A method, device and equipment for predicting activity of chemically modified siRNA

Abstract

This invention discloses a method, apparatus, and device for predicting the activity of chemically modified siRNA, comprising: acquiring the original base sequence and chemical modification information of the target siRNA, and retrieving the physicochemical properties of the target siRNA; encoding the original base sequence, chemical modification information, and physicochemical properties as features; and generating a prediction result of the silencing efficiency of the target siRNA based on the encoded features using a pre-constructed prediction model; wherein the pre-constructed prediction model includes: a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on a cross-attention mechanism; and the classification sub-model is used to generate a prediction result of the silencing efficiency based on the fused features. This method, based on a multi-dimensional, multi-view learning strategy and an attention mechanism fusion model, significantly improves the algorithm's ability to represent chemically modified siRNA data, and, combined with the nonlinear data fitting ability of the deep learning framework, improves the accuracy of the algorithm in predicting the drug activity of chemically modified siRNA.

Description

Technical Field

[0001] This invention relates to the field of siRNA activity prediction technology, and in particular to a method, apparatus and equipment for predicting the activity of chemically modified siRNA. Background Technology

[0002] Small interfering RNA (siRNA) is a double-stranded non-coding RNA of 21-23 bases in length. In organisms, the RNA interference pathway (RNAipathway) mediated by siRNA can specifically degrade mRNA and silence genes. Therefore, siRNA has been designed as nucleotide drugs to specifically target disease-causing genes, achieving therapeutic effects at the mRNA level. Traditional drugs mainly fall into two categories: small molecule drugs and antibody drugs, both of which exert their therapeutic effects by targeting proteins. However, currently, fewer than 700 target proteins have been approved, severely limiting the development of traditional drugs. Compared with traditional drugs, siRNA drugs have advantages such as broad target range, high specificity, and short development cycle. In recent years, five siRNA drugs have been marketed, and several other therapies are in clinical trials. Despite the significant social and strategic importance of siRNA drugs, issues such as drug delivery, off-target effects, and immunotoxicity seriously affect drug activity and limit drug development and clinical application.

[0003] RNA modification refers to the chemical modification of RNA. Increasing evidence suggests that adding chemical modifications to the phosphate, ribose, or base pairs of siRNA can improve the stability, specificity, and safety of siRNA drugs, thereby enhancing their activity. Therefore, appropriate siRNA chemical modification is crucial for ensuring the drug-likeness of siRNA. Early chemical modification templates designed based on the prior knowledge of drug developers had a high failure rate in practical applications, resulting in significant time and financial costs. Utilizing machine learning and other methods to design high-performance models to assist in predicting the activity of chemically modified siRNA drugs has become a real need in drug development.

[0004] Many different methods have been developed to predict or design the activity of chemically modified siRNA drugs. Based on the specific models and strategies employed, existing methods are mainly divided into two categories (e.g., ...). Figure 1(as shown): (1) Rule-based chemical modification siRNA drug activity prediction method; (2) Machine learning-based chemical modification siRNA drug activity prediction method. Rule-based siRNA modification design mainly includes Standard template chemistry (STC), Enhanced stabilization chemistry (ESC), etc. These modification templates are designed by researchers based on modification rules observed from experimental data and are continuously optimized and updated after clinical validation (Friedrich, M. et al., BioDrugs 2022, 36, 549-571). This type of method that follows a priori rules has high targeting and specificity, making the synthesis of siRNA molecules simpler and more efficient. However, rule-based design schemes require complex calculations and analyses, which require a high level of technical expertise from researchers; at the same time, the observed chemical modification rules have cellular or tissue bias, resulting in a high failure rate in practical applications, requiring a large number of experiments and analyses.

[0005] Currently, there are only two machine learning-based methods for predicting siRNA chemical modifications: SMEpred (Dar, S.A. et al., RNABiol 2016, 13, 1144-1151), based on Support Vector Machine (SVM), and Dong et al. (Dong et al., Molecules 2022, 27), based on Partial Least Squares (PLS) regression. These methods learn the relationship between known siRNA molecules and drug activity (i.e., silencing efficiency) and then use this knowledge to predict the silencing efficiency of new siRNA molecules. These machine learning methods are highly flexible, unrestricted by rules, and can be automated, improving research efficiency. The datasets contain experimental data from different tissues and cells, resulting in high model robustness. However, both SMEpred and Dong et al. employ feature extraction to characterize the antisense strand of the siRNA and use it as the initial input layer of the model. SMEpred and Dong et al. considered only a limited set of features during feature extraction, ignoring features that affect drug activity, such as RNA sequence information. Therefore, the information extracted by the corresponding model lacked diversity, and the model effectiveness was limited by the researchers' choice of features. At the same time, the model lacked interpretability and could not capture the potential intrinsic characteristics of siRNA sequences.

[0006] In summary, existing predictive models cannot accurately provide stable, efficient, and safe siRNA drug production protocols. Rule-based methods are costly to research and require large-scale, multi-source wet experiments for analysis and validation; machine learning-based methods are still in their early stages and have limitations in feature encoding, the use and construction of machine learning algorithms, and model interpretability. Summary of the Invention

[0007] The purpose of this invention is to overcome the problem that existing machine learning-based siRNA activity prediction methods only consider a single feature when constructing the model, resulting in insufficient model interpretability and accuracy. This invention provides a method, device, and equipment for predicting the activity of chemically modified siRNA. By acquiring multi-dimensional features of siRNA and performing multi-dimensional feature fusion and joint prediction based on a pre-established prediction model, this invention significantly improves the model's ability to represent chemically modified siRNA data and enhances prediction accuracy.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A method for predicting the activity of chemically modified siRNA, comprising:

[0010] Obtain the original base sequence and chemical modification information of the target siRNA, and retrieve the physicochemical properties of the target siRNA based on the original base sequence;

[0011] The original base sequence, chemical modification information, and physicochemical properties are feature-encoded, and a pre-constructed prediction model is used to generate a prediction result of the silencing efficiency of the target siRNA based on the encoded features;

[0012] The pre-built prediction model includes a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on a cross-attention mechanism; and the classification sub-model is used to generate a silencing efficiency prediction result based on the fused features.

[0013] According to a specific implementation method, in the above-mentioned chemically modified siRNA activity prediction method, the chemical modification information includes: sense strand modification information and antisense strand modification information.

[0014] According to one specific implementation, in the above-mentioned method for predicting the activity of chemically modified siRNA, the physicochemical properties include: molecular weight, XLogP3, number of hydrogen bond donors, number of hydrogen bond acceptors, exact mass, mass of a single isotope, topological surface area, number of heavy atoms, complexity, and number of defined bond stereocenters.

[0015] According to a specific implementation, in the above-mentioned chemically modified siRNA activity prediction method, the classifier sub-model is a two-layer convolutional neural network, including: two convolutional sub-networks, two fully connected layers, and an output layer.

[0016] According to a specific implementation, in the above-mentioned chemically modified siRNA activity prediction method, each convolutional subnetwork includes: a convolutional layer, a ReLU activation function, and a max pooling layer.

[0017] According to one specific implementation, in the above chemically modified siRNA activity prediction method, the convolutional layer uses a 3x3 convolutional kernel with a stride of 1 and a padding of 1 pixel.

[0018] According to a specific implementation, in the above-mentioned method for predicting the activity of chemically modified siRNA, the prediction model is pre-constructed using the following methods:

[0019] An initial model for the prediction model is established, comprising: a feature fusion sub-model and a classification sub-model;

[0020] The original base sequences, chemical modification information, and silencing efficiencies of multiple chemically modified siRNA drugs were obtained from the siRNAmod database; their physicochemical properties were retrieved based on the original base sequence of each chemically modified siRNA drug.

[0021] Data samples were constructed based on the chemical modification information, physicochemical properties, and silencing efficiency of multiple chemically modified siRNA drugs, and a sample dataset was generated.

[0022] The initial model is trained based on the sample dataset, and the prediction model is obtained when the model meets the conditional convergence.

[0023] In another aspect, the present invention provides a device for predicting the activity of chemically modified siRNA, comprising:

[0024] The acquisition unit is used to acquire the original base sequence and chemical modification information of the target siRNA;

[0025] A retrieval unit is used to retrieve the physicochemical properties of the target siRNA based on the original base sequence;

[0026] The encoding unit is used to encode the original base sequence, chemical modification information and physicochemical properties.

[0027] A prediction unit is used to generate a prediction result of the silencing efficiency of the target siRNA based on the encoded features using a pre-built prediction model; wherein the pre-built prediction model includes: a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on a cross-attention mechanism; the classification sub-model is used to generate a prediction result of the silencing efficiency based on the fused features.

[0028] According to one specific implementation, the above-mentioned chemically modified siRNA activity prediction device further includes: a storage unit for storing the physicochemical properties of the base sequence;

[0029] The detection unit is configured to retrieve the physicochemical properties of the target siRNA from the storage unit based on the original base sequence.

[0030] In another aspect, the present invention provides an electronic device including a processor, a network interface, and a memory, wherein the processor, the network interface, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the above-described chemically modified siRNA activity prediction method.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. The method provided in this embodiment of the invention obtains the original base sequence and chemical modification information of the target siRNA, and retrieves the physicochemical properties of the target siRNA based on the original base sequence; then, it encodes the original base sequence, chemical modification information, and physicochemical properties as features, and generates a prediction result of the silencing efficiency of the target siRNA based on the encoded features according to a pre-built prediction model; this method is based on a multi-dimensional, multi-view learning strategy and a fusion model based on a cross-attention mechanism, which significantly improves the algorithm's ability to represent chemically modified siRNA data, and combined with the nonlinear data fitting ability of the deep learning framework, improves the algorithm's prediction accuracy of the drug activity of chemically modified siRNA;

[0033] 2. Based on the multi-dimensional characteristics of siRNA, a cross-attention mechanism model that is compatible with it is selected as the fusion model, which is more effective in realizing the feature fusion of chemically modified siRNA drugs in different views and improving prediction accuracy. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the existing siRNA activity prediction method described in the background section of this invention.

[0035] Figure 2This is a schematic diagram of the chemically modified siRNA activity prediction method in one embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the cross-attention mechanism fusion model structure in one embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the initial architecture of the prediction model in one embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram illustrating the application results of different algorithms on four FDA-approved drugs;

[0039] Figure 6 This is a schematic diagram of an online software tool based on the siRNA activity prediction device provided by the present invention;

[0040] Figure 7 This is a schematic diagram of an electronic device architecture in one embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0042] Example 1

[0043] Figure 2 A method for predicting the activity of chemically modified siRNA, as illustrated in an exemplary embodiment of the present invention, includes:

[0044] S1. Obtain the original base sequence and chemical modification information of the target siRNA, and retrieve the physicochemical properties of the target siRNA based on the original base sequence;

[0045] S2. Encode the original base sequence, chemical modification information, and physicochemical properties using feature encoding.

[0046] S3. Based on the pre-constructed prediction model, the silencing efficiency prediction result of the target siRNA is generated according to the chemical modification information and the physicochemical properties.

[0047] The pre-built prediction model includes a feature fusion sub-model and a classification sub-model. The feature fusion sub-model is used to fuse the chemical modification information with the physicochemical properties based on an attention mechanism. The classification sub-model is used to generate a silencing efficiency prediction result based on the fused features.

[0048] In this embodiment, the original base sequence and chemical modification information of the target siRNA are obtained, and the physicochemical properties of the target siRNA are retrieved based on the original base sequence. Then, the original base sequence, chemical modification information, and physicochemical properties are feature-encoded, and the silencing efficiency prediction result of the target siRNA can be generated based on the pre-built prediction model according to the encoded features. This method is based on a multi-dimensional, multi-view learning strategy and a cross-attention mechanism fusion model, which significantly improves the algorithm's ability to represent chemically modified siRNA data. Combined with the nonlinear data fitting ability of the deep learning framework, the algorithm improves the prediction accuracy of the drug activity of chemically modified siRNA.

[0049] Understandably, common attention mechanisms include: multi-head attention, hard attention, soft attention, key-value pair attention, and structured attention.

[0050] In this embodiment, based on the characteristics of chemically modified siRNA drugs across multiple views, a cross-attention mechanism is selected to construct a fusion sub-model. This more effectively facilitates the feature fusion of chemically modified siRNA drugs across different views, improving prediction accuracy.

[0051] Example 2

[0052] In one possible implementation, the aforementioned chemical modification information includes: chemical modification information of the sense chain and chemical modification information of the antisense chain.

[0053] Extensive experimental research has revealed that the thermal stability between the sense and antisense strands of siRNA, as well as the molecular affinity between the antisense strand and the target mRNA, both affect silencing efficiency. Therefore, in this embodiment, when acquiring chemical modification information, the information on the sense and antisense strand modifications is obtained, enabling accurate prediction of siRNA activity during model fitting based on this information.

[0054] Specifically, based on the chemical modification information of the sense / antense strands, it can be known which chemical modification molecule was added at which base position of the sense / antense strand. The above S2 specifically includes: during encoding, the chemical modification molecule structure can be reserved on the basis of the original base sequence and encoded together into an input matrix.

[0055] In one possible implementation, the aforementioned physicochemical properties include: molecular mass and topological surface area.

[0056] Specifically, in S1 above, the physicochemical properties of the target siRNA are retrieved based on the original base sequence, including: the physicochemical properties of the four ribonucleotides and deoxythymidine corresponding to the original base sequence can be retrieved online through an interface link to the PubChem database.

[0057] And / or, the physicochemical properties of the four ribonucleotides corresponding to the original base sequence and deoxythymidine are pre-searched in the PubChem database; then, based on the search results, a physicochemical property database as shown in Table 1 is pre-constructed in the device to facilitate offline data retrieval.

[0058] Table 1. List of 10 physicochemical properties

[0059]

[0060] Accordingly, in this embodiment, when acquiring chemical modification information, multiple physicochemical properties corresponding to the original sequence are obtained, so that the activity of siRNA can be accurately predicted based on multiple physicochemical properties during model fitting.

[0061] In one possible implementation, S2 specifically includes: employing a multi-view learning strategy to represent the characteristics of cm-siRNA, including the double-stranded sequence (positive and negative strands) of siRNA, chemical modification molecules, and characterization of physicochemical properties.

[0062] The double-stranded sequence: The double-stranded cm-siRNA sequence consists of ribonucleotides (AGUC) and deoxythymine (T). In this embodiment, a method commonly used in natural language processing—one-hot encoding—is used to numerically transform the sequence. In this embodiment, the value of K is set to 5, representing four ribonucleotides and one deoxythymine. Each letter is represented by a five-dimensional vector, where only the position corresponding to the element is set to 1, and all other positions are set to 0. We set the cm-siRNA sequence window to 25, representing the double-stranded siRNA as a 50x5 matrix (each strand is represented by a 25x5 matrix).

[0063] one_hot=SeqEncoder(cm-siRNA)

[0064] In one possible implementation, we use molecular access system software (MACC) to encode the molecular structure of the siRNA forward / reverse strands, capturing information on the RNA sequence's base composition and chemical modifications. We use the rdkit package (version 2022.9.5) in Python to convert the bases and chemical modifications into MACC. Specifically, rdkit matches a dictionary of 166 substructures in the MACC against the target molecule and assigns a binary value (0 / 1) to indicate the presence or absence of each substructure. Therefore, by using rdkit, cm-siRNA is represented as a 166-dimensional binary vector, thereby encoding the chemical modification information:

[0065] MACCS = [x1, x2, x3, ..., x 166 ],x i ∈[0,1]

[0066] In one possible implementation, the physicochemical properties of four ribonucleotides and deoxythymine were obtained beforehand from the PubChem database. A total of ten different physicochemical properties were obtained, such as molecular weight and topological surface area.

[0067] RNA properties =[molecularweight,XLogP3,…HeavyAtom Count]

[0068] We then characterized the physicochemical properties of the cm-siRNA based on the properties retrieved from the original double-stranded sequence. Similar to the double-stranded sequence, we set the cm-siRNA sequence window to 25, thus representing the double-stranded siRNA as a 50x10 matrix (each strand represented by a 25x10 matrix).

[0069] Specifically, the key to cross-attention mechanisms lies in simultaneously learning the relationships between multiple input sources, thereby better utilizing information from different input sources and improving feature fusion performance. In our method, the relationships between two classes of features are calculated by exchanging query vectors, thus obtaining global contextual information. This process can be viewed as an interaction and information exchange between feature matrices. In this way, the cross-attention mechanism can identify and strengthen features that are more important for prediction while weakening features that are less important. Specifically, such as... Figure 3As shown, each Cross-attention module consists of two self-attention modules. In the self-attention module, elements in the sequence are weighted and their importance is evaluated. The final output is generated by summarizing information from different parts of the sequence. Each self-attention module consists of a query vector (Q), a key vector (K), and a value vector (V), as shown in the following formula:

[0070] Q = X i W Q K = X i W K V = X i W V

[0071] in These are linear transformation matrices, representing the query matrix, key matrix, and value matrix, respectively. The formula for calculating the attention module output matrix Z is as follows:

[0072]

[0073] Based on the Cross-attention principle, the query vectors (Q) of the two self-attention layers... m Q o They can exchange information and feature fusion by using the inner product of query vectors, key vectors, and value vectors from different views.

[0074] In one possible implementation, the classification sub-model is a two-layer convolutional neural network, comprising: two convolutional sub-networks, two fully connected layers, and an output layer.

[0075] Specifically, each convolutional sub-network consists of convolutional layers, ReLU activation functions, and max-pooling layers. The first convolutional layer starts with 3 input channels, generating 16 feature maps, while the second convolutional layer expands from 16 input channels to 32 output channels. All convolutional operations use 3x3 kernels with a stride of 1 and 1-pixel padding to preserve the spatial dimension of the feature maps. After each convolutional operation, a ReLU activation function is applied to introduce non-linearity, followed by downsampling through a 2x2 max-pooling operation. Two fully connected layers: The outputs of the convolutional layers are flattened and processed through two fully connected layers for advanced feature learning and final prediction. The first fully connected layer accepts an input of size 3212128 and outputs 128 nodes, then processes it through a ReLU activation function and a 50% dropout layer to increase the model's generalization ability and prevent overfitting. The final output layer has 1 node and uses a sigmoid activation function, ensuring the model's output is between 0 and 1, thus predicting drug activity. The formula is as follows:

[0076]

[0077] P (i) =MaxPool 2*2 (ReLU(C (i) ))

[0078] Where C (i) This is the result of the first convolution operation, where i represents the layer number of the convolutional neural network. P (i) It is the result of the first layer of convolution. ReLU is a non-linear activation function.

[0079] In one possible implementation, the prediction model is pre-built using the following methods:

[0080] An initial model for the prediction model is established, comprising: a feature fusion sub-model and a classification sub-model;

[0081] The original base sequences, chemical modification information, and silencing efficiencies of multiple chemically modified siRNA drugs were obtained from the siRNAmod database; their physicochemical properties were retrieved based on the original base sequence of each chemically modified siRNA drug.

[0082] Data samples were constructed based on the chemical modification information, physicochemical properties, and silencing efficiency of multiple chemically modified siRNA drugs, and a sample dataset was generated.

[0083] The initial model is trained based on the sample dataset, and the prediction model is obtained when the model meets the conditional convergence.

[0084] Specifically, the above sample dataset is generated using the following methods:

[0085] S101. Download the activity data of chemically modified siRNA drugs from the siRNAmod database. After preprocessing, low-quality and unmodified siRNAs are removed, leaving 4278 data entries. The data is divided into 10 bins based on scores. 10% of the data from each bin is randomly selected, resulting in 3850 (Dataset1) and 428 (Dataset2). Dataset1 serves as the benchmark dataset, and Dataset2 as the independent test set. Additionally, four chemically modified siRNA drugs approved by the FDA are identified and compiled into Dataset3.

[0086] S102. Based on the cleaned data described above, perform modeling and prediction, such as... Figure 4 As shown, it mainly consists of three steps: siRNA feature extraction, S-feature fusion, and neural network classification.

[0087] Extract siRNA sequences, chemically modified molecules, and physicochemical properties.

[0088] siRNA Sequence: The raw siRNA sequence data consists of categorical data represented by letters of ribonucleotides and deoxythymine, which needs to be converted into computable data. One-hot encoding, a common encoding method in natural language processing, was used. In this study, K is set to 5, representing 4 ribonucleotides and 1 deoxythymine. Each letter is represented by a 5-dimensional vector, with only the corresponding element having a value of 1 and all other positions having a value of 0. We set the siRNA sequence window to 25, so that each siRNA strand can be represented by a 25×5 matrix (e.g., ...). Figure 4 ).

[0089] Chemically modified molecules: The molecular structure of siRNA positive / antisense strands and mRNA is encoded using a compound molecule encoding strategy, simultaneously encompassing the RNA sequence base composition and chemical modification information. The RDKit tool converts the bases and chemical modifications into binary vectors. Specifically, the dictionary of 166 seed structures in the MACC is compared with the target molecule, displaying only 0 / 1 to indicate whether the target molecule contains a substructure. This is represented as a 166-dimensional binary vector, with each dimension corresponding one-to-one with a specific substructure (e.g., ...). Figure 4 ).

[0090] Physicochemical properties: Physicochemical properties of four ribonucleotides and deoxythymidine were searched on the PubChem database, and 10 different physicochemical properties, such as molecular weight and topological surface area, were obtained. Figure 4 ).

[0091] In one possible implementation, the above-mentioned training of the initial model based on the sample dataset includes: training the model parameters using Dataset1, and comparing and analyzing with other models using Dataset2 and Dataset3 as independent test sets.

[0092] Specifically, the known silencing efficiencies of drugs in the drug database are compared with the data predicted by the model, and 5-fold cross-validation is used to evaluate the performance of the comparative model. Five metrics are used to evaluate the model's performance: Pearson correlation coefficient (PCC), Spearman rank correlation coefficient (SCC), area under the ROC curve (AUC), mean absolute error (MAE), and mean squared error (MSE), as shown in the following formula:

[0093]

[0094]

[0095] MAE=(1 / n)×∑|y i -y t |

[0096] MSE=(1 / n)×∑(y i -y t ) 2

[0097] Where cov is the covariance function, σ X σ is the standard deviation of X. Y Yd is the standard deviation of Y. i =x i -y i In addition, y i y represents the prediction efficiency for a given cm-siRNA. t This indicates the actual observation efficiency of cm-siRNA.

[0098] The performance of the algorithm and the other four methods on dataset1 (training set) is shown in Table 2. Compared with the model using multi-head attention mechanism and the SMEpred algorithm with three different representation strategies, the algorithm shows the best performance in all five indicators of the prediction results.

[0099] To demonstrate the performance and capability improvements of our algorithm, we compared it (cross-attention) with another multi-head attention-based model and the state-of-the-art SMEpred algorithm (which uses different feature descriptions) on dataset 1. We evaluated performance using multiple metrics, including PCC, SCC, AUC, MAE, and MSE. As shown in Table 1, our algorithm outperforms the other algorithms across all metrics, with PCC, SCC, AUC, MAE, and MSE values ​​of 0.8186, 0.8164, 0.9111, 0.1177, and 0.0304, respectively.

[0100] Table 2. Performance comparison of several algorithms on Dataset 1

[0101]

[0102] To evaluate the algorithm's generalization ability, we applied it to an independent dataset (dataset2) and compared it with other methods. To ensure a fair comparison, all models and methods used for comparison were trained independently on dataset1. As shown in Table 3, our algorithm also outperformed other methods on multiple metrics, with PCC, SCC, AUC, MAE, and MSE values ​​of 0.7700, 0.7670, 0.8931, 0.1300, and 0.0379, respectively.

[0103] Table 3. Performance comparison of several algorithms on Dataset 2

[0104]

[0105] To comprehensively evaluate the applicability of the algorithm in real-world research and development scenarios, we applied several methods to four FDA-approved cm-siRNA drugs. See Table 4 and... Figure 5 As shown, compared with other methods, our algorithm outperforms other methods on three drugs (Patisiran (0.7927), Givosiran (0.9122), and Inclisiran (0.8480), only slightly lagging behind multi-head attention methods on Lumasiran (0.6819). Furthermore, for comprehensive metrics, our algorithm also outperforms other methods in both MAE and MSE. The results demonstrate that our algorithm maintains a high level of accuracy and robustness in practical applications, showcasing its potential practical value.

[0106] Table 4. Performance of several algorithms on FDA-approved siRNA drugs

[0107]

[0108] Example 3

[0109] In another aspect, the present invention provides a chemically modified siRNA activity prediction device for implementing the method described in Example 1 or Example 2 above, comprising:

[0110] The acquisition unit is used to acquire the original base sequence and chemical modification information of the target siRNA;

[0111] A retrieval unit is used to retrieve the physicochemical properties of the target siRNA based on the original base sequence;

[0112] The encoding unit is used to encode the original base sequence, chemical modification information and physicochemical properties.

[0113] A prediction unit is used to generate a prediction result of the silencing efficiency of the target siRNA based on the encoded features using a pre-built prediction model; wherein the pre-built prediction model includes: a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on an attention mechanism; the classification sub-model is used to generate a prediction result of the silencing efficiency based on the fused features.

[0114] In a further embodiment of the present invention, such as Figure 6 As shown, based on the aforementioned chemically modified siRNA activity prediction device, a user-friendly interactive online software tool was developed by further optimizing model parameters, encapsulating the model interface, and designing UI controls. The website's front-end is built using Hyper Text Markup Language (HTML) and supports multiple widely used browsers such as Google Chrome, Firefox, and Internet Explorer 11+. According to requirements, client-side task submission is implemented by modifying and embedding corresponding functions and methods of asynchronous JavaScript and XML (Asynchronous JavaScript and XML, Ajax) into the overall code, enabling cross-platform operation on multiple platforms including Linux, Windows, and iOS.

[0115] Example 4

[0116] Another aspect of the present invention, such as Figure 7 As shown, an electronic device is provided, comprising a processor, a network interface, and a memory, wherein the processor, the network interface, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the chemically modified siRNA activity prediction method described in Embodiment 1 or Embodiment 2 above.

[0117] In another aspect, the present invention provides a computer storage medium storing program instructions, which, when executed by at least one processor, are used to implement the above-described intelligent slope and alignment adjustment method.

[0118] In embodiments of the present invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0119] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0120] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0121] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0122] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0123] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0124] It should be understood that the system disclosed in this invention can be implemented in other ways. For example, the division of modules is merely a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the communication connection between modules can be through some interfaces, indirect coupling or communication connections between servers or units, and can be electrical or other forms.

[0125] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one processing unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the activity of chemically modified siRNA, characterized in that, include: Obtain the original base sequence and chemical modification information of the target siRNA, and retrieve the physicochemical properties of the target siRNA based on the original base sequence; The original base sequence, chemical modification information, and physicochemical properties are feature-encoded, and a pre-constructed prediction model is used to generate a prediction result of the silencing efficiency of the target siRNA based on the encoded features; The pre-built prediction model includes a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on a cross-attention mechanism; and the classification sub-model is used to generate a silencing efficiency prediction result based on the fused features.

2. The method for predicting the activity of chemically modified siRNA as described in claim 1, characterized in that, The chemical modification information includes: chemical modification information of the sense chain and chemical modification information of the antisense chain.

3. The method for predicting the activity of chemically modified siRNA as described in claim 1, characterized in that, The physicochemical properties include: molecular weight, XLogP3, number of hydrogen bond donors, number of hydrogen bond acceptors, exact mass, mass of a single isotope, topological surface area, number of heavy atoms, complexity, and number of defined bond stereocenters.

4. The method for predicting the activity of chemically modified siRNA as described in claim 1, characterized in that, The classification sub-model is a two-layer convolutional neural network, including: two convolutional sub-networks, two fully connected layers, and an output layer.

5. The method for predicting the activity of chemically modified siRNA as described in claim 4, characterized in that, Each convolutional subnetwork layer includes: a convolutional layer, a ReLU activation function, and a max pooling layer.

6. The method for predicting the activity of chemically modified siRNA as described in claim 5, characterized in that, The convolutional layer uses a 3x3 convolutional kernel with a stride of 1 and a padding of 1 pixel.

7. The method for predicting the activity of chemically modified siRNA as described in any one of claims 1 to 6, characterized in that, The prediction model is pre-built using the following methods: An initial model for the prediction model is established, comprising: a feature fusion sub-model and a classification sub-model; The original base sequences, chemical modification information, and silencing efficiencies of multiple chemically modified siRNA drugs were obtained from the siRNAmod database; their physicochemical properties were retrieved based on the original base sequence of each chemically modified siRNA drug. Data samples were constructed based on the chemical modification information, physicochemical properties, and silencing efficiency of multiple chemically modified siRNA drugs, and a sample dataset was generated. The initial model is trained based on the sample dataset, and the prediction model is obtained when the model meets the conditional convergence.

8. A device for predicting the activity of chemically modified siRNA, characterized in that, include: The acquisition unit is used to acquire the original base sequence and chemical modification information of the target siRNA; A retrieval unit is used to retrieve the physicochemical properties of the target siRNA based on the original base sequence; The encoding unit is used to encode the original base sequence, chemical modification information and physicochemical properties. A prediction unit is used to generate a prediction result of the silencing efficiency of the target siRNA based on the encoded features using a pre-built prediction model; wherein the pre-built prediction model includes: a feature fusion sub-model and a classification sub-model; the feature fusion sub-model is used to perform feature fusion on the encoded features based on a cross-attention mechanism; the classification sub-model is used to generate a prediction result of the silencing efficiency based on the fused features.

9. The chemically modified siRNA activity prediction device as described in claim 8, characterized in that, It also includes: storage units for storing the physicochemical properties of the base sequence; The retrieval unit is configured to retrieve the physicochemical properties of the target siRNA from the storage unit based on the original base sequence.

10. An electronic device, characterized in that, The device includes a processor, a network interface, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the chemically modified siRNA activity prediction method as described in any one of claims 1 to 8.

Images