Open reading frame prediction method and device, electronic equipment and medium

Abstract

The embodiment of the invention discloses an open reading frame prediction method and device, electronic equipment and a storage medium. The open reading frame prediction method comprises the following steps: acquiring a first modal feature corresponding to a deoxyribonucleic acid sequence and a second sample feature corresponding to a ribosome sequencing signal; constructing a multi-modal feature based on a bidirectional cross attention mechanism, the first modal feature and the second modal feature; according to the multi-modal features and a preset conditional diffusion model, generating an open reading frame sample; adding the open reading frame sample into the original sample set to obtain a balanced training set; performing disturbance processing on the balance training set, and extracting a high-dimensional embedded feature corresponding to each training sample in the processed balance training set through an independent encoder; fusing the high-dimensional embedded features based on a cross-modal consistency loss function to obtain multi-modal coding representation; and training a classification model according to the multi-modal coding representation and a preset target loss function to obtain a target classification model so as to predict the open reading frame.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to an open reading frame prediction method, apparatus, electronic device, and storage medium. Background Technology

[0002] An open reading frame (ORF) is a DNA sequence from the start codon to the stop codon. Small open reading frames (sORFs) are even shorter, less than 300 base pairs in size, and encode micropeptides that play important regulatory roles in life processes. To accurately delineate specific regions of expressed genes and achieve accurate gene annotation, open reading frame prediction involves predicting whether a segment of a DNA sequence is an open reading frame, obtaining a set of open reading frames applicable to biological experiments.

[0003] Existing open reading frame prediction models rely solely on ribosome profiling (Ribo-seq) signal distribution and DNA sequence characteristics for identification. Currently, only two models are specifically designed for prokaryotic prediction: REPARATION and DeepRibo. Compared to the former, the latter offers more automated feature extraction and can uncover latent features within the data.

[0004] However, due to the severe imbalance between the number of positive and negative samples in open reading frames (OPFs), with significantly fewer positive samples, and the DeepRibo model's simple use of a weighted loss function, improper weighting in the case of imbalanced data can easily lead to learning biases, resulting in low prediction accuracy. Furthermore, due to defects such as false negatives, existing prediction schemes perform poorly in predicting open reading frames in prokaryotes, leading to low-quality and unreliable predicted OPFs. Summary of the Invention

[0005] This application provides an open reading frame prediction method, electronic device, apparatus, and storage medium, which can solve the technical problems of low quality and limited experimental feasibility of open reading frames predicted by existing models.

[0006] In a first aspect, embodiments of this application provide an open reading frame prediction method, including:

[0007] Obtain the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal;

[0008] Multimodal features are constructed based on a bidirectional cross-attention mechanism, the first modal feature, and the second modal feature;

[0009] Based on the multimodal features and the preset conditional diffusion model, open reading box samples are generated;

[0010] The open reading frame samples are added to the original sample set to obtain a balanced training set;

[0011] The balanced training set is perturbed, and the high-dimensional embedding features corresponding to each training sample in the balanced training set are extracted by an independent encoder.

[0012] The high-dimensional embedding features are fused based on the cross-modal consistency loss function to obtain a multimodal coded representation;

[0013] The classification model is trained based on the multimodal encoding representation and the preset target loss function to obtain a target classification model for predicting open reading frames.

[0014] Optionally, in some embodiments of this application, the construction of multimodal features based on the bidirectional cross-attention mechanism, the first modal feature, and the second modal feature includes:

[0015] Using the first modal feature as the query, and based on the bidirectional cross-attention mechanism and the second modal feature, a cross-modal feature in the first direction is constructed;

[0016] Using the second modal feature as the query, and based on the bidirectional cross-attention mechanism and the first modal feature, a cross-modal feature in the second direction is constructed;

[0017] Multimodal features are obtained by fusing the cross-modal features in the first direction and the cross-modal features in the second direction.

[0018] Optionally, in some embodiments of this application, generating open reading box samples based on the multimodal features and a preset conditional diffusion model includes:

[0019] The multimodal features are used as conditions to input into a preset conditional diffusion model. Through the forward diffusion process of the conditional diffusion model, Gaussian noise is gradually injected into the original samples to construct a noise sample sequence.

[0020] Based on the noise sample sequence, through the reverse denoising process of the conditional diffusion model, open reading box samples consistent with the distribution of real open reading box data are generated under the conditional constraints of multimodal features.

[0021] Optionally, in some embodiments of this application, the perturbation processing of the balanced training set and the extraction of high-dimensional embedding features corresponding to each training sample in the processed balanced training set by an independent encoder include:

[0022] Biological perturbation of the deoxyribonucleic acid sequence samples in the balanced training set, and;

[0023] Noise processing is performed on the ribosome sequencing signal samples in the balanced training set to generate diverse views;

[0024] Extract the first high-dimensional embedding feature corresponding to the biologically perturbed deoxyribonucleic acid sequence sample, and;

[0025] Extract the second high-dimensional embedding features corresponding to the diverse views.

[0026] Optionally, in some embodiments of this application, the high-dimensional embedding feature mapping includes a first high-dimensional embedding feature and a second high-dimensional embedding feature, and the step of fusing the high-dimensional embedding features based on the cross-modal consistency loss function to obtain a multimodal coding representation includes:

[0027] The first high-dimensional embedding feature and the second high-dimensional embedding feature are mapped to the same embedding space to obtain the first projection vector corresponding to the first high-dimensional embedding feature and the second projection vector corresponding to the second high-dimensional embedding feature.

[0028] A multimodal coded representation is generated based on the cross-modal consistency loss function, the first projection vector, and the second projection vector.

[0029] Optionally, in some embodiments of this application, generating a multimodal coded representation based on the cross-modal consistency loss function, the first projection vector, and the second projection vector includes:

[0030] Obtain the contrastive loss function;

[0031] Based on the cross-modal consistency loss function and the contrast loss function, the first projection vector and the second projection vector are optimized respectively;

[0032] By fusing the optimized first projection vector with the optimized second projection vector, a multimodal coded representation is obtained.

[0033] Optionally, in some embodiments of this application, the step of training the classification model based on the multimodal coding representation and a preset loss function to obtain a target classification model for predicting open reading boxes includes:

[0034] The multimodal encoding representation is input into the classification model to determine the predicted label of each training sample belonging to the open reading frame;

[0035] Based on the true and predicted labels of each training sample, the classification model is trained to obtain the target classification model for predicting open reading frames.

[0036] Secondly, embodiments of this application provide an open reading frame prediction device, comprising:

[0037] The acquisition module is used to acquire the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal;

[0038] A construction module is used to construct multimodal features based on a bidirectional cross-attention mechanism, the first modal feature, and the second modal feature;

[0039] The generation module is used to generate open reading box samples based on the multimodal features and a preset conditional diffusion model;

[0040] An addition module is used to add the open reading frame samples to the original sample set to obtain a balanced training set;

[0041] The extraction module is used to perturb the balanced training set and extract the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder.

[0042] The fusion module is used to fuse the high-dimensional embedding features based on the cross-modal consistency loss function to obtain a multimodal coded representation;

[0043] The prediction module is used to train the classification model based on the multimodal encoding representation and the preset target loss function to obtain a target classification model for predicting open reading boxes.

[0044] Accordingly, this application also provides an electronic device, including a memory, a processor, and a processor program stored in the memory and executable on the processor, wherein the processor executes the program as described in any of the methods above.

[0045] This application also provides a storage medium storing a processor program that, when executed by a processor, implements any of the methods described above.

[0046] This application provides an open reading frame (OPG) prediction method, apparatus, electronic device, and storage medium. After acquiring the first modal features corresponding to the deoxyribonucleic acid (DNA) sequence and the second sample features corresponding to the ribosome sequencing signal, a multimodal feature is constructed based on a bidirectional cross-attention mechanism, the first modal features, and the second modal features. Then, based on the multimodal features and a preset conditional diffusion model, open reading frame samples are generated. Next, the open reading frame samples are added to the original sample set to obtain a balanced training set. Subsequently, the balanced training set is perturbed, and high-dimensional embedding features corresponding to each training sample in the balanced training set are extracted using an independent encoder. Finally, the high-dimensional embedding features are fused based on a cross-modal consistency loss function to obtain a multimodal encoded representation. A classification model is trained based on the multimodal encoded representation and a preset target loss function to obtain a target classification model for predicting open reading frames. In the open reading frame prediction scheme provided in this application, enhanced samples that conform to the true distribution are generated based on multimodal features to alleviate the data imbalance problem. This method generates supplementary positive samples through data generation, avoiding learning bias caused by unreasonable weight settings when using weighted loss in existing prediction models, thus improving prediction accuracy. In the feature learning stage, sample differentiation is achieved through contrastive learning pre-training, avoiding false negatives caused by misclassifying unknown samples as negative samples. Ultimately, this results in the prediction of high-quality, experimentally feasible open reading frames, improving prediction performance. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0048] Figure 1 This is a flowchart illustrating the open reading frame prediction method provided in an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the structure of the data augmentation module based on the multimodal conditional diffusion model provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of the structure of the classification prediction module based on contrastive learning and positive example-unlabeled learning provided in the embodiments of this application;

[0051] Figure 4 This is a schematic diagram of the open reading frame prediction device provided in the embodiments of this application;

[0052] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0053] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0054] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.

[0055] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0056] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustrative purposes and has no specific meaning in itself. Therefore, "module," "part," or "unit" may be used interchangeably.

[0057] The following describes in detail the embodiments involved in this application. It should be noted that the order of description of the embodiments in this application is not intended to limit the priority of the embodiments.

[0058] This application provides an open reading frame prediction method, apparatus, storage medium, and smart terminal. Specifically, the open reading frame prediction method of this application can be executed by a smart terminal or a server, wherein the smart terminal can be a terminal. The terminal can be a smartphone, tablet computer, laptop computer, touch screen, game console, personal computer (PC), personal digital assistant (PDA), or other smart terminal. The terminal may also include a client, which can be a media playback client or an instant messaging client, etc.

[0059] This application provides an open reading box prediction method, which can be executed by an electronic device or a server. This application example illustrates the method executed by an electronic device. The electronic device includes a touchscreen display and a processor. The touchscreen display is used to present a graphical user interface (GUI) and receive user input commands generated by the GUI. When the user operates the GUI through the touchscreen display, the GUI can control the content locally on the electronic device or control the content on the server side in response to the received input commands.

[0060] The open reading frame prediction scheme provided in this application generates augmented samples that conform to the true distribution based on multimodal features, alleviating the data imbalance problem. This method generates supplementary positive samples through data generation, avoiding learning bias caused by unreasonable weight settings when using weighted loss in existing prediction models, thus improving prediction accuracy. In the feature learning stage, sample differentiation is achieved through contrastive learning pre-training, avoiding false negatives caused by misclassifying unknown samples as negative samples. Ultimately, this results in the prediction of high-quality, experimentally feasible open reading frames, improving prediction performance.

[0061] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the priority of the embodiments.

[0062] An open reading frame (OPG) prediction method includes: acquiring first modality features corresponding to deoxyribonucleic acid (DNA) sequences and second sample features corresponding to ribosome sequencing signals; constructing multimodal features based on a bidirectional cross-attention mechanism, the first modality features, and the second modality features; generating OPG samples based on the multimodal features and a pre-defined conditional diffusion model; adding the OPG samples to the original sample set to obtain a balanced training set; perturbing the balanced training set and extracting high-dimensional embedding features corresponding to each training sample in the processed balanced training set using an independent encoder; fusing the high-dimensional embedding features based on a cross-modal consistency loss function to obtain a multimodal encoding representation; and training a classification model based on the multimodal encoding representation and a pre-defined target loss function to obtain a target classification model for predicting OPGs.

[0063] Please see Figure 1 , Figure 1 This application provides a flowchart illustrating the open reading frame prediction method. The specific flow of this open reading frame prediction method is as follows:

[0064] 101. Obtain the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal.

[0065] Deoxyribonucleic acid (DNA) sequences are linear molecular sequences composed of four bases—adenine (A), thymine (T), guanine (G), and cytosine (C)—linked by phosphodiester bonds. They serve as the carriers of genetic information, containing functional elements such as protein-coding regions (e.g., open reading frames) and non-coding regulatory regions. Ribosome sequencing (Ribo-seq) signals are high-throughput sequencing data obtained through ribosome profiling, reflecting the position, density, and residence time of ribosomes on mRNA being translated in cells. They can be used to infer dynamic characteristics such as translation initiation sites, codon bias, and translation efficiency. First-modality features are semantic features extracted from DNA sequences, used to characterize the sequence's structural and functional information, such as promoter motifs, open reading frame (ORF) integrity, GC content, and codon distribution. The second modality feature is the translation dynamics feature extracted from the Ribo-seq signal, which is used to characterize the behavior pattern of ribosomes on mRNA, such as the 3-nt periodicity of ribosome coverage (reflecting the codon reading frame), translation elongation rate, and ribosome hotspot regions.

[0066] Specifically, the DNA sequence is converted into a numerical input (e.g., one-hot encoding, where each base corresponds to a 4-dimensional vector); local features (e.g., short sequence motifs) are extracted using convolutional and pooling layers of a CNN; and a high-dimensional feature vector h is output through a fully connected layer. DNA The dimensions are typically 512-1024. Simultaneously, Ribo-seq signals are converted into time-series data (e.g., using mRNA transcripts as coordinates and arranging ribosome coverage intensity by position); long-range dependencies in the sequence (such as the global dynamics from translation initiation to termination) are captured through the gating mechanism of a gated recurrent unit (GRU, a type of recurrent neural network), outputting a feature vector h. Ribo Dimensions and h DNA Consistent.

[0067] 102. Construct multimodal features based on bidirectional cross-attention mechanism, first modality features, and second modality features.

[0068] Among them, the bidirectional cross-attention mechanism is a cross-modal feature interaction mechanism that achieves intermodal information fusion through bidirectional query-key-value pair interaction: using one modal feature as the query and another modal feature as the key and value, cross-modal attention weights are calculated; the above process is repeated in reverse to achieve bidirectional information flow, and finally the feature representation containing bimodal interaction information is obtained.

[0069] The bidirectional cross-attention mechanism explicitly models the bidirectional dependencies between modalities, enabling the semantic features of DNA sequences and the translational dynamics features of Ribo-seq to guide and enhance each other. The resulting multimodal features retain the unique information of each modality while capturing the synergistic relationship between the two (such as the correspondence between ORF sequences and translation start sites), providing a rich cross-modal semantic foundation for subsequent sample generation and classification tasks.

[0070] Optionally, in some embodiments of this application, the step "constructing multimodal features based on a bidirectional cross-attention mechanism, first modal features, and second modal features" may specifically include:

[0071] Using the first modality feature as the query, and based on the bidirectional cross-attention mechanism and the second modality feature, a cross-modality feature in the first direction is constructed;

[0072] Using the second modality feature as the query, and based on the bidirectional cross-attention mechanism and the first modality feature, a cross-modality feature in the second direction is constructed;

[0073] Multimodal features are obtained by fusing cross-modal features from the first direction and cross-modal features from the second direction.

[0074] For example, specifically, H DNA (First modality feature) as query, H Ribo (Second modal feature) simultaneously serves as both key and value. Cross-modal interaction weights are calculated using scaled dot product attention to generate the first-direction cross-modal feature, reflecting the guiding relationship of the DNA sequence on the Ribo-seq signal; similarly, H... Ribo (Second modality feature) as query, H DNA The first modal feature serves as both a key and a value. Cross-modal interaction weights are calculated using scaled dot product attention to generate a second-direction cross-modal feature that reflects the guiding relationship of the Ribo-seq signal on the DNA sequence. Then, the bidirectional cross-modal features are concatenated along the feature dimension to form a preliminary fusion feature containing bidirectional interaction information. Finally, the concatenated features are globally interacted through the Transformer self-attention mechanism to eliminate redundancy and enhance cross-modal consistency, ultimately outputting a multimodal feature.

[0075] 103. Generate open reading frame samples based on multimodal features and a preset conditional diffusion model.

[0076] Multimodal features are cross-modal representations that integrate semantic features of deoxyribonucleic acid (DNA) sequences with dynamic features of ribosome sequencing (Ribo-seq) signal translation. Conditional diffusion models are generative models based on stepwise noise diffusion and denoising, using multimodal features as conditional constraints to generate samples that conform to the true distribution. Open reading frame (ORF) samples are continuous nucleotide sequences in the genome from a start codon (e.g., AUG) to a stop codon (e.g., UAA / UAG / UGA), which may encode proteins. The generated ORF samples must satisfy the following conditions:

[0077] The sequence length is an integer multiple of 3 (corresponding to the codon reading frame);

[0078] The base distribution conforms to biological laws (such as GC content and codon usage bias);

[0079] The corresponding Ribo-seq signals exhibit translational activity (e.g., 3-nt periodicity, ribosome-covered peaks).

[0080] Conditional diffusion models, through semantic constraints on multimodal features, ensure that the generated ORF samples conform to both the grammatical rules of DNA sequences (such as codon frames) and possess the biological plausibility of translation dynamics (such as ribosome binding signals). Compared to traditional generative models (such as GANs), diffusion models can capture more complex data distributions through progressive denoising, while the introduction of multimodal conditions further enhances the biological interpretability and predictive value of the generated samples.

[0081] Optionally, in some embodiments of this application, the step of "generating open reading box samples based on multimodal features and a preset conditional diffusion model" may specifically include:

[0082] Multimodal features are used as conditional inputs to a pre-defined conditional diffusion model. Gaussian noise is gradually injected into the original samples through the forward diffusion process of the conditional diffusion model to construct a noise sample sequence.

[0083] Based on noisy sample sequences, open reading box samples consistent with the distribution of real open reading box data are generated through the reverse denoising process of the conditional diffusion model under the conditional constraints of multimodal features.

[0084] Known DNA sequences of real ORFs, such as validated short open reading frames (smORFs); multimodal features (Z): cross-modal representations corresponding to the original sample, containing sequence semantics and translation dynamics information.

[0085] The original samples are gradually perturbed with Gaussian noise, and the linear or cosine scheduling increases with the number of steps to ensure that the noise intensity gradually increases. The final result is a noise sample sequence {x1, x2, ..., x...}. T}, where x TThese are samples that are completely noisy (and have no relation to the original data distribution).

[0086] High-noise samples (x) T The initial noise level (x) serves as the starting point for generation and has no specific biological significance. Multimodal features (Z) are combined with a denoising neural network through an embedding layer to guide the generation direction. The denoising neural network predicts the current noise level and calculates the denoised sample x. t Relating Z to x t After splicing, the data is input into a neural network, or an attention mechanism is used to focus the network on key features in Z (such as DNA features corresponding to the start codon). After denoising, a generated sample T is obtained, whose DNA sequence must conform to the ORF format (including start / stop codons), and the corresponding virtual Ribo-seq signal must have real translation activity features (such as ribosome coverage peaks).

[0087] 104. Add the open reading frame samples to the original sample set to obtain a balanced training set.

[0088] The original sample set contains a real dataset of known open reading frames (ORFs), which often suffers from data imbalance—the number of positive samples (true ORFs) is far less than the number of negative samples (non-ORF regions) or unlabeled samples. For example, in short open reading frames (smORF) studies, experimentally validated positive samples are scarce, causing the model training to be biased towards predicting negative samples. Balancing the training set involves using data augmentation techniques (such as generating ORF samples using conditional diffusion models) to increase the number of positive samples in the original sample set, ensuring that the ratio of positive to negative (or unlabeled) samples in the training set is within a reasonable range, thus mitigating the impact of class imbalance on model training.

[0089] Specifically, the generated ORF samples are merged with the positive and negative samples (or unlabeled samples) in the original sample set to form a new training set.

[0090] Positive samples: Original positive samples + generated ORF samples;

[0091] Negative samples: Original non-ORF region samples (such as intron and inter-region sequences);

[0092] Unlabeled samples: Unverified potential ORF regions, treated as candidate positive examples in PU learning.

[0093] The merged training set is randomly divided, usually into a training set, a validation set, and a test set in an 8:1:1 ratio, for model training, hyperparameter tuning, and performance evaluation.

[0094] 105. Perturb the balanced training set and extract the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder.

[0095] Applying diverse perturbations to samples in a balanced training set to generate multiple enhanced views of the same original sample enhances the model's ability to generalize to data distributions. This can specifically include the following:

[0096] Biological perturbation: For deoxyribonucleic acid (DNA) sequence samples, we simulate the mutation process in natural evolution, such as random base mutations and local sequence scrambling, to increase data diversity while preserving the functional semantics of the sequence.

[0097] Technical noise handling: For ribosome sequencing (Ribo-seq) signal samples, simulate technical errors in the experimental data, such as time window shift and signal intensity normalization, to enhance the robustness of translation dynamic features.

[0098] Dedicated feature extraction networks designed for different modalities of data avoid the limitations of single-modal encoders in representing heterogeneous data.

[0099] Convolutional Neural Networks (CNNs): As independent encoders of DNA sequences, they capture local features of the sequence (such as promoter motifs and codon distribution) through convolutional layers and output a high-dimensional semantic feature vector h. DNA Gated Recurrent Unit (GRU): As an independent encoder of the Ribo-seq signal, it captures the temporal dependencies of translation dynamics (such as ribosome coverage periodicity and elongation rate) through a gating mechanism, and outputs a high-dimensional dynamic feature vector h. Ribo .

[0100] Optionally, in some embodiments of this application, the step "perturbing the balanced training set and extracting the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder" may specifically include:

[0101] Biological perturbation is performed on the deoxyribonucleic acid sequence samples in the balanced training set, and noise processing is performed on the ribosome sequencing signal samples in the balanced training set to generate diverse views;

[0102] The first high-dimensional embedding feature corresponding to the biologically perturbed deoxyribonucleic acid sequence sample is extracted, and the second high-dimensional embedding feature corresponding to the diversified views is extracted.

[0103] Biological perturbation refers to interference operations applied to DNA sequences that simulate natural evolution, including:

[0104] Random mutation: Randomly replacing a single base (such as A→T) to mimic a single nucleotide polymorphism (SNP);

[0105] Local sequence shuffling: Randomly rearrange the base sequence within a sliding window, disrupting the surface sequence structure but preserving long-distance functional dependencies.

[0106] Noise processing is the operation applied to the Ribo-seq signal to correct for simulation experimental errors, including:

[0107] Time window offset: shifts the signal time axis to simulate ribosome positioning error;

[0108] Intensity normalization: Adjusting signal intensity to correct sequencing depth bias.

[0109] A diversified view is a collection of enhanced versions of the same original sample after applying different perturbations, such as different noise versions of the same Ribo-seq signal (e.g., noise intensity 1, noise intensity 2).

[0110] Specifically, inputting DNA sequence samples from the balanced training set (including original positive samples, generated positive samples, and negative samples) generates diverse views through random mutation (mutation rate 5%-10%) and local sequence scrambling (window size 5-10 nt). For example, a DNA sequence containing a start codon, after perturbation, may generate a new view that retains the start codon but has a mutation in the middle base, or a view that scrambles bases in non-functional regions. Inputting Ribo-seq signal samples from the balanced training set (including real signals and generated signals) simulates experimental errors in ribosome localization through time sliding window offset (offset amount 1-3 nt), and eliminates sequencing depth differences between different samples through intensity normalization (such as Z-score normalization), generating signal views with different noise levels.

[0111] Local motif features are extracted using convolutional layers in a CNN (e.g., 3 layers with 3×3 kernels and a stride of 1). After dimensionality reduction using max pooling layers, a high-dimensional feature vector h is output by a fully connected layer. DNA For example, CNNs can capture TATA box features in promoter regions or codon frequency patterns in ORF regions. Temporal dependencies can be captured through the hidden layers of a GRU (such as 128-dimensional gated units), with the hidden state at the final time step serving as a high-dimensional feature vector h. Ribo For example, GRU can identify anomalous signal patterns in ribosome-covered signals, such as 3-nt periodicity (a characteristic of translationally active) or stagnant sites.

[0112] 106. Based on the cross-modal consistency loss function, high-dimensional embedding features are fused to obtain a multimodal coding representation.

[0113] The cross-modal consistency loss function measures the degree of semantic alignment between different modal features in the joint embedding space. By explicitly constraining the spatial distance between different modal features, it addresses the semantic gap problem in multimodal data. Combined with contrastive learning, this method not only achieves cross-modal alignment but also enhances the discriminative power between samples. This results in multimodal encoded representations that possess both semantic consistency (cross-modal feature alignment) and discriminative robustness (effective separation between samples), providing a high-quality feature foundation for subsequent biological sequence analysis tasks.

[0114] Optionally, in some embodiments of this application, the high-dimensional embedding feature mapping includes a first high-dimensional embedding feature and a second high-dimensional embedding feature, and the step "fusing high-dimensional embedding features based on cross-modal consistency loss function to obtain multimodal coding representation" may specifically include:

[0115] Map the first high-dimensional embedding feature and the second high-dimensional embedding feature to the same embedding space to obtain the first projection vector corresponding to the first high-dimensional embedding feature and the second projection vector corresponding to the second high-dimensional embedding feature.

[0116] A multimodal coded representation is generated based on the cross-modal consistency loss function, the first projection vector, and the second projection vector.

[0117] In this context, a common embedding space is used to uniformly represent low-dimensional spaces (e.g., 256-dimensional) of features from different modalities. High-dimensional features are mapped to this space via a shared projection head (Multilayer Perceptron, MLP), eliminating semantic gaps between modalities and enabling geometric comparability of features. The first projection vector (zDNA) is the vector obtained by mapping the first high-dimensional embedding feature to the common embedding space through the shared projection head, and its formula is: z... DNA =MLP(h DNA ), representing the semantic encoding of DNA sequences in the joint space.

[0118] Second projection vector (z) Ribo ) is the vector formed by mapping the second high-dimensional embedding feature to the same embedding space through a shared projection head, and the formula is: z Ribo =MLP(h Ribo ), representing the dynamic coding of Ribo-seq signals in the joint space.

[0119] The cross-modal consistency loss function (LC) measures the semantic alignment of cross-modal projection vectors of the same biological sample. By minimizing this loss, the geometric distance between the projection vectors of the DNA sequence and the corresponding Ribo-seq signal in the joint space is minimized, ensuring semantic consistency. The multimodal coded representation (Z) is a joint feature vector that fuses cross-modal information, obtained by concatenating or averaging the optimized first and second projection vectors.

[0120] Specifically, high-dimensional features from different modalities are compressed into the same space to facilitate the calculation of cross-modal similarity. For example, promoter motif features in DNA and translation initiation signal features in Ribo-seq should be adjacent in the joint space. Then, z is calculated for each sample. DNA With z Ribo The squared distances are summed and averaged to obtain the LC (Lower Modal Consistency). Backpropagation is used to update the parameters of the projector head and encoder, ensuring that cross-modal projection vectors of the same biological sample are as close as possible, while projection vectors of different samples are as far apart as possible. Heterogeneous modal features are transformed into comparable vector representations through shared projector head mapping. The cross-modal consistency loss function is used to explicitly align cross-modal information of the same biological sample. The resulting multimodal encoded representation retains the unique semantics of each modality (such as the sequence syntax of DNA and the translation dynamics of Ribo-seq) while capturing the synergistic relationship between the two (such as the correspondence between ORF sequences and translational activity). This process addresses the semantic gap problem in multimodal data, providing a high-quality feature foundation for subsequent sample generation in conditional diffusion models and accurate prediction in classification models.

[0121] Optionally, in some embodiments of this application, the step of "generating a multimodal coded representation based on the cross-modal consistency loss function, the first projection vector, and the second projection vector" may specifically include:

[0122] Obtain the contrastive loss function;

[0123] Based on the cross-modal consistency loss function and the contrastive loss function, the first projection vector and the second projection vector are optimized respectively;

[0124] By fusing the optimized first projection vector with the optimized second projection vector, a multimodal coded representation is obtained.

[0125] The contrastive loss function (NT-Xent loss), also known as the normalized temperature-scaled cross-entropy loss, is used to measure the discriminative power of different sample features within the same modality. Specifically, cosine similarity is calculated for each sample view pair, and the contrastive loss is obtained after normalization using the Softmax function. This forces view features of the same sample to be close in the embedding space, while features of different samples are kept far apart. The parameters of the independent encoder (CNN and GRU) and the shared projector head (MLP) are updated simultaneously using stochastic gradient descent (SGD) or the Adam optimizer.

[0126] NT-Xent loss driver: Clustering of different view features within the same modality (e.g., similar features of different mutated versions of a DNA sequence);

[0127] LC loss-driven: Alignment of features of the same biological sample across modalities (e.g., z-axis alignment of DNA sequences) DNA ) and the corresponding Ribo-seq signal (zRibo (Distance minimization).

[0128] Optimized feature state:

[0129] Intramodal: Different view features of the same sample form tight clusters in the embedding space (e.g., all DNA mutation views belonging to the same ORF and Ribo-seq noise views are clustered separately);

[0130] Cross-modal: z-species of the same biological sample DNA and z Ribo Located near the center of the same cluster, different samples maintain a large distance between clusters.

[0131] 107. Based on the multimodal encoding representation and the preset loss function, the classification model is trained to obtain the target classification model for predicting open reading boxes.

[0132] The pre-defined loss function refers to the imbalanced positive-unlabeled learning (PU) loss function, used to address the problem of extreme imbalance between positive and negative samples in classification tasks (such as the scarcity of positive open reading frames (ORFs)). The classification model is a prediction model built on multimodal encoded representations, used to determine whether a sample is an open reading frame. Specifically, a strategy of fixing the multimodal feature extractor and training a linear classification layer can be used to improve the classifier's semantic understanding of ORFs by utilizing pre-trained cross-modal feature representations (captured sequence-translation association information). At the same time, the PU loss function adaptively adjusts the sample weights to mitigate the prediction bias caused by data imbalance. This method does not require training a complex model from scratch, can efficiently utilize the complementary information of multimodal data, and ultimately generate an open reading frame prediction model with high biological interpretability.

[0133] Optionally, in some embodiments of this application, the step of "training the classification model based on the multimodal coding representation and a preset loss function to obtain a target classification model for predicting open reading boxes" may specifically include:

[0134] The multimodal encoding representation is input into the classification model to determine the predicted label of each training sample belonging to an open reading frame;

[0135] The classification model is trained based on the true and predicted labels of each training sample to obtain the target classification model for predicting open reading frames.

[0136] In this model, the prediction result of the classification model for the training samples, i.e., y', represents the probability that the sample belongs to the ORF. By setting a threshold (such as 0.5), the probability can be converted into a hard label (0 or 1), but the probability value is used directly in the loss function calculation to preserve uncertainty information.

[0137] Authentic Labels

[0138] The actual class label of the training sample, denoted as y.

[0139] Positive samples (ORF):

[0140] y = 1 indicates that the sample is a real open reading frame;

[0141] Unlabeled samples:

[0142] y = 0, containing potential ORF or non-ORF regions, is treated as a candidate positive example in unbalanced positive-unlabeled learning (PU learning).

[0143] Target classification model

[0144] The optimized classification model, after training, possesses the ability to predict open reading frames. Its core advantage lies in leveraging multimodal features to enhance the joint modeling ability of ORF sequence structure and translational activity, while mitigating the data imbalance problem through loss function design.

[0145] Specifically, each sample in the balanced training set corresponds to a multimodal encoded representation Z, which incorporates cross-modal information from DNA and Ribo-seq. Z is used to calculate the predicted label y' through a linear classification layer. Then, model optimization is performed based on the imbalanced positive-unlabeled learning (PU) loss function. By minimizing the imbalanced positive-unlabeled learning (PU) loss function, the model is forced to increase the prediction probability of positive samples (y=1) (y' approaches 1), while decreasing the prediction probability of negative samples in unlabeled samples (y' approaches 0), but retaining the sensitivity to potential positives. This is used to train the classification model to obtain the target classification model for predicting open reading frames.

[0146] To further understand the open reading frame prediction scheme of this application, please refer to [link / reference]. Figure 2 and Figure 3 This application provides a data augmentation module based on a multimodal conditional diffusion model and a classification prediction module based on contrastive learning and positive-unlabeled learning, as detailed below:

[0147] The conditional coding model is used to extract multimodal input information and convert it into a conditional representation of smORF prediction. Multimodal features are utilized through a multi-level attention fusion machine, a feature fusion method based on self-attention mechanism is adopted, and finally data augmentation is achieved through a diffusion model method.

[0148] When processing DNA sequence information, the sequence is converted into a fixed-length embedding representation, with each base mapped to a vector to capture semantic relationships. The custom encoder E_DNA consists of an 8-layer convolutional neural network with skip connections, which can capture local and global relationships, generate latent feature representations, and capture local sequence patterns.

[0149] Ribo-seq signal data processing employs a recurrent neural network encoder, mapping the ribosome coverage signal and location coordinates at each time step into a dense vector. This vector is then concatenated and fused before being input into the recurrent unit (RRU) to capture the patterns of coverage intensity fluctuations and the spatial correlation of translation sites. The final hidden state integrates global sequence information to generate a latent representation. The gated recurrent unit (GRU) captures long-term dependencies in the sequence data, reducing gradient vanishing and aiding in the learning of Ribo-seq signal data.

[0150] In terms of multimodal feature fusion, a bidirectional interaction mechanism is constructed based on the cross-attention mechanism to achieve cross-modal feature alignment and complementary fusion. Query, key, and value vectors are generated through GRU linear projection, and a multi-head cross-attention mechanism promotes interactive learning between the two modalities. DNA features serve as the query vector, while Ribo-seq signal features assume the roles of key and value, generating an attention weight matrix to achieve DNA feature-guided cross-modal enhancement sequences. Conversely, Ribo-seq signals serve as the query vector, and DNA features serve as key-value pairs, generating DNA feature optimization sequences reflecting the influence of translation regulatory signals. The bidirectional cross-attention architecture establishes a deep association between gene sequence features and translation dynamic features, maintaining feature space independence. A Transformer encoder is used to encode the input sequence, with a self-attention mechanism at its core to capture long-distance dependencies. The multi-head self-attention mechanism computes features from different subspaces in parallel, fusing multi-dimensional information. The fused features are stabilized during the training process through residual connections and layer normalization.

[0151] The conditional diffusion model, based on multimodal fusion features, is designed for smORF generation tasks. By using multimodal features as generation conditions, fine-grained conditional injection is implemented at each time step of the diffusion process. Gaussian noise is progressively injected during the forward diffusion process to construct a forward Markov chain. The reverse process recovers the smORF sample data from the Gaussian noise, defined as a Markov chain from xT to x0. The formula is as follows:

[0152]

[0153] The conditional diffusion model learns a backward diffusion process, training the model to predict noise and transforming the noise into the true data distribution during inference. The backward diffusion process is parameterized, aiming to eliminate the Gaussian noise added during the diffusion process. At diffusion step t, the noisy smORF sample is represented by a specific formula. During training, the network parameters are optimized by minimizing the error between predicted and true noise, using mean squared error (MSE) loss to measure the Euclidean distance between the true and predicted noise. The backward diffusion process is as follows:

[0154]

[0155] To address the significant false negative problem inherent in currently used supervised data, this application proposes a detection scheme that combines biological priors, cross-modal self-supervised learning, and semi-supervised classification. The core architecture consists of, for example, […]. Figure 3 The three main modules shown work together to form

[0156] Based on the ribosomal translation process in prokaryotes, which follows a 3-base codon translation unit movement pattern, a local window Fourier transform is performed on the Ribo-seq signal to extract periodic modulation signals related to translation activity. Since the translation process involves movement in 3-nucleotide cycles, the spatial domain signal based on Ribo-seq data exhibits periodic signal intensity variations. However, traditional periodic signal frequency band extraction algorithms struggle to capture these low-frequency periodic features, especially for shorter smORF sequences in the genome. Therefore, this application aims to provide a signal enhancement method based on frequency domain analysis. A symmetrical window extending to the left and right of the target location is used to capture the spatial distribution of signal intensity within the corresponding window. Discrete Fourier transform is then used to convert the signal intensity to the frequency domain, and the amplitude values ​​corresponding to the frequency components of the 3-base cycle are calculated. This approach improves the model's sensitivity to translationally active regions. Furthermore, the full-length count and unit count from the Ribo-seq experiment are used as inputs as two different data streams to more fully characterize the distribution features of ribosome-occupied sites.

[0157] To further uncover the potential semantic connections between DNA sequences and Ribo-seq signals, this application designs two independent encoders to perform biologically relevant enhancement operations (random mutation, local sequence shuffling) on ​​the DNA sequence data, generating semantically irrelevant perturbation samples. Simultaneously, time-window shifting and signal intensity normalization are applied to the Ribo-seq signal to generate semantically irrelevant technical noise. Based on these two different perturbation data, two independent encoders are used to characterize their features, resulting in their respective high-dimensional embedding features. By optimizing the normalized temperature-scaled cross-entropy loss function (NT-Xent), the model is constrained to tightly cluster different enhanced views of the same sample in the embedding space, while the representations of different samples are significantly separated. This application innovatively introduces a cross-modal consistency constraint: by jointly optimizing the projected representations of the sequence encoder and the signal encoder, the semantic features of the DNA sequence and the translation dynamics of the corresponding Ribo-seq signal are forced to maintain geometric consistency in the joint embedding space.

[0158] This application addresses the issue of unlabeled negative samples in practical smORF recognition problems by proposing a training optimization method for a PU-learning classification model. Based on the pre-trained multimodal encoding representation, the classification model updates its parameters by optimizing a modified imbalanced PU loss function. This loss function utilizes the prior probability distribution of positive instances obtained during pre-training to dynamically change the loss weights between positive and unlabeled instances, correcting training bias caused by class imbalance. Simultaneously, a sigmoid function is used to transform the classification boundary into a non-linear function mapping, distinguishing samples with insufficient model confidence.

[0159] Trinucleotide periodicity is a periodicity specific to prokaryotic ribosomal translation, determined by the triplet codons of the genetic code. Due to the periodicity of translation, each ribosome movement spans three bases, resulting in periodic changes in translational activity along the nucleic acid sequence. Ribo-seq signals exhibit a 3-nt periodic modulation phenomenon, which is an important biological characteristic of functional smORFs. However, conventional sequence analysis and traditional signal detection methods struggle to directly capture this specific periodic pattern, posing a particular challenge for detecting shorter ORFs in prokaryotic genomes.

[0160] Based on this, this application adopts a prokaryotic three-period extraction scheme based on Fourier transform, and combines it with multi-channel Ribo-seq signal input to enhance the predictive ability of the model. In the specific implementation, this application collects data from Ribo-seq experiments and designs full-length counting and unit counting respectively to form two input signal channels: (1) full-length counting; (2) unit counting. In this way, this application obtains dynamic information of ribosome translation, providing the model with multi-dimensional input features.

[0161] To further obtain the three-periodity information of the Ribo-seq signal itself, this application utilizes a Fourier transform method based on a local window. The Fourier transform is a mathematical technique that converts a signal in time (or space) to frequency space, separating its different frequency components. By using the Fourier transform technique to transform the sequence signal from time space to the frequency domain, the intensity of the component signals is obtained, thereby improving the model's sensitivity to the three-period perception of translation. This application uses the calculated three-period signal intensity as one of the input features of the model.

[0162] Contrastive learning pre-training is a contrastive learning strategy based on unsupervised learning. Its key techniques mainly include four modules: data augmentation methods to generate diverse sample views, a contrastive loss function to optimize the embedding space distribution, a strategy for constructing positive and negative sample pairs, and encoder architecture design. Data augmentation achieves semantic invariance through appropriate transformations, and the loss function is designed to maximize the similarity of positive samples and minimize the metric function of negative samples, thereby improving the model's discriminative ability.

[0163] For DNA sequence data, perturbation samples are generated using biologically reasonable enhancement methods such as random mutation and local sequence substitution. The loss function is calculated as follows:

[0164]

[0165] For Ribo-Seq signals, perturbation samples are generated using enhancement methods such as signal scaling and time-domain shifting. To maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs, NT-Xent (normalized temperature-scaled cross-entropy loss) is used as the distance function. The loss function is calculated as follows:

[0166]

[0167] The overall pre-training loss consists of three parts: sequence contrast loss to enhance sequence perturbation invariance, signal contrast loss to improve signal perturbation robustness, and consistency loss to maintain cross-domain consistency. As shown in the following equation:

[0168]

[0169] Positive-unlabeled learning (PU learning) is a semi-supervised learning technique that addresses the problem of incompletely labeled samples. In classic supervised learning models, the dataset typically contains explicitly labeled positive and negative samples. However, in many practical problems, only positive samples may be labeled, while the labels for negative samples are either missing or nonexistent. PU learning provides a method to predict potential positive samples within the negative class by training only with positive and unlabeled samples.

[0170] Early literature on Purpose-Based Learning (PU) explored methods for solving the PU learning problem, including 2-stage methods and methods utilizing class prior information. Accurate estimation of the positive class prior is crucial for training PU learning-based models. Some works have provided unbiased estimates of the positive class prior and optimized the training risk based on the assumption of known positive class priors. In deep learning scenarios, further estimation techniques can achieve better overfit protection. Meanwhile, with increasing attention to imbalanced data and sample bias problems, PU learning methods have been continuously advanced and improved, enabling them to handle more complex scenarios.

[0171] The PU learning classifier proposed in this chapter solves the classification problem of positive and unlabeled samples by simultaneously optimizing the feature extractor and the classifier. It optimizes the imbalanced nPU learning problem by minimizing the imbalanced nPU loss. To calculate the loss function, the common sigmoid loss function is used to measure the classification error of each sample.

[0172] During training, the feature extractor undergoes a pre-training phase, optimizing representation learning by minimizing the biased contrastive loss. The training phase consists of two steps: first, a biased feature representation is pre-trained, and then a linear classifier is optimized on this feature representation as the final classification task. The relationship between classifier risk and the balance loss is also explored.

[0173] The open reading frame (OPG) prediction method of this application, after obtaining the first modality features corresponding to the deoxyribonucleic acid (DNA) sequence and the second sample features corresponding to the ribosome sequencing signal, constructs multimodal features based on a bidirectional cross-attention mechanism, the first modality features, and the second modality features. Then, based on the multimodal features and a preset conditional diffusion model, OPG samples are generated. Next, the OPG samples are added to the original sample set to obtain a balanced training set. Subsequently, the balanced training set is perturbed, and high-dimensional embedding features corresponding to each training sample in the processed balanced training set are extracted using an independent encoder. Finally, the high-dimensional embedding features are fused based on a cross-modal consistency loss function to obtain a multimodal encoded representation. Based on the multimodal encoded representation and a preset target loss function, a classification model is trained to obtain a target classification model for predicting OPGs. In the OPG prediction scheme provided in this application, the generation of augmented samples that conform to the true distribution is guided by multimodal features, alleviating the data imbalance problem. This method generates supplementary positive samples through data generation, avoiding learning bias caused by unreasonable weight settings when existing prediction models use weighted loss, thus improving prediction accuracy. In the feature learning stage, sample differentiation is achieved through contrastive learning pre-training, avoiding the false negative problem caused by misclassifying unknown samples as negative samples, and finally predicting high-quality, experimentally feasible open reading frames, thereby improving the prediction effect.

[0174] To facilitate better implementation of the open reading frame prediction method of this application, this application also provides an open reading frame prediction device. The meanings of the terms used are the same as in the open reading frame prediction system described above, and specific implementation details can be found in the description of the system embodiments.

[0175] Please see Figure 4 , Figure 4This is a schematic diagram of the structure of the open reading frame prediction device provided in an embodiment of this application. The open reading frame prediction device may specifically include an acquisition module 201, a construction module 202, a generation module 203, an addition module 204, an extraction module 205, a fusion module 206, and a prediction module 207, as follows:

[0176] The acquisition module 201 is used to acquire the first modality features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal;

[0177] Module 202 is used to construct multimodal features based on a bidirectional cross-attention mechanism, first modality features, and second modality features;

[0178] The generation module 203 is used to generate open reading box samples based on multimodal features and a preset conditional diffusion model;

[0179] Add module 204 to add open reading frame samples to the original sample set to obtain a balanced training set;

[0180] The extraction module 205 is used to perturb the balanced training set and extract the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder.

[0181] Fusion module 206 is used to fuse high-dimensional embedding features based on cross-modal consistency loss function to obtain multimodal coded representation;

[0182] The prediction module 207 is used to train the classification model based on the multimodal encoding representation and the preset target loss function to obtain the target classification model for predicting open reading boxes.

[0183] This application provides an open reading frame (OPG) prediction device. After the acquisition module 201 acquires the first modal features corresponding to the deoxyribonucleic acid (DNA) sequence and the second sample features corresponding to the ribosome sequencing signal, the construction module 202 constructs multimodal features based on a bidirectional cross-attention mechanism, the first modal features, and the second modal features. Next, the generation module 203 generates OPG samples based on the multimodal features and a preset conditional diffusion model. Then, the addition module 204 adds the OPG samples to the original sample set to obtain a balanced training set. Following this, the extraction module 205 perturbs the balanced training set and extracts the high-dimensional embedding features corresponding to each training sample in the balanced training set using an independent encoder. Finally, the fusion module 206 fuses the high-dimensional embedding features based on a cross-modal consistency loss function to obtain a multimodal encoding representation. The prediction module 207 trains a classification model based on the multimodal encoding representation and a preset target loss function to obtain a target classification model for predicting OPGs. In the open reading frame prediction scheme provided in this application, enhanced samples that conform to the true distribution are generated based on multimodal features to alleviate the data imbalance problem. This method generates supplementary positive samples through data generation, avoiding learning bias caused by unreasonable weight settings when using weighted loss in existing prediction models, thus improving prediction accuracy. In the feature learning stage, sample differentiation is achieved through contrastive learning pre-training, avoiding false negatives caused by misclassifying unknown samples as negative samples. Ultimately, this results in the prediction of high-quality, experimentally feasible open reading frames, improving prediction performance.

[0184] Furthermore, embodiments of this application also provide an electronic device, such as... Figure 5 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0185] The electronic device may include components such as a processor 301 with one or more processing cores, a memory 302 with one or more processor-readable storage media, a power supply 303, and an input unit 304. Those skilled in the art will understand that... Figure 5 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0186] The processor 301 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 302, and by calling data stored in the memory 302, it performs various functions and processes data, thereby providing overall monitoring of the electronic device. Optionally, the processor 301 may include one or more processing cores; preferably, the processor 301 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 301.

[0187] The memory 302 can be used to store software programs and modules. The processor 301 executes various functional applications and open reading frame prediction methods by running the software programs and modules stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 302 may also include a memory controller to provide the processor 301 with access to the memory 302.

[0188] The electronic device also includes a power supply 303 that supplies power to various components. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 303 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0189] The electronic device may also include an input unit 304, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0190] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 301 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 302 according to the following instructions, and the processor 301 runs the applications stored in the memory 302 to realize various functions, as follows:

[0191] The process involves obtaining the first modality features corresponding to the deoxyribonucleic acid (DNA) sequence and the second sample features corresponding to the ribosome sequencing signal. Multimodal features are constructed based on a bidirectional cross-attention mechanism, the first modality features, and the second modality features. Open reading frame (OPG) samples are generated based on the MPG features and a pre-defined conditional diffusion model. These OPG samples are added to the original sample set to obtain a balanced training set. The balanced training set is perturbed, and high-dimensional embedding features corresponding to each training sample in the processed balanced training set are extracted using an independent encoder. The high-dimensional embedding features are fused based on a cross-modal consistency loss function to obtain a multimodal encoding representation. A classification model is trained based on the multimodal encoding representation and a pre-defined target loss function to obtain a target classification model for predicting OPGs.

[0192] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0193] In this embodiment, after obtaining the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal, a multimodal feature is constructed based on a bidirectional cross-attention mechanism, the first modal feature, and the second modal feature. Then, based on the multimodal feature and a preset conditional diffusion model, open reading box samples are generated. Next, the open reading box samples are added to the original sample set to obtain a balanced training set. Subsequently, the balanced training set is perturbed, and high-dimensional embedding features corresponding to each training sample in the processed balanced training set are extracted using an independent encoder. Finally, the high-dimensional embedding features are fused based on a cross-modal consistency loss function to obtain a multimodal encoded representation. Based on the multimodal encoded representation and a preset target loss function, a classification model is trained to obtain a target classification model for predicting open reading boxes. In the open reading box prediction scheme provided in this application, the generation of augmented samples that conform to the true distribution is guided by multimodal features, alleviating the data imbalance problem. This method generates supplementary positive samples through data generation, avoiding learning bias caused by unreasonable weight settings when existing prediction models use weighted loss, thus improving prediction accuracy. In the feature learning stage, sample differentiation is achieved through contrastive learning pre-training, avoiding the false negative problem caused by misclassifying unknown samples as negative samples, and finally predicting high-quality, experimentally feasible open reading frames, thereby improving the prediction effect.

[0194] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a processor-readable storage medium and loaded and executed by a processor.

[0195] To this end, embodiments of this application provide a storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the open reading frame prediction methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0196] The process involves obtaining the first modality features corresponding to the deoxyribonucleic acid (DNA) sequence and the second sample features corresponding to the ribosome sequencing signal. Multimodal features are constructed based on a bidirectional cross-attention mechanism, the first modality features, and the second modality features. Open reading frame (OPG) samples are generated based on the MPG features and a pre-defined conditional diffusion model. These OPG samples are added to the original sample set to obtain a balanced training set. The balanced training set is perturbed, and high-dimensional embedding features corresponding to each training sample in the processed balanced training set are extracted using an independent encoder. The high-dimensional embedding features are fused based on a cross-modal consistency loss function to obtain a multimodal encoding representation. A classification model is trained based on the multimodal encoding representation and a pre-defined target loss function to obtain a target classification model for predicting OPGs.

[0197] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0198] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0199] Since the instructions stored in the storage medium can execute the steps of any of the open reading frame prediction methods provided in the embodiments of this application, the beneficial effects that any of the open reading frame prediction methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0200] The above provides a detailed description of an open reading frame prediction method, apparatus, electronic device, and storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An open reading frame prediction method, characterized in that, include: Obtain the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal; Multimodal features are constructed based on a bidirectional cross-attention mechanism, the first modal feature, and the second modal feature; Based on the multimodal features and the preset conditional diffusion model, open reading box samples are generated; The open reading frame samples are added to the original sample set to obtain a balanced training set; The balanced training set is perturbed, and the high-dimensional embedding features corresponding to each training sample in the balanced training set are extracted by an independent encoder. The high-dimensional embedding features are fused based on the cross-modal consistency loss function to obtain a multimodal coded representation; The classification model is trained based on the multimodal encoding representation and the preset target loss function to obtain a target classification model for predicting open reading frames.

2. The open reading frame prediction method according to claim 1, characterized in that, The construction of multimodal features based on the bidirectional cross-attention mechanism, the first modal feature, and the second modal feature includes: Using the first modal feature as the query, and based on the bidirectional cross-attention mechanism and the second modal feature, a cross-modal feature in the first direction is constructed; Using the second modal feature as the query, and based on the bidirectional cross-attention mechanism and the first modal feature, a cross-modal feature in the second direction is constructed; Multimodal features are obtained by fusing the cross-modal features in the first direction and the cross-modal features in the second direction.

3. The open reading frame prediction method according to claim 1, characterized in that, The step of generating open reading box samples based on the multimodal features and a preset conditional diffusion model includes: The multimodal features are used as conditions to input into a preset conditional diffusion model. Through the forward diffusion process of the conditional diffusion model, Gaussian noise is gradually injected into the original samples to construct a noise sample sequence. Based on the noise sample sequence, through the reverse denoising process of the conditional diffusion model, open reading box samples consistent with the distribution of real open reading box data are generated under the conditional constraints of multimodal features.

4. The open reading frame prediction method according to claim 1, characterized in that, The process of perturbing the balanced training set and extracting the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder includes: Biological perturbation of the deoxyribonucleic acid sequence samples in the balanced training set, and; Noise processing is performed on the ribosome sequencing signal samples in the balanced training set to generate diverse views; Extract the first high-dimensional embedding feature corresponding to the biologically perturbed deoxyribonucleic acid sequence sample, and; Extract the second high-dimensional embedding features corresponding to the diverse views.

5. The open reading frame prediction method according to claim 1, characterized in that, The high-dimensional embedding feature mapping includes a first high-dimensional embedding feature and a second high-dimensional embedding feature. The process of fusing the high-dimensional embedding features based on a cross-modal consistency loss function to obtain a multimodal coding representation includes: The first high-dimensional embedding feature and the second high-dimensional embedding feature are mapped to the same embedding space to obtain the first projection vector corresponding to the first high-dimensional embedding feature and the second projection vector corresponding to the second high-dimensional embedding feature. A multimodal coded representation is generated based on the cross-modal consistency loss function, the first projection vector, and the second projection vector.

6. The open reading frame prediction method according to claim 5, characterized in that, The step of generating a multimodal coded representation based on the cross-modal consistency loss function, the first projection vector, and the second projection vector includes: Obtain the contrastive loss function; Based on the cross-modal consistency loss function and the contrast loss function, the first projection vector and the second projection vector are optimized respectively; By fusing the optimized first projection vector with the optimized second projection vector, a multimodal coded representation is obtained.

7. The open reading frame prediction method according to claim 1, characterized in that, The step of training the classification model based on the multimodal encoding representation and a preset loss function to obtain a target classification model for predicting open reading boxes includes: The multimodal encoding representation is input into the classification model to determine the predicted label of each training sample belonging to the open reading frame; Based on the true and predicted labels of each training sample, the classification model is trained to obtain the target classification model for predicting open reading frames.

8. An open reading frame prediction device, characterized in that, include: The acquisition module is used to acquire the first modal features corresponding to the deoxyribonucleic acid sequence and the second sample features corresponding to the ribosome sequencing signal; A construction module is used to construct multimodal features based on a bidirectional cross-attention mechanism, the first modal feature, and the second modal feature; The generation module is used to generate open reading box samples based on the multimodal features and a preset conditional diffusion model; An addition module is used to add the open reading frame samples to the original sample set to obtain a balanced training set; The extraction module is used to perturb the balanced training set and extract the high-dimensional embedding features corresponding to each training sample in the balanced training set through an independent encoder. The fusion module is used to fuse the high-dimensional embedding features based on the cross-modal consistency loss function to obtain a multimodal coded representation; The prediction module is used to train the classification model based on the multimodal encoding representation and the preset target loss function to obtain a target classification model for predicting open reading boxes.

9. An electronic device, characterized in that, include: A memory, a processor, and a processor program stored in the memory and executable on the processor, wherein the processor executes the program as steps of the open reading frame prediction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The computer processing program is stored and can be loaded by a processor to execute the open reading frame prediction method as described in any one of claims 1 to 7.

Images