Method for predicting the targeting efficacy of a crisper-cas13d system

Abstract

This application relates to the field of bioinformatics technology, and particularly to a method for predicting the targeted efficacy of CRISPR-Cas13d-like systems. The method includes: mapping a first feature to a second feature containing physical interaction information via a learnable perceptual matrix of a first machine learning model; running the encoder portion of the first machine learning model based on the second feature to obtain local features; running the second machine learning model based on the first feature to obtain global features; inputting the fused features to a bidirectional cross-attention layer of a third machine learning model, performing cross-attention calculations in both directions to obtain bidirectional attention features; and inputting the bidirectional attention features to the regression portion of the third machine learning model to obtain the targeted efficacy prediction result. This invention improves the perception and prediction accuracy of dynamic RNA structural changes and spatial mismatch gradients, providing computational engine support for high-precision RNA-targeted therapy and efficient screening of gene editing guide sequences.

Description

Technical Field

[0001] This application belongs to the field of bioinformatics technology, and in particular relates to methods, devices, equipment, media and programs for predicting the targeted efficacy of CRISPR-Cas13d-like systems. Background Technology

[0002] With the widespread application of the CRISPR-Cas13 system in RNA-targeted intervention, accurately predicting the targeting efficacy of guide RNAs (gRNAs) has become a core challenge in transcriptome engineering and molecular diagnostics. Unlike the Cas9 system, which targets DNA, the editing efficiency of Cas13d RNA is profoundly influenced by the highly complex and dynamically changing secondary structure of RNA. In early research stages, gRNA design heavily relied on large-scale in vitro and in vivo experimental screening and heuristic rules based on small samples. This not only consumed a large amount of computational and experimental resources, but also often resulted in poor generalization ability due to limited understanding of the molecular mechanisms of the Cas13 system. Although traditional machine learning algorithms (such as random forests and XGBoost) have been gradually introduced into this field in recent years, these models mostly rely on shallow artificial feature engineering such as k-mer frequencies and local complementarity, which limits their feature representation capabilities and cannot automatically capture deep sequence-structure hierarchical relationships and complex biophysical dynamics.

[0003] In terms of deep learning models and feature extraction strategies, existing technologies still have limitations in handling the dynamic interactions and spatial conformations of gRNA and target RNA. For example, some models based on CNN-RNN fusion architectures (such as DeepCas13), while introducing parameters such as hybridization free energy, primarily focus on complete gRNA sequences in their input design, making it difficult to systematically address and quantify mismatches or insertions / deletions. Another type of cutting-edge deep learning model (such as TIGER), while effectively capturing position dependence and dose control through large-scale mismatch datasets, relies excessively on convolutional networks to model local mismatch locations, thus severely neglecting the global structural context of the target RNA. High-throughput experimental studies have shown that the design efficiency of gRNA is highly dependent on the local structural features of the target RNA. For example, the stem-loop structure of the 5' proximal flanking sequence (PFS) of the target can cause a sharp drop of more than 50% in the binding affinity of Cas13d, and mismatches at different distances can trigger drastically different spatial inhibition mechanisms. Existing models generally lack the ability to adaptively fuse long-range structural information and cannot effectively address the synergistic destructive effects of complex multi-point mismatches in spatial topology.

[0004] Therefore, there is an urgent need for a efficacy prediction method that integrates physical perception interaction rules, dynamic extraction of structural context, and long-range spatial dependency modeling to overcome the limitations of existing technologies in RNA targeted sequence analysis. Summary of the Invention

[0005] This application provides a method and apparatus for predicting the targeting efficacy of a CRISPR-Cas13d-like system, which can overcome the limitations of existing technologies in RNA targeting sequence analysis.

[0006] In a first aspect, embodiments of this application provide a method for predicting the targeting efficacy of a CRISPR-Cas13d-like system, comprising: The first RNA features are learned through the perceptual matrix of the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing first training on the first sample. The physical interaction information refers to the physical interaction and stability differences of all possible base pair combinations between two RNA sequences based on preset thermodynamic pairing prior rules. Based on the second RNA characteristic Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; Based on the first RNA characteristics Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; Based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; Input the fusion feature To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. Input the bidirectional attention feature , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

[0007] In one possible implementation of the first aspect, the first RNA features are processed via a learnable perceptual matrix of a first machine learning model. Mapping to second RNA features containing physical interaction information The steps include: Obtain the original RNA sequence of the CRISPR-Cas13d system Regarding the original RNA sequence Perform a truncation or padding operation to obtain a product with a fixed length. Fixed-length RNA sequences ; Encoding the fixed-length RNA sequence The first RNA feature was obtained. ; For the learnable perceptual matrix and the first RNA feature Second RNA features were obtained by performing physical interaction operations. The physical interaction operation includes broadcast dot product and outer product mapping executed sequentially.

[0008] In one possible implementation of the first aspect, the second RNA feature is used. Running the encoder part of the first machine learning model yields local features. The steps include: The second RNA feature is generated using a projection operation with a bias term. Mapping to a feature space with hidden dimensions and injecting positional encoding yields the hidden dimension input features. The location encoding is injected through feature superposition and / or feature space transformation. Input features using the hidden dimension The encoder portion of the first machine learning model is used as input to obtain local features. .

[0009] In one possible implementation of the first aspect, the second machine learning model is obtained by performing a second training on a second sample, the second training preceding the first training, during which the parameters of the second machine learning model are frozen.

[0010] In one possible implementation of the first aspect, the bidirectional cross-attention layer includes a bias matrix, the bias matrix being used to adjust the bias based on any two positions in the RNA sequence. , relative distance between Set the bias coefficient; Wherein, the bias coefficient is the relative distance A function is used to simulate distance-based spatial mismatch tolerance, and the bias coefficient is relative to the relative distance. The function includes a non-linear component.

[0011] In one possible implementation of the first aspect, the regression component is a multilayer perceptron; The input of the bidirectional attention feature , The steps for obtaining the targeted efficacy prediction results in the regression part of the third machine learning model include: splicing the bidirectional attention features , This yields the total feature vector that includes all bidirectional interactive information; The total feature vector is input into the multilayer perceptron, and the dimension is compressed layer by layer. The output is the log-folded change value (LFC) used to normalize the target efficacy prediction score.

[0012] The beneficial effects of the above methods include: First, the above method constructs a learnable physical perception embedding matrix and combines it with the structural semantic features extracted by the RNA Large Language Model (RNA-FM) to achieve the fusion modeling of local sequence features and global structural information, thereby improving the model's ability to represent RNA target accessibility.

[0013] Secondly, the above method employs a bidirectional positional bias cross-attention mechanism and combines it with a piecewise positional bias function to model spatial mismatch relationships at different distances, thereby enhancing the model's ability to represent the features of mismatches at different positions and their interactions.

[0014] Finally, compared to models that can only assess the relative loss of activity of variants, the above method directly predicts the logarithmic fold change (LFC) value of gene targeting efficacy. It achieved an AUROC of 0.91 and a PCC of 0.79 in benchmark tests covering hundreds of thousands of systematic mismatches. It showed good predictive performance in both perfect match and complex multi-point mismatch scenarios, and can provide a reference for the design of guide RNAs for CRISPR-RNA editing systems.

[0015] Overall, the above methods integrate biophysical prior learnable interaction matrices, pre-trained RNA large language models (such as RNA-FM), and bidirectional position-biased cross-attention mechanisms to achieve joint modeling of local sequence information and global structural information. This improves the model's ability to represent the dynamic structure and spatial mismatch features of RNA and can be applied to CRISPR-Cas13d targeting efficacy prediction and related guide RNA design tasks.

[0016] Secondly, embodiments of this application provide a targeting efficacy prediction device similar to a CRISPR-Cas13d system, comprising: The physical perception interaction module is used to: process the first RNA features via a learnable perception matrix from the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing a first training on a first sample, and the physical interaction information refers to the physical interaction of all possible base pair combinations between two RNA sequences. Local feature module, used for: based on the second RNA feature Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; The global feature module is used to: based on the first RNA feature Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; The fusion module is used for: based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; A bidirectional cross-attention module is used to: input the fused features To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. The regression module is used to: input the bidirectional attention features , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

[0017] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the targeting efficacy prediction method for a CRISPR-Cas13d-like system as described in any of the first aspects above.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the targeting efficacy prediction method for a CRISPR-Cas13d-like system as described in any of the first aspects above.

[0019] Fifthly, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute the targeting efficacy prediction method of the CRISPR-Cas13d-like system described in any of the first aspects.

[0020] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the targeting efficacy prediction method for a CRISPR-Cas13d-like system provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the targeting efficacy prediction device of the CRISPR-Cas13d-like system provided in the embodiments of this application; Figure 3This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application; Figure 4 A schematic diagram of the overall workflow of the CRISPR-Cas13d targeting efficacy prediction method based on deep learning and large language model provided in the embodiments of this application; Figure 5 A schematic diagram of the CRISPR-Cas13d targeting mechanism provided in the embodiments of this application; Figure 6 A diagram illustrating the one-hot coding rules of nucleotides provided in the embodiments of this application; Figure 7 The weight heatmap of the learnable physical perception interaction matrix provided in the embodiments of this application after model training convergence.

[0023] Figure label: Physical perception interaction module 201; local feature module 202; global feature module 203; fusion module 204; bidirectional cross-attention module 205; regression module 206; terminal device 30; processor 301; memory 302; computer program 303. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0025] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0026] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0028] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0030] The first embodiment of this application provides a method for predicting the targeting efficacy of a CRISPR-Cas13d-like system, such as... Figure 1 As shown, it includes: Step 102: The first RNA features are processed through the learnable perceptual matrix of the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing first training on the first sample. The physical interaction information refers to the physical interaction and stability differences of all possible base pair combinations between two RNA sequences based on preset thermodynamic pairing prior rules. Step 104, based on the second RNA feature Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; In step 104 of this embodiment, "based on the second RNA feature" The "encoder part of running the first machine learning model" refers to the second RNA feature. It can be the direct input data of the first machine learning model, or it can be the second RNA feature. After pre-processing, it is input into the first machine learning model.

[0031] In this embodiment, the encoder of the first machine learning model essentially constitutes a two-stream (or more-stream) symmetric encoder, with the first branch and the second branch forming a two-stream encoder (if more class index subscripts are introduced). The value of can be used to add encoder branches to form a multi-stream encoder. Symmetry means that the two branches (e.g., the first branch and the second branch) are completely symmetrical in terms of network architecture (e.g., number of layers, in-layer components), dimensionality rules (e.g., dimensionality increase / decrease rules, positional encoding), operation logic (e.g., weight operation), and parameter scale. The features of the two branches obtained at any position are alignable, interactive, comparable, and can be understood as being in the same feature space.

[0032] It is worth noting that symmetry does not imply complete weight / parameter sharing. Two branches can share some weights / parameters (e.g., weights / parameters based on combined considerations), but considering... and The different types of information flows play different biological roles and have different functions in the system. They need to learn their own characteristics and rules independently. Symmetrical design can ensure that the learning mechanism and logic of the two are consistent, and it is neither necessary nor necessary to ensure that the learning results of the two are consistent.

[0033] Step 106, based on the first RNA feature Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; In step 106 of this embodiment, the hidden state output refers to the features passed between different layers of the machine learning model (especially neural networks), i.e., intermediate features (or at least a portion of intermediate features).

[0034] Step 108, based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; Step 110, input the fusion feature To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. Step 112, input the bidirectional attention features , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

[0035] Although the concepts of the embodiments of this application originated from the CRISPR-Cas13d system, the embodiments of this application can be applied to CRISPR-Cas13d-like systems.

[0036] CRISPR-Cas13d-like systems refer to biological systems that possess characteristics similar to those of CRISPR-Cas13d systems. These characteristics are characterized by highly complex and dynamically changing secondary structures.

[0037] More specifically, a CRISPR-Cas13d-like system can be defined as satisfying at least one of the following: 1. RNA molecules have a secondary structure, and the spatial structure of this secondary structure changes spontaneously with changes in environmental factors; among which, environmental factors include at least one of temperature, ion concentration, nucleic acid / protein binding, and RNA epigenetic modification.

[0038] 2. RNA molecules have a secondary structure, and the spatial structure of this secondary structure changes spontaneously during molecular binding.

[0039] 3. RNA molecules have a secondary structure, and this secondary structure has multiple metastable spatial structures with similar energies; among them, metastable state refers to the state in which the spatial structure spontaneously changes under perturbation.

[0040] 4. RNA molecules have at least two secondary structures.

[0041] The aforementioned secondary structures include stem rings, internal rings, convex rings, multi-branched rings, and pseudoknots.

[0042] Furthermore, although the concept of the embodiments of this application originates from the two-way information flow of guide RNA and target RNA to predict targeting efficacy, the fusion idea of ​​the embodiments of this application can be extended to other information.

[0043] In other words, in a preferred embodiment, the category index subscript (category) includes, but is limited to, guide RNA subscripts and target RNA (also referred to as target RNA for readability) subscripts. In this preferred embodiment, , These are the guide RNA subscript and the target RNA subscript, respectively. , The corresponding information streams are guide RNA information and target RNA information, respectively.

[0044] In some alternative implementations, , Or other category index subscripts can be: Nucleic acid entity subscripts, such as RNA flanking sequences (PFS); Subscripts for biological / physical / chemical modalities, such as the thermodynamic features of RNA secondary structure and the features of epitranscriptome modifications; Subscripts for biological macromolecular entities, such as protein sequence features.

[0045] In terms of implementation effectiveness alone, the combination of guide RNA features and target RNA features provides the most essential information and theoretically yields the best prediction results. Therefore, the following will be based on " , These are the guide RNA subscript and the target RNA subscript, respectively. , The corresponding information streams are guide RNA information and target RNA information, providing more preferred implementation methods.

[0046] The preferred embodiment of step 102 is as follows: via tensor (The learnable perceptual matrix) maps the first guide RNA feature and the first target RNA feature to the second guide RNA feature and the second target RNA feature, respectively; the tensor Used to introduce physical interaction information, which refers to the physical interactions between all possible base pair combinations between the guide sequence and the target sequence; the tensor It is at least a part of a first machine learning model obtained by performing first training on a first sample, and the tensor The parameters are corrected through the first training; Furthermore, based on the above preferred embodiments, the preferred embodiment of step 104 is as follows: Based on the second guide RNA features and the second target RNA features, the guide flow branch and target flow branch of the dual-stream encoder are run respectively to obtain the guide local features. Target local features The dual-stream encoder is at least a part of the first machine learning model, and the guide flow branch is symmetrical to the target flow branch. In the above preferred embodiments, , They were specifically defined as , .

[0047] Furthermore, based on the above preferred embodiments, the preferred embodiment of step 106 is as follows: Using the first guide RNA feature and the first target RNA feature as inputs, respectively, the second machine learning model is run to obtain the global guide RNA feature. Targeting global features The second machine learning model is a pre-trained biological domain semantic understanding model, and the guide global features... and the targeted global features All are extracted from the hidden state output matrix of the second machine learning model; In the above preferred embodiments, , They were specifically defined as , .

[0048] Furthermore, based on the above preferred embodiments, the preferred embodiment of step 108 is as follows: Based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ,in ; Alternatively, the preferred embodiment of step 108 can be more accurately described as follows: Based on the local features of the wizard Construct the first query tensor based on the global features of the wizard. Construct a first key tensor and a first value tensor. Then, perform an attention-based fusion operation using the first query tensor, the first key tensor, and the first value tensor to obtain the guided fusion feature. ; Based on the target local features Construct a second query tensor based on the targeted global features. Construct a second key tensor and a second value tensor. Perform attention-based fusion using the second query tensor, the second key tensor, and the second value tensor to obtain targeted fusion features. ; In the above preferred embodiments, , They were specifically defined as , .

[0049] Step 108 can be further specified in practice as follows: Introduce a cross-modal attention layer to integrate local and global features, and set modal indicator variables. (Here, "modal indicator variable" is a term used in the context of software programming. In this embodiment, modal variables need to be specified in the program during the fusion process. In the overall scheme of this application, it can be understood as being related to the category index subscript.) (having equal status), the local sequence features output by the Transformer encoder As a query vector, the global semantic features after dimensionality reduction As a key-value vector, the fused features The calculation formula is: ,in , , The learning weight matrix is ​​independent for each modality. Scaling the dimensions for attention.

[0050] More preferably, in the preceding steps, the attention scaling dimension of the cross-modal fusion mechanism... From the dimension of local hidden features Divide by the number of attention heads generated, where the number of attention heads is set to 4, and calculate the attention scaling dimension. The value is 32.

[0051] Based on the first embodiment described above, a second embodiment is provided for step 102 as follows.

[0052] The learnable perceptual matrix via the first machine learning model will incorporate the first RNA features. Mapping to second RNA features containing physical interaction information The steps include: Obtain the original RNA sequence of the CRISPR-Cas13d system Regarding the original RNA sequence Perform a truncation or padding operation to obtain a product with a fixed length. Fixed-length RNA sequences ; Encoding the fixed-length RNA sequence The first RNA feature was obtained. ; For the learnable perceptual matrix and the first RNA feature Second RNA features were obtained by performing physical interaction operations. The physical interaction operation includes broadcast dot product and outer product mapping executed sequentially.

[0053] Referring to the preferred embodiment of step 102 in the first embodiment, this embodiment further provides the following preferred embodiment of step 102.

[0054] via tensor The steps of mapping the first guide RNA feature and the first target RNA feature to the second guide RNA feature and the second target RNA feature, respectively, include: Obtain the original guide RNA sequence and the original target RNA sequence of a CRISPR-Cas13d-like system, and perform truncation or padding operations on the original guide RNA sequence and the original target RNA sequence to obtain a sequence with a fixed length. Guide RNA input sequence and target RNA input sequence ; The guide RNA input sequences are encoded respectively. and the target RNA input sequence The characteristics of the first guide RNA and the characteristics of the first target RNA are obtained. For the tensor Performing physical interaction operations with the first guide RNA feature yields the second guide RNA feature, and then performing physical interaction operations on the tensor. The second target RNA feature is obtained by performing a physical interaction operation with the first target RNA feature; The physical interaction operation includes broadcast dot product and outer product mapping executed sequentially.

[0055] In practical operation, the solution in this embodiment can be specifically as follows: Obtaining the CRISPR-Cas13d guide RNA input sequence (Original guide RNA sequence) and target RNA input sequence (The original target RNA sequence) is uniformly truncated or padded to a fixed length. And obtain its corresponding original label value for targeted efficacy. (For training purposes, the aforementioned reasoning process is not involved); to eliminate the distribution bias caused by the different biological importance of target genes, [the following is done]... Gene-level Z-score normalization was performed, and the calculation formula is as follows: ,in and The normalized label is obtained by representing the mean and standard deviation of the true power of the target gene to which the sequence belongs. ; More preferably, in the preceding step, the fixed length L of the fixed-length sequence is set to 23, and any part of the sequence that is less than this length is padded with the padding character 'N'.

[0056] The sequence after the fixed-length operation and Perform one-hot encoding and construct a dimension of Dynamically learnable physical perception interaction matrix (i.e., learn the perception matrix / tensor) This is used to map all possible base pair combinations between the guide sequence and the target sequence; through matrix mapping, discrete one-heat encodings are transformed into initial embedding representations that integrate biophysical properties and thermodynamic stability. and (First RNA characteristics) ); More preferably, in the preceding steps, the physical perception interaction matrix Dimensions The value of is 4, and its internal elements contain initial weight parameters based on biophysical priors. The initial weight value of CG pairing is 3.0, the initial weight value of AU pairing is 2.0, the initial weight value of GU swing pairing is 1.0, and the initial weight value of other non-complementary mismatched combinations is uniformly -1.0.

[0057] Based on the first embodiment described above, a third embodiment is provided for step 104 as follows.

[0058] The second RNA feature Running the encoder part of the first machine learning model yields local features. The steps include: The second RNA feature is generated using a projection operation with a bias term. Mapping to a feature space with hidden dimensions and injecting positional encoding yields the hidden dimension input features. The location encoding is injected through feature superposition and / or feature space transformation. Input features using the hidden dimension The encoder portion of the first machine learning model is used as input to obtain local features. .

[0059] Referring to the preferred embodiment of step 104 in the first embodiment, this embodiment further provides the following preferred embodiment of step 104.

[0060] Based on the second guide RNA features and the second target RNA features, the guide flow branch and target flow branch of the dual-stream encoder are run respectively to obtain guide local features. Target local features The steps include: The second guide RNA feature and the second target RNA feature are mapped to a feature space with hidden dimensions using a linear projection operation with a bias term, and positional encoding is injected to obtain guide flow input features and target flow input features; wherein, the positional encoding is injected through feature superposition and / or feature space transformation. Using the guide flow input features and the target flow input features as inputs respectively, the guide flow branch and the target flow branch are run to obtain the guide local features. The targeted local features .

[0061] In this embodiment, the position encoding can be Rotary Position Embedding (RoPE), Sine / Cosine Position Encoding (or Complex Position Encoding based on Trigonometric Functions, such as writing Trigonometric Functions in Complex Form), Piecewise Function Encoding (encoding based on position intervals rather than absolute positions), Single Function Encoding (e.g., implementing linear distance penalty through Linear Attention with Linear Biases), Binary Encoding / Integer Position Encoding / Exponential Decay Encoding, or any combination thereof.

[0062] In practical operation, the solution in this embodiment can be specifically as follows: Construct a two-stream symmetric Transformer encoder and use linear projection to transform the initial embedding representation. and Mapped to hidden dimension The dense feature space is injected with sine and cosine position codes. For the th in the sequence The position of the first The position encoding formulas for even and odd dimensions are as follows:

[0063]

[0064] in The base is constant; subsequently, a multi-layer multi-head self-attention mechanism is used to extract position-aware local sequence feature representations. and ; More preferably, in the preceding steps, the hidden dimension The value is set to 128, which is the constant base in the sine and cosine position encoding. The value is set to 10000, and the dual-stream symmetrical Transformer encoder consists of four network stacked layers.

[0065] Based on the first embodiment described above, a fourth embodiment is provided for step 106 as follows.

[0066] The second machine learning model is obtained by performing a second training on a second sample. The second training is earlier than the first training. During the execution of the first training, the parameters of the second machine learning model are frozen.

[0067] In practical operation, the solution in this embodiment can be specifically as follows: The input sequence is processed using a pre-trained RNA large language model, and the model's first... The hidden state of the layer is used as a dimension. The language feature matrix encapsulates global sequence context and evolutionary constraint information; subsequently, a linear projection layer smoothly reduces the dimensionality of the language features to [a specific value]. This yields global semantic features aligned with the main sequence encoding. and ; Based on the first embodiment described above, a fifth embodiment is provided for step 110 as follows.

[0068] The bidirectional cross-attention layer includes a bias matrix, which is used to adjust the bias based on any two positions in the RNA sequence. , relative distance between Set the bias coefficient; Wherein, the bias coefficient is the relative distance A function is used to simulate distance-based spatial mismatch tolerance, and the bias coefficient is relative to the relative distance. The function includes a non-linear component.

[0069] In practical operation, the solution in this embodiment can be specifically as follows: Calculate the relative positional distance between corresponding bases in the sequence. And design a piecewise position offset function. To simulate the tolerance of long- and short-range spatial mismatches.

[0070] when hour:

[0071] when hour:

[0072] in , , These are the learning parameters that control the linear slope, piecewise distance boundary, and logarithmic curvature, respectively; the attention matrix for guiding to the target is calculated. and the opposite Their scoring formulas all include a bias term:

[0073] in This is the bias weight matrix; More preferably, in the preceding step, the segment distance boundary parameter in the segmented position offset function The initial value is set to 8, which controls the parameter of the linear slope. The initial value is set to 0.1, which controls the parameter of logarithmic curvature. The initial value is set to 0.5.

[0074] Based on the first embodiment described above, a sixth embodiment is provided for step 112 as follows.

[0075] The regression component is a multilayer perceptron; The input of the bidirectional attention feature , The steps for obtaining the targeted efficacy prediction results in the regression part of the third machine learning model include: splicing the bidirectional attention features , This yields the total feature vector that includes all bidirectional interactive information; The total feature vector is input into the multilayer perceptron, and the dimension is compressed layer by layer. The output is the log-folded change value (LFC) used to normalize the target efficacy prediction score.

[0076] In the practical training process, the solution in this embodiment can be specifically as follows: The features output from the bidirectional attention matrix are concatenated and input into a multilayer perceptron regression network to output a targeting efficacy prediction score. Construct a multi-objective joint loss function. Perform network reverse propagation, in which For predicted score With normalized labels Mean squared error loss between them For weight parameters, The symmetry constraint loss between the two attention matrices is expressed by the following formula:

[0077] More preferably, in the preceding steps, the weight parameter λ_sym in the multi-objective joint loss function is set to 0.1, the batch size is set to 256, the initial learning rate is set to 1e-4, and the random dropout rate of each layer of the network is set to 0.3.

[0078] Based on any of the foregoing embodiments, a seventh embodiment with specific parameters and performance comparisons will be provided below. This seventh embodiment is implemented based on a narrowly defined CRISPR-Cas13d system (rather than a CRISPR-Cas13d-like system). , They were specifically defined as , .

[0079] In particular, in this embodiment, the symbols for some parameters differ from those in the previous embodiments, but terms with the same name are used to refer to the same parameters (even if the symbols are different).

[0080] This embodiment presents a CRISPR-Cas13d targeting efficacy prediction method based on deep learning and large language models, which relates to the fields of bioinformatics and gene editing-assisted design technology, and is particularly optimized for the complex spatial dependence of RNA sequence-structure and specific local mismatch features.

[0081] This method addresses the problem that existing prediction tools struggle to capture the complex spatial dependencies between sequence composition and dynamic RNA structure, and thus cannot achieve accurate efficacy prediction. It proposes a prediction framework that integrates a pre-trained RNA large language model and a two-stream Transformer architecture.

[0082] This method first performs one-hot encoding on the sequence and introduces a dynamically fine-tunable physical-aware interaction matrix to capture biophysical base pairing features. Then, it extracts position-aware local features through a two-stream Transformer and dynamically fuses the global structural context extracted by a large language model using a cross-modal attention layer. Finally, it accurately quantifies spatial gradient differences and mismatch tolerance through a bidirectional cross-attention mechanism with segmented position bias, and outputs a prediction score by a regression network.

[0083] This embodiment effectively integrates local thermodynamic constraints with long-range structural features, improving the accuracy of the CRISPR-Cas13d system in predicting the targeting efficacy in perfect matching and various mismatch scenarios.

[0084] Figure 4 This document demonstrates the main workflow of the method described in this embodiment for predicting the targeting efficacy of CRISPR-Cas13d.

[0085] First, the obtained guide RNA (gRNA) and target RNA sequences are length-aligned and one-hot encoded, and then mapped to a learnable physical-aware embedding matrix to capture biophysical pairing features. Subsequently, a deep feature processing container was constructed, and a two-stream Transformer encoder and a pre-trained RNA large language model (RNA-FM) were used in parallel to extract local features with location awareness and deep semantic features containing global structure. In the feature fusion stage, the sequence spatial breakage effect and mismatch tolerance gradient are dynamically captured by using a cross-modal attention layer and introducing a bidirectional cross-attention mechanism with segmented position bias. Finally, the bidirectional interaction features are concatenated and input into the multilayer perceptron (MLP) regression network. During the training phase, a symmetry-constrained multi-target joint loss is used to optimize the parameters, ultimately outputting a high-precision target efficacy prediction score.

[0086] The input data for the method in this embodiment mainly includes the CRISPR-Cas13d system (the CRISPR-Cas13d targeting process, such as...). Figure 5 The guide RNA sequence file (FASTA format) and the corresponding targeted knockout experimental tag file (CSV format) are shown.

[0087] The FASTA file contains a unique sequence ID and the corresponding guide and target nucleotide sequences (including A, U, C, and G), with the sequence length uniformly set to 23 nt after preprocessing. The tag file contains the sequence ID and the corresponding experimentally measured raw value of the log-Fold Change (LFC). After standardization and normalization preprocessing of the input file, the final output file of this method is a prediction result file (CSV format), whose main fields include: sequence_ID (sequence identifier) ​​and predicted_LFC (predicted targeting efficacy score). This output file can be directly used for virtual screening of downstream high-activity guide sequences and guide the design of guide libraries.

[0088] This embodiment proposes a CRISPR-Cas13d targeting efficacy prediction method based on deep learning and a large language model. It captures biophysical base pairing features by one-hot encoding the RNA sequence and introducing a dynamically fine-tunable physical-aware interaction matrix. It integrates a two-stream Transformer architecture with a pre-trained RNA large language model (RNA-FM), dynamically fusing global structural context and local sequence features using a cross-modal attention layer. Simultaneously, it accurately quantifies the interaction and mismatch tolerance of the spatial gradient between the guide RNA (gRNA) and target RNA through a bidirectional cross-attention mechanism with segmented positional bias. Finally, a multilayer perceptron regression network outputs the targeting efficacy prediction score. Implementation results demonstrate that this method achieves good prediction performance in both perfect matching and complex multi-point mismatch scenarios.

[0089] The identification method in this embodiment can be divided into the following steps: 1. Dataset Acquisition and Standardization Preprocessing The system first imports a systematic CRISPR-Cas13d high-throughput screening dataset covering 16 essential human genes (such as CD9 and EEF2) using a parallel I / O interface. This dataset contains 118,267 samples, including a pool of 10,000 perfectly matched guide RNA sequences and 108,600 mismatched mutant sequences covering single-point, double-point, and random multi-point distributions. In the preprocessing stage, the data cleaning module first removes low-confidence samples with raw sequencing depths below a threshold. Assuming the system reads a gRNA sequence targeting the EEF2 gene, due to the truncation design in the original library, its length is only 20 nt (e.g., the sequence 5'-GACGUAAUGCUACUGGCUAA-3'). The preprocessing tensor module automatically fills its 3' end with three 'N' characters representing unknown nucleotides using a zero-padded strategy, strictly spatially aligning and normalizing it to a fixed-length 23-nt sequence.

[0090] For the raw log-fold change (LFC) values ​​in the tag file, the system calls the ComBat algorithm package to perform batch effect elimination within an empirical Bayesian framework to smooth out systematic variance between different cell infection replicates. Assume the mean LFC for EEF2 gene knockout survival in a genome-wide screening context. Standard deviation The original measurement value of the current detection sequence is... The system performs element-wise calculations using the Z-score normalization function to obtain the true predicted label tensor. This operation mathematically maps the distribution of the efficacy of genes of different survival importance to a standard normal distribution space. Internally, it completely blocks the interference of the biological necessity of the gene itself with variance optimization.

[0091] 2. Dynamic gradient fine-tuning of the physical perception interaction matrix The above 23-nt gRNA and the corresponding target RNA sequence are input into the sequence embedding module.

[0092] The system first establishes a one-hot encoding dictionary based on the alphabet of A, U, C, G, N. The specific matrix transformation rules for its nucleotides can be found in [reference needed]. Figure 6 The diagram illustrates how discrete characters are mapped to an initial floating-point tensor of dimension (Batch_Size, 23, 4).

[0093] To overcome the static limitation of traditional deep learning models that treat nucleotides as independent symbols without physical meaning, the system adds a dimension in the GPU memory. Physical perception tensor The tensor is not initialized with a random Gaussian distribution during the weight initialization phase. Instead, it is assigned directional values ​​strictly according to the prior of the thermodynamic pairing free energy of RNA double strands: the CG pair, representing three hydrogen bonds and the most stable structure, is initialized to 3.0; the AU pair, representing two hydrogen bonds, is initialized to 2.0; the GU wobbling pair, which is crucial in RNA dynamic folding, is initialized to 1.0; and all non-complementary mismatched positions (such as AC and GG), representing steric hindrance and thermodynamic bubbles, are initialized to -1.0.

[0094] In the actual forward propagation matrix calculation, the feature tensors of the guide and target sequences are compared with... Broadcast dot product and outer product mapping are performed to generate initial interaction features of dimension (Batch_Size, 23, 16). These interaction features effectively integrate the biophysical pairing information and thermodynamic stability of sequence combinations, and... As leaf node tensors tracked by PyTorch's require_grad=True, they are fine-tuned during backpropagation using the gradient descent algorithm.

[0095] For example, after training converges for 150 epochs, combined with Figure 7 The weight heatmap shown after model training convergence reveals that the weights of CG pairs monitored internally by the system may adaptively fine-tune to 3.12, while the weights of highly destructive purine-purine mismatches (such as AG) may drop to -1.65, proving that the model can autonomously learn the nonlinear mismatch penalties under different biochemical microenvironments.

[0096] 3. Local Feature Extraction using Two-Stream Transformer To represent sequence features in a higher-dimensional latent space, the system projects the low-dimensional physical embedding tensor output from step 2 into a dense space with a hidden dimension of 128 through a fully connected nn.Linear(16, 128) layer configured with a bias term, outputting a continuous feature tensor with dimensions (Batch_Size, 23, 128). To compensate for the inherent lack of sequence temporal awareness in the Transformer architecture, the system generates a fixed sine and cosine positional encoding matrix in GPU memory. .

[0097] For absolute spatial location in the sequence Its even-numbered dimensions with odd dimensions Using formulas respectively and Perform calculations and The tensor is added element-wise to the sequence-dense tensor. The resulting tensor is then fed into a two-stream symmetric Transformer encoder consisting of four cascaded network layers.

[0098] Within each layer, the system activates four attention heads in parallel, each of which projects the input 128-dimensional features through an independent projection matrix. Mapping to a subspace, the processed feature sub-dimension is Attention is calculated by scaling the dot product. Subsequently, the multi-head features are concatenated and passed through a feedforward network configured as 128->512->128 with the GELU activation function. After supplementing with residual connections and layer normalization to prevent gradient vanishing in deep networks, the guiding flow and targeting flow output high-level local representation tensors with deep position-dependent awareness, respectively. and .

[0099] 4. Deep semantic extraction from RNA large language model In the macro-semantic extraction branch, which runs parallel to the Transformer encoder, the original 23-nt plain text sequence is input into an RNA-FM large language model with 650 million parameters that has been pre-trained on 23 million cross-species non-coding RNA sequences for masked language modeling.

[0100] To prevent catastrophic forgetting and overfitting for specific downstream knockout tasks, all network weights of the RNA-FM model are frozen in the training graph (requires_grad=False). The system uses a pre-defined hook function to extract the hidden state output of the penultimate layer (layer 11 of 12) of the large model. The last layer is discarded because it retains the purest evolutionary constraints and global structural topology information, without being overly interfered with by the classification head of the pre-trained task.

[0101] This operation generates a massive, ultra-high-dimensional semantic tensor with dimensions (Batch_Size, 23, 1200). To achieve dimension alignment in subsequent cross-modal tensor computations, the system utilizes a weight matrix... A linear dimensionality reduction layer smoothly compresses the ultra-high-dimensional tensor while preserving most of the principal component variances, ultimately outputting a large-scale model macroscopic projection feature tensor with dimensions (Batch_Size, 23, 128). and The specific parameters for configuring the extraction layers and feature dimensions of the pre-trained model are detailed in Table 1.

[0102] Table 1. Configuration and Feature Extraction Parameter Settings for Pre-trained RNA-FM Model

[0103] 5. Cross-modal feature dynamic attention fusion The local perceptual tensor output from step 3 (carrying base proximity interaction information) and the macroscopic projection tensor output from step 4 (carrying global structural folding prior) are simultaneously fed into the core cross-modal attention fusion operator layer of the system, provided they have identical dimensions (both Batch_Size, 23, 128). Taking guided feature fusion as an example, the system will convert the local features output by the Transformer... Multiplying by the weight matrix maps to a query tensor, while simultaneously incorporating the global features output by the large language model. The mapping is into keys and values.

[0104] The system executes the standard scaled dot product attention mechanism, whose mathematical fusion equation is defined as:

[0105] The physical mapping of this architecture at the bioinformatics level is equivalent to this: when the main sequence extraction network scans each base pair, it dynamically and selectively "queries" the RNA large language model to see if the core position (e.g., the Cas13d seed region at positions 15-21) is in a potentially high-strength stem-loop or tight pseudo-knot structure, based on the thermodynamic instability caused by local mismatches. This adaptively determines whether to amplify the activity inhibition weight at that position, ultimately outputting a high-dimensional feature representation that unifies microscopic and macroscopic aspects. and .

[0106] 6. Segmented bias and bidirectional cross-attention interaction To address the interaction penalty modeling of mismatch distances in three-dimensional space for nucleic acid chains, the system introduces a bidirectional position bias cross-attention calculation mechanism and network architecture, with independently running segmented position bias calculation kernels.

[0107] The system initializes a distance matrix with dimensions (23, 23). Calculate any two interaction positions in the sequence. relative distance System calls are defined as linear slopes with learnable parameters. Segment boundary and logarithmic curvature Construct a nonlinear piecewise bias matrix .

[0108] For example, when position 4 interacts with position 7, the distance Its bias penalty value is When the first and last positions 1 and 22 interact, the distance... Its bias value enters the logarithmic decay region and is calculated as follows: Subsequently, the system constructs a bidirectional cross-attention computation graph of "guide-guided target" and "target-guided guide". When generating the attention score matrix using tensor multiplication, the system utilizes PyTorch's broadcast mechanism to transmit the calculated bias matrix. Multiply by the adaptive weight matrix Then, it is directly injected into the attention calculation formula as a bias term. This mechanism perfectly mimics the biological reality in long RNA where the breakage effect caused by local mutations decays non-linearly (logarithmically) with increasing nucleotide distance, effectively avoiding the attention collapse of conventional attention mechanisms in the face of distant noise.

[0109] 7. Regression Prediction of Multi-Objective Joint Loss After undergoing the aforementioned complex feature engineering, the bidirectional cross-attention layer ultimately outputs two highly interactive tensors with the same dimensions (Batch_Size, 23, 128). and .

[0110] The system flattens the sequence along the hidden dimension to (Batch_Size, 2944) using the torch.flatten(start_dim=1) operator, and then concatenates it along the feature dimension (torch.cat) to form a huge one-dimensional feature vector with a dimension as high as (Batch_Size, 5888), which is used to capture the nonlinear combination of all dimensions.

[0111] The vector is input into a deep multilayer perceptron (MLP) network. In this embodiment, the number of channels in the MLP layers is strictly configured as 5888 -> 1024 -> 256 -> 64 -> 1. The LeakyReLU activation function (with a negative slope of 0.01) is used between hidden layers to avoid neuron death, and a Dropout layer with a dropout rate of 0.3 is configured after the first two layers.

[0112] The final layer of the MLP uses an activation-free linear mapping to directly output the predicted continuous LFC power score. In the network backpropagation optimization settings, the batch size is set to 256, and the AdamW optimizer is used. The weight decay factor is set to 1e-4), and the initial learning rate is set to... It also employs a learning rate dynamic decay strategy based on cosine annealing.

[0113] To constrain the topological consistency of the network in bidirectional spatial modeling, the loss function incorporates the mean squared error used for regression prediction. And the bidirectional attention difference symmetry penalty computed via the Frobenius norm. In total losses Driven by this, the gradient descent guides the entire parameter space toward the global optimum.

[0114] 8. Comparison of Predictive Performance Evaluation and Experimental Results In the training and testing phases, to ensure rigorous evaluation of the model's generalization ability, the system performed 16-fold cross-validation based on a massive dataset covering 16 target genes. This involved completely isolating all sequences of a target gene (including perfect matches and all mutants) as an unknown test set in each fold, and training the model using the remaining 15 genes to simulate real-world clinical scenarios targeting novel targets. To comprehensively quantify the model's classification and regression prediction performance when facing unknown targets, this embodiment introduced four key evaluation metrics in each fold validation: Area Under the Receiver Operating Characteristic (AUROC), Area Under the Precision-Recall (AUPRC), Pearson Correlation Coefficient (PCC), and Spearman Rank Correlation Coefficient (SCC). AUROC and AUPRC are used to evaluate the model's binary classification ability against high / low activity guide sequences (the former reflects overall ranking ability, while the latter is more sensitive to imbalanced data). Their continuous integral mathematical definitions are as follows:

[0115]

[0116] In the formula The true positive rate, The false positive rate, For accuracy, This refers to the recall rate.

[0117] PCC is used to measure the absolute value of predictive power ( ) and the true logarithmic folding change label ( The degree of linear correlation between the two is calculated using the following formula:

[0118] In the formula and These are the sample means of the actual value and the predicted value, respectively.

[0119] SCC is used to assess the monotonic rank correlation between the distribution of predicted and actual values. The calculation formula is:

[0120] In the formula For the first The difference in rank between a sample in the true value sequence and the predicted value sequence. This represents the total number of samples in the test set.

[0121] Since this method strictly implements 16-fold cross-validation to correspond to 16 independent and non-overlapping target gene test sets, in order to objectively reflect the global stability and performance fluctuation of the model, the results of each evaluation index are aggregated and presented in the statistical form of "mean ± standard deviation (Mean ± SD)". To verify the technical progress of the present invention, this embodiment, based on the above rigorous evaluation system, first compares the overall performance of the present invention (LJDeep) with the current mainstream cutting-edge deep learning models (TIGER, DeepCas13) and traditional machine learning models (random forest, XGBoost, Adaboost). The specific test results are shown in Table 2.

[0122] Table 2. Overall performance evaluation comparison between LJDeep and existing models in perfect matching (PM) scenarios.

[0123] As shown in Table 2, the LJDeep model in this embodiment achieved optimal performance in both classification and regression tasks. In particular, in terms of the Pearson correlation coefficient (PCC) (0.79), which reflects the linear correlation of targeting efficacy, this invention outperforms the TIGER model (0.77), which fails to integrate the global structural context, and traditional models that rely on manual features, demonstrating the overall effectiveness of the cross-modal large model feature fusion mechanism.

[0124] To further demonstrate the technical contributions and necessity of each core module in the architecture of this invention, this embodiment designed and executed an ablation experiment of the system. The three core innovative modules of this invention (learnable physical perception embedding, RNA-FM large language model fusion, and bidirectional position-biased cross-attention) were removed or replaced respectively, and tested under the same training parameters. The results of the ablation experiment are shown in Table 3.

[0125] Table 3. Ablation experimental performance verification of the model in this embodiment under different network structure configurations.

[0126] As shown in Table 3, the ablation experiment data shows that the model performance significantly decreased when any of the core modules claimed in this invention were removed: (1) When degenerated to traditional one-hot encoding, the PCC plummeted to 0.62, indicating that learnable physical perception embedding plays an important role in model performance. (2) After stripping the features of the pre-trained large language model, the overall indicators declined across the board, confirming the gains of the "cross-modal attention fusion" in this embodiment in overcoming the blind spots of traditional manual feature engineering and understanding the spatial conformation and long-range folding relationships of complex RNAs. (3) When conventional unidirectional or multi-head self-attention is used to replace the cross-attention of this invention, the prediction correlation (PCC) is as low as 0.55 due to the lack of spatial location guidance, indicating that the "bidirectional cross-attention mechanism with segmented bias" in this embodiment plays an important role in simulating the dynamic recognition of gRNA and target RNA and accurately quantifying the long- and short-range spatial mismatch interaction gradients.

[0127] Furthermore, to verify the robustness of the method of the present invention in the face of complex RNA spatial breaks and large gradient mismatch environments, this embodiment tested the performance degradation gradient of the model under different mismatch complexities, and the results are shown in Table 4.

[0128] Table 4 Performance gradient verification of the LJDeep model under different mismatch complexities in this embodiment.

[0129] As shown in Table 4, thanks to the bidirectional position-biased cross-attention mechanism, this invention maintains extremely high prediction performance (e.g., PCC reaches 0.78 in the SM scenario) even when facing scenarios where local thermodynamic bubbles are generated, such as single-point (SM) and continuous two-point mismatch (DM). Although the performance decreases when dealing with global conformational collapse caused by random multi-point discrete mismatch (RDM / RTM), its decay gradient is gentle and the overall performance baseline is still higher than that of existing mainstream frameworks, indicating that the method of this invention has stability and application value.

[0130] The comprehensive implementation results confirm that the prediction architecture of this invention achieves good prediction performance. It not only performs well on perfectly matched data but also maintains extremely high prediction stability in complex multi-point mismatch scenarios. Furthermore, it demonstrates its effectiveness in prediction tasks, outperforming comparative models that do not incorporate structural context information.

[0131] The beneficial effect of this embodiment is that it realizes end-to-end prediction of the targeting efficacy of CRISPR-Cas13d system guide RNA through deep semantic feature fusion and dynamic bidirectional attention mechanism, effectively overcoming the limitations of traditional shallow machine learning models that rely on handcrafted features or existing deep learning algorithms that only focus on local sequence mismatches.

[0132] First, the method in this embodiment breaks through the bottleneck of shallow sequence complementarity matching. By constructing a dynamically fine-tunable physical perception embedding matrix and introducing the RNA large language model (RNA-FM), it not only accurately captures the local thermodynamic features of base pairing, but also integrates the long-range global structural context of the target RNA (such as stem-loop structure or potential folding pattern), thereby significantly improving the model's prediction accuracy for target accessibility.

[0133] Secondly, the innovative bidirectional positional bias cross-attention mechanism in this embodiment uses a piecewise function to accurately quantify the spatial mismatch tolerance at different distances. This mechanism can not only automatically focus on the core "seed region" (positions 15-21) that determines the binding and cutting activity of Cas13d, but also accurately model the synergistic destructive effect and structural feedback caused by multi-point spatial distribution mismatch.

[0134] Finally, compared to models that can only assess the relative loss of activity of variants, this embodiment directly predicts the logarithmic fold change (LFC) value of gene targeting efficacy, achieving an AUROC of 0.91 and a PCC of 0.79 in benchmark tests covering hundreds of thousands of systematic mismatches. It demonstrates the best generalization performance to date in perfect matching and various complex multi-point mismatch scenarios, providing a computationally valuable framework for efficient de novo guide sequence design for CRISPR-RNA editing systems.

[0135] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0136] Corresponding to the targeting efficacy prediction method of the CRISPR-Cas13d-like system described in the above embodiments, Figure 2 The diagram shows a structural block diagram of a targeting efficacy prediction device for a CRISPR-Cas13d-like system provided in an embodiment of this application. For ease of explanation, only the parts relevant to the embodiments of this application are shown.

[0137] Reference Figure 2 The device includes: The physical perception interaction module 201 is used to: process the first RNA features via a learnable perception matrix from the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing first training on the first sample. The physical interaction information refers to the physical interaction and stability differences of all possible base pair combinations between two RNA sequences based on preset thermodynamic pairing prior rules. Local feature module 202, used for: based on the second RNA feature Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; Global feature module 203 is used to: based on the first RNA feature Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; fusion module 204 is used for: based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; Bidirectional cross-attention module 205, used for: inputting the fused features To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. Regression module 206 is used to: input the bidirectional attention features , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

[0138] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0139] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0140] This application also provides a terminal device, such as... Figure 3 As shown, the terminal device 30 includes: at least one processor 301, a memory 302, and a computer program 303 stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the steps in any of the above-described method embodiments.

[0141] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0142] This application provides a computer program product that, when run on a mobile terminal, enables the mobile terminal to implement the steps described in the above-described method embodiments.

[0143] 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, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0144] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0145] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0146] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0147] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting the targeted efficacy of a CRISPR-Cas13d-like system, characterized in that, include: The first RNA features are learned through the perceptual matrix of the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing first training on the first sample. The physical interaction information refers to the physical interaction and stability differences of all possible base pair combinations between two RNA sequences based on preset thermodynamic pairing prior rules. Based on the second RNA characteristic Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; Based on the first RNA characteristics Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; Based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; Input the fusion feature To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. Input the bidirectional attention feature , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

2. The method for predicting the targeting efficacy of a CRISPR-Cas13d-like system as described in claim 1, characterized in that, The learnable perceptual matrix via the first machine learning model will incorporate the first RNA features. Mapping to second RNA features containing physical interaction information The steps include: Obtain the original RNA sequence of the CRISPR-Cas13d-like system Regarding the original RNA sequence Perform a truncation or padding operation to obtain a product with a fixed length. Fixed-length RNA sequences ; Encoding the fixed-length RNA sequence The first RNA feature was obtained. ; For the learnable perceptual matrix and the first RNA feature Second RNA features were obtained by performing physical interaction operations. The physical interaction operation includes broadcast dot product and outer product mapping executed sequentially.

3. The method for predicting the targeting efficacy of a CRISPR-Cas13d-like system as described in claim 1, characterized in that, The second RNA feature Running the encoder part of the first machine learning model yields local features. The steps include: The second RNA feature is generated using a projection operation with a bias term. Mapping to a feature space with hidden dimensions and injecting positional encoding yields the hidden dimension input features. The location encoding is injected through feature superposition and / or feature space transformation. Input features using the hidden dimension The encoder portion of the first machine learning model is used as input to obtain local features. .

4. The method for predicting the targeting efficacy of a CRISPR-Cas13d-like system as described in claim 1, characterized in that, The second machine learning model is obtained by performing a second training on a second sample. The second training is earlier than the first training. During the execution of the first training, the parameters of the second machine learning model are frozen.

5. The method for predicting the targeting efficacy of a CRISPR-Cas13d-like system as described in claim 1, characterized in that, The bidirectional cross-attention layer includes a bias matrix, which is used to adjust the bias based on any two positions in the RNA sequence. , relative distance between Set the bias coefficient; Wherein, the bias coefficient is the relative distance A function is used to simulate distance-based spatial mismatch tolerance, and the bias coefficient is relative to the relative distance. The function includes a non-linear component.

6. The method for predicting the targeting efficacy of a CRISPR-Cas13d-like system as described in claim 1, characterized in that, The regression component is a multilayer perceptron; The input of the bidirectional attention feature , The steps for obtaining the targeted efficacy prediction results in the regression part of the third machine learning model include: splicing the bidirectional attention features , This yields the total feature vector that includes all bidirectional interactive information; The total feature vector is input into the multilayer perceptron, and the dimensions are compressed layer by layer. The output is the log-folded change value (LFC) used to normalize the target efficacy prediction score.

7. A targeting efficacy prediction device similar to a CRISPR-Cas13d system, characterized in that, include: The physical perception interaction module is used to: process the first RNA features via a learnable perception matrix from the first machine learning model. Mapping to second RNA features containing physical interaction information Among them, the category index subscript It has at least two values and The first machine learning model is obtained by performing a first training on a first sample, and the physical interaction information refers to the physical interaction of all possible base pair combinations between two RNA sequences. Local feature module, used for: based on the second RNA feature Running the encoder part of the first machine learning model yields local features. The encoder includes... For input, get The first branch and with For input, get The second branch, the first branch is symmetrical to the second branch; The global feature module is used to: based on the first RNA feature Run the second machine learning model to obtain global features. The second machine learning model is a pre-trained biological domain semantic understanding model, and the global features... The hidden state output matrix is ​​extracted from the second machine learning model; The fusion module is used for: based on local features Constructing query tensors based on global features Construct key and value tensors, and perform attention-based fusion using the query tensor, key tensor, and value tensor to obtain fused features. ; A bidirectional cross-attention module is used to: input the fused features To the bidirectional cross-attention layer of the third machine learning model, respectively along to , to Cross-attention calculation is performed in both directions to obtain bidirectional attention features. , The third machine learning model is obtained by performing a third training operation using a third sample. The regression module is used to: input the bidirectional attention features , The regression part of the third machine learning model is used to obtain the targeted efficacy prediction results.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

10. A computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images