A method and device for constructing a LNP transfection efficiency prediction model

Abstract

The present invention relates to the field of deep learning technology, and discloses a method and device for constructing a LNP transfection efficiency prediction model, the method comprising: obtaining an ionizable lipid molecule data set and constructing a molecule pair; determining a differential structure mask corresponding to each molecule in the generated molecule pair; performing word segmentation encoding processing on each molecule in the molecule pair, and determining a molecular feature vector corresponding to each molecule; determining a differential structure feature vector corresponding to each molecule based on the differential structure mask and the molecular feature vector; obtaining a differential structure feature vector between two molecules in the molecule based on the differential structure feature vector; training the model based on the molecular feature vector corresponding to each molecule in the molecule pair and the differential structure feature vector between the two molecules in the molecule pair, and constructing a LNP transfection efficiency prediction model. The model constructed by the present invention fully considers the particularity and similarity of ionizable lipid molecules, and effectively improves the LNP transfection efficiency prediction accuracy and generalization ability of the model.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and device for constructing a LNP transfection efficiency prediction model. Background Art

[0002] mRNA (messenger ribonucleic acid) plays a key role in organisms and is an important bridge connecting genes and protein synthesis. mRNA-based technologies have shown broad application prospects in the fields of viral vaccines, protein replacement therapy, cancer immunotherapy, genome editing, etc.

[0003] However, achieving targeted delivery and releasing mRNA from endosomes remains a challenge for mRNA delivery systems. Lipid nanoparticles (LNP) are a nanoscale delivery system for carrying biomolecules such as mRNA. LNPs are composed of four components: ionizable lipids, phospholipids, cholesterol, and polyethylene glycol. Among these components, ionizable lipids have the highest molar ratio, determine the delivery efficiency and stability of the formulation, and are the core structure of LNPs. Therefore, the key to selecting efficient LNPs lies in selecting suitable ionizable lipids. Although component chemistry methods based on three-component reactions (3-CR) and four-component reactions (4-CR) can synthesize a variety of ionizable lipids, experimental methods to test the transfection efficiency of each synthetic lipid are time-consuming and expensive.

[0004] Some studies have shown that artificial intelligence technologies such as machine learning and deep learning can predict LNP transfection efficiency. For example, four machine learning methods, including support vector machine, random forest, extreme gradient boosting and multi-layer perceptron, were used to classify the transfection efficiency of 572 LNPs into two categories, achieving a classification accuracy of 98%. Although the above research has made significant progress, the inventors found that the current prediction of LNP transfection efficiency based on machine learning and deep learning still has some significant limitations, such as ignoring the particularity of ionizable lipid molecules and ignoring the similarity of ionizable lipid molecules. These limitations affect the accuracy and generalization ability of the model, resulting in the prediction accuracy of LNP transfection efficiency still needs to be improved. Summary of the invention

[0005] In view of this, the present invention provides a method and device for constructing a LNP transfection efficiency prediction model to solve the problem that the method for predicting LNP transfection efficiency based on deep learning in the prior art has low accuracy.

[0006] In a first aspect, the present invention provides a method for constructing a LNP transfection efficiency prediction model, the method comprising:

[0007] Acquire an ionizable lipid molecule dataset, and construct molecule pairs based on the ionizable lipid molecule dataset;

[0008] Determine the common structure and the difference structure between two molecules in the molecular pair, and generate a one-to-one difference structure mask for each molecule in the molecular pair;

[0009] Perform word segmentation encoding processing on each molecule in the molecular pair to determine the molecular feature vector corresponding to each molecule;

[0010] Based on the one-to-one corresponding difference structure mask and the molecular feature vector of each molecule, determining the difference structure feature vector corresponding to each molecule;

[0011] Based on the one-to-one correspondence of the differential structural feature vectors of each molecule in the molecular pair, the differential structural feature vectors between the two molecules in the molecule are obtained;

[0012] The model was trained based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair to construct a prediction model for LNP transfection efficiency.

[0013] In an optional embodiment, the model is trained based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair to construct a LNP transfection efficiency prediction model, including:

[0014] Based on the molecular feature vectors corresponding to each molecule in the molecular pair, the predicted value of LNP transfection efficiency corresponding to each molecule is predicted;

[0015] Obtain the true value of LNP transfection efficiency corresponding to each molecule in the molecular pair;

[0016] Based on the predicted LNP transfection efficiency value corresponding to each molecule in the molecular pair and the actual value of LNP transfection efficiency, the LNP transfection efficiency loss corresponding to the molecular pair is calculated;

[0017] Based on the differential structural feature vectors between the two molecules in the molecular pair, the differential value of LNP transfection efficiency corresponding to the molecular pair is predicted;

[0018] Obtain the true difference value of LNP transfection efficiency corresponding to the molecular pair;

[0019] Based on the predicted difference value of LNP transfection efficiency corresponding to the molecular pair and the actual difference value of LNP transfection efficiency, the difference value loss of LNP transfection efficiency corresponding to the molecular pair is calculated;

[0020] The LNP transfection efficiency loss and the LNP transfection efficiency difference loss corresponding to each molecule pair are input into the pre-established model loss function and the hyperparameters in the model loss function are adjusted to obtain the model loss value. The hyperparameters include the feature vector dimension, the number of encoding layers, and the number of heads in the attention network.

[0021] When the model loss value is the smallest, the corresponding model is determined to be the LNP transfection efficiency prediction model.

[0022] In an optional implementation, each molecule in the molecule pair is subjected to word segmentation encoding processing to determine a molecular feature vector corresponding to each molecule, including:

[0023] The AIS word segmentation method is used to segment each molecule in the molecule pair to obtain the AIS word segmentation vector corresponding to each molecule;

[0024] Each AIS word segmentation vector is randomly embedded to obtain the AIS word segmentation vector matrix;

[0025] Perform position encoding on the AIS word segmentation vector matrix to obtain the AIS word segmentation vector embedding matrix;

[0026] The AIS word segmentation vector embedding matrix is ​​used as the input of the bidirectional conversion encoder and passes through multiple encoding layers in sequence to obtain the molecular feature vector corresponding to the molecule.

[0027] In an optional embodiment, the bidirectional conversion encoder is composed of a stack of encoder layers, each encoder including two sub-layer connection structures; wherein the first sub-layer connection structure includes a multi-head self-attention layer, a normalization and a residual connection layer; and the second sub-layer connection structure includes a feedforward fully connected layer, a normalization and a residual connection layer.

[0028] In an optional embodiment, determining the differential structural feature vector corresponding to each molecule based on the differential structural mask corresponding to each molecule and the molecular feature vector includes: multiplying the differential structural mask corresponding to each molecule by the molecular feature vector to obtain the differential structural feature vector corresponding to each molecule;

[0029] Based on the one-to-one difference structure feature vectors of each molecule in the molecule pair, the difference structure feature vectors between two molecules in the molecule are obtained, including: subtracting the one-to-one difference structure feature vectors of the two molecules in the molecule pair to obtain the difference structure feature vectors between the two molecules in the molecule.

[0030] In an optional implementation, the model loss function is:

[0031] Loss = (1-λ) Loss single +λLoss pair

[0032] Among them, Loss is the model loss value, λ is the hyperparameter, and Loss single is the LNP transfection efficiency loss corresponding to the molecular pair, Loss pair It is the difference loss of LNP transfection efficiency corresponding to the molecular pair.

[0033] In a second aspect, the present invention provides a device for constructing a LNP transfection efficiency prediction model, the device comprising:

[0034] An acquisition module, used for acquiring an ionizable lipid molecule dataset and constructing a molecule pair based on the ionizable lipid molecule dataset;

[0035] An intermolecular difference structure extraction module is used to determine the common structure and difference structure between two molecules in a molecular pair, and generate a one-to-one difference structure mask for each molecule in the molecular pair;

[0036] A molecular feature extraction module is used to perform word segmentation encoding processing on each molecule in the molecular pair to determine a molecular feature vector corresponding to each molecule;

[0037] An intermolecular difference feature calculation module, used to determine the difference structure feature vector corresponding to each molecule based on the one-to-one difference structure mask and the molecular feature vector of each molecule;

[0038] The LNP transfection efficiency prediction module is used to train the model based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair, so as to construct the LNP transfection efficiency prediction model.

[0039] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the method for constructing a LNP transfection efficiency prediction model of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0040] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the method for constructing a LNP transfection efficiency prediction model according to the first aspect or any corresponding embodiment thereof.

[0041] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions, wherein the computer instructions are used to enable a computer to execute the method for constructing a LNP transfection efficiency prediction model according to the first aspect or any corresponding embodiment thereof.

[0042] The beneficial effects of the present invention are as follows:

[0043] According to a method for constructing a LNP transfection efficiency prediction model provided by the present invention, a deep learning model for predicting LNP transfection efficiency based on intermolecular differential structure comparison learning is constructed. The model uses the AIS word segmentation method to segment ionizable lipid molecules, so that the chemical environment information of the branched atoms of the ionizable lipid molecules is fully characterized. In addition, the model can extract the differential structure between ionizable lipid molecules, learn the relationship between the molecular structure itself and the LNP transfection efficiency, and the relationship between the differential structure and the LNP transfection efficiency difference. The model fully considers the particularity and similarity of ionizable lipid molecules, and effectively improves the model's LNP transfection efficiency prediction accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 is a schematic diagram of a process for constructing a LNP transfection efficiency prediction model according to an embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of a flow chart of difference structure mask calculation according to an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of an AIS word segmentation method according to an embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of determining a difference structure feature vector according to an embodiment of the present invention;

[0049] Figure 5 is a structural block diagram of a device for constructing a prediction model for LNP transfection efficiency according to an embodiment of the present invention;

[0050] Figure 6 is a schematic diagram of the structure of a bidirectional Transformer encoder according to an embodiment of the present invention;

[0051] Figure 7 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0053] Compared with other small molecule compounds, ionizable lipid molecules have longer side chains and fewer backbone structures. Their side chain structures are mainly aliphatic hydrocarbons. Different atoms in the side chains have similar chemical environments. The common SMILES code representation method is difficult to represent the differences in atomic environments in different side chains. In addition, in general, ionizable lipid molecules are generated by component chemistry methods such as three-component reaction (3-CR) and four-component reaction (4-CR), which makes different ionizable lipid molecules have relatively similar structures. The difference between different lipid molecular structures has become the key to the difference in transfection efficiency.

[0054] However, previous studies have not conducted in-depth research on these differential structures. How to construct a suitable deep learning or machine learning model based on the structural characteristics of ionizable lipid molecules to improve the prediction of LNP transfection efficiency is an urgent problem to be solved in this field.

[0055] In view of this, according to an embodiment of the present invention, an embodiment of a method for constructing a LNP transfection efficiency prediction model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0056] In this embodiment, a method for constructing a LNP transfection efficiency prediction model is provided, which can be executed by a server, a terminal, a mobile terminal, etc. Figure 1 : is a flow chart of a method for constructing a LNP transfection efficiency prediction model according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0057] Step S101, obtaining an ionizable lipid molecule dataset, and constructing molecule pairs based on the ionizable lipid molecule dataset.

[0058] As mentioned above, the structural characteristics of ionizable lipid molecules have a very important influence on the transfection efficiency of LNP. This embodiment takes the ionizable lipid molecules in the AGILE (a database for studying and optimizing LNP design) data set as an example. The AGILE data set can be divided into a training set, a validation set, and a test set at a ratio of 7:2:1 using a random partitioning method, and randomly constructs molecular pairs in the training set.

[0059] Step S102, determining the common structure and the difference structure between two molecules in the molecule pair, and generating a one-to-one difference structure mask for each molecule in the molecule pair.

[0060] The tools for calculating the common structure of a molecular pair include but are not limited to open source toolkits such as RDKit. The common structure and differential structure of a molecular pair are calculated, and a differential structure mask is generated for each molecule in the molecular pair. The calculation process of the differential structure mask between molecule 1 and molecule 2 in the molecular pair is as follows: Figure 2 As shown. The differential structures of the two molecules in the molecular pair are calculated and annotated, and then a differential structure mask is generated for each molecule. The length of the differential structure mask is consistent with the length of the AIS word segmentation vector of the corresponding molecule. The mask corresponding to the common structure is 0, and the mask corresponding to the differential structure is 1.

[0061] Step S103, performing word segmentation coding processing on each molecule in the molecule pair to determine a one-to-one corresponding molecule feature vector for each molecule.

[0062] In this embodiment, the AIS (atomin SMILES) word segmentation method can be used to perform word segmentation processing on each molecule in the molecule pair, and a bidirectional conversion encoder (bidirectional Transformer encoder) is used to encode the molecule into a molecule feature vector.

[0063] Among them, the AIS word segmentation method is to divide molecules into atoms and other symbols according to SMILES (a linear representation method for describing chemical molecules), and encode the chirality information of atoms, the number of connected hydrogen atoms and the connected atoms in atoms. Other symbols include chemical bonds, branch chain information, etc. Figure 3 As shown, the AIS segmentation method is used to segment the ionizable lipid molecules. Compared with the traditional SMILES segmentation method, the chemical environment information of the branched atoms of the ionizable lipid molecules is more fully characterized.

[0064] Step S104, determining the difference structure feature vector corresponding to each molecule based on the one-to-one difference structure mask and the molecule feature vector of each molecule.

[0065] Reference Figure 4As shown, each word segmented by the AIS word segmentation method is used as the input string of the bidirectional conversion encoder. After the bidirectional conversion encoder obtains the molecular feature vector of each molecule, the molecular feature vector of each molecule is multiplied by the difference structure mask to obtain the corresponding difference structure feature vector.

[0066] Step S105, based on the one-to-one corresponding difference structure feature vectors of each molecule in the molecule pair, obtain the difference structure feature vectors between the two molecules in the molecule pair.

[0067] The calculation of the difference characteristics between molecules can be carried out by subtraction, that is, the numerical values ​​of the corresponding positions of the difference structure feature vectors are subtracted, so as to determine the difference structure feature vector between the two molecules and determine the difference between the molecular structure feature vectors.

[0068] Step S106, training the model based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair to construct a LNP transfection efficiency prediction model.

[0069] In this embodiment, the LNP transfection efficiency prediction module can be used to calculate the LNP transfection efficiency prediction value according to the molecular feature vector corresponding to each molecule, and calculate the LNP transfection efficiency prediction difference value according to the difference structural feature vector between the two molecules in the molecule pair. Then, the model is trained according to the LNP transfection efficiency prediction value and the LNP transfection efficiency prediction difference value.

[0070] Among them, the LNP transfection efficiency prediction module adopts a fully connected neural network, the input size of the fully connected neural network is the feature vector dimension, the output size is 1, and the fully connected neural network consists of a linear layer, a GLUE activation function, and a Dropout layer. During the training process, the LNP transfection efficiency prediction module optimizer can use an Adam optimizer (Adaptive MomentEstimation, referred to as Adam optimizer).

[0071] After the above training is completed, the validation set and test set can be further used for verification and testing. The optimal model is selected and saved according to the loss value of the validation set. Finally, in the model inference stage, the data of the test set is input into the optimal model for final evaluation, thereby obtaining the optimal LNP transfection efficiency prediction model.

[0072] The present invention provides a deep learning model for predicting LNP transfection efficiency based on comparative learning of intermolecular difference structures. The model uses the AIS segmentation method to segment ionizable lipid molecules, so that the chemical environment information of the branched atoms of the ionizable lipid molecules is fully characterized. In addition, the model can extract the difference structure between ionizable lipid molecules, learn the relationship between the molecular structure itself and the LNP transfection efficiency, and the relationship between the difference structure and the difference in LNP transfection efficiency. The model fully considers the particularity and similarity of ionizable lipid molecules, and effectively improves the prediction accuracy and generalization ability of the model for LNP transfection efficiency.

[0073] In some optional embodiments, the above step S106, i.e., training the model based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair to construct the LNP transfection efficiency prediction model, comprises:

[0074] Step S1061, based on the molecular feature vector corresponding to each molecule in the molecular pair, predict the LNP transfection efficiency prediction value corresponding to each molecule.

[0075] Step S1062, obtaining the true value of the LNP transfection efficiency corresponding to each molecule in the molecule pair.

[0076] Step S1063, based on the predicted LNP transfection efficiency corresponding to each molecule in the molecule pair and the actual LNP transfection efficiency, calculate the LNP transfection efficiency loss corresponding to the molecule pair.

[0077] After obtaining the molecular feature vector of each molecule, the molecular feature vector of each molecule is respectively input into the LNP transfection efficiency prediction module, and the LNP transfection efficiency prediction module is composed of a fully connected neural network. In this embodiment, the LNP transfection efficiency prediction module is composed of 3 linear layers, and the GLUE activation function and the Dropout layer are respectively connected after the first linear layer and the second linear layer. The input dimension of the first linear layer is set to the dimension of the feature vector, and the input dimension is 1024 dimensions. The input dimension and output dimension of the second linear layer are both 1024 dimensions. The input dimension of the last linear layer is 1024 dimensions, and the input dimension is 1. The output value of the LNP transfection efficiency prediction module is the LNP transfection efficiency prediction value.

[0078] Furthermore, the LNP transfection efficiency loss (Loss single ), where the loss function is as follows:

[0079] Loss single =||y i -p i || 2+||y j -p j || 2

[0080] Among them, y i and j are the true values ​​of LNP transfection efficiency of the two molecules in the molecular pair, p i and p j They are the predicted values ​​of LNP transfection efficiency of the two molecules in the molecular pair.

[0081] Step S1064, based on the difference structural feature vectors between the two molecules in the molecule pair, predict the difference value of LNP transfection efficiency corresponding to the molecule pair.

[0082] Step S1065, obtaining the true difference value of LNP transfection efficiency corresponding to the molecule pair.

[0083] Step S1066, based on the predicted difference value of LNP transfection efficiency corresponding to the molecule pair and the actual difference value of LNP transfection efficiency, calculate the difference value loss of LNP transfection efficiency corresponding to the molecule pair.

[0084] After obtaining the difference structural feature vectors between the molecules in the molecular pair, the difference structural feature vectors between the molecules are also input into the LNP transfection efficiency prediction module to obtain the predicted difference value of the LNP transfection efficiency. Further, the loss of the difference value of the LNP transfection efficiency between the predicted difference value of the LNP transfection efficiency and the actual difference value of the LNP transfection efficiency is calculated according to the loss function. pair ), where the loss function is as follows:

[0085] Loss pair =||Δy ij -Δp ij || 2

[0086] Among them, Δy ij is the true difference in LNP transfection efficiency, Δp ij Predict differential values ​​for LNP transfection efficiency.

[0087] Step S1067, input the LNP transfection efficiency loss and the LNP transfection efficiency difference loss corresponding to each molecule pair into the pre-established model loss function and adjust the hyperparameters in the model loss function to obtain the model loss value, where the hyperparameters include the feature vector dimension, the number of encoding layers, and the number of heads of the attention network.

[0088] The model loss function in this embodiment includes the LNP transfection efficiency loss of each molecule (Loss single ) and the difference in LNP transfection efficiency between the two molecular pairs (Loss pair). In some optional implementations, the model loss function is:

[0089] Loss = (1-λ) Loss single +λLoss pair

[0090] Among them, Loss is the model loss value, λ is the hyperparameter, and Loss single is the LNP transfection efficiency loss corresponding to the molecular pair, Loss pair The difference in LNP transfection efficiency between the two molecular pairs.

[0091] Among them, λ is a hyperparameter that needs to be optimized and determined. The hyperparameters that need to be optimized include the feature vector dimension, the number of encoding layers in the bidirectional Transformer encoder, and the number of heads in the attention network. In this embodiment, the learning rate is set in a gradually decaying manner. As the number of training rounds increases, the learning rate gradually decays. The specific decay strategy should be optimized according to the specific task. The model is optimized according to the loss function using the Adam optimizer. In this embodiment, the initial learning rate is set to 0.005. As the number of training rounds increases, the learning rate gradually decays, and it is set to decay to 0.97 times the original learning rate every 5 training rounds. Use the training set to optimize the model parameters, and use the validation set to verify the model training effect after each training round.

[0092] Step S1068, when the model loss value is the smallest, determine the corresponding model as the LNP transfection efficiency prediction model.

[0093] Reference Figure 5 As shown, in this embodiment, a deep learning method for predicting LNP transfection efficiency based on molecular difference structure contrast learning includes a model training stage and a model reasoning stage, wherein the model reasoning stage also includes: molecular feature extraction, using the AIS word segmentation method to process the test set molecules, each word after word segmentation is used as the input of the bidirectional Transformer encoder, and the molecular feature vector of the molecule is obtained through the bidirectional Transformer encoder. Then, the LNP transfection efficiency prediction module is used to calculate the LNP transfection efficiency according to the molecular feature vector. Through model reasoning, the generalization ability and practical application performance of the model are further evaluated to ensure the accuracy and reliability of the model. The LNP transfection efficiency prediction module in the model reasoning stage is consistent with the LNP transfection efficiency prediction module in the model training stage.

[0094] The deep learning model for predicting LNP transfection efficiency based on molecular structure contrast learning provided in this embodiment can extract the difference structure between ionizable lipid molecules, learn the relationship between the molecular structure and LNP transfection efficiency, and the relationship between the difference structure and the difference in LNP transfection efficiency. The model fully considers the particularity and similarity of ionizable lipid molecules, and effectively improves the accuracy and generalization ability of the model's prediction of LNP transfection efficiency.

[0095] In some optional embodiments, the above step S103, i.e., performing word segmentation encoding processing on each molecule in the molecule pair to determine a one-to-one corresponding molecule feature vector for each molecule, includes:

[0096] Step S1031, using the AIS word segmentation method to perform word segmentation processing on each molecule in the molecule pair, and obtain an AIS word segmentation vector corresponding to each molecule.

[0097] Step S1032: randomly embed each AIS word segmentation vector to obtain an AIS word segmentation vector matrix.

[0098] Step S1033, position encoding is performed on the AIS word segmentation vector matrix to obtain an AIS word segmentation vector embedding matrix.

[0099] When encoding positions, the length of the molecule is not limited, which is the maximum length of the input molecule. For molecules shorter than the maximum length, the input word vector can be padded with the special number 0, and the attention weight of the special number 0 is set to 0. The dimension of the input feature vector is not limited, and the optimal feature vector dimension is finally obtained through optimization for a specific task.

[0100] Step S1034, the AIS word segmentation vector embedding matrix is ​​used as the input of the bidirectional conversion encoder, and passes through multiple encoding layers in sequence to obtain the molecule feature vector corresponding to the molecule.

[0101] In some optional embodiments, the bidirectional conversion encoder is composed of a stack of encoder layers, each encoder including two sub-layer connection structures; wherein the first sub-layer connection structure includes a multi-head self-attention layer, a normalization and residual connection layer; and the second sub-layer connection structure includes a feed-forward fully connected layer, a normalization and residual connection layer. It should be noted that the number of encoder layers and the number of heads in the multi-head attention layer should be optimized according to the specific task.

[0102] The structural diagram of the bidirectional Transformer encoder is shown in the figure below. Figure 6As shown, the input is the AIS word segmentation vector, using a random embedding method, and adding position encoding to the random embedding. After passing through the encoding layer consisting of a head self-attention layer, a normalization and residual connection layer, a feedforward full connection layer, a normalization and residual connection layer, multiple encoding layers are stacked to form an encoder. In this embodiment, the number of encoding layers is 4, the number of heads in the multi-head self-attention layer is 4, the feature vector dimension is 64, and the output of the encoder is a molecular feature vector. The AIS word segmentation method, the bidirectional Transformer encoder and the feature vector are consistent.

[0103] The AIS word segmentation method is different from the SMILES word segmentation method. In the SMILES word segmentation method, the representation of each atom is only related to the type of atom. In the AIS word segmentation method, the representation of each atom is related to the type of atom, the number of hydrogen atoms connected to the atom, the chirality of the atom, and the atoms connected to the atom. The SMILES word segmentation method cannot distinguish between different chemical environments of atoms of the same type.

[0104] In this embodiment, the AIS word segmentation method is used to segment the ionizable lipid molecules so that the chemical environment information of the branched atoms of the ionizable lipid molecules can be more fully characterized. Using the AIS word segmentation vector as the input of the bidirectional conversion encoder can fully capture the contextual information of the molecular structure and improve the accuracy of feature representation. The stacking of multiple encoding layers enhances the expressive power of the model, enabling it to better learn complex molecular features, thereby improving the predictive performance of the model.

[0105] In some optional embodiments, determining the differential structural feature vector corresponding to each molecule based on the differential structural mask corresponding to each molecule and the molecular feature vector includes: multiplying the differential structural mask corresponding to each molecule by the molecular feature vector to obtain the differential structural feature vector corresponding to each molecule;

[0106] Based on the one-to-one difference structure feature vectors of each molecule in the molecule pair, the difference structure feature vectors between two molecules in the molecule are obtained, including: subtracting the one-to-one difference structure feature vectors of the two molecules in the molecule pair to obtain the difference structure feature vectors between the two molecules in the molecule.

[0107] In this example, in order to verify the feasibility and excellence of the deep learning model and method for predicting LNP transfection efficiency based on molecular difference structure comparative learning of the present invention, the following experiments were performed:

[0108] In this experiment, the model of the present invention is compared with other models for predicting LNP transfection efficiency in related technologies. Other models and methods include traditional random forest model (RF), LightGBM model, AGILE model and TransLNP model. Among them, the random forest model (RF) and LightGBM model use machine learning algorithms and take molecular descriptors as input. The AGILE model uses a pre-trained graph neural network algorithm to encode molecules into molecular graphs. TransLNP is a deep learning model based on Transformer, which learns the three-dimensional structural features of molecules and uses data balancing and other means to achieve a mean square error of 5 on the AGILE data set. It is currently the model with the best performance in predicting LNP transfection efficiency.

[0109] The model of the present invention and other models are evaluated in the AGILE dataset. The evaluation indicators include mean square error (MES) and Pearson correlation coefficient (PCC). The results are shown in Table 1. The model of the present invention has the lowest MSE and the highest PCC value, which is significantly improved compared with the existing models.

[0110] Table 1 Comparison of results of different models

[0111] Model MSE↓ PCC↑ RF 5.73 0.66 LightGBM 7.10 0.56 AGELE 7.41 0.51 TransLNP 5.17 0.69 The present invention 4.81 0.73

[0112] Note: ↓ means the lower the index is, the higher the model performance is; ↑ means the higher the index is, the higher the model performance is.

[0113] In this embodiment, in order to test the contribution of the AIS word segmentation method and the molecular difference structure comparison learning method to the model of the present invention, the following ablation test was performed:

[0114] The model based on intermolecular difference structure contrast learning and using the AIS word segmentation method was used as the baseline model to test the effects of removing different modules (AIS word segmentation method and contrast learning method based on contrast intermolecular difference structure) on the model performance. The results are shown in Table 2. Without using the AIS word segmentation method, the model's root mean square error RMSE increased to 2.748 and PCC decreased to 0.524. Without contrast learning, the model's RMSE increased to 2.281 and PCC decreased to 0.700. This indicates that the use of the AIS word segmentation method and the contrast learning method based on intermolecular difference structure proposed in the present invention both contribute to the improvement of the model.

[0115] Table 2 Model ablation experiment results

[0116] Model RMSE PCC Baseline model (AIS word segmentation method + contrastive learning) 2.194±0.097 0.730±0.023 SMILES word segmentation method + contrastive learning 2.748±0.139 0.524±0.058 AIS word segmentation method (no contrastive learning) 2.281±0.068 0.700±0.030

[0117] In this embodiment, a device for constructing a prediction model for LNP transfection efficiency is also provided, and the device is used to implement the above-mentioned embodiments and preferred embodiments, and the descriptions that have been made will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0118] This embodiment provides a device for constructing a LNP transfection efficiency prediction model. Figure 5 As shown, the device comprises:

[0119] An acquisition module, used for acquiring an ionizable lipid molecule dataset and constructing a molecule pair based on the ionizable lipid molecule dataset;

[0120] An intermolecular difference structure extraction module is used to determine the common structure and difference structure between two molecules in a molecular pair, and generate a one-to-one difference structure mask for each molecule in the molecular pair;

[0121] A molecular feature extraction module is used to perform word segmentation encoding processing on each molecule in the molecular pair to determine a molecular feature vector corresponding to each molecule;

[0122] An intermolecular difference feature calculation module, used to determine the difference structure feature vector corresponding to each molecule based on the one-to-one difference structure mask and the molecular feature vector of each molecule;

[0123] The LNP transfection efficiency prediction module is used to train the model based on the molecular feature vectors corresponding to each molecule in the molecular pair and the differential structural feature vectors between the two molecules in the molecular pair, so as to construct the LNP transfection efficiency prediction model.

[0124] In some optional embodiments, the LNP transfection efficiency prediction module comprises:

[0125] The LNP transfection efficiency prediction unit is used to predict the predicted value of LNP transfection efficiency corresponding to each molecule in the molecular pair based on the molecular feature vector corresponding to each molecule in the molecular pair; obtain the true value of LNP transfection efficiency corresponding to each molecule in the molecular pair; calculate the loss of LNP transfection efficiency corresponding to the molecular pair based on the predicted value of LNP transfection efficiency corresponding to each molecule in the molecular pair and the true value of LNP transfection efficiency; predict the predicted difference value of LNP transfection efficiency corresponding to the molecular pair based on the difference structural feature vector between two molecules in the molecular pair; obtain the true difference value of LNP transfection efficiency corresponding to the molecular pair; calculate the loss of LNP transfection efficiency difference value corresponding to the molecular pair based on the predicted difference value of LNP transfection efficiency corresponding to the molecular pair and the true difference value of LNP transfection efficiency; input the LNP transfection efficiency loss corresponding to each molecular pair and the LNP transfection efficiency difference value loss into a pre-established model loss function and adjust the hyperparameters in the model loss function to obtain the model loss value, the hyperparameters including the feature vector dimension, the number of encoding layers and the number of heads of the attention network; when the model loss value is minimized, determine that the corresponding model is the LNP transfection efficiency prediction model.

[0126] In an optional embodiment, the molecular feature extraction module includes:

[0127] The molecule feature extraction unit is used to perform word segmentation processing on each molecule in the molecule pair by using the AIS word segmentation method to obtain the AIS word segmentation vector corresponding to each molecule; each AIS word segmentation vector is randomly embedded to obtain the AIS word segmentation vector matrix; the AIS word segmentation vector matrix is ​​positionally encoded to obtain the AIS word segmentation vector embedding matrix; the AIS word segmentation vector embedding matrix is ​​used as the input of the bidirectional conversion encoder, and is sequentially passed through multiple encoding layers to obtain the molecule feature vector corresponding to the molecule.

[0128] The LNP transfection efficiency prediction model construction device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0129] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0130] The embodiment of the present invention also provides a computer device having the above Figure 5 The apparatus for building a model for predicting LNP transfection efficiency is shown.

[0131] See also Figure 7 , Figure 7 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 7As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 7 A processor 10 is taken as an example.

[0132] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0133] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0134] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0135] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0136] The computer device further comprises a communication interface 30 for the computer device to communicate with other devices or a communication network.

[0137] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0138] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0139] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for constructing a LNP transfection efficiency prediction model, characterized in that: The method comprises: Acquiring an ionizable lipid molecule dataset, and constructing molecule pairs based on the ionizable lipid molecule dataset; Determine the common structure and the difference structure between the two molecules in the molecular pair, and generate a one-to-one difference structure mask for each molecule in the molecular pair; Performing word segmentation encoding processing on each molecule in the molecule pair to determine a molecular feature vector corresponding to each molecule; Determine the difference structure feature vector corresponding to each molecule based on the difference structure mask and the molecule feature vector corresponding to each molecule; Based on the one-to-one corresponding difference structural feature vectors of each molecule in the molecular pair, obtaining a difference structural feature vector between two molecules in the molecule; The model is trained based on the molecular feature vector corresponding to each molecule in the molecular pair and the differential structural feature vector between the two molecules in the molecular pair to construct a LNP transfection efficiency prediction model.

2. The method according to claim 1, characterized in that The method of training the model based on the molecular feature vector corresponding to each molecule in the molecular pair and the differential structural feature vector between the two molecules in the molecular pair to construct a LNP transfection efficiency prediction model includes: Based on the molecular feature vectors corresponding to each molecule in the molecular pair, predicting the LNP transfection efficiency prediction value corresponding to each molecule; Obtaining the true value of LNP transfection efficiency corresponding to each molecule in the molecular pair; Based on the predicted LNP transfection efficiency value corresponding to each molecule in the molecule pair and the actual value of the LNP transfection efficiency, calculating the LNP transfection efficiency loss corresponding to the molecule pair; Based on the difference structural feature vectors between the two molecules in the molecular pair, predicting the difference value of LNP transfection efficiency corresponding to the molecular pair; Obtaining the true difference value of LNP transfection efficiency corresponding to the molecular pair; Based on the predicted difference value of the LNP transfection efficiency corresponding to the molecule pair and the actual difference value of the LNP transfection efficiency, calculating the difference value loss of the LNP transfection efficiency corresponding to the molecule pair; Input the LNP transfection efficiency loss and the LNP transfection efficiency difference value loss corresponding to each of the molecular pairs into a pre-established model loss function and adjust the hyperparameters in the model loss function to obtain a model loss value, wherein the hyperparameters include the feature vector dimension, the number of encoding layers, and the number of heads of the attention network; When the model loss value is the smallest, the corresponding model is determined to be the LNP transfection efficiency prediction model.

3. The method according to claim 1, characterized in that The performing word segmentation encoding processing on each molecule in the molecule pair to determine a one-to-one corresponding molecule feature vector for each molecule includes: Using the AIS word segmentation method to perform word segmentation processing on each molecule in the molecule pair, and obtain an AIS word segmentation vector corresponding to each molecule; Each of the AIS word segmentation vectors is randomly embedded to obtain an AIS word segmentation vector matrix; Performing position encoding on the AIS word segmentation vector matrix to obtain an AIS word segmentation vector embedding matrix; The AIS word segmentation vector embedding matrix is ​​used as the input of a bidirectional conversion encoder, and passes through multiple encoding layers in sequence to obtain the molecule feature vector corresponding to the molecule.

4. The method according to claim 3, characterized in that The bidirectional conversion encoder is composed of a stack of encoder layers, each encoder including two sub-layer connection structures; wherein the first sub-layer connection structure includes a multi-head self-attention layer, a normalization and residual connection layer; and the second sub-layer connection structure includes a feedforward fully connected layer, a normalization and residual connection layer.

5. The method according to claim 1, characterized in that The determining the difference structure feature vector corresponding to each molecule based on the difference structure mask corresponding to each molecule and the molecule feature vector comprises: multiplying the difference structure mask corresponding to each molecule by the molecule feature vector to obtain the difference structure feature vector corresponding to each molecule; The method of obtaining the difference structural feature vector between two molecules in the molecule based on the one-to-one difference structural feature vector of each molecule in the molecule pair comprises: subtracting the one-to-one difference structural feature vectors of the two molecules in the molecule pair to obtain the difference structural feature vector between the two molecules in the molecule.

6. The method according to claim 2, characterized in that The model loss function is: Loss=(1-λ)Loss single +λLoss pair Among them, Loss is the model loss value, λ is the hyperparameter, and Loss single is the LNP transfection efficiency loss corresponding to the molecule pair, Loss pair It is the difference value loss of LNP transfection efficiency corresponding to the molecular pair.

7. A device for constructing a LNP transfection efficiency prediction model, characterized in that: The device comprises: An acquisition module, used for acquiring an ionizable lipid molecule dataset and constructing a molecule pair based on the ionizable lipid molecule dataset; An intermolecular difference structure extraction module, used to determine the common structure and difference structure between two molecules in the molecular pair, and generate a one-to-one difference structure mask for each molecule in the molecular pair; A molecular feature extraction module is used to perform word segmentation coding processing on each molecule in the molecular pair to determine a molecular feature vector corresponding to each molecule; An intermolecular difference feature calculation module, used to determine the difference structure feature vector corresponding to each molecule based on the difference structure mask and the molecular feature vector corresponding to each molecule; The LNP transfection efficiency prediction module is used to train the model based on the molecular feature vector corresponding to each molecule in the molecular pair and the differential structural feature vector between the two molecules in the molecular pair to construct an LNP transfection efficiency prediction model.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for constructing a LNP transfection efficiency prediction model according to any one of claims 1 to 6 by executing the computer instructions.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method for constructing a LNP transfection efficiency prediction model according to any one of claims 1 to 6.

10. A computer program product, characterized in that The method comprises computer instructions, wherein the computer instructions are used to enable a computer to execute the method for constructing a LNP transfection efficiency prediction model according to any one of claims 1 to 6.

Images