Construction and training method, device, medium and program product of cationic lipid screening model

Abstract

The present application relates to a method, device, medium and program product for constructing and training a cationic lipid screening model. A first data set and a second data set containing the transfection efficiency of cationic lipids on tissues / cells are obtained, and a cationic lipid screening model including a molecular information encoding unit and a deep learning model is constructed. The molecular information encoding unit is used to generate a molecular structure word unit dictionary and an encoding vector of cationic lipid molecular information, and the encoding vector is input into the deep learning model; the molecular information encoding unit and the deep learning model are first jointly trained using the first data set so that the model learns the general rules of cationic lipid transfection efficiency; the second data set is used in combination with the first data set for a second joint training until a molecular structure word unit dictionary that is substantially the same as the molecular structure word unit dictionary after the first joint training is generated. The model of the present application has a stronger generalization ability, a more accurate prediction of transfection efficiency, a more targeted cationic lipid design, and a higher screening efficiency.

Description

Technical Field

[0001] The present application relates to the field of biomedical technology, and further to the field of cationic lipid design technology, and in particular to a method, device, medium and program product for constructing and training a cationic lipid screening model. Background Art

[0002] Cationic lipids are the decisive factors affecting mRNA transfection efficiency. The traditional method of using experiments to screen cationic lipids with high transfection efficiency is time-consuming, labor-intensive and costly.

[0003] By establishing an artificial intelligence deep learning model, it is expected to achieve rapid and high-throughput screening of cationic lipids. However, since the transfection efficiency of cationic lipids with different molecular structures on tissues / cells is also different, how to construct and train the artificial intelligence model in an appropriate way so that the trained model can be applied to a wider range of cationic lipid molecular structures and have more accurate prediction results for the transfection efficiency of cationic lipids, thereby effectively assisting in the screening of cationic lipids with high transfection efficiency, is a problem that the existing technology has not been able to solve. Summary of the invention

[0004] The present application is provided to solve the above-mentioned defects existing in the prior art. A method, device, medium and program product for constructing and training a cationic lipid screening model is needed, which can make the trained cationic lipid screening model applicable to a wider range of cationic lipid molecular structures and have more accurate prediction results for the transfection efficiency of cationic lipids of various molecular structures, thereby effectively assisting in screening out cationic lipids with high transfection efficiency.

[0005] According to the first scheme of the present application, a method for constructing and training a cationic lipid screening model is provided, comprising: obtaining a first data set comprising a first cationic lipid, wherein the first cationic lipid in the first data set is labeled with a transfection efficiency for a first tissue / cell; obtaining a second data set comprising a second cationic lipid, wherein the second cationic lipid in the second data set is labeled with a transfection efficiency for a second tissue / cell; and the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different; constructing a cationic lipid screening model comprising a cascaded molecular information encoding unit and a deep learning model, wherein the molecular information encoding unit is used to generate a molecular structure word unit dictionary, and based on the generated molecular structure word unit dictionary, a cationic lipid screening model is constructed. The method converts cationic lipid molecular information into a coding vector so that the converted coding vector can be used as input information of a deep learning model; using the first data set, a first joint training is performed on the molecular information coding unit and the deep learning model, and the accuracy of the transfection efficiency of the first cationic lipid to the first tissue / cell predicted by the deep learning model after the first joint training is higher than a first threshold; using the second data set, in combination with the first data set, a second joint training is performed on the molecular information coding unit and the deep learning model after the first joint training, and the deviation between the molecular structure word unit dictionary generated by the molecular information coding unit after the second joint training and the molecular structure word unit dictionary generated by the molecular information coding unit after the first joint training is lower than a second threshold.

[0006] According to the second scheme of the present application, a device for constructing and training a cationic lipid screening model is provided, comprising an interface and at least one processor. The interface is configured to obtain a first data set containing a first cationic lipid, wherein the first cationic lipid in the first data set is annotated with a transfection efficiency of a first tissue / cell; and, to obtain a second data set containing a second cationic lipid, wherein the second cationic lipid in the second data set is annotated with a transfection efficiency of a second tissue / cell, and the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different. At least one processor is configured to implement the construction and training methods of the cationic lipid screening model as described in each embodiment of the present application.

[0007] According to the third embodiment of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the construction and training methods of the cationic lipid screening model as described in the various embodiments of the present application.

[0008] According to the fourth embodiment of the present application, a computer program product is provided, comprising computer instructions, wherein the computer instructions are used to enable a computer to execute the method for constructing and training a cationic lipid screening model as described in the various embodiments of the present application.

[0009] The construction and training methods, devices, media and program products of the cationic lipid screening model provided in each embodiment of the present application construct a cationic lipid screening model including a cascaded molecular information encoding unit and a deep learning model, and successively use the first data set and the second data set to perform the first joint training and the second joint training of the molecular information encoding unit and the deep learning model on the constructed cationic lipid screening model, wherein after completing the first joint training, the cationic lipid screening model has successfully learned the relevant general rules of cationic lipid transfection efficiency, and thus can make a relatively accurate prediction of the transfection efficiency of the cationic lipids in the first data set. On this basis, the second data set with different cationic lipid molecular structures is used to improve the generalization ability of the model in terms of molecular structure and expand the scope of application of the model. However, the present application is different from other prior arts. In the second joint training process, attention is not only paid to the transfection efficiency value output by the deep learning model, but also to whether the molecular structure word-element dictionary used to encode the cationic lipid molecular information learned by the molecular information encoding unit is basically the same as the molecular structure word-element dictionary adopted by the molecular information encoding unit after the first joint training. Thus, through the training of the first joint training and the second joint training, not only the accuracy of the screening model in predicting the transfection efficiency of cationic lipids of various molecular structures is guaranteed, but also the intrinsic relationship between the molecular structure of cationic lipids and the transfection efficiency is learned. The trained cationic lipid screening model adopts a unified and stable molecular structure word-element dictionary for different cationic lipids, with better generalization ability and higher credibility of the prediction results, so that it can more effectively and reliably assist in screening out cationic lipids with high transfection efficiency from a variety of designed cationic lipids, and reduce the risk and cost of subsequent cationic lipid synthesis and testing.

[0010] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0011] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0013] Figure 1 A schematic diagram showing the process of constructing and training a cationic lipid screening model according to an embodiment of the present application.

[0014] Figure 2 A schematic diagram showing the composition of a molecular information encoding unit according to an embodiment of the present application is shown.

[0015] Figure 3 The first 20 initial segmentation word units and the corresponding initial segmentation word unit codes according to the embodiment of the present application are shown.

[0016] Figure 4 A schematic diagram showing concatenated word encoding from input SMILES representation to output molecular structure according to an embodiment of the present application.

[0017] Figure 5 A schematic diagram of the composition structure of a cationic lipid screening model according to an embodiment of the present application is shown.

[0018] Figure 6 A schematic diagram showing the calculation results of the explainability analysis according to an embodiment of the present application.

[0019] Figure 7 A schematic diagram showing the interpretability analysis results of the cationic lipid screening model according to an embodiment of the present application after the first joint training is completed.

[0020] Figure 8 A schematic diagram showing some components of a cationic lipid screening model and a training device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiment of the present application clearer, the technical solution of the embodiment of the present application will be clearly and completely described in conjunction with the drawings of the embodiment of the present application. Obviously, the described embodiment is a part of the embodiment of the present application, not all of the embodiments. Based on the described embodiment of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0022] Unless otherwise defined, the technical terms or scientific terms used in this application shall have the common meanings understood by persons with ordinary skills in the field to which this application belongs. "Include" or "comprising" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, but do not exclude other elements or objects.

[0023] The words "first", "second" and similar words used in this application do not indicate any order, quantity or importance, but are only used to distinguish. Words such as "include" or "comprise" and similar words mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also covering other elements. The execution order of the various steps in the method described in this application in conjunction with the accompanying drawings is not intended to be limiting. As long as it does not affect the logical relationship between the various steps, several steps can be integrated into a single step, a single step can be decomposed into multiple steps, and the execution order of the various steps can be changed according to specific needs.

[0024] It should also be understood that the term "and / or" in this application is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this application generally indicates that the associated objects before and after are in an "or" relationship.

[0025] In order to keep the following description of the embodiments of the present application clear and concise, the present application omits detailed descriptions of known functions and known components.

[0026] RNA transfection technology is the process of introducing RNA molecules such as mRNA into cells and expressing them. RNA transfection technology is mainly used for functional research of RNA, and the purpose of introducing mRNA into cells is to express encoded proteins and study gene functions and regulation. In this application, transfection efficiency refers to the transfection efficiency of the chemical structure of cationic lipids on mRNA in tissues or cells, wherein tissues may specifically include liver, lungs, spleen, etc., and cells may include antigen presenting cells, such as B cells, pDC cells, cDC cells, macrophages, etc., which are not listed here one by one.

[0027] Figure 1 A schematic diagram of the construction and training method of the cationic lipid screening model according to an embodiment of the present application is shown. Figure 1 As shown, in step 101, a first data set including a first cationic lipid may be first obtained, wherein the first cationic lipid in the first data set is annotated with a transfection efficiency for a first tissue / cell. In step 101, a second data set including a second cationic lipid may be further obtained, wherein the second cationic lipid in the second data set is annotated with a transfection efficiency for a second tissue / cell. In some embodiments, the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different.

[0028] In an embodiment of the present application, the data in the first data set is derived from data in the literature, including 1200 cationic lipid molecules and their transfection efficiency information (reference: Xu Y, Ma S, Cui H, et al. AGILEplatform: a deep learning powered approach to accelerate LNP development formRNA delivery[J]. Nature Communications, 2024, 15(1): 6305.), and the second data set may be, for example, other data sets different from the above data sources, or data accumulated through laboratory experiments. The present application does not limit this, as long as the cationic lipid molecules in the second data set are not included in the first data set.

[0029] Next, in step 102, a cationic lipid screening model is constructed including a cascaded molecular information encoding unit and a deep learning model, wherein the molecular information encoding unit is used to generate a molecular structure word-unit dictionary, and based on the generated molecular structure word-unit dictionary, the cationic lipid molecular information is converted into a coding vector, so that the converted coding vector can be used as input information of the deep learning model.

[0030] In some embodiments, the molecular information encoding unit may be integrated with a molecular structure word unit generation algorithm and a coding network capable of encoding molecular information based on molecular structure words, wherein the coding network may, for example, adopt any applicable neural network, a network structure based on Transformer, etc., which is not specifically limited in this application. The deep learning model may, for example, adopt a network structure based on Transformer or other any applicable convolutional neural network, which is not specifically limited in this application.

[0031] It is worth noting that the molecular information encoding unit and deep learning model in the embodiments of the present application both have trainable parameters, and their internal structures can also be adjusted as needed during the training process of the cationic lipid screening model.

[0032] Then, in step 103, the molecular information encoding unit and the deep learning model are first jointly trained using the first data set, and the accuracy of the transfection efficiency of the first cationic lipid to the first tissue / cell predicted by the deep learning model after the first joint training is higher than a first threshold.

[0033] In an embodiment of the present application, the main goal of the first joint training is to enable the cationic lipid screening model to learn that cationic lipids have a general internal law of different transfection efficiencies on tissues / cells. More specifically, it is expected that through the first joint training, the cationic lipid screening model can accurately establish the association between various words and transfection efficiency in the cationic lipid molecular structure, so that in the application stage of the cationic lipid screening model, the transfection efficiency of the cationic lipid with a new molecular structure designed for the user can be accurately predicted. In some embodiments, the transfection efficiency of the cationic lipid predicted by the model output can be expressed as a percentage of continuous values, or the transfection efficiency can be divided into different levels or categories from the perspective of practical application, for example, it can be divided into two categories of high transfection efficiency and low transfection efficiency, or it can be divided into multiple levels such as high transfection efficiency, predicted transfection efficiency, and predicted transfection efficiency. The application does not limit this, and it can be set as needed. In addition, the first threshold value should be set accordingly according to the specific setting of the transfection efficiency index, and the specific requirements for accuracy under the corresponding index system, which will not be repeated here.

[0034] In some embodiments, the amount of data in the first data set may be larger than the amount of data in the second data set, for example, to ensure that the cationic lipid screening model after the first joint training has the ability to accurately predict the transfection efficiency as described above.

[0035] Furthermore, in step 104, the second data set can be used in combination with the first data set to perform a second joint training on the molecular information encoding unit and the deep learning model after the first joint training, and the deviation between the molecular structure word unit dictionary generated by the molecular information encoding unit after the second joint training and the molecular structure word unit dictionary generated by the molecular information encoding unit after the first joint training is lower than a second threshold.

[0036] As mentioned above, the second cationic lipid in the second data set has a different molecular structure from the first cationic lipid in the first data set. Therefore, through the second joint training, the applicability of the cationic lipid screening model to cationic lipids with different molecular structures can be effectively improved, so that the trained model has stronger generalization ability in terms of cationic lipid molecular structure.

[0037] It is worth noting that, as mentioned above, during the second joint training process, the parameters and even the network structure of the molecular information encoding unit and the deep learning model may change, which may lead to unacceptable changes in the performance of the cationic lipid screening model on the first data set, or / in addition, it may also lead to the use of different molecular structure word-element dictionaries for the cationic lipids in the two data sets. This is an undesirable situation, because the present application does not want the constructed cationic lipid screening model to only be able to output the predicted transfection efficiency in a black box manner, but intends to construct an interpretable cationic lipid screening model so that it can understand the semantics represented by the cationic lipid molecular structure and gain insight into the influence and internal mechanism of different words in the molecular structure and the way they constitute a specific molecular structure on the transfection efficiency prediction results. Therefore, the inventors of the present application creatively proposed that in the second joint training, not only the second data set should be used, The first data set should also be considered in combination. For example, after performing a specified number of rounds of training using the second data set, the first data set can be used to perform a performance test on the cationic lipid screening model in the training process; further, not only should the performance of the model on the second data set be concerned (for example, so that the accuracy of the transfection efficiency of the second cationic lipid predicted by the deep learning model for the second tissue / cell is higher than the third threshold), but the molecular structure word unit dictionary generated by the molecular information encoding unit after the second joint training should also be ensured to be completely identical or sufficiently close to the molecular structure word unit dictionary generated by the molecular information encoding unit after the first joint training, for example, the deviation between the two is lower than the second threshold, so that the accuracy performance of the model can be ensured while not losing its interpretability, which can improve the robustness of the model to a certain extent, so that when it is applied to cationic lipids with new molecular structures, it can output more stable and reasonable prediction results. Among them, the calculation of the deviation between the molecular structure word unit dictionaries can be, for example, any applicable method such as cosine distance, and the setting of the second threshold can be set accordingly according to the deviation calculation method and the specific requirements of the model application, which are not listed here one by one.

[0038] According to the construction and training method of the cationic lipid screening model of the embodiment of the present application, by using the first data set and the second data set to perform the first joint training and the second joint training on the cationic lipid screening model, the trained cationic lipid screening model can not only predict the transfection efficiency of cationic lipids with different molecular structures, but also enable the model to maintain an understanding of the intrinsic relationship between different words in the molecular structure and the way in which they constitute a specific molecular structure and the transfection efficiency prediction results, that is, the model has good interpretability. Since the trained cationic lipid screening model uses a unified and stable molecular structure word dictionary for different cationic lipids, the generalization ability is better and the credibility of the prediction results is higher, so that it can more effectively and reliably assist in screening out cationic lipids with high transfection efficiency from the designed multiple cationic lipids, reducing the risks and costs of subsequent cationic lipid synthesis and testing.

[0039] Since the efficiency of some cationic lipids in transfecting mRNA in tissues / cells varies depending on the tissue / cell type, when the transfection efficiency of cationic lipids in different tissues / cells is significantly different, the trained cationic lipid screening model can also be made to have better generalization ability in tissues / cells by setting the training data set. For example, for the first data set and the second data set, in addition to the different cationic lipid molecular structures of the first cationic lipid and the second cationic lipid, the tissue / cell types of the first tissue / cell and the second tissue / cell in the data set may also be different. In this case, it is necessary to add a module for encoding the tissue / cell type, and then, after splicing / fusing the encoding vector corresponding to the cationic lipid molecule information with the encoding vector corresponding to the tissue / cell type, the method combined with Figure 1 A method similar to the construction and training method of the cationic lipid screening model described above can make the trained cationic lipid screening model have good generalization ability in both cationic lipid molecular structure and tissue / cell type, thereby further expanding the scope of application of the model.

[0040] Combine the following Figure 2 , Figure 3 and Figure 4 , which introduces in detail the composition and workflow of the molecular information encoding unit.

[0041] Figure 2 FIG. 2 shows a schematic diagram of the composition of the molecular information encoding unit according to an embodiment of the present application. Figure 2 As shown, in some embodiments, the molecular information encoding unit 200 includes a dictionary 1, a dictionary 2, and a dictionary 3 connected in series, and a coding network 202, and a molecular structure word unit generation algorithm is integrated in the molecular information encoding unit 200.

[0042] In some embodiments, the dictionary 1 is used to complete a one-way mapping from initial segmentation word units to initial segmentation word unit codes. That is, the dictionary 1 contains a one-to-one correspondence between initial segmentation word units and mapping numbers as word unit codes, wherein the key of the dictionary 1 is the initial segmentation word unit and the value is the mapping number (that is, the initial segmentation word unit code). In some embodiments, when constructing the dictionary 1, for example, the word units in the first data set and the second data set may be considered but not limited to.

[0043] Figure 3 The first 20 initial segmentation words and the corresponding initial segmentation word encodings according to the embodiment of the present application are shown. Figure 3 In the , the different initial segmentation word units obtained by segmentation are represented by the SMILES symbol system. Dictionary 1 also contains special completion word units <pad>and start word <cls>, where the special completion word <pad>The corresponding mapping number is 0, the starting word <cls>The corresponding mapping number is 1, the mapping number corresponding to the molecular structure CCCCCCCC is 2, and so on, which are not listed here one by one. By mapping the initial segmentation tokens to numbers, it is easier to generate spliced ​​tokens with longer lengths and capable of representing richer molecular structures in subsequent steps. In some embodiments, when selecting data in the second data set, it can be considered to make the initial segmentation tokens obtained based on the second cationic lipid segmentation and the initial segmentation tokens obtained based on the first cationic lipid segmentation have a large difference, so that after the second joint training, the cationic lipid screening model can have a wider applicability in terms of molecular structure.

[0044] Figure 2 The molecular structure word unit generation algorithm shown is used to splice words, that is, initial segmentation words or splicing words to generate new splicing words. The molecular structure word unit generation algorithm in the embodiment of the present application can be newly built from scratch, or an existing word unit merging algorithm can be used, etc. This application is not limited to this, as long as it can complete the word unit splicing generation function, and its parameters can be adjusted and optimized with the joint training of the cationic lipid screening model. As an example only, the molecular structure word unit generation algorithm in the embodiment of the present application can, for example, be based on the results of the dictionary 1 built by the first data set and the second data set, uniformly count the number of word unit occurrences in the two sets of data, and by setting the number of merging, continuously merge the corresponding numbers of the existing word units to generate new numbers.

[0045] Figure 2 The dictionary 2 shown, based on the molecular structure word unit generation algorithm, completes the one-way mapping of the initial segmentation word unit code or the splicing word unit code to the new splicing word unit. In other words, the dictionary 2 actually undertakes the splicing word unit interpretation function of merging and splicing the mapping numbers of the word units to generate new mapping numbers, and interpreting the newly generated mapping numbers one by one as new splicing word units. Through the dictionary 2, the fragmented molecular structure word units can be transformed into more continuous and more complex new word units, thereby better capturing the cationic lipid molecular structure, which is used for the subsequent interpretation of the relationship between the cationic lipid molecular structure word units and the transfection efficiency, and helps to design and optimize cationic lipid molecules with high transfection efficiency.

[0046] Figure 2 The dictionary 3 shown is used to complete the one-way mapping of the spliced ​​word to the spliced ​​word encoding, and the dictionary 3 will be used as the molecular structure word dictionary finally generated by the molecular information encoding unit, so that the molecular structure word dictionary can be used for the subsequent model processing of the semantic information between the cationic lipid molecular structures, thereby helping to accurately predict the transfection efficiency of cationic lipids. In addition, in the molecular structure word dictionary, the spliced ​​words as dictionary keys are sorted from long to short according to the length of the spliced ​​words.

[0047] Figure 4 Schematic diagram showing the concatenation word encoding from input SMILES representation to output molecular structure according to an embodiment of the present application. Figure 4 As shown, the left part is the input SMILES representation using Figure 2 The dictionary 1 shown maps numbers ( Figure 4 Marked as before merging), then the dictionary 2 uses the molecular structure word unit generation algorithm to perform word unit splicing, and the number corresponding to the newly generated spliced ​​word unit code is ( Figure 4 in the figure).

[0048] In some embodiments, the molecular information encoding unit further includes an encoding network, which is used to complete the unidirectional mapping of splicing words in the cationic lipid molecule information to splicing word encoding in order of priority from long to short according to the sorting of dictionary keys in dictionary 3, and generate an encoding vector of the cationic lipid molecule information. Among them, according to the order of the dictionary keys in dictionary 3, the unidirectional mapping of the spliced ​​words in the cationic lipid molecule information to the spliced ​​word encoding is completed in sequence according to the priority order of the spliced ​​words from long to short. In this way, the long words can be replaced with numbers first, and the short words can be replaced with numbers later. By replacing the long words first and then the short words, the words with complex and complete molecular structures can be identified and converted into numbers first, which is beneficial for the subsequent model to extract the semantic relationships between complex and complete molecular structure words, improve the model analysis and mining of the relationship between cationic lipid molecular structure words, and identify longer (semantically richer) words with positive / negative meanings for transfection efficiency in a more efficient way, which can better support the interpretation of the relationship between cationic lipid molecular structure and transfection efficiency, thereby greatly improving the efficiency and credibility of model-assisted cationic lipid molecule design.

[0049] In some embodiments, the word table containing the initial segmented word-grams can be constructed from scratch, or an existing public word-gram table can be used. If a public word-gram table is used, it is necessary to ensure that the encoding numbers corresponding to the newly generated concatenated word-grams in dictionary 2 and dictionary 3 are treated as an independent new word-gram during the interpretable analysis process, and cannot be segmented again into two or more word-grams, so as to ensure the accuracy of the model's interpretable analysis results.

[0050] Combine the following Figure 5 The structure of the cationic lipid screening model is further explained. Figure 5 The schematic diagram of the composition structure of the cationic lipid screening model according to the embodiment of the present application is shown. Figure 5 As shown, the cationic lipid screening model 500 includes at least a molecular information encoding unit 501 and a deep learning model 502, wherein the molecular information encoding unit 501 has been combined with Figure 2 It has been explained in detail and will not be repeated here.

[0051] Just as an example, Figure 5 The SMILES characterization of the chemical structure of the cationic lipids input into the cationic lipid screening model 500 is:

[0052] CCCCCCCCCCCCCCNC(=O)C(CCCCCOC(=O)CCCCCCCCC#C)NCCN(C(C)C)C(C)C, can be converted into the following 7 words:

[0053] ' <cls>', 'CCCCCCCCCCCCCCNC(=O)C(CCCCCOC(=O)', 'CCCCCCCCC#C)N ', 'CCN(', 'C(C)C)', 'C(C)C', ' <end>'.

[0054] As an example only, the deep learning model 502 can be constructed based on the Transformer encoder + classifier, for example. However, the difference from other Transformer-based deep learning models is that, in the present application, the normalization layer in the Transformer encoder is set before the multi-head self-attention layer + residual connection layer, and the normalization layer is also set before the fully connected layer + residual connection layer of the model, thereby greatly improving the stability of model training and alleviating the occurrence of gradient disappearance or gradient explosion. In addition, the deep learning model 502 further includes a hidden layer, which is connected to the classifier through the hidden layer, wherein the classifier includes an output layer for calculating and outputting the cationic lipid transfection efficiency prediction result. In the deep learning model 502, the Transformer encoder can be repeated multiple times, and the classifier after the Transformer encoder can be a convolutional neural network structure, or a fully connected layer structure or a recurrent neural network, etc., and the classifier can also be repeated multiple times, and the present application does not limit this, and can be set as needed.

[0055] In some embodiments, the cationic lipid screening model 500 may further include position embedding, so as to combine the position of each word in the cationic lipid molecular structure to learn a more accurate and diversified association between the word and the transfection efficiency.

[0056] In some embodiments, when the molecular information encoding unit and the deep learning model are subjected to the second joint training, the parameters and structure of the cationic lipid screening model after the first joint training can be modified. It is worth noting that when the screening model needs to be structurally modified, the model structure and model parameters of the deep learning model should be modified layer by layer from the output layer of the deep learning model, and the accuracy change of the screening model on the first data set and the second data set after each modification is tested, and finally the optimal modified model structure is selected as the structure of the screening model after the second joint training. Thus, it is ensured that after the second joint training, the cationic lipid screening model has the optimal model structure and model parameters.

[0057] The embodiments of the present application also provide an interpretable analysis of the cationic lipid screening model. After the first joint training is completed, for example, based on the parameters of the molecular information encoding unit and the deep learning model after the first joint training, and the category of the transfection efficiency prediction result output by the given deep learning model, the corresponding first analysis model can be used to perform an interpretable analysis on the relationship between different word units and the transfection efficiency of the first cationic lipid predicted by the deep learning model on the first tissue / cell, so as to obtain the analysis result of the first analysis model.

[0058] As an example only, when the transfection efficiency prediction results output by the deep learning model are divided into two categories: high transfection efficiency and low transfection efficiency: in the first analysis result obtained by using the first analysis model corresponding to the high transfection efficiency, the various words are sorted from high to low according to the degree of contribution to the high transfection efficiency of the cationic lipid; in the second analysis result obtained by using the first analysis model corresponding to the low transfection efficiency, the various words are sorted from high to low according to the degree of contribution to the low transfection efficiency of the cationic lipid; and, in combination with the sorting of the various words in the first analysis result and the second analysis result, the words in the new cationic lipid molecular structure are optimized and selected. The above-mentioned interpretable analysis model can not only intuitively show the degree of influence of different words (characterizing different cationic lipid molecular structures) on the transfection efficiency prediction results, but also optimize the design and structural modification of cationic lipids according to the results of the interpretable analysis, thereby helping to improve the design quality of cationic lipid molecules.

[0059] As an example only, the first analysis model adopted in the above-mentioned explainable analysis can be constructed based on the Shap framework. More specifically, the shap.Explainer function of the Shap framework can be used to calculate the SHAP value, and then the average value of the SHAP value can be calculated. The matplotlib tool can be used to draw a graph to show the effect of different terms in the cationic lipid chemical structure on the transfection efficiency.

[0060] Figure 6 A schematic diagram showing the calculation results of the explainability analysis according to an embodiment of the present application. Figure 6 The examples in the same Figure 5 The same cationic lipid chemical structure as Figure 6 The .base_values ​​and .values ​​shown in represent the SHAP values ​​corresponding to each word in the first analysis result calculated based on the 7 words using the first analysis model corresponding to high transfection efficiency.

[0061] Figure 7 A schematic diagram showing the interpretability analysis results of the cationic lipid screening model according to an embodiment of the present application after the first joint training is completed. Figure 7 The figure shows the influence of 10 cationic lipid molecular structure words on the high transfection efficiency in the cationic lipids in the first data set after the first joint training. The words are ranked from high to low in importance, which are: CCCC1, CCOCC1, C(C)C, CCCCC1, CCN(C)CC1, N1CCOCC1, CCN), N=, CC), and all other remaining words. The above analysis results provide a useful reference for users to design or adjust the molecular structure of cationic lipids in an intuitive way. For example, when generating new cationic lipid molecules, a higher selection weight is set for the top words. Generally, it can be predicted that such cationic lipids will also have a higher probability of having a higher transfection efficiency. In addition, an interpretable analysis can be provided from another perspective, that is, the second analysis result is obtained by using the first analysis model corresponding to the low transfection efficiency. Referring to the contribution of each word in the second analysis result to the low transfection efficiency of cationic lipids, when generating new cationic lipid molecules, these top words are set with a lower selection weight, and even these unfavorable words can be avoided, which may also provide a valuable reference for the design of cationic lipids.

[0062] Similarly, after the second joint training is completed, based on the parameters of the molecular information encoding unit and the deep learning model after the second joint training, and the category of the transfection efficiency prediction results output by the given deep learning model, the corresponding second analysis model can be used to perform an interpretable analysis on the relationship between different word units and the transfection efficiency of the second cationic lipid predicted by the deep learning model on the second tissue / cell, so as to obtain the analysis results of the second analysis model. It can be understood that the cationic lipid screening model after the second joint training has a stronger generalization ability, especially when the tissues / cells transfected by the first data set and the second data set are different, the interpretable analysis results performed at this time will also have reference value in a wider range, and are more suitable for high-throughput, multiple molecular structures, and the screening of cationic lipids for transfection of different tissues / cells.

[0063] It can be seen that the construction and training method of the cationic lipid screening model according to the embodiment of the present application is based on the interpretable analysis model, so that the cationic lipid screening model can not only accurately predict whether the new cationic lipid has a high transfection efficiency, but also reveal the relationship between the molecular structure of the cationic lipid and the transfection efficiency, and provide a reliable reference at the word level for users to design new cationic lipid molecules. In addition, in view of the principle of replacing long words first when using dictionary 3 to encode cationic lipid molecule information, the analysis results of the interpretable analysis model of the present application must also contain more long words, which will have a higher reference value for the design of cationic lipid molecules.

[0064] According to an embodiment of the present application, a device for constructing and training a cationic lipid screening model is also provided. Figure 8 A schematic diagram showing some components of a cationic lipid screening model and a training device according to an embodiment of the present application is shown.

[0065] like Figure 8 As shown, the device 800 at least includes an interface 801 and at least one processor 802 .

[0066] Among them, interface 801 can be configured, for example, to: obtain a first data set containing a first cationic lipid, wherein the first cationic lipid in the first data set is marked with a transfection efficiency for a first tissue / cell; and obtain a second data set containing a second cationic lipid, wherein the second cationic lipid in the second data set is marked with a transfection efficiency for a second tissue / cell; and the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different.

[0067] At least one processor 802, for example, can be configured to implement the steps in the method for constructing and training a cationic lipid screening model as described in the various embodiments of the present application as described above.

[0068] In some embodiments, at least one processor 802 may be a processing device including one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor running other instruction sets, or a processor running a combination of instruction sets. The processor may also be one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a system on a chip (SoC), etc.

[0069] According to an embodiment of the present application, a computer-readable storage medium is also provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the steps in the method for constructing and training a cationic lipid screening model as described in the various embodiments of the present application.

[0070] The above-mentioned computer-readable storage medium may be, for example, read-only memory (ROM), random access memory (RAM), phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), electrically erasable programmable read-only memory (EEPROM), other types of random access memory (RAM), flash disks or other forms of flash memory, cache, registers, static memory, compact disk read-only memory (CD-ROM), digital versatile disks (DVD) or other optical storage, cassettes or other magnetic storage devices, or any other possible non-transitory medium used to store information or instructions that can be accessed by a computer device.

[0071] According to an embodiment of the present application, a computer program product is also provided, which includes computer instructions, and the computer instructions are used to enable a computer to execute the steps in the method for constructing and training a cationic lipid screening model as described in each embodiment of the present application.

[0072] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present application with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims will be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the execution of the application, and the examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the claims and the full scope of their equivalents.

[0073] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more schemes thereof) may be used in combination with each other. For example, a person of ordinary skill in the art may use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features may be grouped together to simplify the present application. This should not be interpreted as an intention that a disclosed feature that is not required to be protected is necessary for any claim. On the contrary, the subject matter of the present application may be less than all the features of a particular disclosed embodiment. Thus, the claims are incorporated herein into the specific embodiments as examples or embodiments, wherein each claim is independently a separate embodiment, and it is considered that these embodiments may be combined with each other in various combinations or arrangements. The scope of the present application should be determined with reference to the claims and the full scope of equivalent forms granted by these claims.

[0074] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present application.< / end> < / cls> < / cls> < / pad> < / cls> < / pad>

Claims

1. Construction and training method of a cationic lipid screening model, It is characterized in that include: Acquire a first data set comprising a first cationic lipid, wherein the first cationic lipid in the first data set is annotated with a transfection efficiency for a first tissue / cell; acquire a second data set comprising a second cationic lipid, wherein the second cationic lipid in the second data set is annotated with a transfection efficiency for a second tissue / cell; and the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different; Constructing a cationic lipid screening model comprising a cascaded molecular information encoding unit and a deep learning model, wherein the molecular information encoding unit is used to generate a molecular structure word unit dictionary, and convert the cationic lipid molecular information into a coding vector based on the generated molecular structure word unit dictionary, so that the converted coding vector is used as input information of the deep learning model; Using the first data set, performing a first joint training on the molecular information encoding unit and the deep learning model, and making the accuracy of the transfection efficiency of the first cationic lipid on the first tissue / cell predicted by the deep learning model after the first joint training higher than a first threshold; Using the second data set and in combination with the first data set, the molecular information encoding unit and the deep learning model after the first joint training are subjected to a second joint training, and the deviation between the molecular structure word unit dictionary generated by the molecular information encoding unit after the second joint training and the molecular structure word unit dictionary generated by the molecular information encoding unit after the first joint training is lower than a second threshold.

2. The construction and training method according to claim 1, It is characterized in that The first tissue / cell is of a different tissue / cell type than the second tissue / cell.

3. The construction and training method according to claim 1 or 2, It is characterized in that The amount of data in the first data set is greater than the amount of data in the second data set.

4. The construction and training method according to claim 1 or 2, It is characterized in that The molecular information encoding unit includes a dictionary 1, a dictionary 2 and a dictionary 3 connected in series, and an encoding network, wherein: The dictionary 1 is used to complete the one-way mapping from the initial segmentation word unit to the initial segmentation word unit encoding; The dictionary 2 completes a one-way mapping of the initial segmentation word unit code or the splicing word unit code to the new splicing word unit based on the molecular structure word unit generation algorithm, wherein the molecular structure word unit generation algorithm is used to splice the initial segmentation word unit or the splicing word unit to generate a new splicing word unit; The dictionary 3 is used to complete the one-way mapping of concatenated word units to concatenated word unit codes, and serves as a molecular structure word unit dictionary generated by the molecular information encoding unit, and in the molecular structure word unit dictionary, the concatenated word units as dictionary keys are sorted from long to short according to the length of the concatenated word units; The encoding network is used to complete the unidirectional mapping of the splicing word units in the cationic lipid molecule information to the splicing word unit encoding in sequence according to the order of the dictionary keys in the dictionary 3 and the priority order of the splicing word units from long to short, and generate the encoding vector of the cationic lipid molecule information.

5. The construction and training method according to claim 1 or 2, It is characterized in that The deep learning network is constructed based on a Transformer encoder + a classifier, wherein a normalization layer in the Transformer encoder is set before a multi-head self-attention layer + a residual connection layer, and a normalization layer is set before a fully connected layer + a residual connection layer of the deep learning model.

6. The construction and training method according to claim 1 or 2, It is characterized in that The construction and training method further comprises: When the molecular information encoding unit and the deep learning model are subjected to the second joint training, the model structure and model parameters of the deep learning model are modified layer by layer from the output layer of the deep learning model upward, so that the deep learning model structure has the optimal model structure and model parameters after the second joint training is completed.

7. The construction and training method according to claim 1 or 2, It is characterized in that The construction and training method further comprises: Based on the parameters of the molecular information encoding unit and the deep learning model after the first joint training, and the category of the transfection efficiency prediction result output by the given deep learning model, using the corresponding first analysis model, an interpretable analysis is performed on the relationship between different word units and the transfection efficiency of the first cationic lipid on the first tissue / cell predicted by the deep learning model to obtain the analysis result of the first analysis model; Based on the parameters of the molecular information encoding unit and the deep learning model after the second joint training, and the category of the transfection efficiency prediction results output by the given deep learning model, the corresponding second analysis model is used to perform an interpretable analysis on the relationship between different word units and the transfection efficiency of the second cationic lipid on the second tissue / cell predicted by the deep learning model to obtain the analysis results of the second analysis model.

8. The construction and training method according to claim 7, It is characterized in that When the transfection efficiency prediction results output by the deep learning model are divided into two categories: high transfection efficiency and low transfection efficiency: In the first analysis result obtained by using the first analysis model corresponding to high transfection efficiency, each word is sorted from high to low according to the contribution degree to the high cationic lipid transfection efficiency; In the second analysis result obtained by using the first analysis model corresponding to low transfection efficiency, each word is sorted from high to low according to the contribution degree to the low transfection efficiency of the cationic lipid; and, The word units in the new cationic lipid molecular structure are optimized and selected by combining the ranking of each word unit in the first analysis result and the second analysis result.

9. Construction and training device of a cationic lipid screening model, It is characterized in that include: An interface is configured to obtain a first data set containing a first cationic lipid, wherein the first cationic lipid in the first data set is annotated with a transfection efficiency for a first tissue / cell; and obtain a second data set containing a second cationic lipid, wherein the second cationic lipid in the second data set is annotated with a transfection efficiency for a second tissue / cell; and the cationic lipid molecular structures of the first cationic lipid and the second cationic lipid are different; At least one processor is configured to implement the construction and training method of the cationic lipid screening model as described in any one of claims 1-8.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the construction and training method of the cationic lipid screening model as described in any one of claims 1-8.

11. A computer program product, It is characterized in that The method comprises computer instructions for causing a computer to execute the method for constructing and training a cationic lipid screening model as described in any one of claims 1 to 8.

Images