Method and device for predicting particle size of mRNA lipid nanoparticle vaccine

Abstract

The application discloses a kind of mRNA lipid nano particle vaccine particle size prediction method and device, it is related to particle size prediction field, after obtaining mRNA lipid nano particle prescription, extract the particle size characteristics therein, respectively input into second preset machine learning model and multiple first preset machine learning model, obtain the particle size prediction result of this prescription and each first preset machine learning model for the particle size influence value of particle size characteristics, finally output multiple particle size influence values and particle size prediction result to be referenced by staff, these machine learning models are different types of models.Determine particle size influence value can let staff simply determine the influence degree of each particle size characteristic on particle size;Particle size can be simply predicted by machine learning model prediction, and this is used as a reference to optimize the prescription, and a reasonable prescription can be simply obtained, without staff manually determining the influence degree of particle size characteristic and predicting particle size.

Description

Technical Field

[0001] This invention relates to the field of particle size prediction, and in particular to a method and apparatus for predicting the particle size of mRNA lipid nanoparticle vaccines. Background Technology

[0002] Liposomes (LNPs) are a common delivery system for mRNA lipid nanoparticle vaccines. When designing mRNA-LNP formulations, the design and optimization of both liposomes and the formulation depend primarily on the selection and proportioning of materials. Among these factors, some also influence the particle size of the mRNA-LNP vaccine. Since the particle size is a crucial indicator of the quality and stability of the developed product, it is necessary to predict the particle size based on the actual designed formulation to determine its rationality. Current techniques for predicting mRNA-LNP vaccine particle size typically rely on the experience of the personnel involved. The formulation is then continuously refined through experimentation based on the predicted particle size, ultimately yielding a more suitable formulation. However, this method requires a high level of experience from staff. For staff with less experience, it may be unclear which formulations and factors affect the particle size of mRNA lipid nanoparticle vaccines, making it impossible to accurately predict the particle size of mRNA lipid nanoparticle vaccines. In the process of formulation research and optimization, there is a lack of targets and references, making it difficult to obtain reasonable formulations. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for predicting the particle size of mRNA lipid nanoparticle vaccines, so that the formulation can be improved based on the target and reference, and a reasonable formulation can be obtained more easily without the need for staff to determine the degree of influence of particle size characteristics and predict the particle size themselves.

[0004] To address the aforementioned technical problems, this invention provides a method for predicting the particle size of mRNA lipid nanoparticle vaccines, comprising:

[0005] Obtain mRNA lipid nanoparticle formulation;

[0006] The particle size characteristics of the mRNA lipid nanoparticle formulation were extracted;

[0007] The particle size features are respectively input into n first preset machine learning models to obtain n particle size influence values ​​of the particle size features. The n first preset machine learning models are all pre-trained with a preset number of the particle size features, and n is a positive integer.

[0008] The particle size features are input into a second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation. The second preset machine learning model is trained in advance by a preset number of the particle size features.

[0009] Output n particle size influence values ​​and the particle size prediction results;

[0010] When n is greater than 2, the n first preset machine learning models are all machine learning models of different types.

[0011] Preferably, when the number of extracted particle size features is not less than 2, the particle size features are respectively input into n first preset machine learning models, including:

[0012] Each of the aforementioned particle size features is converted into parameter features of the same order of magnitude as the preset order of magnitude;

[0013] All the aforementioned parameter features are input into n first preset machine learning models respectively.

[0014] Preferably, after extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes:

[0015] Determine whether the number of extracted particle size features is greater than a first preset number;

[0016] If the number is greater than the first preset number, then proceed to the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset number;

[0017] If the number is not greater than the first preset number, determine whether the number of extracted particle size features is greater than the second preset number.

[0018] If the number is not greater than the second preset number, then the missing particle size feature in the current mRNA lipid nanoparticle formulation is determined, and the missing particle size feature in the current mRNA lipid nanoparticle formulation is filled based on the corresponding particle size feature in other mRNA lipid nanoparticle formulations, and then the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset order of magnitude is entered.

[0019] Wherein, the first preset quantity is greater than the second preset quantity.

[0020] Preferably, after determining whether the number of extracted particle size features is greater than a second preset number, the method further includes:

[0021] If the number is greater than the second preset number, the calculation method of the first preset machine learning model is set to a calculation method that does not consider the particle size features missing in the current mRNA lipid nanoparticle formulation.

[0022] Preferably, converting each of the particle size features into parameter features of the same order of magnitude as a preset order of magnitude includes:

[0023] When the particle size feature is a numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude, based on the position of the value of the particle size feature within the preset range corresponding to the particle size feature.

[0024] When the particle size feature is a non-numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude using a preset conversion method corresponding to the particle size feature.

[0025] Preferably, the first preset machine learning model is either a tree model or a LASSO model with L1 penalty.

[0026] Preferably, the second preset machine learning model is a Gaussian process model. The particle size features are input into the second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation, including:

[0027] In the Gaussian process model, the distance between the particle size feature and each training set is determined;

[0028] The particle size prediction results of the mRNA lipid nanoparticle formulation are determined based on each of the aforementioned distances.

[0029] Preferably, after extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes:

[0030] The particle size features are input into a third preset machine learning model to determine the particle size distribution value of the mRNA lipid nanoparticle formulation. The third preset machine learning model is pre-trained from a preset number of the particle size features.

[0031] Output n particle size influence values ​​and the particle size prediction results, including:

[0032] Output n values ​​representing the particle size influence, the particle size prediction result, and the particle size distribution value.

[0033] Preferably, when two mRNA lipid nanoparticle formulations are obtained simultaneously, after inputting the particle size features into a second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation, the method further includes:

[0034] In the second preset machine learning model, the size prediction results of the two mRNA lipid nanoparticle formulations are compared, and the comparison results are output.

[0035] This application also provides a device for predicting the particle size of mRNA lipid nanoparticle vaccines, comprising:

[0036] Memory, used to store computer programs;

[0037] A controller is used to execute the computer program to implement the steps of the method for predicting the particle size of mRNA lipid nanoparticle vaccines as described above.

[0038] This application provides a method and apparatus for predicting the particle size of mRNA lipid nanoparticle vaccines. After obtaining the mRNA lipid nanoparticle formulation, the particle size features in the formulation are extracted. These particle size features are then input into multiple first and second preset machine learning models to obtain the particle size influence value of each first preset machine learning model on the particle size features, and the particle size prediction result of the second preset machine learning model for the formulation. These machine learning models are all pre-trained with a preset number of particle size features. Finally, multiple particle size influence values ​​and particle size prediction results are output. These first preset machine learning models are of different types. By determining the influence value of each particle size feature on the particle size through multiple different machine learning models, researchers can easily determine the degree of influence of each particle size feature in the formulation on the particle size of the mRNA lipid nanoparticle vaccine, identify important features in the formulation that affect the particle size, and accurately predict the particle size through machine learning models. Using these parameters as targets and references, researchers can easily obtain reasonable formulations when researching and optimizing formulations without having to determine the degree of influence of particle size features or predict the particle size themselves. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the prior art and embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart illustrating a method for predicting the particle size of an mRNA lipid nanoparticle vaccine provided in this application;

[0041] Figure 2 A schematic diagram illustrating the number of times the branching feature is selected in a tree model provided in this application;

[0042] Figure 3 This application provides a distribution map of the matching results of prescription comparisons;

[0043] Figure 4A distribution map of particle size distribution for a training set provided in this application;

[0044] Figure 5 A distribution map of the proportion of ionizable lipids in a training set provided for this application;

[0045] Figure 6 A distribution diagram of phospholipid ratio in a training set provided for this application;

[0046] Figure 7 A distribution map of cholesterol lipids in a training set provided for this application;

[0047] Figure 8 A distribution map of PEG lipids in a training set provided for this application;

[0048] Figure 9 This application provides a distribution map showing the matching of particle size distribution comparison results;

[0049] Figure 10 This is a schematic diagram of the structure of a device for predicting the particle size of an mRNA lipid nanoparticle vaccine provided in this application. Detailed Implementation

[0050] The core of this invention is to provide a method and apparatus for predicting the particle size of mRNA lipid nanoparticle vaccines, which allows for formulation improvement based on targets and references, and makes it easier to obtain a reasonable formulation without requiring personnel to determine the degree of influence of particle size characteristics or predict the particle size themselves.

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The development of nucleic acid drugs faces three main challenges: the instability of nucleic acid molecules in vivo; the potential side effects of nucleic acid molecules; and the difficulty in developing nucleic acid drug delivery systems. Among these, the development of delivery systems is crucial. After injection, nucleic acid drugs must remain in the body long enough to accurately target the lesion site while avoiding damage to normal cells. This necessitates the development of highly efficient and safe drug delivery systems to address challenges related to nucleic acid drug delivery, stability, and off-target effects.

[0053] Lipid nanoparticles (LNPs) are one of the most widely used delivery systems in nucleic acid drug research, as they can deliver nucleic acids safely and efficiently. Nucleic acid drugs using LNPs have already been approved.

[0054] Therefore, to improve drug utilization, liposomes are commonly used as a vaccine delivery system. Liposomes are lipid vesicles formed by phospholipids linked together to create a double membrane. Regardless of their size or structure, liposomes can encapsulate other substances. Liposomes protect their internal contents from contact with human enzymes or acids, preventing their influence. Furthermore, because the structure of liposomes is identical to that of human cell membranes, they can directly fuse with cell membranes to deliver their contents directly into the bloodstream, thus significantly improving drug utilization.

[0055] However, when using liposomes to transport mRNA lipid nanoparticle vaccines, the particle size is affected by the structure and composition of the liposomes themselves. Therefore, when designing liposomes for transporting mRNA lipid nanoparticle vaccines (i.e., designing mRNA-LNP formulations), it is necessary to consider whether the final designed liposomes will affect the particle size and efficacy of the mRNA lipid nanoparticle vaccine. Current technologies typically rely on the experience of the personnel involved, which demands a high level of technical expertise and is time-consuming. Furthermore, since the particle size of mRNA lipid nanoparticle vaccines affects vaccine efficacy and product stability, it is also necessary to predict the particle size of mRNA lipid nanoparticle vaccines to aid in the design of mRNA lipid nanoparticle formulations. This also currently relies heavily on the experience of the personnel involved, requiring a high level of technical expertise.

[0056] To solve the above technical problems, please refer to Figure 1 , Figure 1 A flowchart of a method for predicting the particle size of mRNA lipid nanoparticle vaccines provided in this application includes:

[0057] S1: Obtain the mRNA lipid nanoparticle formulation;

[0058] S2: Extracting particle size characteristics from mRNA lipid nanoparticle formulations;

[0059] S3: Input the particle size features into n first preset machine learning models respectively, so as to obtain n particle size influence values ​​of the particle size features respectively. The n first preset machine learning models are all pre-trained from a preset number of particle size features, and n is a positive integer.

[0060] S4: Input the particle size features into the second preset machine learning model in order to determine the particle size prediction result of the mRNA lipid nanoparticle formulation. The second preset machine learning model is pre-trained from a preset number of particle size features.

[0061] S5: Output n particle size influence values ​​and particle size prediction results;

[0062] When n is greater than 2, the n first preset machine learning models are all different types of machine learning models.

[0063] When staff design a new mRNA lipid nanoparticle formulation (hereinafter referred to as the formulation), since there are some aspects of the initial formulation that need to be corrected and improved, a formulation usually needs to be optimized several times to obtain a reasonable and usable formulation. In order to reduce the technical experience threshold for staff and reduce their manual workload, machine learning models can be used to replace manual work in formula verification, specifically to verify whether the particle size of the designed / improved formulation is reasonable.

[0064] Specifically, when training the machine learning model, existing formulations can be used as training and testing sets. The model's accuracy is improved by utilizing the data features described in these formulations that are necessary for practical applications. These data features can be broadly categorized into three types: characteristics of the mRNA lipid nanoparticle vaccine itself, liposome composition, and preparation process parameters. Please refer to Table 1, which shows the particle size characteristics of existing formulations. Table 1 specifically lists 299 formulations from 39 studies. These 299 formulations describe 21 variables (particle size characteristics), and the final particle size of the mRNA lipid nanoparticles from these formulations is used as a reference for training. This accurately determines the influence of each particle size characteristic on the particle size, allowing researchers to identify which features significantly affect the particle size of the formulation. This information is then used to predict the particle size value of the formulation. When using these formulations to train the machine learning model, they can be divided into training and testing sets in a certain ratio, such as an 8:2 ratio.

[0065]

[0066] Table 1: Particle size characteristics in existing formulations

[0067] In practical applications, when a staff member designs a new prescription, the aforementioned data features are extracted and input into a first machine learning model and a second machine learning model. The first machine learning model identifies the key features influencing the particle size of the prescription, while the second machine learning model calculates the distance between the new prescription and previously collected prescription samples. Samples that are closer or more similar contribute more to the prescription prediction, thus predicting the particle size. For prescription improvement, staff members can make different adjustments to the prescription based on the key features influencing the particle size identified by the first machine learning model, making it more reasonable. Based on the above, when designing and improving prescriptions, staff members can adjust the prescription based on the difference between the current particle size prediction and the ideal particle size, and based on the influence of each particle size feature, thereby making the particle size more reasonable.

[0068] Furthermore, since different models may output different results when determining the degree of influence of particle size features on particle size, there may be situations where different models believe that the same particle size feature has different degrees of influence on particle size. Therefore, in order to accurately determine the actual degree of influence of each particle size feature on particle size, multiple machine learning models with different structures can be set up, and the actual degree of influence of each particle size feature on particle size can be comprehensively considered based on the output results of multiple machine learning models.

[0069] In summary, after obtaining the mRNA lipid nanoparticle formulation, the particle size characteristics of the formulation are extracted. These characteristics are then input into multiple first and second pre-defined machine learning models to obtain the particle size influence values ​​of each first-defined machine learning model on the particle size characteristics, and the particle size prediction results of the second-defined machine learning model for the formulation. These machine learning models are all pre-trained with a predetermined number of particle size characteristics. Finally, multiple particle size influence values ​​and particle size prediction results are output. These first-defined machine learning models are of different types. By determining the influence values ​​of each particle size characteristic on the particle size through multiple different machine learning models, researchers can easily determine the degree of influence of each particle size characteristic in the formulation on the particle size of the mRNA lipid nanoparticle vaccine, identify important characteristics affecting particle size in the formulation, and accurately predict particle size using machine learning models. Using these parameters as targets and references, researchers can easily obtain a reasonable formulation when improving it, without needing to determine the degree of influence of particle size characteristics or predict particle size themselves.

[0070] Based on the above embodiments:

[0071] In a preferred embodiment, when the number of extracted particle size features is not less than 2, the particle size features are respectively input into n first preset machine learning models, including:

[0072] Each particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude;

[0073] All parameters and features are input into n pre-defined machine learning models.

[0074] To accurately determine the influence of each particle size feature on particle size, this application typically extracts multiple particle size features when extracting them from the formulation. However, considering the differences in orientation, representation, and order of magnitude of each feature, directly inputting these features into the model might lead to inaccurate judgments due to unit differences. Therefore, before inputting the particle size features into the model, they need to be standardized. This involves converting each feature, which originally differed in order of magnitude and representation, into a single value within the same range. The maximum and minimum values ​​of this range can be set according to the specific application, for example, a range of 0 to 1. By standardizing the particle size features, the influence caused by feature differences is eliminated, achieving the goal of accurately determining the influence of each particle size feature on particle size.

[0075] As a preferred embodiment, after extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes:

[0076] Determine whether the number of extracted particle size features is greater than the first preset number;

[0077] If the number is greater than the first preset number, proceed to the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset number;

[0078] If the number is not greater than the first preset number, determine whether the number of extracted particle size features is greater than the second preset number.

[0079] If the number is not greater than the second preset number, the missing particle size feature in the current mRNA lipid nanoparticle formulation is determined, and the missing particle size feature in the current mRNA lipid nanoparticle formulation is filled based on the corresponding particle size feature in other mRNA lipid nanoparticle formulations. Then, the process proceeds to the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset number of magnitude.

[0080] The first preset quantity is greater than the second preset quantity.

[0081] To more accurately determine the influence of each particle size feature on particle size, this application considers that the types and numbers of features described in different formulations may vary. However, the trained model will calculate the formulation based on all types of features when put into use. If the formulation only contains some features, it may lead to errors in the model's output. For example, the particle size features described in formulation A include q / w, while those in formulation B include w / e / r. It is evident that the types and numbers of features in these two formulations are different. When using the machine learning model, it will calculate based on features such as q / w / e / r. For both formulations A and B, some features are missing, which may lead to certain errors in the model's calculations for both formulations. Therefore, before sending particle size features from a prescription to a machine learning model, it is first necessary to determine the number of missing features in the prescription based on the types and quantities of features covered by the model. If the number of missing features is small and will not produce significant errors, or even if there are no missing features, they can be directly input into the machine learning model. If the number of missing features is large but within an acceptable range, to avoid the missing features affecting the model's calculation results, imputation values ​​for these missing features can be calculated based on previously collected samples. For example, when many particle size features are missing in this prescription but the number is within an acceptable range, for each missing particle size feature, since the name and code of the particle size feature itself are known, the mode of that particle size feature in other prescriptions that describe it can be used as the imputation value. This is because the mode indicates that the particle size feature is frequently chosen as this value, indirectly indicating that the particle size feature at this value has little impact on the particle size. Based on this, by imputing the missing features in the prescription, the degree of influence of each particle size feature on the particle size can be further accurately determined.

[0082] In a preferred embodiment, after determining whether the number of extracted particle size features is greater than a second preset number, the method further includes:

[0083] If the number exceeds the second preset quantity, the calculation method of the first preset machine learning model will be set to a calculation method that does not consider the particle size characteristics missing in the current mRNA lipid nanoparticle formulation.

[0084] To reduce errors, in this application, if the number and variety of features described in the input prescription are too small, resulting in too many missing features exceeding the acceptable range, the machine learning model cannot obtain a relatively accurate particle size prediction result and the influence value of each particle size feature on the particle size based on the previously trained results. Therefore, to reduce model calculation errors, when it is determined that a prescription is missing too many particle size features, the model can be set to only consider the existing particle size features in the prescription to calculate the particle size prediction result and influence value of the prescription. For example, the model would normally calculate a prescription based on four features: q / w / e / r. However, if the input prescription only includes q / w, the missing features exceed the acceptable range. In this case, the model is set to only consider the q / w features and ignore the e / r features to calculate the prescription, thereby reducing errors.

[0085] As a preferred embodiment, converting each particle size characteristic into parameter characteristics of the same order of magnitude as a preset order of magnitude includes:

[0086] When the particle size feature is a numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude, based on the position of the particle size feature value itself within the preset range corresponding to the particle size feature.

[0087] When the particle size feature is a non-numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude using the preset conversion method corresponding to the particle size feature.

[0088] To simplify the standardization of different particle size characteristics, this application categorizes particle size characteristics into numerical and non-numerical features. For example, the proportion of ionizable lipids and phospholipids are numerical features, while the types of ionizable lipids and phospholipids are non-numerical features. To convert these particle size characteristics into parametric features within the same numerical range, specifically, for numerical particle size characteristics, the conversion is based on the position of the characteristic's value within its corresponding theoretical range. For instance, if the range of a parametric feature is 0–1, and the actual value of a numerical feature is 10, while its theoretical range is 0–100, then this numerical feature can be converted to the parametric feature 0.1. For non-numerical particle size characteristics, the conversion is based on the types of particle size characteristics themselves. Specifically, it's based on the similarity between different types of the particle size characteristic. One type of the particle size characteristic can be designated as a baseline type, and other types can be expanded into new features, with the similarity between these types used as the parametric feature. For example, when the parameter feature range is 0 to 1, phospholipid type, as a non-numerical feature, includes five types: DOPC, DOPE, DPPC, DSPC, and DSPE. During transformation, these need to be converted into four-dimensional indicator variables as parameter features. Assuming DOPC is the baseline type, when the actual input phospholipid type is DOPC, the values ​​of these four indicator variables are 0, 0, 0, and 0, respectively; when the actual input phospholipid type is DPPC, the values ​​are 0, 0, 1, and 0, respectively, and so on, to transform these non-numerical particle size features. Based on this, different particle size features can be easily standardized.

[0089] As a preferred embodiment, the first preset machine learning model is either a tree model or a LASSO model with L1 penalty.

[0090] The tree model is a greedy algorithm. When encountering a decision branch point, it randomly selects a subset of particle size features from all training samples. Within this subset, it iterates through each feature, taking values ​​to determine which feature has the greatest impact on particle size, and selects the most influential feature. After traversing all decision branch points in the tree model, it calculates the number of times each feature is selected and sorts them. The feature selected more frequently has the greatest impact on particle size. Please refer to [reference needed]. Figure 2 , Figure 2 This diagram illustrates the number of times the branching feature is selected in a tree model provided in this application. Figure 2The horizontal axis represents the number of times each particle size feature was selected, and the vertical axis represents each particle size feature. As for the specific tree model used, it could be the LightGBM model or a random forest model, etc.; this application does not specify this limitation.

[0091] The LASSO (Least Absolute Shrinkage and Selection Operator) model reduces the parameter values ​​of certain particle size features in a formulation to zero by applying an L1 penalty term. The L1 penalty term is set based on the degree of influence of a particle size feature on particle size. If a particle size feature itself has a very small parameter value in the LASSO model, it indicates that the particle size feature has a small influence on particle size. To focus more on particle size features with a greater influence, these low-influence parameters are set to 0 to ignore them. Particle size features with non-zero parameters are considered to have a greater influence, thus achieving the purpose of feature selection. Let y i To predict the particle size of an object, the particle size feature x i As the input vector, n represents the number of samples in the training set. The LASSO model finds the optimal parameters w by reducing a portion of the particle size features to zero, thereby minimizing the following loss function C1:

[0092]

[0093] Here, γ is a penalty term for the complexity of the LASSO model, which can be pre-selected through five-fold cross-validation. In actual use, please refer to Tables 1 and 2 above. Table 1 shows the particle size characteristics in existing formulations, and Table 2 shows the non-zero particle size characteristics selected by the LASSO model. It can be seen that the LASSO model further selected 12 particle size characteristics from Table 1, including 5 liposome formulation characteristics, 3 experimental condition characteristics, 1 mRNA-related characteristic, and 2 liposome species characteristics. According to Table 2, the parameters and particle size characteristics far from 0 have a greater impact on particle size.

[0094]

[0095]

[0096] Table 2: Non-zero particle size characteristics selected in the LASSO model

[0097] In summary, by using multiple machine learning models to calculate each particle size feature in the prescription from different perspectives, and combining the calculation results of multiple models, the degree of influence of particle size features on particle size can be accurately determined.

[0098] In a preferred embodiment, the second preset machine learning model is a Gaussian process model. Particle size features are input into the second preset machine learning model to determine the predicted particle size of the mRNA lipid nanoparticle formulation, including:

[0099] In the Gaussian process model, the distance between the particle size feature and each training set is determined;

[0100] The particle size prediction results of the mRNA lipid nanoparticle formulation were determined based on various distances.

[0101] A Gaussian process is an infinite-dimensional Gaussian distribution whose mean is a function m(x) of each particle size feature in the prescription. The particle size features are input into the Gaussian process model. The covariance is defined as the kernel function k(x1, x2) of the same particle size feature between any two prescriptions, where x1 is the particle size feature of the input prescription, and x2 can be the same particle size feature from any sample prescription containing that particle size feature in the model's previous training. The kernel function used in this application is the Radial Basis Function (RBF kernel), with the following functional form:

[0102]

[0103] Here, σ and l are hyperparameters of the radial basis function. During training, the Gaussian process model continuously learns and optimizes these two hyperparameters to obtain their optimal values ​​for predicting particle size. The model also updates the kernel function k(x1, x2) and the posterior distribution of the particle size feature function m(x) based on the results of each training iteration to improve the model's accuracy in particle size prediction. For example, when training the Gaussian process model using 299 prescriptions, the final RMSE (Root Mean Squared Error) is 0.32 and R0 is... 2 Training accuracy with a coefficient of determination of 0.62.

[0104] As a preferred embodiment, after extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes:

[0105] The particle size features are input into a third preset machine learning model in order to determine the particle size distribution value of the mRNA lipid nanoparticle formulation. The third preset machine learning model is pre-trained from a preset number of particle size features.

[0106] Output n particle size influence values ​​and particle size prediction results, including;

[0107] Output n particle size influence values, particle size prediction results, and particle size distribution values.

[0108] Given the inherent instability of mRNA, a balance needs to be struck between particle size and size distribution, as well as encapsulation efficiency, transfection efficiency, and stability when designing formulations. In this application, when collecting literature formulations as the training set, the particle size distribution described in those literature formulations can be used as a reference. Please refer to... Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 , Figure 4 This application provides a distribution map of the particle size distribution of a training set. Figure 5 This application provides a distribution map of the proportion of ionizable lipids in a training set. Figure 6 This application provides a distribution map of the phospholipid ratio in a training set. Figure 7 This application provides a distribution map of cholesterol lipids in a training set. Figure 8 A distribution map of PEG lipids in a training set provided for this application. Figure 4 It can be seen that the particle size distribution in most of the literature formulations is between 0.1 and 0.2. The minimum particle size distribution in these literature formulations is 0.029, and the maximum is 0.89. It can be understood that a particle size distribution of less than 0.25 is more reasonable and beneficial to the stability of the formulation. Similarly, since both particle size distribution and particle size are affected by the various parameters in Table 1 above, after knowing that the particle size distribution in the literature formulations is usually between 0.1 and 0.2, a reasonable particle size distribution can be obtained by further selecting the various lipid ratios in the literature formulations. Figure 5 It can be seen that the proportion of ionizable lipids fluctuates considerably. Therefore, when designing formulations, it is advisable to prioritize the proportion of ionizable lipids selected from formulations with a large number of literature studies. According to... Figure 6 It can be seen that the phospholipid ratio in most literature prescriptions is below 30%, which confirms that a phospholipid ratio of less than 30% is more reasonable; according to Figure 6 It can be seen that the cholesterol-phospholipid ratio in most literature prescriptions is between 40% and 50%, which confirms that a cholesterol-phospholipid ratio of 40% to 50% is reasonable; according to Figure 8 It can be seen that most literature formulations have a PEG lipid ratio between 1.5 and 2.5, but many formulations also use a ratio between 2.5 and 5. Therefore, a PEG lipid ratio between 1.5 and 5 is considered more reasonable. In summary, because the adjustment range for various lipids and particle size distributions is quite wide, relying solely on the experience of technical personnel to adjust these ratios is difficult to obtain the most suitable ratio. Therefore, particle size characteristics are sent to a third-party machine learning model to help determine the complex relationship between particle size distribution and formulation.

[0109] In determining the particle size distribution, the third machine learning model is also a Gaussian process model. However, unlike the Gaussian process model described above, which outputs the exact particle size value, the third machine learning model outputs the range of particle size distribution values. To better predict the particle size distribution, the third machine learning model uses the Automatic Relevance Determination Kernel (ARDkernel). Its mean is a function m(x) of each particle size feature in the prescription, and its covariance is defined as the kernel function j(y1, y2) of the same particle size feature between any two prescriptions. y1 is the particle size feature of the input prescription, and y2 can be the same particle size feature from any sample prescription containing that particle size feature that the model has previously trained on. The functional form of the Automatic Relevance Determination Kernel is as follows:

[0110]

[0111] Here, σ and Θ are hyperparameters of the automatic association decision function. Θ is a diagonal matrix, and different weights are assigned to different dependent variables based on their diagonal elements. The third machine learning model continuously learns and optimizes these two hyperparameters during training to obtain their optimal values ​​for predicting particle size distribution. For example, when training the third machine learning model using 105 recipes, an RMSE of 0.63 and R0.05 can be obtained. 2 =0.6 training accuracy.

[0112] As a preferred embodiment, when two mRNA lipid nanoparticle formulations are obtained simultaneously, after inputting the particle size features into a second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation, the method further includes:

[0113] In the second preset machine learning model, the size prediction results of the two mRNA lipid nanoparticle formulations are compared and the comparison results are output.

[0114] To simplify prescription optimization, this application typically involves comparing two prescriptions in practical applications to determine which prescription performs better. Specifically, the current prescription is compared to a pre-defined benchmark prescription to improve the current prescription. Therefore, two prescriptions can be input simultaneously, and particle size prediction can be performed on both prescriptions separately. Then, the larger and smaller particle size prediction results for each prescription can be determined. Please refer to... Figure 3 , Figure 3 This application provides a distribution map of the matching results of prescription comparisons. Figure 3The vertical axis represents the number of comparisons, and the horizontal axis represents the absolute value of the difference between the particle sizes of the two formulations. The gray area represents the part where the comparison result is inconsistent with the actual result, and the black area represents the part where the comparison result is consistent with the actual result. It can be seen that the greater the difference between the two formulations, the higher the accuracy of the particle size prediction results. If the Gaussian process model of the above embodiment is used for comparison, its accuracy in relative conclusion judgment can reach over 86.1%. Based on the comparison between the two formulations, the formulation can be easily optimized.

[0115] Similarly, the third machine learning model can also compare the particle size distribution between two prescriptions; please refer to [reference needed]. Figure 9 , Figure 9 This application provides a distribution map showing the matching results of particle size distribution comparison. Figure 9 The vertical axis represents the number of comparisons, and the horizontal axis represents the absolute value of the difference in particle size distribution between the two prescriptions. The gray area represents the part where the comparison result is inconsistent with the true result, and the black area represents the part where the comparison result is consistent with the true result. It can be seen that the greater the difference between the two prescriptions, the higher the accuracy of comparing the particle size distributions of the two prescriptions. If the third machine learning model in the above embodiment is used for comparison, its accuracy in relative conclusion judgment can reach over 75%. Based on the comparison between the two prescriptions, prescriptions with smaller particle size distributions can be accurately screened, facilitating prescription optimization by staff.

[0116] To verify the predictive ability of a model, the model output can be compared with the results of a wet experiment. Specifically, the following methods can be used for verification:

[0117] I. The formulation mainly consists of ionizable lipids MC3, DSPC, cholesterol Cholesterol, and DMG-PEG2000. EGFP mRNA was used, and the citrate buffer for dissolving the mRNA was pH 4.0 with an N / P ratio of 3. The volume ratio of ethanol after mixing was 25%. To compare the particle size distribution of LNP-mRNA with different molar ratios of the same lipid material, experiments were conducted according to the formulation table below.

[0118] MC3 Ionizable lipids 40 10 DSPC assist lipids 10 60 Cholesterol cholesterol 48 28 DMG-PEG2000 PEGylated phospholipids 2 2

[0119] Table 3: First Prescription Table

[0120] Preparation of LNP solution: Dissolve all lipid materials in 1 ml of anhydrous ethanol solution (99.7%) at a certain molar ratio.

[0121] Preparation of mRNA solution: Dissolve mRNA in 50mM citrate buffer to prepare a volume of 3mL.

[0122] LNP-mRNA preparation method: LNP-mRNA was prepared using the ethanol injection method. The ethanol solution of lipids was drawn into a syringe and slowly added dropwise to rapidly stirred mRNA-citrate buffer. The prepared LNP-mRNA was then added to a dialysis bag, which was placed in PBS and dialyzed for 12 hours. After sample collection, ultrafiltration was performed using a Milpore 30KD ultrafiltration tube, centrifuged at 3000g for 20 minutes, and then stored in pH 7.4 PBS solution at 4°C.

[0123] Particle size determination: The sample was diluted with purified water and the particle size was determined using a Malvern Zetasizer Nano. The particle size distribution results are shown in Table 4.

[0124]

[0125] Table 4: Comparison of Particle Size Distribution Prediction Results and Experimental Results

[0126] The model prediction results show that the lipid material is composed of MC3, DSPC, cholesterol (Cholesterol), and DMG-PEG2000. With an N / P ratio of 3, the particle size distribution of 40 / 10 / 48 / 2 is smaller and more uniform than that of 10 / 60 / 28 / 2. The wet test results are consistent with the prediction results.

[0127] II. The formulation mainly consists of ionizable lipids DLinDMA, DSPC, cholesterol Cholesterol, and DMG-PEG2000. EGFP mRNA was used, and the citrate buffer for dissolving the mRNA was pH 4.0 with an N / P ratio of 3. The volume ratio of ethanol after mixing was 25%. The particle size distribution of LNP-mRNA was compared with the same lipid material at different molar ratios. The experiment was conducted according to the formulation table below.

[0128] DLinDMA Ionizable lipids 40 10 DSPC assist lipids 10 60 Cholesterol cholesterol 48 28 DMG-PEG2000 PEGylated phospholipids 2 2

[0129] Table 5: The preparation method of the second formulation table is the same as above, and the particle size distribution results are shown in Table 6.

[0130]

[0131] Table 6: Comparison of Particle Size Distribution Prediction Results and Experimental Results

[0132] The model prediction results show that the lipid material is composed of DLinDMA, DSPC, cholesterol Cholesterol and DMG-PEG2000. With an N / P ratio of 3, the particle size distribution of 40 / 10 / 48 / 2 is smaller and more uniform than that of 10 / 60 / 28 / 2. The wet test results are consistent with the prediction results.

[0133] III. The formulation mainly consists of ionizable lipids MC3, DSPC, cholesterol Cholesterol, and DMG-PEG2000. EGFP mRNA was used, and the citrate buffer for dissolving the mRNA was pH 4.0 with an N / P ratio of 6. The volume ratio of ethanol after mixing was 25%. The particle size distribution of LNP-mRNA was compared with the same lipid material at different molar ratios. The experiment was conducted according to the formulation table below.

[0134]

[0135]

[0136] Table 7: Third Prescription Table

[0137] The preparation method is the same as above, and the particle size distribution results are shown in Table 8.

[0138]

[0139] Table 8: Comparison of Particle Size Distribution Prediction Results and Experimental Results

[0140] The model prediction results show that the lipid material is composed of MC3, DSPC, cholesterol (Cholesterol), and DMG-PEG2000. With an N / P ratio of 6, the particle size distribution of the molar ratio 40 / 10 / 48 / 2 is not much different from that of the particle size distribution of 50 / 10 / 38.5 / 1.5. The wet test results are consistent with the prediction results.

[0141] IV. The formulation mainly consists of ionizable lipids MC3, DSPC, cholesterol Cholesterol, and DMG-PEG2000. EGFP mRNA was used, and the citrate buffer for dissolving the mRNA was pH 4.0 with an N / P ratio of 6. The volume ratio of ethanol after mixing was 25%. The particle size distribution of LNP-mRNA was compared with the same lipid material at different molar ratios. The particle size of LNP-mRNA was predicted using a model and verified by wet experiments. The experiment was conducted according to the formulation table below.

[0142] MC3 Ionizable lipids 60 35 DSPC assist lipids 10 16 Cholesterol cholesterol 29.5 46.5 DMG-PEG2000 PEGylated phospholipids 4.5 2.5

[0143] Table 9: Fourth Prescription Table

[0144] The preparation method is the same as above, and the particle size distribution results are shown in Table 10.

[0145]

[0146] Table 10: Comparison of Particle Size Distribution Prediction Results and Experimental Results

[0147] The model prediction results show that the lipid material is composed of MC3, DSPC, cholesterol (Cholesterol), and DMG-PEG2000. With an N / P ratio of 6, the particle size distribution of the molar ratio 60 / 10 / 29.5 / 4.5 is larger than that of the particle size distribution of 35 / 16 / 46.5 / 2.5. The wet test results are consistent with the prediction results.

[0148] Please refer to Figure 10 , Figure 10 A schematic diagram of a device for predicting the particle size of mRNA lipid nanoparticle vaccines provided in this application includes:

[0149] Memory 21 is used to store computer programs;

[0150] The controller 22 is used to execute a computer program to implement the steps of the method for predicting the particle size of mRNA lipid nanoparticle vaccines as described above.

[0151] For a detailed description of the device for predicting the particle size of mRNA lipid nanoparticle vaccines provided in this application, please refer to the embodiments of the above-described method for predicting the particle size of mRNA lipid nanoparticle vaccines, which will not be repeated here.

[0152] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0153] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0154] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the particle size of mRNA lipid nanoparticle vaccines, characterized in that, include: Obtain mRNA lipid nanoparticle formulation; The particle size characteristics of the mRNA lipid nanoparticle formulation were extracted; The particle size features are respectively input into n first preset machine learning models to obtain n particle size influence values ​​of the particle size features. The n first preset machine learning models are all pre-trained with a preset number of particle size features, and n is a positive integer. The first preset machine learning model is either a tree model or a LASSO model with L1 penalty. The particle size features are input into a second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation. The second preset machine learning model is trained in advance by a preset number of the particle size features. Output n particle size influence values ​​and the particle size prediction results; When n is greater than 2, the n first preset machine learning models are all machine learning models of different types.

2. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 1, characterized in that, When the number of extracted particle size features is not less than 2, the particle size features are respectively input into n first preset machine learning models, including: Each of the aforementioned particle size features is converted into parameter features of the same order of magnitude as the preset order of magnitude; All the aforementioned parameter features are input into n first preset machine learning models respectively.

3. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 2, characterized in that, After extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes: Determine whether the number of extracted particle size features is greater than a first preset number; If the number is greater than the first preset number, then proceed to the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset number; If the number is not greater than the first preset number, determine whether the number of extracted particle size features is greater than the second preset number; If the number is not greater than the second preset number, then the missing particle size feature in the current mRNA lipid nanoparticle formulation is determined, and the missing particle size feature in the current mRNA lipid nanoparticle formulation is filled based on the corresponding particle size feature in other mRNA lipid nanoparticle formulations, and then the step of converting the particle size feature into a parameter feature of the same order of magnitude as the preset order of magnitude is entered. Wherein, the first preset quantity is greater than the second preset quantity.

4. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 3, characterized in that, After determining whether the number of extracted particle size features is greater than a second preset number, the method further includes: If the number is greater than the second preset number, the calculation method of the first preset machine learning model is set to a calculation method that does not consider the particle size features missing in the current mRNA lipid nanoparticle formulation.

5. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 2, characterized in that, Converting each of the aforementioned particle size features into parameter features of the same order of magnitude as a preset order of magnitude includes: When the particle size feature is a numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude, based on the position of the value of the particle size feature within the preset range corresponding to the particle size feature. When the particle size feature is a non-numerical feature, the particle size feature is converted into a parameter feature of the same order of magnitude as the preset order of magnitude using a preset conversion method corresponding to the particle size feature.

6. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 1, characterized in that, The second preset machine learning model is a Gaussian process model. The particle size features are input into the second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation, including: In the Gaussian process model, the distance between the particle size feature and each training set is determined; The particle size prediction results of the mRNA lipid nanoparticle formulation are determined based on each of the aforementioned distances.

7. The method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in claim 1, characterized in that, After extracting the particle size characteristics of the mRNA lipid nanoparticle formulation, the method further includes: The particle size features are input into a third preset machine learning model to determine the particle size distribution value of the mRNA lipid nanoparticle formulation. The third preset machine learning model is pre-trained from a preset number of the particle size features. Output n particle size influence values ​​and the particle size prediction results, including: Output n values ​​representing the particle size influence, the particle size prediction result, and the particle size distribution value.

8. The method for predicting the particle size of mRNA lipid nanoparticle vaccines according to any one of claims 1 to 7, characterized in that, When two mRNA lipid nanoparticle formulations are obtained simultaneously, after inputting the particle size features into a second preset machine learning model to determine the particle size prediction result of the mRNA lipid nanoparticle formulation, the method further includes: In the second preset machine learning model, the size prediction results of the two mRNA lipid nanoparticle formulations are compared, and the comparison results are output.

9. A device for predicting the particle size of mRNA lipid nanoparticle vaccines, characterized in that, include: Memory, used to store computer programs; A controller is configured to execute the computer program to implement the steps of the method for predicting the particle size of mRNA lipid nanoparticle vaccines as described in any one of claims 1 to 8.

Images