Prediction device, particle size distribution prediction method for prediction device, and particle size distribution prediction program

Abstract

A prediction device (100) comprises: an information acquisition unit (102) that acquires a condition for preparing lipid particles that are composed of any of a phospholipid, an anionic lipid, and a cationic lipid, and a constituent lipid of the lipid particles; a trained model (103) that predicts the particle size distribution of the lipid particles from the acquired condition for preparing the lipid particles and the acquired constituent lipid of the lipid particles; and an output unit (104) that outputs the particle size distribution of the lipid particles predicted by the trained model (103).

Description

Prediction device, particle size distribution prediction method for prediction device, and particle size distribution prediction program

[0001] The present invention relates to a prediction device, a particle size distribution prediction method for the prediction device, and a particle size distribution prediction program.

[0002] Lipid particles prepared by mixing phospholipids or other lipids (e.g., ionized lipids, PEG-modified lipids, and cholesterol) are primarily used as nanomedicines due to their relatively high safety and biocompatibility. Lipid-encapsulated formulations are bioinspired nanocarriers that can protect drugs from degradation, control drug release, target drug delivery, modify biodistribution, and enhance bioavailability.

[0003] Furthermore, lipid particles are one of the most popular cell-mimetic compartments used to create artificial cells, and as such, they are not only used in drug delivery systems (DDSs), cosmetics, and agrochemicals, but also play a central role in the bottom-up creation of bacterial-sized artificial cells.

[0004] The cell-mimicking compartmental functions of lipid particles include substrate uptake, protein trafficking, phospholipid biosynthesis, cell division and related processes, and membrane protein evolution. The key parameters of lipid particles are their size, lipid composition, and lamellae, which determine their physicochemical properties and affect compound encapsulation efficiency, cellular uptake efficiency, and target tissue delivery.

[0005] Furthermore, in bottom-up construction of artificial cells, the shape and size of lipid particles significantly alter the behavior of cytoskeletal proteins oriented inside them.

[0006] Therefore, controlling the size and lipid composition of lipid particles plays an important role in a wide range of fields, such as the construction of artificial cells using bacteria-sized lipid particles.

[0007] Currently, there are two main strategies for producing lipid particles: one is a top-down approach that involves the formation of large vesicles followed by size reduction processes (extruders, high shear stress, sonication), and the other is a bottom-up approach that forms vesicles by promoting self-assembly from individual lipid monomers.

[0008] In recent years, fabrication methods using microfluidic devices have become common as bottom-up methods. The main advantages of using microfluidic channels for lipid particle fabrication are the high reproducibility of the size distribution of the resulting lipid particles and the scalability achieved by parallelizing the microfluidic channels, which allows for stable customization of the production scale. Here, the inventors discovered that by using lipids with a composition equivalent to that of E. coli, it is possible to fabricate 1-μm liposomes using microfluidic channels.

[0009] In microchannel-mediated lipid particle assembly, the size of lipid particles can be controlled by adjusting the total flow rate and the ratio of the organic and aqueous phases. Furthermore, the intrinsic properties of the lipids used (e.g., charge, phase transition temperature, molecular weight), the chip structure, and the mixing temperature are important factors that determine the size and polydispersity of the resulting lipid particles.

[0010] Here, existing microfluidic devices can control the particle size distribution of lipid particles within the range of approximately 100 nm to 300 nm.

[0011] However, because multiple factors, such as the flow rate ratio of the buffer solution to the lipid, the lipid composition, and the pH of the buffer solution, affect each other, it is difficult to accurately control the distribution using a microchannel device, especially when creating lipid particles of around 1000 nm.

[0012] To analyze such complex factors, predictions are made using machine learning. For example, gradient boosting methods such as LightGBM (Light Gradient Boosting Machine) can accurately predict the objective variable and evaluate the contribution rate of the explanatory variable (see, for example, Patent Documents 1 and 2).

[0013] JP 2023-022170 A JP 2021-085849 A

[0014] Although gradient boosting is a useful method, the accuracy improvement achieved by adjusting hyperparameters in gradient boosting is small, and additional features are required. Therefore, it has been suggested that gradient boosting requires a considerable amount of experimental data. Therefore, it is difficult to predict the size of lipid particles using gradient boosting.

[0015] Therefore, an object of the present invention is to provide a predictor, a particle size distribution prediction method for a prediction device, and a particle size distribution prediction program that can accurately predict the size distribution of lipid particles including those with sizes of 100 to 1000 nm.

[0016] That is, the above-mentioned object of the present invention is achieved by the following configuration. (1) A prediction device comprising: an information acquisition unit that acquires production conditions for lipid particles composed of any one of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; a trained model that predicts the particle size distribution of lipid particles from the acquired production conditions for lipid particles and the constituent lipids of the lipid particles; and an output unit that outputs the particle size distribution of lipid particles predicted by the trained model. (2) The prediction device described in (1), wherein the trained model is a trained model trained by machine learning using the production conditions for lipid particles and the constituent lipids of the lipid particles as explanatory variables and the particle size distribution of the lipid particles as a target variable. (3) The prediction device described in (1) or (2), wherein the particle size distribution of lipid particles is a curve representing the abundance ratio of lipid particles at each particle size. (4) The prediction device described in (3), wherein each particle size is in the range of 0 [nm] to 10,000 [nm]. (5) The prediction device according to (2), wherein the trained model is machine-learned using the explanatory variables or the response variables that have been preprocessed. (6) The prediction device according to (5), wherein the explanatory variables are obtained by acquiring or predicting lipid physical properties in the preprocessing, and a weighted average is added according to the lipid proportion of each lipid particle. (7) The prediction device according to (5), wherein the explanatory variables are obtained by adding a lipid diversity index in the preprocessing. (8) The prediction device according to (5), wherein the response variable is experimental data from which aggregated lipid particle samples have been removed in the preprocessing. (9) The prediction device according to (8), wherein the experimental data is obtained by removing experimental conditions under which phosphatidylethanolamine exceeds 60% with respect to the lipid weight constituting the lipid particles, or experimental conditions under which the lipid flow rate relative to the total flow rate of the lipid and buffer exceeds 0.6. (10) The prediction device according to (1), wherein the conditions for producing the lipid particles include at least one selected from the flow rate ratio, pH, and lipid composition of the composition solution used for producing the lipid particles, or a combination thereof.(11) A particle size distribution prediction method for a prediction device that executes the following steps: (1) an information acquisition unit acquires production conditions for lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; (2) a trained model predicts the particle size distribution of lipid particles from the acquired production conditions for lipid particles and the constituent lipids of the lipid particles, and (3) an output unit outputs the particle size distribution of lipid particles predicted by the trained model. (12) A particle size distribution prediction program that executes the following steps: (1) having a computer acquire production conditions for lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; (2) having a computer predict the particle size distribution of lipid particles from the acquired production conditions for lipid particles and the constituent lipids of the lipid particles, and (3) outputting the particle size distribution of lipid particles predicted by the trained model.

[0017] According to the present invention, it is possible to predict an accurate size distribution of lipid particles, including sizes of 100 to 1000 nm.

[0018] 3A 。 FIG. 3B is a diagram illustrating an example of the configuration of a prediction device according to the present embodiment. FIG. 3C is a flowchart illustrating particle size distribution prediction processing executed by the prediction device according to the present embodiment. FIG. 3D is a graph showing the prediction accuracy of particle size distribution before excluding the processing of steps S003 and S005 as preprocessing. FIG. 3E is a graph showing the prediction accuracy of particle size distribution as a result of excluding the processing of steps S003 and S005 from the processing that generated the result shown in the graph of FIG. 3A. FIG. 3F is an explanatory diagram showing the concept of obtaining physical property values ​​shown in step S007 and predicting physical property values ​​shown in step S009. FIG. 3G is a graph showing the particle size distribution of a predetermined lipid predicted by the prediction device (part 1). FIG. 3H is a graph showing actual measured values ​​of lipid particles actually produced using a microchannel system (part 1). FIG. 3I is a graph showing the particle size distribution of a predetermined lipid predicted by the prediction device (part 2). FIG. 3I is a graph showing actual measured values ​​of lipid particles actually produced using a microchannel system (part 2).

[0019] The following describes in detail embodiments of the present invention. Note that the embodiments described below are examples for realizing the present invention, and should be appropriately modified or changed depending on the configuration of the device to which the present invention is applied and various conditions. Therefore, the present invention is not limited to the following embodiments. Furthermore, the present invention may be configured by appropriately combining parts of the embodiments described below. Note that the same components are given the same reference numerals, and descriptions thereof will be omitted as appropriate.

[0020] <Present Embodiment> [Overall Configuration of Image Processing Apparatus] Fig. 1 is a diagram illustrating an example of the configuration of a prediction apparatus according to this embodiment. The prediction apparatus 100 according to this embodiment is configured to include a CPU 101, an information acquisition unit 102, a trained model 103, an output unit 104, a display unit 105, a ROM (Read Only Memory) 106, a RAM (Random Access Memory) 107, and an external storage device 108. The external storage device 108 is configured to include a particle size distribution prediction program 109.

[0021] 1, CPU 101 is an arithmetic processing unit that controls the entire prediction apparatus 100. CPU 101 performs various controls by executing programs stored in ROM 106 or external storage device 108. For example, CPU 101 executes particle size distribution prediction program 109 stored in external storage device 108 to perform particle size distribution prediction processing of the flowchart shown in FIG. 2, which will be described later.

[0022] The information acquisition unit 102 acquires the conditions for creating lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, as well as the constituent lipids of the lipid particles. The information acquisition unit 110 is composed of, for example, operating members such as a keyboard, a mouse, various switches, buttons, and a touch panel that accept various operations from a user. Note that the lipid particles refer to vesicles composed of at least one of ionized (pH-sensitive) lipids, PEGylated lipids, cholesterol, and phospholipids, or a combination thereof.

[0023] The trained model 103 predicts the particle size distribution of lipid particles based on the acquired lipid particle production conditions and the constituent lipids of the lipid particles. The lipid particle production conditions include at least one selected from the flow rate ratio, pH, and lipid composition of the composition solution used to produce the lipid particles, or a combination thereof. The constituent lipids of the lipid particles are composed of ionized lipids, PEGylated lipids, phospholipids, and cholesterol. The predicted particle size distribution of the lipid particles is a curve representing the abundance ratio of lipid particles at each particle size. Furthermore, each particle size of the lipid particles ranges from 0 [nm] to 10,000 [nm].

[0024] The trained model 103 is composed of a trained model that has been machine-learned using the lipid particle production conditions and the constituent lipids of the lipid particles as explanatory variables and the particle size distribution of the lipid particles as a target variable. The trained model 103 may be machine-learned using preprocessed explanatory variables or target variables. For example, the explanatory variables may be obtained or predicted in preprocessing by acquiring or predicting the physical properties of lipids, and a weighted average may be added according to the lipid proportion of each lipid particle. Furthermore, a lipid diversity index may be added to the explanatory variables in preprocessing.

[0025] The objective variable may be experimental data obtained by removing (excluding) aggregated lipid particle samples during pretreatment. In this case, the experimental data may be, for example, experimental conditions in which phosphatidylethanolamine exceeds 60% relative to the lipid weight constituting the lipid particles. Furthermore, the experimental data may be, for example, experimental conditions in which the lipid flow rate relative to the total flow rate of the lipid and buffer exceeds 0.6. Note that these numerical values ​​are merely examples and are not limited to these. For example, experimental conditions in which the ratio of phosphatidylethanolamine to lipid weight exceeds 75% may be removed (excluding).

[0026] The trained model 103 includes a pre-trained classifier and stores various parameters. Note that the trained model 103 only needs to be able to predict the particle size distribution of lipid particles from the acquired lipid particle production conditions and the constituent lipids of the lipid particles, and may, for example, apply a different predetermined training model.

[0027] The output unit 104 outputs the particle size distribution of lipid particles predicted by the trained model 103. The output unit 104 outputs to, for example, a printer, a display, or an information processing device.

[0028] The display unit 105 is configured by, for example, a display device such as a liquid crystal display, an organic EL display, or a printer, or software for displaying (a viewer).

[0029] The ROM 106 stores, for example, a control program that controls the prediction device 100. The RAM 107 is configured, for example, by a dynamic random access memory (DRAM), and functions as a working memory that temporarily stores data required for the CPU 101 to execute the particle size distribution prediction program 109.

[0030] The external storage device 108 is configured by, for example, a hard disk drive (HDD), and stores necessary data in association with each other.

[0031] [Particle Size Distribution Prediction Process of Prediction Device] Next, the particle size distribution prediction process of lipid particles executed by the prediction device 100 configured as described above will be described using a flowchart with reference to FIG.

[0032] 2 is a flowchart showing the particle size distribution prediction process executed by the prediction device according to this embodiment. Steps S001 to S011 correspond to pre-processing.

[0033] First, the prediction device 100 accepts a user operation and acquires, from the information acquisition unit 102, data on the conditions for creating lipid particles composed of either phospholipids, anionic lipids, or cationic lipids, and data on the constituent lipids of the lipid particles (step S001).

[0034] Next, the prediction device 100 performs a setting to remove (exclude) experimental conditions in which phosphatidylethanolamine exceeds 75% relative to the weight of lipids constituting the lipid particles (step S003). Note that the experimental conditions in the exclusion setting are not limited to experimental conditions in which phosphatidylethanolamine exceeds 75%, and phosphatidylethanolamine may be configured at other ratios.

[0035] Furthermore, the prediction device 100 performs a setting to remove (exclude) experimental conditions in which the lipid flow rate relative to the total flow rate of the lipid and the buffer exceeds 0.6 (step S005). Note that the experimental conditions in the exclusion setting are not limited to experimental conditions in which the lipid flow rate relative to the total flow rate of the lipid and the buffer exceeds 0.6.

[0036] 3A is a graph showing the prediction accuracy of particle size distribution before the pre-processing steps S003 and S005 are excluded. In the graph of FIG. 3A, the horizontal axis represents prediction and the vertical axis represents accuracy.

[0037] 3A shows that there is a large variation in the prediction accuracy 300 indicated by the straight line, and that the accuracy is reduced. Therefore, the prediction device 100 is set to exclude samples with low prediction accuracy of particle size distribution in the processes of steps S003 and S005.

[0038] Fig. 3B is a graph showing the results of measuring the particle size distribution obtained by excluding the processes of steps S003 and S005 from the processes used to measure the results shown in the graph of Fig. 3A. As with Fig. 3A, the horizontal axis of the graph of Fig. 3B also represents predictions, and the vertical axis represents the correct answer.

[0039] As shown in Figure 3B, the plots are closer to the straight line representing the prediction accuracy of 300, and the variance is reduced. This suggests that excluding samples with poor accuracy from the experimental conditions significantly improves the accuracy of the data and makes it possible to predict the particle size distribution of lipid particles.

[0040] Returning to FIG. 2 , the explanation will be continued. The prediction device 100 receives input of the names of lipid compounds used in the experiment via the information acquisition unit 102. The prediction device 100 acquires the physical property values ​​of the lipids from external data (e.g., a compound information database) (acquisition of physical property values) (step S007). Then, the prediction device 100 adds a weighted average of the physical property values ​​corresponding to the proportion of lipids contained in the particles as an explanatory variable (step S008).

[0041] The prediction device 100 also receives input of the SMIES (Simplified Molecular Input Line Entry System) notation for the lipids used in the experiment via the information acquisition unit 102. The SMIES notation is a method of unambiguously expressing the structure of a molecule in a string of alphanumeric characters in ASCII code (expressed in text format). The prediction device 100 predicts the physical property values ​​of the lipids by applying a natural language processing model to the string expressed in SMIES notation (prediction of physical property values) (step S009). The prediction device 100 then adds a weighted average of the predicted lipid physical property values, based on the proportion of lipids contained in the particle, to the explanatory variables (step S010).

[0042] FIG. 4 is an explanatory diagram showing the concept of obtaining physical property values ​​in step S007 and predicting physical property values ​​in step S009.

[0043] 4, the prediction device 100 acquires the physical property values ​​of the lipid from the compound information database by receiving an input of the lipid compound name in the information acquisition unit 102. Then, in step S008, the prediction device 100 calculates a weighted average of the acquired lipid physical property values ​​and adds the weighted average as an explanatory variable.

[0044] In this case, for example, the physical property values ​​of the lipid to be obtained include logP, mass, number of hydrogen bond donors, number of hydrogen bond acceptors, and the like.

[0045] 4, the prediction device 100 predicts the physical property values ​​of the lipids from the natural language processing model by receiving input of the SMILES notation of the lipids in the information acquisition unit 102. Then, in step S010, the prediction device 100 adds a weighted average of the predicted physical property values ​​(predicted lipid property values) that is adjusted to the proportion of lipids contained in the particles to the explanatory variables.

[0046] In this case, for example, Balaban J, Labute ASA, solubility, number of valence electrons, etc., are predicted physical property values.

[0047] Returning to Fig. 2, the description will continue. Next, the prediction device 100 adds the diversity index of each particle as an explanatory variable based on the ratio of lipids that make up the lipid particle (step S011).

[0048] As an example of a diversity index, the Simpson's diversity index to be applied is shown below in equation (1).

[0049]

[0050] The prediction device 100 executes (sets) the preprocessing from step S001 to step S011, and predicts the particle size distribution of the lipid particles from the acquired lipid particle production conditions and the constituent lipids of the lipid particles (step S013).

[0051] Fig. 5A is a graph showing the particle size distribution of a given lipid predicted by the prediction device 100. On the other hand, Fig. 5B is a graph showing the actual measured values ​​of lipid particles actually produced using a microfluidic device. In Figs. 5A and 5B, the horizontal axis represents size [nm] and the vertical axis represents intensity [%].

[0052] As shown in the graph in Figure 5A, prediction models 501 to 503 show predictions for each flow rate ratio between the lipid flow rate and the buffer flow rate. By controlling the flow rate ratio, the prediction device 100 can adjust the flow rate in the microchannel and control the size of the lipid particles. The particle size distribution of the lipid particles is a curve that represents the abundance ratio of lipid particles at each particle size.

[0053] Prediction model 501 shows a case where the lipid and buffer flow rates are 0.3, prediction model 502 shows a case where the lipid and buffer flow rates are 0.4, and prediction model 503 shows a case where the lipid and buffer flow rates are 0.5.

[0054] The prediction models 501 to 503 are all from 0 [nm] to about 10 ４ The prediction covers the entire range of [nm], and in particular, ２ [nm] to about 10 ３ The size distribution in [nm] is predicted.

[0055] 5B, actual measurement models 551 to 553 show actual measurement values ​​for each flow rate ratio between the lipid flow rate and the buffer flow rate. That is, these actual measurement models 551 to 553 represent the actual results of lipid particles produced using a microfluidic device.

[0056] The actual measurement model 551 shows a case where the lipid and buffer flow rates are 0.3, the actual measurement model 552 shows a case where the lipid and buffer flow rates are 0.4, and the actual measurement model 553 shows a case where the lipid and buffer flow rates are 0.5.

[0057] The actual measurement models 551 to 553 all have curves with similar characteristics to the prediction models 501 to 503. This indicates that the prediction models 501 to 503 produced by the prediction device 100 can predict the results of lipid particle production by inputting predetermined experimental conditions. In particular, the prediction models 501 to 503 produced by the prediction device 100 have a curve with a similar characteristic to that of the prediction models 501 to 503 produced by the prediction device 100. ２ [nm] to about 10 ３ It covers sizes ranging from [nm].

[0058] Similarly, we will change the experimental conditions and verify whether the size of lipid particles can be predicted under other experimental conditions.

[0059] Fig. 6A is a graph showing the particle size distribution of a given lipid predicted by the prediction device 100. On the other hand, Fig. 6B is a graph showing the actual measured values ​​of lipid particles actually produced using a microfluidic device. In Figs. 6A and 6B, the horizontal axis represents size [nm] and the vertical axis represents intensity [%].

[0060] As shown in the graph in Figure 6A, prediction models 601-603 show predictions for each flow rate ratio of lipid flow rate to buffer flow rate. Prediction model 601 shows the case where the lipid and buffer flow rates are 0.1. Prediction model 602 shows the case where the lipid and buffer flow rates are 0.2. Prediction model 603 shows the case where the lipid and buffer flow rates are 0.4.

[0061] The prediction models 601 to 603 are all from 0 [nm] to about 10 ４ The prediction covers the entire range of [nm], and in particular, ２ [nm] to about 10 ３ The size distribution in [nm] is predicted.

[0062] 6B, actual measurement models 661 to 663 show actual measurement values ​​for each flow rate ratio between the lipid flow rate and the buffer flow rate. Similar to actual measurement models 5511 to 553, these actual measurement models 661 to 663 are the results of actual production of lipid particles using a microfluidic device.

[0063] The actual measurement model 661 shows a case where the lipid and buffer flow rates are 0.1, the actual measurement model 662 shows a case where the lipid and buffer flow rates are 0.2, and the actual measurement model 663 shows a case where the lipid and buffer flow rates are 0.4.

[0064] The actual measurement models 661 to 663 all have curves with similar characteristics to the prediction models 601 to 603. This indicates that the prediction models 601 to 603 produced by the prediction device 100 can predict the results of lipid particle production in the same way as the prediction models 501 to 503, by inputting predetermined experimental conditions. In particular, the prediction models 501 to 503 produced by the prediction device 100 have a curve with a magnitude of about 10 ２ [nm] to about 10 ３ It covers sizes ranging from [nm].

[0065] As described above, the prediction device 100 according to this embodiment is configured to include an information acquisition unit 102, a trained model 103, and an output unit 104. The information acquisition unit 102 acquires the conditions for creating lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles. The trained model 103 predicts the particle size distribution of the lipid particles from the acquired conditions for creating lipid particles and the constituent lipids of the lipid particles. The output unit 104 outputs the particle size distribution of the lipid particles predicted by the trained model 103.

[0066] As a result, the prediction device 100 according to this embodiment can predict the distance between 100 and 1000 [nm] (10 ３ The exact size distribution of lipid particles, including their size (nm), can be predicted.

[0067] The prediction device 100 according to this embodiment is also configured to include a trained model 103. The trained model 103 may be a trained model that has been machine-learned using the lipid particle production conditions and the constituent lipids of the lipid particles as explanatory variables and the particle size distribution of the lipid particles as a target variable.

[0068] With this configuration, the trained model 103 can perform data analysis called supervised learning and classify explanatory variables according to the target variable, thereby enabling high-accuracy classification of the lipid particle production conditions and the constituent lipids of the lipid particles.

[0069] Furthermore, in the prediction device 100 according to this embodiment, the particle size distribution of lipid particles may be a curve representing the abundance ratio of lipid particles at each particle size.

[0070] With this configuration, the prediction models 501-503 and 601-603 in the prediction device 100 can adjust the flow rate in the microchannel and control the size of the lipid particles, for example, by controlling the flow rate ratio, which allows the prediction models 501-503 and 601-603 to form curves that represent the abundance ratio of lipid particles at each particle size.

[0071] In addition, in the prediction device 100 according to this embodiment, each particle size may be set to a range of 0 [nm] to 10,000 [nm].

[0072] This configuration makes it possible to predict particle sizes larger than those of lipid particles created using existing microchannel devices, thereby promoting the use of microchannel devices and enabling stable customization of production scale.

[0073] Furthermore, in the prediction device 100 according to this embodiment, the trained model 103 may be machine-learned using explanatory variables or target variables that have been preprocessed.

[0074] According to this configuration, the prediction device 100 can improve accuracy and reliability as well as increase versatility by using machine learning based on the objective variable or explanatory variable using the trained model 103.

[0075] Furthermore, in the prediction device 100 according to this embodiment, the physical properties of lipids may be acquired or predicted in preprocessing, and a weighted average may be added according to the lipid proportion of each lipid particle. In this case, a lipid diversity index may be added as an explanatory variable in preprocessing. Furthermore, experimental data from which aggregated lipid particle samples have been removed in preprocessing may be used as a response variable.

[0076] In particular, experimental data may exclude (exclude) experimental conditions in which phosphatidylethanolamine exceeds 60% relative to the lipid weight constituting the lipid particle, or in which the lipid flow rate relative to the sum of the lipid and buffer flow rates exceeds 0.6.

[0077] According to this configuration, the prediction device 100 performs such preprocessing, thereby significantly improving the accuracy of the data and enabling highly accurate prediction of size distribution.

[0078] In particular, the prediction device 100 can dramatically improve accuracy by setting experimental conditions to be excluded from the experimental data. For example, in Fig. 3A, the prediction accuracy was 0.678 with a predetermined coefficient, but in Fig. 3B, by excluding certain experimental conditions, the prediction accuracy can be improved to 0.884 with a predetermined coefficient.

[0079] Furthermore, the prediction device 100 according to this embodiment may include, as lipid particle production conditions, at least one selected from the flow rate ratio, pH, and lipid composition of the composition solution used to produce the lipid particles, or a combination thereof.

[0080] With this configuration, the prediction device 100 can predict what size lipid particles can be created under predetermined creation conditions (experimental conditions).

[0081] REFERENCE SIGNS LIST 100 Prediction device 101 CPU 102 Information acquisition unit 103 Trained model 104 Output unit 105 Display unit 106 ROM 107 RAM 108 External storage device 109 Particle size distribution prediction program

Claims

1. A prediction device comprising: an information acquisition unit that acquires the conditions for creating lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; a trained model that predicts the particle size distribution of lipid particles from the acquired conditions for creating lipid particles and the constituent lipids of the lipid particles; and an output unit that outputs the particle size distribution of lipid particles predicted by the trained model.

2. The prediction device according to claim 1, wherein the trained model is a trained model that has been machine-learned using the conditions for producing the lipid particles and the constituent lipids of the lipid particles as explanatory variables and the particle size distribution of the lipid particles as a target variable.

3. The prediction device according to claim 1 or 2, wherein the particle size distribution of the lipid particles is a curve representing the abundance ratio of lipid particles at each particle size.

4. The prediction device according to claim 3, wherein each particle size is in the range of 0 nm to 10,000 nm.

5. The prediction device according to claim 2, wherein the trained model is machine-trained using the explanatory variables or the target variables that have been preprocessed.

6. The prediction device according to claim 5, wherein the explanatory variables are obtained by acquiring or predicting the physical properties of lipids in the preprocessing and adding a weighted average according to the lipid proportion of each lipid particle.

7. The prediction device according to claim 5, wherein a lipid diversity index is added to the explanatory variables in the preprocessing.

8. The prediction device according to claim 5, wherein the dependent variable is experimental data from which aggregated lipid particle samples have been removed in the preprocessing.

9. The prediction device described in claim 8, wherein the experimental data excludes experimental conditions in which phosphatidylethanolamine exceeds 60% relative to the lipid weight constituting the lipid particles, or experimental conditions in which the lipid flow rate relative to the total flow rate of the lipid and buffer exceeds 0.

6.

10. The prediction device according to claim 1, wherein the conditions for producing the lipid particles include at least one selected from the flow rate ratio, pH, and lipid composition of the composition solution used for producing the lipid particles, or a combination thereof.

11. A particle size distribution prediction method for a prediction device that executes the steps of: an information acquisition unit acquiring conditions for creating lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; a trained model predicting the particle size distribution of lipid particles from the acquired conditions for creating lipid particles and the constituent lipids of the lipid particles; and an output unit outputting the particle size distribution of lipid particles predicted by the trained model.

12. A particle size distribution prediction program that causes a computer to execute the following steps: acquiring conditions for creating lipid particles composed of any of phospholipids, anionic lipids, and cationic lipids, and the constituent lipids of the lipid particles; having a trained model predict the particle size distribution of lipid particles from the acquired conditions for creating lipid particles and the constituent lipids of the lipid particles; and outputting the particle size distribution of lipid particles predicted by the trained model.

Images