A cross-modal targeted capture probe design method and system based on physical information prior and domain adversarial network

Abstract

The application provides a cross-modal targeted capture probe design method and system based on physical information priori and field adversarial network, wherein the method comprises the following steps: acquiring targeted capture data sets of different experimental modes under different sequencing platforms, each piece of data in the targeted capture data set comprising a probe sequence and a real sequencing depth of the probe sequence in a target region, and assigning a field label and a regression label to each piece of data based on the target region and the real sequencing depth; calculating and extracting multi-dimensional core physical characteristics of the probe sequence through a thermodynamic parameter calculation engine; constructing a hybrid neural network of a Transformer and a CNN, extracting local k-mer modal characteristics of the probe sequence through the CNN, capturing global dependency of the probe sequence through the Transformer, fusing the local k-mer modal characteristics and the global dependency, and obtaining sequence coding characteristics. Through the method and the corresponding system, the capture efficiency of the probe can be accurately predicted.

Description

Technical Field

[0001] This invention proposes a cross-modal targeted capture probe design method and system based on physical information priors and domain adversarial networks, which relates to the field of molecular biology detection. Background Technology

[0002] With the rapid development of next-generation sequencing technology, targeted capture sequencing, as a technique for high-depth sequencing of specific genomic regions, has been widely used in clinical genetic disease screening, tumor liquid biopsy, and population genetics research. In targeted capture experiments, the quality of probe design directly determines the capture efficiency, data uniformity, and sequencing cost of the target region.

[0003] Existing deep learning-based next-generation sequencing probe design tools often treat probe design as a purely sequence extraction problem, neglecting objective thermodynamic physical laws in nucleic acid hybridization, such as Tm values, free energy, hairpin structures, and dimers. This leads to a significant drop in prediction accuracy in regions with extreme GC content or those prone to secondary structure formation. Severe platform and modality biases also exist, as different commercial kits, such as IDT and Twist, have different hybridization buffer formulations and synthesis processes. Existing models often overfit to specific training data, resulting in probes designed for platform A performing poorly on platform B. Furthermore, current technologies cannot simultaneously design probes compatible with both standard DNA sequencing and sulfite-converted methylated sequencing data. Existing tools typically only allow for fixed-length (e.g., 120 nt) sequence evaluation and cannot dynamically find the most efficient and cost-effective probe length combinations based on the complex sequence environment surrounding the target site. Summary of the Invention

[0004] This invention provides a cross-modal targeted capture probe design method and system based on physical information priors and domain adversarial networks to solve the aforementioned problems: This invention proposes a cross-modal targeted capture probe design method based on physical information priors and domain adversarial networks, the method comprising: Targeted capture datasets for different experimental modalities under different sequencing platforms are obtained. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a neighborhood label and a regression label are assigned to each data point. The multidimensional core physical features of the probe sequence are calculated and extracted using a thermodynamic parameter calculation engine. A hybrid neural network of Transformer and CNN is constructed. The local k-mer modal features of the probe sequence are extracted by CNN, and the global dependencies of the probe sequence are captured by Transformer. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. In the back end of the hybrid neural network, a parallel regression prediction branch and a domain adversarial branch are connected. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch, and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. The target genome region input by the user is obtained, and the target genome region is traversed using a variable-length sliding window to generate multi-scale candidate probe sequences. Extract the fusion feature vector of each candidate probe sequence in the multi-scale candidate probe sequence, input the fusion feature vector into the hybrid neural network model, obtain the capture efficiency prediction efficiency of the candidate probe sequence, and output the probe combination and its coordinates sorted by prediction efficiency.

[0005] Furthermore, the multidimensional core physical characteristics include: dissolution temperature, GC content percentage, heterodimer Gibbs free energy, normalized theoretical yield, hairpin structure Tm value, homodimer Tm value, GC skewness, Shannon entropy, 5' end Gibbs free energy, 3' end Gibbs free energy, normalized length, and collision penalty.

[0006] Furthermore, the back-end parallel regression prediction branch of the hybrid neural network includes: The regression prediction branch consists of multiple fully connected layers, which are used to map the input fused feature vector to the capture efficiency prediction score of the probe sequence. The capture efficiency prediction score and the regression label are substituted into the regression loss function to calculate the regression loss.

[0007] Furthermore, the back-end parallel domain adversarial branch of the hybrid neural network includes: The domain adversarial branch consists of a gradient inversion layer and a domain classifier; The domain classifier is used to map the input fused feature vector to a domain label prediction value, and to substitute the domain label prediction value and the domain label into the domain classification loss function to calculate the domain classification loss. The gradient inversion layer inverts the gradient of the domain classifier and passes it to the hybrid neural network during backpropagation.

[0008] Furthermore, the physical characteristic risks include at least one of the following: the hairpin structure Tm value exceeds a preset threshold, the homodimer Tm value exceeds a preset threshold, the heterodimer Gibbs free energy is lower than a preset threshold, the GC content percentage exceeds a preset range, and the normalized length exceeds a preset boundary.

[0009] Furthermore, the total loss function is: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

[0010] Furthermore, the length of the variable-length sliding window is calculated as follows: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

[0011] This invention proposes a cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks. The system includes: The label assignment module is used to obtain targeted capture datasets for different experimental modalities under different sequencing platforms. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a domain label and a regression label are assigned to each data point. A multi-dimensional core physical feature extraction module is used to calculate and extract the multi-dimensional core physical features of the probe sequence through a thermodynamic parameter calculation engine. A module for obtaining fused feature vectors is used to construct a hybrid neural network of Transformer and CNN. The CNN extracts the local k-mer modal features of the probe sequence, and the Transformer captures the global dependencies of the probe sequence. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. The training model module is used to connect the regression prediction branch and the domain adversarial branch at the back end of the hybrid neural network. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. A candidate probe sequence generation module is used to obtain the target genome region input by the user, traverse the target genome region with a variable-length sliding window, and generate multi-scale candidate probe sequences. The output probe combination module is used to extract the fusion feature vector of each candidate probe sequence in the multi-scale candidate probe sequence, input the fusion feature vector into the hybrid neural network model, obtain the capture efficiency prediction efficiency of the candidate probe sequence, and output the probe combination and its coordinates sorted by prediction efficiency.

[0012] Furthermore, the training model module includes: Construct the total loss function module; the total loss function is: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

[0013] Furthermore, the candidate probe sequence generation module includes: A module for calculating the length of a variable-length sliding window is used to calculate the length of the variable-length sliding window according to the following method: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

[0014] The beneficial effects of this invention are as follows: By fusing prior features of nucleic acid thermodynamic physics with a deep learning model (Transformer-CNN) and introducing contrastive learning and domain adversarial neural networks (DANN), this invention can accurately predict probe capture efficiency; effectively eliminate systematic biases caused by experimental platforms and sequencing modalities, and achieve robust design that is universal across platforms / modalities with a single training; support dynamic length (e.g., 60-120 nt) sliding window optimization, automatically avoid high-risk secondary structures, and achieve perfect coverage of clinical mutation hotspots (e.g., TP53 gene sequence). Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in this invention; Figure 2 This image shows the practical application effect of the cross-modal targeted capture probe design method based on physical information prior and domain adversarial network described in this invention on the clinically important oncogene TP53 Exon 7. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Numerous specific details are set forth in the following description to provide a thorough understanding of the invention. The described embodiments are only a part of, and not all, of the embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0019] Example 1, refer to Figure 1 As shown, a cross-modal targeted capture probe design method based on physical information priors and domain adversarial networks is presented, the method comprising: Targeted capture datasets for different experimental modalities under different sequencing platforms are obtained. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a neighborhood label and a regression label are assigned to each data point. The multidimensional core physical features of the probe sequence are calculated and extracted using a thermodynamic parameter calculation engine. A hybrid neural network of Transformer and CNN is constructed. The local k-mer modal features of the probe sequence are extracted by CNN, and the global dependencies of the probe sequence are captured by Transformer. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. In the back end of the hybrid neural network, a parallel regression prediction branch and a domain adversarial branch are connected. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch, and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. The target genome region input by the user is obtained, and the target genome region is traversed using a variable-length sliding window to generate multi-scale candidate probe sequences. Extract the fused feature vector of each candidate probe sequence from the multi-scale candidate probe sequences, input the fused feature vector into the hybrid neural network model, obtain the capture efficiency and prediction efficiency of the candidate probe sequence, and output the probe combinations and their coordinates sorted by prediction efficiency, such as... Figure 2 As shown, the best candidate probe (Rank 4, shown by the red band in the figure) performs as follows: Prediction efficiency: The AI ​​efficiency score Z-score given by the model reaches 0.7430, which is of extremely high confidence; Coverage of key regions: The probe not only completely covers the target exon region (shown by the orange box), but also accurately covers the clinical high-frequency mutation hotspot R248 (the purple dashed area).

[0020] The working principle and effects of the above technical solution are as follows: It not only extracts the sequence information of the probe but also obtains the core physical characteristics of the probe (such as thermal stability and free energy) through a thermodynamic parameter calculation engine. In the feature extraction layer, a CNN (Convolutional Neural Network) is used to capture local k-mer modal features (such as base motifs), and a Transformer is used to capture long-distance global dependencies. Finally, the physical features and sequence coding features are concatenated into vectors to form a fused feature vector that combines biochemical mechanisms and sequence semantic information. To address data bias caused by different sequencing platforms or experimental modalities, the solution introduces a domain adversarial branch in the backend of the model. This branch learns how to blur the boundaries between different domain labels, forcing the feature extraction layer to extract platform-independent general features. This multi-task learning strategy (regression prediction + domain classification adversarial) enables the model to learn the essential laws of capture efficiency, rather than noise from specific experimental environments. In the application, the system is not limited to fixed-length sequences but uses a variable-length sliding window to traverse the target genome region. This approach simulates the possible variations in probe length in real experiments, generating multi-scale candidate probe sequences and ensuring the diversity and coverage integrity of the candidate pool. The trained hybrid neural network model, acting as an end-to-end evaluator, can directly output a capture efficiency prediction value from the input fused feature vector. By scoring and ranking all candidate sequences, the system can automatically select the optimal probe combination and its precise coordinates, achieving automated transformation from genomic regions to high-performance probe design.

[0021] Through a hybrid architecture combining CNN and Transformer, the model can simultaneously understand the local motif and overall structural features of probes. Combined with thermodynamic physical features, this significantly improves the accuracy of predicting capture efficiency for complex regions (such as high GC content and repetitive sequence regions). The introduction of domain adversarial branches effectively suppresses experimental batch effects and platform differences. This means that after training on one dataset, the model can quickly transfer and accurately predict probe performance on other sequencing systems, reducing the cost of retraining the model for new platforms. It significantly improves the uniformity and coverage of sequencing data. Because it can accurately predict the true performance of probes, the system outputs probe combinations sorted by predicted efficiency, effectively avoiding cold spots (insufficient coverage) and hot spots (over-sequencing) in target regions, thereby improving the uniformity of sequencing data and reducing subsequent sequencing costs. The design of multi-scale sliding windows and multi-dimensional feature fusion allows it to adapt to probe selection requirements of different lengths and experimental conditions, providing an efficient tool for customized targeted capture chip design in scientific research and clinical applications.

[0022] In one embodiment of the present invention, the multidimensional core physical features include: dissolution temperature, GC content percentage, heterodimer Gibbs free energy, normalized theoretical yield, hairpin structure Tm value, homodimer Tm value, GC skewness, Shannon entropy, 5' end Gibbs free energy, 3' end Gibbs free energy, normalized length, and collision penalty.

[0023] In specific applications, dissolution temperature (Tm) refers to the temperature at which 50% of the DNA double helix dissociates into single strands during the denaturation process, expressed in degrees Celsius. Tm directly affects the specificity of probe-target binding. If Tm is too low, the probe may not bind effectively to the target; if Tm is too high, it may bind to non-specific sequences. Typically, probe Tm values ​​should be set between 55-70°C, and the Tm values ​​of upstream and downstream probes or primers should be as close as possible. GC content percentage refers to the percentage of guanine (G) and cytosine (C) bases in the probe sequence relative to the total number of bases. Heterodimer Gibbs free energy refers to the change in Gibbs free energy when two different probe molecules interact and form a dimeric complex. Normalized theoretical yield refers to the probability or efficiency that a probe can theoretically bind successfully to the target and generate an effective signal under ideal hybridization conditions; it is a dimensionless value between 0 and 1 obtained after normalization. A hairpin structure refers to a local stem-loop secondary structure formed by the self-folding of the probe sequence. Its Tm value is the temperature at which 50% of the hairpin structure unfolds into a random single-stranded state. A homodimer is a dimer complex formed between two identical probe molecules through complementary base pairing. The Tm value of a homodimer is the temperature at which 50% of the homodimer dissociates into monomers. GC skew is an indicator that measures the relative difference in the number of G and C bases in a sequence, used to describe the base composition asymmetry of the probe sequence along the 5′→3′ direction. Shannon entropy, derived from information theory, is used here to quantify the uncertainty of the base composition at each site in the probe sequence, i.e., the information content of the sequence site. The 5′ Gibbs free energy refers to the change in free energy of the base pairing structure formed in the local region of the 5′ end of the probe (usually the first 5-8 bases) when participating in target hybridization. The 3' Gibbs free energy refers to the change in free energy of the base pairing structure formed in a local region at the 3' end of the probe (usually the last 5-8 bases) during target hybridization. Normalized length is the dimensionless value obtained by normalizing the physical length (number of bases) of the probe. Collision penalty is an empirical regularization feature used to penalize probes that overlap, compete with, or sterically conflict with other candidate probes on the target genome. Its value is typically a positive number between 0 and 1; the higher the collision risk, the larger the penalty value.

[0024] In one embodiment of the present invention, the back-end parallel regression prediction branch of the hybrid neural network includes: The regression prediction branch consists of multiple fully connected layers, which are used to map the input fused feature vector to the capture efficiency prediction score of the probe sequence. The capture efficiency prediction score and the regression label are substituted into the regression loss function to calculate the regression loss.

[0025] In one embodiment of the present invention, the back-end parallel domain adversarial branch of the hybrid neural network includes: The domain adversarial branch consists of a gradient inversion layer and a domain classifier; The domain classifier is used to map the input fused feature vector to a domain label prediction value, and to substitute the domain label prediction value and the domain label into the domain classification loss function to calculate the domain classification loss. The gradient inversion layer inverts the gradient of the domain classifier and passes it to the hybrid neural network during backpropagation.

[0026] The working principle and effect of the above technical solution are as follows: The regression prediction branch utilizes the high-order nonlinear fitting capability of multiple fully connected layers to map the high-performance fused feature vector extracted from the front end, which contains both physical and sequence features, to a continuous numerical space. Each fully connected layer performs weighted combination and nonlinear activation of the features through a weight matrix, ultimately outputting a prediction score representing the capture efficiency. The difference between the predicted value and the true label is measured by a regression loss function (such as mean squared error MSE), guiding the model to learn the key weights affecting the capture efficiency. The fused feature vector is fed into the domain classifier, attempting to identify which sequencing platform or experimental modality the data comes from, i.e., predicting the domain label. During backpropagation, the gradient inversion layer (GRL) multiplies the gradient generated by the domain classifier by a negative constant (inverts it). This means that while the domain classifier is trying to learn how to distinguish the domain, the signal received by the front-end hybrid neural network (feature extractor) is how to extract features that the classifier cannot distinguish the domain. The model minimizes both the regression loss and the domain classification loss during training. Under the influence of GRL, the model is forced to remove platform-specific information carried in the features (adversarial branching objective) while ensuring the accuracy of prediction capture (regression branching objective), thereby extracting common essential features across platforms and modalities.

[0027] In one embodiment of the present invention, the physical characteristic risk includes at least one of the following: the hairpin structure Tm value exceeds a preset threshold, the homodimer Tm value exceeds a preset threshold, the heterodimer Gibbs free energy is lower than a preset threshold, the GC content percentage exceeds a preset range, and the normalized length exceeds a preset boundary.

[0028] In specific application scenarios, the above rules can be used individually or in combination. In the actual process, the candidate probes output by the model are first sorted by prediction efficiency, and then their physical characteristics are checked to see if they violate any risk rules. If they do, their ranking is lowered or they are directly eliminated, and finally, a combination of efficient probes that meets physical feasibility is output.

[0029] One embodiment of the present invention provides a cross-modal targeted capture probe design method based on physical information priors and domain adversarial networks, wherein the total loss function is: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

[0030] In specific application scenarios, the goal of contrastive learning is to bring positive sample pairs (such as the feature representation of the same probe sequence under different experimental modalities, or different probes with similar capture efficiencies under the same modality) closer together in the feature space, and to push away negative sample pairs (such as probes with large differences in capture efficiencies, or samples from different fields but with unrelated efficiencies).

[0031] in, and This represents a positive sample pair, which consists of feature vectors from two samples that belong to the same data source but have undergone different data augmentation or processing. This represents the other samples, in a batch of 2N samples, besides... It itself, and all other samples All are assumed to be negative samples, This represents the temperature coefficient.

[0032] One embodiment of the present invention provides a cross-modal targeted capture probe design method based on physical information priors and domain adversarial networks, wherein the length of the variable-length sliding window is calculated according to the following method: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. A positive value indicates stability; the original calculated value is usually negative, but here we use the absolute value or a positive value directly. The larger the value, the more stable the hairpin structure, but the higher the risk. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

[0033] The working principle and effects of the above technical solution are as follows: In this study, when the GC content is 50%, the factor is 1 (the longest value is taken); when the GC content is 0% or 100%, the factor is 0 (the shortest value is taken). Sequences with GC content extremely below 50% exhibit thermodynamic abnormalities (too high a value easily leads to non-specific binding, too low a value results in weak binding), so short probes are recommended to reduce risk. The DNA sequence within the window (composed of the four bases A, T, G, and C) is selected. The frequency of each of the four bases appearing within this window is then statistically analyzed. , , In this context, Shannon entropy H(i) measures the information content / complexity of a sequence. Higher entropy indicates a more random sequence with better specificity, allowing for longer probes; low entropy (such as repetitive sequences) indicates poor specificity, requiring the use of short probes.

[0034] Example 2: A cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks, the system comprising: The label assignment module is used to obtain targeted capture datasets for different experimental modalities under different sequencing platforms. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a domain label and a regression label are assigned to each data point. A multi-dimensional core physical feature extraction module is used to calculate and extract the multi-dimensional core physical features of the probe sequence through a thermodynamic parameter calculation engine. A module for obtaining fused feature vectors is used to construct a hybrid neural network of Transformer and CNN. The CNN extracts the local k-mer modal features of the probe sequence, and the Transformer captures the global dependencies of the probe sequence. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. The training model module is used to connect the regression prediction branch and the domain adversarial branch at the back end of the hybrid neural network. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. A candidate probe sequence generation module is used to obtain the target genome region input by the user, traverse the target genome region with a variable-length sliding window, and generate multi-scale candidate probe sequences. The output probe combination module is used to extract the fusion feature vector of each candidate probe sequence in the multi-scale candidate probe sequence, input the fusion feature vector into the hybrid neural network model, obtain the capture efficiency prediction efficiency of the candidate probe sequence, and output the probe combination and its coordinates sorted by prediction efficiency.

[0035] One embodiment of the present invention provides a cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks, wherein the training model module includes: The module for constructing the total loss function is as follows: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

[0036] One embodiment of the present invention provides a cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks, wherein the candidate probe sequence generation module includes: A module for calculating the length of a variable-length sliding window is used to calculate the length of the variable-length sliding window according to the following method: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

[0037] The cross-modal targeted capture probe design system based on physical information prior and domain adversarial network described in this embodiment of the invention is used to implement the cross-modal targeted capture probe design method based on physical information prior and domain adversarial network in Embodiment 1 above. Both are based on the same inventive concept and have the same technical effect, which will not be repeated here.

[0038] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0039] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes. Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for designing cross-modal targeted capture probes based on physical information priors and domain adversarial networks, characterized in that, The method includes: Targeted capture datasets for different experimental modalities under different sequencing platforms are obtained. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a neighborhood label and a regression label are assigned to each data point. The multidimensional core physical features of the probe sequence are calculated and extracted using a thermodynamic parameter calculation engine. A hybrid neural network of Transformer and CNN is constructed. The local k-mer modal features of the probe sequence are extracted by CNN, and the global dependencies of the probe sequence are captured by Transformer. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. In the back end of the hybrid neural network, a parallel regression prediction branch and a domain adversarial branch are connected. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch, and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. The target genome region input by the user is obtained, and the target genome region is traversed using a variable-length sliding window to generate multi-scale candidate probe sequences. Extract the fusion feature vector of each candidate probe sequence in the multi-scale candidate probe sequence, input the fusion feature vector into the hybrid neural network model, obtain the capture efficiency prediction efficiency of the candidate probe sequence, and output the probe combination and its coordinates sorted by prediction efficiency.

2. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The multidimensional core physical characteristics include: dissolution temperature, GC content percentage, heterodimer Gibbs free energy, normalized theoretical yield, hairpin structure Tm value, homodimer Tm value, GC skewness, Shannon entropy, 5' end Gibbs free energy, 3' end Gibbs free energy, normalized length, and collision penalty.

3. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The back-end parallel regression prediction branch of the hybrid neural network includes: The regression prediction branch consists of multiple fully connected layers, which are used to map the input fused feature vector to the capture efficiency prediction score of the probe sequence. The capture efficiency prediction score and the regression label are substituted into the regression loss function to calculate the regression loss.

4. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The back-end parallel domain adversarial branch of the hybrid neural network includes: The domain adversarial branch consists of a gradient inversion layer and a domain classifier; The domain classifier is used to map the input fused feature vector to a domain label prediction value, and to substitute the domain label prediction value and the domain label into the domain classification loss function to calculate the domain classification loss. The gradient inversion layer inverts the gradient of the domain classifier and passes it to the hybrid neural network during backpropagation.

5. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The physical characteristic risks include at least one of the following: the hairpin structure Tm value exceeds a preset threshold, the homodimer Tm value exceeds a preset threshold, the heterodimer Gibbs free energy is lower than a preset threshold, the GC content percentage exceeds a preset range, and the normalized length exceeds a preset boundary.

6. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The total loss function is: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

7. The cross-modal targeted capture probe design method based on physical information prior and domain adversarial network as described in claim 1, characterized in that, The length of the variable-length sliding window is calculated as follows: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

8. A cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks, characterized in that, The system includes: The label assignment module is used to obtain targeted capture datasets for different experimental modalities under different sequencing platforms. Each data point in the targeted capture dataset includes a probe sequence and its actual sequencing depth in the target region. Based on the target region and the actual sequencing depth, a domain label and a regression label are assigned to each data point. A multi-dimensional core physical feature extraction module is used to calculate and extract the multi-dimensional core physical features of the probe sequence through a thermodynamic parameter calculation engine. A module for obtaining fused feature vectors is used to construct a hybrid neural network of Transformer and CNN. The CNN extracts the local k-mer modal features of the probe sequence, and the Transformer captures the global dependencies of the probe sequence. The local k-mer modal features and the global dependencies are fused to obtain sequence encoding features. Then, the multidimensional core physical features and the sequence encoding features are concatenated to obtain the fused feature vector. The training model module is used to connect the regression prediction branch and the domain adversarial branch at the back end of the hybrid neural network. The fused feature vector is simultaneously input into the two branches to construct a total loss function. The optimization objective is to minimize the total loss function. The parameters of the hybrid neural network, the regression prediction branch and the domain classifier are iteratively updated using the backpropagation algorithm to obtain a trained hybrid neural network model. A candidate probe sequence generation module is used to obtain the target genome region input by the user, traverse the target genome region with a variable-length sliding window, and generate multi-scale candidate probe sequences. The output probe combination module is used to extract the fusion feature vector of each candidate probe sequence in the multi-scale candidate probe sequence, input the fusion feature vector into the hybrid neural network model, obtain the capture efficiency prediction efficiency of the candidate probe sequence, and output the probe combination and its coordinates sorted by prediction efficiency.

9. The cross-modal targeted capture probe design system based on physical information prior and domain adversarial network as described in claim 8, characterized in that, The training model module includes: The module for constructing the total loss function is as follows: in, Represents the total loss function. Represents the regression loss function. This represents the domain classification loss function. This represents the ranking-comparative learning loss function. , and This represents the weighting coefficient.

10. The cross-modal targeted capture probe design system based on physical information priors and domain adversarial networks according to claim 8, characterized in that, The candidate probe sequence generation module includes: A module for calculating the length of a variable-length sliding window is used to calculate the length of the variable-length sliding window according to the following method: in, Indicates the length of the sliding window. This indicates the preset minimum probe length. This indicates the preset maximum probe length. This represents the Shannon entropy of the probe sequence within the sliding window. This represents the maximum Shannon entropy of the preset probe sequence. This represents the free energy of the most stable hairpin structure within the sliding window. Indicated by The GC content of the probe sequence centered on the current candidate length window. This represents the preset threshold value for the free energy of the hairpin structure.

Images