are weighted and summed to obtain the training loss value. The conditional variational autoencoder is then optimized with the training loss value as the target to obtain the trained conditional variational autoencoder.

6. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 5, characterized in that, In step S3, during inference, the phenotypic condition vector corresponding to the phenotypic target vector is used as a condition. Multiple sets of latent variables are sampled from a standard normal distribution and decoded by the decoder to obtain a set of candidate culture medium formulations, including: The phenotypic target vector and the cell type embedding vector are concatenated by dimension to obtain the phenotypic condition vector for inference. Based on the inference phenotypic conditional vector, multiple sets of latent variables are independently sampled from the standard normal distribution. Each set of latent variables is concatenated with the inference phenotypic conditional vector according to the dimension and then input into the decoder of the trained conditional variational autoencoder. Each concatenated vector is decoded to obtain multiple sets of reconstructed formula vectors. The reconstructed formulation vectors are denormalized to restore each set of reconstructed formulation vectors to the true concentration range formulations, thus obtaining a set of candidate culture medium formulations.

7. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 1, characterized in that, Step S4 involves inputting the candidate culture medium formulation set into the positive validator, including: The verifier input vector is obtained by concatenating each candidate formulation in the candidate culture medium formulation set with the cell type embedding vector according to the dimension. The validator input vector is input into the forward validator, and phenotypic prediction processing is performed on the validator input vector to obtain the predicted phenotypic vector corresponding to each candidate formulation.

8. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 7, characterized in that, The positive validator consists of a four-layer fully connected network. The input layer receives the input vector of the validator, the activation function of each hidden layer is ReLU, and the activation function of the output layer is Sigmoid. The input vector of the validator is mapped to obtain the predicted phenotypic vector corresponding to each candidate formulation.

9. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 8, characterized in that, Based on the preset safe concentration range of each component, the concentration of each component in each candidate formulation in the candidate culture medium formulation set is subjected to feasibility filtering, and candidate formulations with any component concentration exceeding the safe concentration range are eliminated to obtain a set of feasible candidate formulations.

10. The method for predicting serum-free culture medium components based on artificial intelligence according to claim 9, characterized in that, Based on the predicted phenotypic vector and the phenotypic target vector, the Euclidean distance between the predicted phenotypic vector and the phenotypic target vector is calculated for each candidate formulation in the feasible candidate formulation set to obtain the phenotypic matching error value of each candidate formulation. The candidate formulation with the smallest phenotypic matching error value is then de-standardized to obtain the recommended culture medium formulation.

Images