Essential gene prediction method based on DNA large model and time-frequency domain deep learning fusion

Abstract

The invention belongs to the technical field of essential gene prediction, and particularly relates to an essential gene prediction method based on DNA large model and time-frequency domain deep learning fusion, and the method comprises the steps: taking a domain DNA large model as a core representation layer, and obtaining special gene representation through cross-species corpus pre-training and task fine tuning; a T-Block and F-Block dual-channel time-frequency fusion structure is adopted, and the local dependence and long-range regulation relation of a gene sequence is synchronously captured by expanding DFT (Discrete Fourier Transform), complex value attention and iDFT (Initial Discrete Fourier Transform) conversion; designing an efficient modeling reasoning scheme of sliding window slices and gene-level aggregation aiming at an ultra-long sequence; in combination with class imbalance and a noise robust training strategy, cross-cell line / cross-platform transferable threshold output is realized through temperature scaling calibration, an uncertainty quantization and structured interface is matched, and drug target screening and experimental design decision are supported. The system supports the realization of multiple programming languages, and can complete low-delay end-to-end reasoning in a conventional hardware environment.

Description

Technical Field

[0001] This invention belongs to the field of essential gene prediction technology, and in particular relates to an essential gene prediction method based on the fusion of DNA large model and time-frequency domain deep learning. Background Technology

[0002] Essential genes are generally defined as genes that are indispensable for the survival or reproduction of an organism. The absence of their function will lead to cell death or the organism's inability to complete the normal life cycle. In basic research and biomedical applications, accurately identifying these genes helps to elucidate the origin and evolution of life, define the minimum set of genes that maintain cell survival, and provide important clues for the discovery of pathogen drug targets and the mining of tumor vulnerability. Therefore, they have significant value at both the methodological and engineering implementation levels.

[0003] In existing technologies, the main approaches to identifying essential genes are divided into experimental and computational methods. Experimental approaches include RNA interference, CRISPR-based genome editing, systematic gene knockout, and transposon mutagenesis. These methods can provide relatively reliable results under given conditions, but they are often costly, time-consuming, and technically challenging, especially in complex organisms. This has objectively driven the development of computational prediction methods, represented by machine learning and deep learning. Regarding computational methods, early research often combined high-throughput omics features with traditional machine learning for essential gene prediction in model organisms or mammals. Subsequently, research expanded to cross-species migration and large-scale gene function annotation. With the advent of deep learning, models integrating multimodal features such as sequence and protein-protein interaction networks (PPIs) have shown superior prediction accuracy compared to traditional methods. For example, DeepHE achieved significant improvements in human essential gene tasks by integrating sequence data with PPI topological information. Furthermore, cell-line specific frameworks (such as DeepCellEss, which combines convolution and self-attention) have begun to emphasize interpretability and adaptability in different cellular contexts. Meanwhile, pre-trained models for genome sequences (DNA "big language models") such as DNABERT and its subsequent versions have been used as general sequence encoders and have shown good characterization capabilities in many types of genome tasks, but there is still a lack of specialized adaptation and threshold / uncertainty calibration mechanisms for the "essential gene prediction" task.

[0004] The existing implementations most similar to this invention can be summarized into two technical routes. One is a route primarily based on sequence statistical features and traditional machine learning: In this route, researchers first collect "essential / non-essential" tags at the human or model organism level, then extract statistical features such as nucleotide composition, k-mer frequency, GC content, and upstream / downstream associations from gene or neighboring sequences. These features are then trained and cross-validated using supervised learners such as logistic regression, support vector machines, or random forests, and classification inference is completed by setting thresholds based on the validation set. This route is simple to implement and has good interpretability, but it has limited characterization of long-range dependencies and cross-domain distribution shifts, and its expressive and transfer capabilities are insufficient when faced with multi-source heterogeneous information.

[0005] The second approach involves integrating multimodal features and employing deep learning. A typical workflow includes integrating essential tags with transcriptome expression matrices, protein-protein interaction networks, and other data; performing one-hot or k-mer embeddings on sequences; normalizing and batch correcting expression data; and extracting node degree, betweenness, or learning graph embeddings from the network side. The model structure utilizes a convolutional / multilayer perceptron-based framework with the potential to incorporate attention mechanisms, using weighted loss or resampling to mitigate class imbalance, and improving generalization through early stopping and regularization. The inference stage outputs gene essentiality scores and binary classification results. Works like DeepHE outperform traditional models in accuracy, while cell-line specific deep frameworks (such as DeepCellEss) further enhance adaptability to different cell backgrounds. However, there is still room for improvement in unified modeling of long sequence dependencies, cross-platform threshold transferability, and engineering inference efficiency.

[0006] Therefore, this invention aims to provide a necessary gene prediction method based on the fusion of DNA large model and time-frequency domain deep learning. It takes pre-trained DNA representation as the core, combines multimodal (sequence, physicochemical features, and large language model information) feature fusion and long-range dependency modeling, and adopts noise robustness and class balance training strategies to output a necessary score and transferable threshold with high accuracy and strong generalization. It also provides low-latency inference and standardized interface to support early optimization and accurate decision-making in drug target screening and experimental design. Summary of the Invention

[0007] The purpose of this invention is to provide an essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning, in order to solve several core technical problems existing in the current field of essential gene prediction: First, insufficient generalization and transferability across species, cell lines, and experimental platforms, with the same model exhibiting large performance fluctuations in different CRISPR screening data and tissue backgrounds; Second, insufficient characterization of long-range regulation and contextual dependence, making it difficult to simultaneously model promoter / enhancer synergy, 3D genome structure, and chromatin state, resulting in unstable identification of boundary cases and conditionally essential genes; Third, insufficient utilization of multi-source heterogeneous information: DNA sequence information, protein sequence information, physicochemical characteristics, etc., are difficult to model uniformly in spatiotemporal context, resulting in incomplete feature coverage and insufficient discrimination of conditionally essential genes.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The essential gene prediction method based on the fusion of large DNA models and time-frequency domain deep learning includes the following steps: Step 1: Construct datasets S1-S3 covering the annotation of essential human genes and essential gene sequences across species for CRISPR screening in multiple cell lines. Standardize all sequences, remove non-standard characters, and divide them into training / validation sets according to a preset ratio. Step 2: Extract multiple types of sequence physicochemical features, including K-mer, PseKNC, cumulativeskew, Z-Curve, GC content, AT / GC ratio, and λ-Z-curve, and use two-level recursive feature elimination to obtain the feature set with the optimal dimensions; Step 3: Construct a time-frequency network with a dual-channel structure including T-Block and F-Block. The T-Block generates the time domain output; the F-Block maps the sequence to the frequency domain through extended discrete Fourier transform. After processing by the complex-valued Transformer encoder, it is converted back to the time domain through inverse discrete Fourier transform to generate the frequency domain output. The time-frequency fusion feature is obtained by weighted summation of the two channel outputs. Step 4: Construct a large DNA model, use 6-mer segmentation to process DNA sequences, implement a sliding window slicing strategy for ultra-long sequences, and obtain unified gene characterization through gene-level aggregation; Step 5: Integrate time-frequency fusion features and unified gene representation as model input to train the model; Step 6: Based on the trained DNA large model, inference is performed on the test set sequences. First, the prediction scores are calibrated by temperature scaling. Then, the transferable decision threshold is determined based on the optimal point of Youden's J or F1 on the validation set. The prediction results of the test sequences are aggregated at the gene level, and finally, the necessity score, binary classification conclusion and uncertainty quantification results are output.

[0009] Preferably, in step 1: Dataset S1 covers 11 human cancer cell lines: A375, DLD1, GBM, HAP1, HCT116, HeLa, Jiyoye, KBM7, K562, Raji, and RPMI-8226. A majority voting strategy was used to define core essential genes: the same gene was marked as positive when it was determined to be essential in more than 6 cell lines. Genes not included in the essential set were considered non-essential after HGNC annotation, forming a clear negative sample set. Dataset S2 contains 1640 gene sequences from 323; Dataset S3 was used to train a large DNA model for essential genes, bringing together essential gene sequences from multiple sources across species, including DEG, CEG, ePath, and OGEE3. All sequences were converted to uppercase before entering pre-training, and only the four standard nucleotide characters A / G / C / T were retained, while spaces, newlines, and other non-standard characters were removed. The training and validation sets were then split at a ratio of 0.95:0.05 to perform the MLM task. S3 contains 478,431 DNA sequences.

[0010] Preferably, the specific process of step 2 is as follows: Extraction of multiple types of sequence physicochemical features: Based on the processed standardized DNA sequence, seven types of sequence physicochemical features were systematically extracted, and the parameters and dimensions of each type of feature were clarified; Perform two-level recursive feature elimination screening: Level 1 RFE screening: Grouping: The 85,995 λ-Z-curve features with the highest dimension are divided into 9 groups of equal size, with each group containing 9,555 features.

[0011] Feature concatenation: Each set of λ-Z-curve features is concatenated with the other 6 types of features, including K-mer and PseKNC, to form 9 sets of sub-features; Independent screening: The RFE algorithm was executed on each of the 9 sub-feature sets, and the top 1000 candidate features with the highest contribution were selected from each set; Merging and deduplication: Collect 9 groups of candidate features, merge them and remove duplicates to obtain an 8263-dimensional intermediate feature set; Second-level RFE screening: Optimal size search: Using an 8263-dimensional intermediate feature set as input, the RFE algorithm is executed again to iterate through the model performance corresponding to different feature dimensions; Determine the optimal dimension: Establish 4800 dimensions as the final input feature dimension.

[0012] Reuse of screening results: This optimal 4800-dimensional feature set is reused uniformly in the subsequent model training of all datasets in S1-S2.

[0013] Preferably, the specific process of constructing a time-frequency network with a dual-channel structure including T-Block and F-Block in step 3, and generating time-domain output through T-Block, is as follows: Preprocessing of the input sequence: Taking the DNA sequence corresponding to the 4800-dimensional optimal feature set obtained through screening as input, one-hot encoding is performed on the sequence to convert the nucleotide characters into a numerical time series S∈ R L , where L is the length of the encoded sequence; Perform T-Block time-domain processing to generate time-domain output. S o (t) ∈ R T .

[0014] Preferably, T-Block time-domain processing is performed to generate time-domain output. S o (t) ∈ R T The specific process is as follows: Sequence block partitioning: dividing the one-hot encoded time series S∈ R L Divided into multiple small blocks of fixed length X p ∈ R P×N Where P is the length of each small block, N is the number of small blocks, and N = [ L / P If the sequence length is not divisible by P, the last block is zero-padded to length P. Block feature embedding: Each small block is embedded through a trainable embedding layer. X p The dimensions are mapped from P to the model's preset feature dimensions to ensure that the block information is adapted to the input requirements of the subsequent Transformer encoder.

[0015] Local dependency capture: All embedded block feature vectors are sequentially fed into the Transformer encoder. The multi-head attention mechanism is used to mine the nucleotide position correlations within each block. The LayerNorm layer and residual connections are combined to alleviate gradient vanishing and enhance local feature preservation. Temporal output generation: The block feature vectors processed by the Transformer encoder are integrated and mapped to a unified dimension through a linear projection layer to generate the T-Block temporal output. S o (t) ∈ RT T is the preset output length, expressed by the formula: .

[0016] Preferably, in step 3, F-Block maps the sequence to the frequency domain through extended discrete Fourier transform, processes it through a complex-valued Transformer encoder, and then transforms it back to the time domain through inverse discrete Fourier transform to generate a frequency domain output. The specific process of obtaining the time-frequency fusion feature by weighted summation of the two channel outputs is as follows: Extended DFT frequency domain mapping: This involves using the extended discrete Fourier transform to map the time series... S∈R L Mapping to the frequency domain yields an extended spectrum of length L+T. The formula is: F=ExtendedDFT ( S ); ; Utilizing the conjugate symmetry of the spectrum of a real-valued sequence, only the first... Individual spectral values ​​reduce computational overhead; Complex-valued spectrum preprocessing: [This refers to the preprocessing of the spectrum.] Normalization is performed, followed by training-compatible linear projection. Embed it as a complex value vector ; Complex-valued spectral attention computation: An improved multi-head attention mechanism is used to process complex-valued vectors. F d : For each attention head h=1,2,...,H, through learnable parameters: ,Will F d Projection yields: Query Q h = F d T W h Q ; key K h = F d T W h K ; value V h = F d TW h V ; Perform complex-valued dot product attention calculation: ; Stitch together all attention head outputs and pass through linear projection The complex-valued spectral attention output is obtained: ; Encoder and Inverse Transform Processing: The complex-valued spectral attention output is fed into a Transformer encoder with LayerNorm, feedforward network, and complex domain residual connections, and the output is... F z Obtained through linear mapping ; By using the inverse discrete Fourier transform (iDFT) to F o Returning to the time domain, the last T time points are extracted as the frequency domain output of the F-Block. The formula is expressed as: ; Dual-channel adaptive weighted fusion generates time-frequency fusion features ; Adaptive weight calculation: Calculate the T-Block output weights based on the fundamental frequency sequence energy of the input sequence S. w t With F-Block output weights w f The formula is: w t , w f =Weighting(S); Time-frequency fusion feature generation: time domain output S o (t) With frequency domain output S o (f) ; .

[0017] Preferably, the specific process of step 4 is as follows: DNA large-scale model backbone initialization: The NucleotideTransformer-v2 model was pre-trained using the S3 unsupervised corpus as training data. The pre-training task was MLM. DNA sequence preprocessing and segmentation: Input sequence acquisition: DNA sequences from the standardized S1-S2 dataset were selected as input data for model fine-tuning and inference; 6-mer segmentation: The DNA sequence is split using a 6-mer segmentation strategy, with 6 consecutive nucleotides forming a segmentation unit. The segmentation unit is encoded using a vocabulary of approximately 4107 words and converted into a token sequence that the model can recognize. Ultra-long sequence sliding window slicing: Length judgment: Set the maximum input sequence length of the model to 2048 tokens. For ultra-long sequences whose length exceeds the upper limit after word segmentation, execute the sliding window slicing strategy.

[0018] Slicing parameters: Set the slice window size to 2048 tokens and the step size to 1024 tokens. Subsequent slices overlap with the previous slice by 1024 tokens until the entire long sequence is completely covered, resulting in multiple sequence segments that do not omit core information. Segment Vector Extraction: Sequence segments obtained from all slices are fed into the pre-trained NTv2 model for forward encoding to obtain the vector representation of each segment. Gene-level aggregation generates a unified gene representation: For all fragment vectors corresponding to the same DNA sequence, perform gene-level aggregation operation: integrate all fragment vectors to generate a unique unified representation vector for the gene.

[0019] Preferably, the specific process of step 5 is as follows: Feature dimension alignment: The time-frequency fusion features output by the time-frequency network and the unified gene representation vector generated by the DNA large model are mapped to the same dimension through linear projection; Feature fusion and concatenation: The two types of features after alignment are fused by concatenation to form a multimodal joint input feature vector; Training parameters and optimizer configuration: The AdamW optimizer is used to update model parameters, and a linear warmup → linear decay strategy is adopted. Training process execution: Training objectives: For the binary classification task of essential genes, train the model to output probability values ​​in the interval [0,1]; for the regression task of "degree of necessity", output a continuous necessity score; Iterative training: Input multimodal joint features and corresponding supervision labels in batches, calculate the loss value through backpropagation, update all trainable parameters of the model using the AdamW optimizer, and repeat the iteration until the early stopping condition is triggered or the preset maximum number of training rounds is reached.

[0020] The beneficial effects of this invention include: To address the challenges of essential gene modeling in eukaryotic genomes, such as those of humans, where sequences are significantly longer and more complex, existing methods struggle to balance fine-grained context with global dependencies, leading to extrapolation bias and misclassification of boundary samples. This invention constructs a long-sequence modeling framework for essential gene tasks, using specialized large-scale model representation as its core, coupled with efficient long-range dependency learning and inference strategies, ensuring stable discrimination without losing crucial context. It maintains stable recognition capabilities on large-scale and more heterogeneous datasets, improving the reliability of discrimination under extreme GC content, repetitive regions, and long-range regulatory backgrounds. Overall AUC / AUPR metrics are significantly improved and remain stable across multiple datasets.

[0021] While existing DNA pre-trained models are effective on general genomics tasks, they lack specialized design for "essential gene prediction," leading to insufficient task alignment, weak threshold and uncertainty calibration, and limited stability when transferred to different cell lines / platforms. This invention proposes a domain-specific large-scale model and fine-tuning workflow for essential genes. It combines task priors and high-quality annotations to construct dedicated representations and introduces calibration and domain adaptation mechanisms to mitigate the impact of distribution shifts. Under the same experimental settings and data partitioning, it achieves significant improvements compared to representative methods and maintains its advantage and robustness on more challenging datasets, demonstrating the value of domain-specific representations in improving discriminative power and stability.

[0022] Given that the sample size of essential genes is far smaller than that of non-essential genes, conventional stratified partitioning can introduce imbalanced distributions into the test set, weakening the true evaluation power for minority essential genes. Simultaneously, noise introduced by different experimental platforms / batch effects further exacerbates overfitting and evaluation bias. Therefore, a training and validation strategy robust to minority classes and noise is adopted to ensure that the evaluation criteria are more "sensitive" and fairer for essential genes. This significantly improves recall and overall performance metrics for minority classes and exhibits more stable generalization curves on multi-cell line / tissue data, validating the suppression effect on imbalance and noise.

[0023] Despite existing methods integrating sequence, expression, and network information, performance still declines in scenarios with larger scales and greater tissue diversity, indicating room for improvement in cross-cell line / tissue transferability and consistency. This invention uses a dedicated large-scale model representation as a common semantic space to uniformly align and fuse multimodal data such as sequence, expression, and network. Domain-robust training ensures that data from different sources / platforms are comparable and transferable within this space. Maintaining advantages and stability on datasets with broader tissue and cell line coverage demonstrates that unified representation and domain-robust training can effectively mitigate performance fluctuations caused by distribution shifts. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning, as described in this invention. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 The present invention will be further described in detail below: Example 1 See appendix Figure 1 As shown, the essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning is characterized by the following steps: Step 1: Data Collection and Processing: Construct datasets S1-S3 covering the annotation of essential human genes for CRISPR screening across multiple cell lines and essential gene sequences across species. S1-S2 are used for classification and regression prediction tasks, while S3 is used as unsupervised corpus for pre-training large DNA models. All sequences are standardized by converting them to uppercase and retaining only the four nucleotide characters A / G / C / T. After removing non-standard characters, the training and validation sets are divided according to a preset ratio.

[0026] Step 2: Feature Extraction and Two-Level Screening: Based on the standardized sequences obtained in Step 1, K-mer (k=1,2,3), PseKNC (k=3), cumulativeskew, Z-Curve, GC content, AT / GC ratio, and λ-Z-curve physicochemical features of multiple sequences are extracted to form an initial feature set of 86,151 dimensions. A two-level recursive feature elimination (RFE) process is used to screen the initial feature set. First, the λ-Z-curve features are divided into 9 equal groups. Each group is concatenated with other features and RFE is performed independently to select the top 1,000 candidates. After merging and deduplication, RFE is performed a second time to determine 4,800 dimensions as the final input features.

[0027] Step 3: Construct a time-frequency network, including a dual-channel structure of T-Block and F-Block. T-Block processes the one-hot encoded sequence corresponding to the feature set in the time domain to capture local dependencies and generate time domain output.

[0028] F-Block maps the sequence to the frequency domain using the extended Discrete Fourier Transform (DFT). After processing by the complex-valued Transformer encoder, it is transformed back to the time domain using the inverse Discrete Fourier Transform (iDFT) to generate the frequency domain output. Then, adaptive weights are determined based on the fundamental sequence energy, and the time-frequency fusion characteristics are obtained by weighted summation of the two channel outputs.

[0029] Step 4: Construct a large DNA model: Using NucleotideTransformer-v2 pre-trained on the unsupervised corpus of step S3 in step S1 as the backbone, the DNA sequences obtained in step S1 are processed by 6-mer word segmentation. For ultra-long sequences, a sliding window slicing strategy (window size 2048, step size 1024) is implemented. Fragment vectors are obtained through [CLS] representation or average pooling, and unified gene representation is obtained through gene-level aggregation.

[0030] Step 5: Joint Training Optimization: The time-frequency fusion features and the DNA large model representation are used as model inputs. The AdamW optimizer and linear warmup → linear decay learning rate schedule are used for training. The class imbalance problem is handled by weighted BCE / Focal loss and resampling strategy. Early stopping and regularization mechanisms are combined to improve the model's generalization ability.

[0031] Step 6: Prediction and Decision Output: Based on the trained DNA large model, inference is performed on the test set sequences. First, the prediction score is calibrated by temperature scaling. Then, the transferable decision threshold is determined based on the optimal point of Youden's J or F1 on the validation set. The prediction results of the test sequences are aggregated at the gene level. Finally, the necessity score, binary classification conclusion and uncertainty quantification results are output to support drug target screening and experimental design decisions.

[0032] In step 1: Dataset S1 is derived from the baseline set constructed by Guo et al. in 2017, which integrates the results of three independent studies and covers 11 human cancer cell lines, namely A375, DLD1, GBM, HAP1, HCT116, HeLa, Jiyoye, KBM7, K562, Raji, and RPMI-8226.

[0033] The core essential genes are defined using a "majority voting" strategy: the same gene is marked as positive when it is determined to be essential in more than 6 cell lines; genes not included in the essential set are considered non-essential after HGNC annotation, forming a clear negative sample set; the dataset provides both DNA and protein sequence information, which facilitates multimodal feature modeling.

[0034] Dataset S2 is derived from large dependency datasets published by ProjectScore and Pacini, containing 1,640 gene sequences from 323, covering a broader cellular genetic background, and helps to assess the robustness of the model under different tissue origins and experimental conditions.

[0035] Dataset S3 is used to train large-scale DNA models for essential genes. It aggregates essential gene sequences from multiple sources, including DEG, CEG, ePath, and OGEE3, across species. All sequences are converted to uppercase before pre-training, and only the four standard nucleotide characters A / G / C / T are retained, while spaces, newlines, and other non-standard characters are removed. The training and validation sets are then split at a ratio of 0.95:0.05 to perform MLM tasks. S3 contains 478,431 DNA sequences, providing ample unsupervised corpus for downstream tasks.

[0036] The specific process of step 2 is as follows: Extraction of physicochemical features of multiple sequences: Based on the standardized DNA sequences processed in step 1, the system extracts 7 types of sequence physicochemical characteristics and clarifies the parameters and dimensions of each type of characteristic: K-mer features: K values ​​were set to 1, 2, and 3, and a total of 84 features were extracted to characterize the local nucleotide composition pattern of the sequence.

[0037] PseKNC features: setting k to 3, extracting 65 features, and fusing information on the correlation between local composition and global physicochemical properties of the sequence.

[0038] Cumulative skew feature: Two features are extracted to reflect the cumulative skewness of nucleotide distribution in the sequence.

[0039] Z-Curve features: Extract three features and characterize the global geometric structure of the sequence through three-dimensional spatial mapping.

[0040] GC content characteristics: One feature was extracted to characterize the ratio of guanine (G) to cytosine (C) in the sequence; AT / GC ratio feature: One feature is extracted, which reflects the ratio of adenine + thymine (AT) to GC in the sequence.

[0041] λ-Z-curve features: Extracts 85,995 features, providing high-precision global geometric characterization information for the sequence.

[0042] Feature set integration: The above 7 types of features are merged to form an initial feature set of 86,151 dimensions, covering multi-scale sequence information from local composition to global geometry.

[0043] Two-level recursive feature elimination (RFE) screening: Using the S1 dataset from step 1 as the feature selection development set, efficient selection of high-dimensional redundant features is achieved through a two-stage RFE: Level 1 RFE screening: Grouping: The 85,995 λ-Z-curve features with the highest dimension are divided into 9 groups of equal size, with each group containing 9,555 features.

[0044] Feature concatenation: Each λ-Z-curve feature is concatenated with the other 6 types of features, such as K-mer and PseKNC (a total of 84+65+2+3+1+1=156 items), to form 9 sets of sub-features. Each set of sub-features has a dimension of 9555+156=9711.

[0045] Independent screening: The RFE algorithm was executed on each of the 9 sub-feature sets, and the top 1000 candidate features with the highest contribution were selected from each set.

[0046] Merging and deduplication: Collect 9 groups of candidate features, merge them and remove duplicates to obtain an 8263-dimensional intermediate feature set.

[0047] Second-level RFE screening: Optimal size search: Using the 8263-dimensional intermediate feature set as input, the RFE algorithm is executed again to iterate through the model performance corresponding to different feature dimensions.

[0048] Determining the optimal dimension: Experimental verification revealed that the model achieved the best overall performance (such as AUPR and F1) in classification and regression tasks when the feature dimension was 4800. Therefore, 4800 was established as the final input feature dimension.

[0049] Reuse of screening results: The optimal 4800-dimensional feature set was reused in the subsequent model training of all datasets in S1-S2. Among them, 4730 features came from the λ-Z-curve, which confirms its core role in the identification of essential genes.

[0050] Example 2 Based on Example 1, step 3 constructs a time-frequency network comprising a dual-channel structure of T-Block and F-Block. The T-Block processes the one-hot encoded sequence corresponding to the feature set in the time domain to capture local dependencies. The specific process for generating the time domain output is as follows: Preprocessing of the input sequence: Taking the DNA sequence corresponding to the 4800-dimensional optimal feature set obtained through screening as input, one-hot encoding is performed on the sequence to convert the nucleotide characters (A / G / C / T) into a numerical time series S∈ R L , where L is the length of the encoded sequence, providing a unified input format for dual-channel processing; Perform T-Block time-domain processing to generate time-domain output. S o (t) ∈ R T .

[0051] Perform T-Block time-domain processing to generate time-domain output. S o (t) ∈ R T The specific process is as follows: Sequence block partitioning: dividing the one-hot encoded time series S∈ R L Divided into multiple small blocks of fixed length X p ∈ R P×N Where P is the length of each small block, N is the number of small blocks, and N = [ L / P If the sequence length is not divisible by P, the last block is zero-padded to length P. Block feature embedding: Each small block is embedded through a trainable embedding layer. X p The dimensions are mapped from P to the model's preset feature dimensions to ensure that the block information is adapted to the input requirements of the subsequent Transformer encoder.

[0052] Local dependency capture: All embedded block feature vectors are sequentially fed into the Transformer encoder. The multi-head attention mechanism is used to mine the nucleotide position correlations within each block. The LayerNorm layer and residual connections are combined to alleviate gradient vanishing and enhance local feature preservation. Temporal output generation: The block feature vectors processed by the Transformer encoder are integrated and mapped to a unified dimension through a linear projection layer to generate the T-Block temporal output. S o (t) ∈ R T T is the preset output length, expressed by the formula: .

[0053] In step 3, F-Block maps the sequence to the frequency domain using the extended Discrete Fourier Transform (DFT). After processing by the complex-valued Transformer encoder, it is transformed back to the time domain using the inverse Discrete Fourier Transform (iDFT) to generate the frequency domain output. Then, adaptive weights are determined based on the fundamental sequence energy, and the time-frequency fusion features are obtained by weighted summation of the two channel outputs. The specific process is as follows: Extended DFT frequency domain mapping: The extended discrete Fourier transform (DFT) is used to map the time series... S∈R L Mapping to the frequency domain yields an extended spectrum of length L+T. The spectrum is aligned with the DFT frequency scale of the entire sequence, as shown in the formula: F=ExtendedDFT ( S ); ; Utilizing the conjugate symmetry of the spectrum of a real-valued sequence, only the first... Individual spectral values ​​reduce computational overhead; Complex-valued spectrum preprocessing: [This refers to the preprocessing of the spectrum.] Perform normalization (subtract the mean and divide by) The standard deviation is calculated, and then passed through a trainable linear projection. Embed it as a complex value vector (No position embedding required); Complex-valued spectral attention computation: An improved multi-head attention mechanism is used to process complex-valued vectors. F d : For each attention head h=1,2,...,H, through learnable parameters: ,Will F d Projection yields: Query Q h = F d T W h Q ; key K h = F d T W h K ; value V h = F d T W h V ; Perform complex-valued dot product attention calculation: ; Stitch together all attention head outputs and pass through linear projection The complex-valued spectral attention output is obtained: ; Encoder and Inverse Transform Processing: The complex-valued spectral attention output is fed into a Transformer encoder with LayerNorm, feedforward network, and complex domain residual connections, and the output is...F z Obtained through linear mapping ; By using the inverse discrete Fourier transform (iDFT) to F o Returning to the time domain, the last T time points are extracted as the frequency domain output of the F-Block. The formula is expressed as: ; Dual-channel adaptive weighted fusion generates time-frequency fusion features ; Adaptive weight calculation: Calculate the T-Block output weights based on the fundamental frequency sequence energy of the input sequence S. w t With F-Block output weights w f The formula is: w t , w f =Weighting(S); Time-frequency fusion feature generation: time domain output S o (t) With frequency domain output S o (f) ; .

[0054] The specific process of step 4 is as follows: DNA large-scale model backbone initialization: Using the S3 unsupervised corpus constructed in step 1 as training data, the NucleotideTransformer-v2 (NTv2) model was pre-trained. The pre-training task was MLM (masked language model), which enabled the model to learn the general representation ability of essential gene sequences across species, forming a basic backbone model for essential gene prediction tasks. DNA sequence preprocessing and segmentation: Input sequence acquisition: Select the DNA sequences of the S1-S2 dataset after standardization in step 1 as input data for model fine-tuning and inference.

[0055] 2.6-mer segmentation: The DNA sequence is split using a 6-mer segmentation strategy, with 6 consecutive nucleotides forming a segmentation unit. The segmentation unit is encoded using a vocabulary of approximately 4107 words and converted into a token sequence that the model can recognize. Processing of ultra-long sequence sliding window slices: Length judgment: Set the maximum input sequence length of the model to 2048 tokens. For excessively long sequences whose length exceeds this limit after word segmentation, execute the sliding window slicing strategy.

[0056] Slicing parameters: Set the slice window size to 2048 tokens and the step size to 1024 tokens. That is, starting from the beginning of the sequence, each slice is cut into 2048 tokens as a segment. Subsequent slices overlap with the previous slice by 1024 tokens until the entire long sequence is completely covered, resulting in multiple sequence segments that do not omit the core information. Fragment vector extraction: Each sequence segment obtained from all slices is fed into the pre-trained NTv2 model for forward encoding, and the vector representation of each segment is obtained through one of the following two methods: [CLS] representation: Take the vector of the [CLS] token at the beginning of the word segmentation sequence in the output of the last layer of the NTv2 model, and use it as the core representation vector of the segment.

[0057] Average pooling: Calculate the mean of all token vectors output by the last layer of the NTv2 model to obtain the global average representation vector of that segment (average pooling is used by default, but can be replaced by max pooling or attention pooling). Gene-level aggregation generates unified gene representation: For all fragment vectors corresponding to the same DNA sequence, perform gene-level aggregation operation: integrate all fragment vectors through average pooling (or a preset aggregation strategy) to eliminate local differences caused by fragmentation and generate a unique unified representation vector for the gene, which is used for subsequent joint training with time-frequency fusion features.

[0058] Example 3 Based on Example 1 or Example 2, the specific process of step 5 is as follows: Multimodal feature fusion preprocessing: Feature dimension alignment: The time-frequency fusion features output by the time-frequency network and the unified gene representation vector generated by the DNA large model are mapped to the same dimension through linear projection (the preset fusion dimension is adapted to the input of the subsequent layers of the model). Feature fusion and concatenation: The two types of features after alignment are fused by concatenation (Concat) to form a multimodal joint input feature vector. This vector contains both sequence local dependency, long-range correlation and time-frequency domain feature information, providing comprehensive representation support for model training. Training parameters and optimizer configuration: Optimizer selection: The AdamW optimizer is used to update model parameters, and the weight decay coefficient is set to 0.01 to alleviate overfitting by suppressing excessive parameter growth.

[0059] Learning rate schedule design: Employing a "linear warmup → linear decay" strategy. Warm-up phase: In the early stages of training, the learning rate is gradually increased in a linear fashion to avoid model instability caused by an initial high learning rate.

[0060] Decay phase: After the warm-up, the learning rate gradually decreases according to a linear law to ensure stable convergence of the model in the later stage; Training efficiency optimization: Enable AMP half-precision training mode to improve computational throughput efficiency and shorten the training cycle while ensuring training accuracy; set Dropout probability to 0.1 to randomly deactivate some neurons to enhance model robustness. Handling class imbalance issues: For class imbalance scenarios where the sample size of essential genes is far less than that of non-essential genes, a dual strategy of "loss weighting + resampling" is adopted: Weighted loss function: In classification tasks, weighted BCE (binary cross-entropy) loss or Focal loss is used to assign higher loss weights to minority class (essential gene) samples, thereby increasing the model's attention to the minority class.

[0061] Data resampling: During training, samples are dynamically resampled. By increasing the sampling probability of minority class samples or decreasing the sampling proportion of majority class samples, the distribution of the two classes of samples in the training set is balanced, and the model is prevented from being biased towards the majority class prediction. Generalization capability enhancement mechanism: Early stopping strategy: Use the validation set AUPR (area under the precision-recall curve) or F1 score as the early stopping criterion. When the validation set criterion no longer improves within several consecutive training cycles, stop training and save the current optimal model parameters to prevent overtraining from causing a decline in generalization ability.

[0062] Regularization mechanism: In addition to weight decay and Dropout, LayerNorm layers are introduced between the layers of the model for normalization processing, stabilizing the data distribution, and mitigating the gradient vanishing problem in conjunction with residual connections, further enhancing the model's ability to adapt to unseen data.

[0063] Training process execution: Training objectives: For the binary classification task of essential genes, train the model to output probability values ​​in the interval [0,1]; for the regression task of "degree of necessity", output continuous necessity scores. The two types of tasks can share an encoder for separate optimization, or adopt a multi-task joint training mode.

[0064] Iterative training: Input multimodal joint features and corresponding supervision labels ("required / non-required labels" or necessity scores of the S1-S2 dataset) in batches, calculate the loss value through backpropagation, update all trainable parameters of the model using the AdamW optimizer, and repeat the iteration until the early stopping condition is triggered or the preset maximum number of training rounds is reached.

[0065] In another implementation, regarding the encoder and representation layers, the domain-specific DNA pre-trained model can be replaced with other large sequence models (such as the DNABERT series, GROVER, other versions of Nucleotide Transformer, Gena-LM-BERT, etc.), or a de novo trained long-sequence Transformer / state-space model can be used as the backbone. Tokenization can be changed from 6-mer to 3–8-mer, a character-level dictionary, or a BPE / Unigram vocabulary. Ambiguous bases (N) can be replaced by rule-based substitution or learnable embeddings. Fine-tuning methods, in addition to full-parameter fine-tuning, can include parameter-efficient strategies such as LoRA / QLoRA, Adapter, and linear probes. For long-sequence modeling, besides sliding window-overlapping slicing, sparse attention (Longformer, BigBird, Performer), hierarchical / pyramid pooling, dilated convolution (Temporal CNN), or a state-space / convolutional hybrid (S4 / H3 / Hyena / RetNet) can be used, and hierarchical pooling or Set-Transformer can be introduced for gene-level aggregation. Regarding the frequency and time domain paths, the frequency domain branch can be replaced by extended DFT with DCT, Discrete Wavelet Transform (DWT), learnable Fourier features, or Hilbert transform envelope; complex-valued attention can be replaced with real-valued spectral attention or band convolution; alternatively, a "time-only" single-channel structure can be adopted, and local and long-range dependencies can be coupled through cross-layer gating / residuals, ultimately performing late fusion at the output or mid-level fusion through cross-attention. For class imbalance and noise robustness, weighted BCE / Focal loss can be replaced with LDAM-DRW, Class-Balanced Loss (based on the number of effective samples), Focal-Tversky, or noise reduction training paradigms such as positive and negative sampling reweighting, PU-Learning (positive sample-unlabeled learning), bootstrap, co-teaching / small loss priority, and consistency regularization can be used. Regarding training paradigms and data augmentation, in addition to regular supervision, self-supervision (MLM / contrastive learning / sequential prediction) can be performed first, followed by a small amount of supervision for fine-tuning; augmentation strategies can include reverse complementarity, random occlusion / replacement, k-mer dithering, noise injection, and test-time augmentation (TTA), supplemented by hierarchical learning rate decay and cosine or multi-stage learning rate schedules.

[0066] The specific process of step 6 is as follows: First, switch the end-to-end model trained in step 5 to inference mode, and perform the same preprocessing and sliding window slicing as during training on each gene sequence in the test set: Let the gene... The segmented sequence is divided into Overlapping segments The model outputs the uncalibrated logit of the classification head for each segment. (Binary classification, output is calibrated logit) and continuous score of the regression head. (Continuity score of regression output).

[0067] The classified output is then calibrated using temperature scaling. For any fragment, logit... Introducing temperature parameters , obtain the calibrated probability in For the sigmoid function, It is an exponential function. This is the uncalibrated logit, where T is the temperature parameter. temperature Fitting only on the validation set to minimize the negative log-likelihood (NLL), thereby improving the probabilistic interpretability and threshold transferability across cell lines / platforms: Let the validation set sample index be... , Real Labels The corresponding logit is ,but in The optimal temperature parameters are obtained by fitting the data on the validation set. This represents selecting parameters that minimize the objective function, and constrains them. This indicates that the temperature must be positive; For the validation set sample index, This represents the summation over all samples in the validation set. get Then, fix this parameter and apply the following formula to the probabilities of all segments in the test set: The calculation involves calculating the temperature scaling threshold. Since temperature scaling is a monotonic transformation of the logit, it does not change the sample ranking but significantly improves the calibration of scores, making the same threshold more stable across different distributions. The transferable decision threshold is then determined on the validation set. For any candidate threshold Define prediction And the sensitivity and specificity are obtained from the confusion matrix: in The true positive rate represents the percentage of positive results within a certain threshold. The proportion of samples that are actually positive that are correctly predicted as positive. The true negative rate represents the percentage of true negatives within a certain threshold. The proportion of samples that are actually negative but are correctly predicted as negative; , , , Thresholds The counts of true positives, false negatives, true negatives, and false positives are as follows: The total number of true positive classes. This represents the total number of true negative classes.

[0068] Youden's Defined as choose This approach achieves the overall optimal separation between sensitivity and specificity; for those placing greater emphasis on identifying minority groups (essential genes), a different approach is used. Best of all: and take The optimal decision threshold is selected on the validation set. This threshold is determined solely by the validation set and consistently used on the test set to avoid information leakage and ensure the definition of a "transferable threshold" holds. Fragment-level predictions are then aggregated at the gene level to obtain the final "necessity score." The aggregation of classification probabilities employs a "logit aggregation followed by calibration" approach to maintain linear additivity. in This is the gene-level aggregated logit obtained by averaging the logit values ​​of all fragments of the gene; As genes The necessity score. The final binary classification conclusion is... in This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. For the regression head, gene-level aggregation is also performed, for example... And can be with The combined output is compatible with the dual-channel decision-making of "scoring + judgment".

[0069] Finally, the uncertainty quantification result is output to indicate the risk of boundary samples and cross-domain drift. Uncertainty consists of two parts: "intra-segment consistency" (reflecting local differences / noise in the sequence) and "model cognitive uncertainty" (reflecting parameter uncertainty / insufficient generalization). Intra-segment consistency can be measured by variance. in For genes Intra-segment consistency uncertainty is used to measure the consistency of predicted probabilities among different sliding window segments of the same gene; Indicates the fragment index variance calculation For genes No. The calibration probabilities of each fragment after temperature scaling. Higher variance indicates more inconsistent evidence provided by different windows for the gene, suggesting stronger influence from boundary samples or local sequence differences / noise. Higher variance also indicates more inconsistent evidence provided by different windows for the same gene; model cognitive uncertainty is obtained through MC Dropout, allowing for multi-model maintenance. The next forward pass yields... Then the predicted mean and variance are: in The number of forward samples performed on the same gene under random submodel conditions; For the first Genes obtained from the next random forward pass Calibration probability prediction; for The mean of the predictions from each sampling is used as the average prediction for that gene. For genes Cognitive uncertainty, using The sample variance characterization of the second prediction, where These are the normalized coefficients with unbiased variance. Confidence intervals based on quantiles can also be provided. Meanwhile, the overall uncertainty is characterized by binary classification prediction entropy: The system ultimately outputs structured results for each gene, thus providing risk warnings and interpretable confidence information that can be used for downstream drug target prioritization and experimental validation site selection, while giving "necessary / non-necessary" conclusions.

[0070] In summary, the essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning provided in this invention constructs an essential gene prediction system for long sequences and cross-scenario applications. It uses a domain-specific large DNA model as the core representation layer, continuing pre-training and task-specific fine-tuning on cross-species / cross-source corpora. On the predictor side, a dual-channel fusion design (T-Block and F-Block) in parallel time and frequency domains is employed. Extended DFT and complex-valued attention are used for frequency domain modeling, followed by iDFT back to the time domain and adaptive weighted fusion with the time branch, thereby simultaneously capturing long-distance dependencies and local discriminative patterns. For ultra-long DNA sequences, an engineered pipeline of sliding window-overlapping slicing and gene-level aggregation is used to achieve high throughput and low memory usage for end-to-end training and inference. At the decision level, class imbalance and noise-robust training, cross-cell line / cross-platform domain robust alignment are introduced, and scores are calibrated using methods such as temperature scaling / order preservation, providing thresholds and uncertainty quantification outputs that can be transferred across datasets, facilitating the practical implementation of drug target prioritization and experimental site selection. Essential gene prediction schemes based on domain-specific DNA large-scale models and time-frequency dual-channel fusion (including extended DFT, complex-valued attention, iDFT backtracking, and adaptive fusion weights); modeling and inference schemes for sliding window overlapping slicing and gene-level aggregation for long sequences; robust training and score calibration for cross-cell line / cross-platform generalization and threshold transferability decision schemes; uncertainty quantification and structured output interface schemes for screening applications.

Claims

1. A method for predicting essential genes based on the fusion of a large DNA model and time-frequency domain deep learning, characterized in that, Includes the following steps: Step 1: Construct datasets S1-S3 covering the annotation of essential human genes and essential gene sequences across species for CRISPR screening in multiple cell lines. Standardize all sequences, remove non-standard characters, and divide them into training / validation sets according to a preset ratio. Step 2: Extract physicochemical features of multiple sequences, and use a two-level recursive feature elimination filter to obtain the feature set with the optimal dimension; Step 3: Construct a time-frequency network with a dual-channel structure including T-Block and F-Block. The T-Block generates the time domain output; the F-Block maps the sequence to the frequency domain through extended discrete Fourier transform. After processing by the complex-valued Transformer encoder, it is converted back to the time domain through inverse discrete Fourier transform to generate the frequency domain output. The time-frequency fusion feature is obtained by weighted summation of the two channel outputs. Step 4: Construct a large DNA model, use 6-mer segmentation to process DNA sequences, implement a sliding window slicing strategy for ultra-long sequences, and obtain unified gene characterization through gene-level aggregation; Step 5: Integrate time-frequency fusion features and unified gene representation as model input to train the model; Step 6: Based on the trained DNA large model, inference is performed on the test set sequences. First, the prediction scores are calibrated by temperature scaling. Then, the transferable decision threshold is determined based on the optimal point of Youden's J or F1 on the validation set. The prediction results of the test sequences are aggregated at the gene level, and finally, the necessity score, binary classification conclusion and uncertainty quantification results are output.

2. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 1, characterized in that, In step 1: Dataset S1 covers 11 human cancer cell lines; a majority voting strategy is used to define core essential genes: the same gene is marked as positive when it is determined to be essential in more than 6 cell lines; genes not included in the essential set are considered non-essential after HGNC annotation, forming a clear negative sample set. Dataset S2 contains 1640 gene sequences from 323; Dataset S3 is used to train large DNA models for essential genes. It aggregates essential gene sequences from multiple sources across species, including DEG, CEG, ePath, and OGEE3. All sequences are converted to uppercase before entering pre-training, and only the four standard nucleotide characters A / G / C / T are retained, while spaces, newlines, and other non-standard characters are removed. Then, the training / validation sets are divided according to a specified ratio to perform MLM tasks.

3. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 1, characterized in that, The specific process of step 2 is as follows: Extraction of multiple types of sequence physicochemical features: Based on the processed standardized DNA sequence, seven types of sequence physicochemical features were systematically extracted, and the parameters and dimensions of each type of feature were clarified; Perform two-level recursive feature elimination screening: Level 1 RFE screening: Grouping: The 85,995 λ-Z-curve features with the highest dimension are divided into 9 equal groups; Feature concatenation: Each set of λ-Z-curve features is concatenated with the other 6 types of features, including K-mer and PseKNC, to form 9 sets of sub-features; Independent screening: The RFE algorithm was executed on each of the 9 sub-feature sets, and the top 1000 candidate features with the highest contribution were selected from each set; Merging and deduplication: Collect 9 groups of candidate features, merge them and remove duplicates to obtain an intermediate feature set; Second-level RFE screening: Optimal size search: Using the intermediate feature set as input, the RFE algorithm is executed again to iterate through the model performance corresponding to different feature dimensions; Determine the optimal dimension: Establish 4800 dimensions as the final input feature dimension; Reuse of screening results: This optimal 4800-dimensional feature set is reused uniformly in the subsequent model training of all datasets in S1-S2.

4. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 3, characterized in that, Step 3 involves constructing a time-frequency network with a dual-channel structure including T-Block and F-Block. The specific process of generating the time domain output using T-Block is as follows: Preprocessing of the input sequence: Taking the DNA sequence corresponding to the 4800-dimensional optimal feature set obtained through screening as input, one-hot encoding is performed on the sequence to convert the nucleotide characters into a numerical time series S∈ R L , where L is the length of the encoded sequence; Perform T-Block time-domain processing to generate time-domain output. S o (t) ∈ R T .

5. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 4, characterized in that, Perform T-Block time-domain processing to generate time-domain output. S o (t) ∈ R T The specific process is as follows: Sequence block partitioning: dividing the one-hot encoded time series S∈ R L Divided into multiple small blocks of fixed length X p ∈ R P ×N Where P is the length of each small block, N is the number of small blocks, and N = [ L / P ]; Block feature embedding: Each small block is embedded through a trainable embedding layer. X p The dimension is mapped from P to the feature dimension preset by the model; Local dependency capture: All embedded block feature vectors are sequentially fed into the Transformer encoder. The multi-head attention mechanism is used to mine the nucleotide position correlations within each block. The LayerNorm layer and residual connections are combined to alleviate gradient vanishing and enhance local feature preservation. Temporal output generation: The block feature vectors processed by the Transformer encoder are integrated and mapped to a unified dimension through a linear projection layer to generate the T-Block temporal output. S o (t) ∈ R T T is the preset output length.

6. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 4, characterized in that, In step 3, F-Block maps the sequence to the frequency domain through extended discrete Fourier transform. After processing by a complex-valued Transformer encoder, it is transformed back to the time domain through inverse discrete Fourier transform to generate the frequency domain output. The specific process of obtaining the time-frequency fusion feature by weighted summation of the two channel outputs is as follows: Extended DFT frequency domain mapping: This involves using the extended discrete Fourier transform to map the time series... S∈R L Mapping to the frequency domain yields an extended spectrum of length L+T. The formula is: F=ExtendedDFT ( S )； ； Utilizing the conjugate symmetry of the spectrum of real-valued sequences, only the first... Individual spectral values ​​reduce computational overhead; Complex-valued spectrum preprocessing: For the spectrum Normalization is performed, followed by training-compatible linear projection. Embed it as a complex value vector ; Complex-valued spectral attention computation: An improved multi-head attention mechanism is used to process complex-valued vectors. F d : For each attention head h=1,2,...,H, through learnable parameters: ,Will F d Projection yields: Query Q h = F d T W h Q ; key K h = F d T W h K ; value V h = F d T W h V ; Perform complex-valued dot product attention calculation: ； Stitch together all attention head outputs and pass through linear projection The complex-valued spectral attention output is obtained: ； Encoder and Inverse Transform Processing: The complex-valued spectral attention output is fed into a Transformer encoder with LayerNorm, feedforward network, and complex domain residual connections, and the output is... F z Obtained through linear mapping ; By using the inverse discrete Fourier transform (iDFT) to F o Returning to the time domain, the last T time points are extracted as the frequency domain output of the F-Block. The formula is expressed as: ； Dual-channel adaptive weighted fusion generates time-frequency fusion features ; Adaptive weight calculation: Calculate the T-Block output weights based on the fundamental frequency sequence energy of the input sequence S. w t With F-Block output weights w f The formula is: w t , w f =Weighting(S); Time-frequency fusion feature generation: time domain output S o (t) With frequency domain output S o (f) ; 。 7. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 1, characterized in that, The specific process of step 4 is as follows: DNA large-scale model backbone initialization: The NucleotideTransformer-v2 model was pre-trained using the S3 unsupervised corpus as training data. The pre-training task was MLM. DNA sequence preprocessing and segmentation: Input sequence acquisition: DNA sequences from the standardized S1-S2 dataset were selected as input data for model fine-tuning and inference; 6-mer segmentation: The DNA sequence is split using a 6-mer segmentation strategy, with 6 consecutive nucleotides forming a segmentation unit. The segmentation unit is encoded using a vocabulary of approximately 4107 words and converted into a token sequence that the model can recognize. Handling of ultra-long sequences by sliding window slicing: Length judgment: Set the maximum input sequence length of the model to 2048 tokens, and execute the sliding window slicing strategy for ultra-long sequences whose length exceeds the upper limit after word segmentation; Slicing parameters: Set the slice window size to 2048 tokens and the step size to 1024 tokens. Subsequent slices overlap with the previous slice by 1024 tokens until the entire long sequence is completely covered, resulting in multiple sequence segments that do not omit core information. Segment Vector Extraction: Sequence segments obtained from all slices are fed into the pre-trained NTv2 model for forward encoding to obtain the vector representation of each segment. Gene-level aggregation generates a unified gene representation: For all fragment vectors corresponding to the same DNA sequence, perform gene-level aggregation operation: integrate all fragment vectors to generate a unique unified representation vector for the gene.

8. The essential gene prediction method based on the fusion of a large DNA model and time-frequency domain deep learning according to claim 1, characterized in that, The specific process of step 5 is as follows: Feature dimension alignment: The time-frequency fusion features output by the time-frequency network and the unified gene representation vector generated by the DNA large model are mapped to the same dimension through linear projection; Feature fusion and concatenation: The two types of features after alignment are fused by concatenation to form a multimodal joint input feature vector; Training parameters and optimizer configuration: The AdamW optimizer is used to update model parameters, and a linear warmup → linear decay strategy is adopted. Training process execution: Training objectives: For the binary classification task of essential genes, train the model to output probability values ​​in the interval [0,1]; for the "degree of necessity" regression task, output a continuous necessity score; Iterative training: Input multimodal joint features and corresponding supervision labels in batches, calculate the loss value through backpropagation, update all trainable parameters of the model using the AdamW optimizer, and repeat the iteration until the early stopping condition is triggered or the preset maximum number of training rounds is reached.

Images