Ribosome stagnation relative strength prediction method of MambaPlus and Transform parallel architecture

Abstract

The invention discloses a ribosome stagnation relative strength prediction method of a MambaPlus and Transform parallel architecture, and belongs to the crossing field of bioinformatics and artificial intelligence. According to the model, firstly, a DNA sequence is converted into multi-dimensional feature representation through K-mer and One-hot composite coding, a local sequence pattern is extracted by adopting a parallel multi-scale convolutional neural network, and long-range dependency and global context information are modeled respectively in combination with a MambaPlus encoder and a Transform encoder; multi-modal features are integrated through an adaptive weighted fusion strategy, dynamic aggregation is carried out on sequences in combination with attention pooling, and finally prediction of ribosome stagnation relative strength is realized in combination with biophysical features. The error of the depth model for predicting the relative strength of ribosome stagnation is smaller than 0.27, and compared with a traditional model, the depth model has higher accuracy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics, artificial intelligence and computational biology, and specifically relates to a method for predicting the relative intensity of ribosome stagnation using a parallel architecture of MambaPlus and Transformer. Background Technology

[0002] Ribosomal stagnation refers to the biological phenomenon in which ribosomes, during mRNA translation, significantly slow down or even briefly halt their movement due to encountering rare codons, mRNA secondary structures, or regulatory elements. It is a crucial mechanism for regulating translation efficiency in cells. The relative intensity of ribosomal stagnation is an indicator used to quantify the degree of stagnation. By comparing the residence time or movement efficiency of ribosomes in the target sequence and the reference sequence, it reflects the relative ability of a specific sequence to induce ribosomal stagnation. The relative intensity of ribosomal stagnation is negatively correlated with protein expression levels; the higher the stagnation intensity, the slower the ribosomes move on the mRNA, and the fewer protein molecules are translated per unit time. Appropriately regulating the stagnation intensity can balance translation quality and efficiency, avoiding insufficient protein synthesis due to excessive stagnation. Therefore, predicting the relative intensity of ribosomal stagnation is a key issue in synthetic biology and metabolic engineering. Accurately predicting the mapping relationship from DNA sequence to the relative intensity of ribosomal stagnation is of great significance for optimizing mRNA stability and protein synthesis efficiency. Traditional methods, mainly based on sequence features and simple machine learning models, suffer from the following technical bottlenecks:

[0003] 1. Insufficient recognition of multi-scale regulatory elements: The spatial differences of regulatory elements affecting ribosome arrest in DNA sequences are significant, and single-scale feature extraction methods cannot simultaneously capture the synergistic effect of codon bias and regulatory element combination patterns;

[0004] 2. Lack of modeling for bidirectional transcriptional regulation mechanisms: DNA sequences exhibit both forward transcription and reverse complementary strand regulation, and existing unidirectional sequence modeling methods cannot recreate the biological characteristics of this bidirectional synergistic regulation;

[0005] 3. Coarse cross-modal integration of sequence and biophysical features: Ribosome arrest is affected by the synergistic influence of sequence logic and biophysical features. Most existing methods process sequence features and biophysical features independently, without fully considering the synergistic mechanism between the two types of features;

[0006] 4. Imbalance between long-range dependence and local patterns in biological equilibrium: Ribosome arrest depends on both short-range local patterns and long-range regulatory relationships. Existing methods cannot capture long-range dependence, consume excessive computing power, and are not sensitive enough to local biological patterns. Summary of the Invention

[0007] This invention aims to overcome the technical limitations of existing methods for predicting the relative intensity of ribosome stagnation in long sequence modeling, multi-scale feature extraction, and multi-source feature fusion.

[0008] Based on the above-mentioned technical problems, this invention proposes a method for predicting the relative intensity of ribosome stagnation using a parallel architecture of MambaPlus and Transformer. The invention employs the following technical solution:

[0009] 1. Using a K-mer and One-Hot serial compound encoding method, the DNA sequence is first scanned in parallel using multiple K-mer sliding windows, and then each K-mer mapping is expanded using One-hot encoding to convert the DNA sequence of length L into a feature matrix. ,in This yields a composite encoding matrix vector, which serves as the input to the subsequent multi-scale CNN convolutional neural network.

[0010] 2. After obtaining the enhanced DNA local feature dimensionality information from step 1, the information is fed into a parallel multi-scale CNN using 3 / 5 / 7-mer convolutional kernels to extract local sequence patterns at different granularities. This network employs a parallel multi-branch architecture, using one-dimensional convolutional kernels of different sizes to simultaneously extract features from the encoded DNA sequence. Through this multi-scale design, the model can accurately capture multi-level collaborative regulatory signals in microscopic codon usage bias, mesoscopic transcription factor binding, and macroscopic ribosome binding regions, overcoming the limitations of single-scale feature methods in terms of biological feature coverage.

[0011] 3: Input the output of step 2 into the MambaPlus branch. The forward MambaPlus module simulates the 5'→3' transcription process, and the backward MambaPlus module simulates the regulation of the reverse complementary strand. The discretized parameter Δ is adaptively matched to the differences in transcription / translation timescales. In this way, the natural characteristics of bidirectional regulation of DNA are restored, which solves the problem of missing reverse regulatory signals in the existing unidirectional modeling methods and improves the accuracy of identifying reverse regulatory elements.

[0012] 4: Dynamically weightedly fuse the MambaPlus features and the feature outputs of the Transformer branch, and highlight the contribution of key DNA sequence fragments through the attention pooling mechanism to solve the problem of coarse multimodal fusion.

[0013] 5: The sequence depth features after dynamic weighted fusion in step 4 are spliced ​​with the standardized biophysical features to solve the problems of coarse fusion of existing radioactive multi-source features and neglect of biological synergistic effects, so that the prediction results are more consistent with the biological processes in the real cellular environment.

[0014] Step 1 specifically includes:

[0015] 1-1. In the initial stage of DNA sequence encoding, the system employs the K-mer sliding window technique to achieve comprehensive capture of local sequence features. Key regulatory elements in the DNA sequence that affect the relative strength of ribosome stagnation, including ribosome binding sites, transcription factor binding sites (RBS), and mRNA secondary structure sensitive regions, typically exist in the form of short sequence motifs. The lengths of these functional units range from a few bases to tens of bases. Based on this biological characteristic, the model is configured with a parameter of K=3, and the original DNA sequence is scanned through a sliding window with a fixed step size. Each window covers three consecutive bases, which corresponds exactly to the length of a complete codon. This window design can systematically cover all possible local combinations in the sequence, while the overlapping scans between windows ensure the continuity of feature extraction.

[0016] 1-2. Based on the single-window scanning mechanism established in 1-1, the model further utilizes a parallel multi-window strategy to deeply explore the complex regulatory patterns within DNA sequences. In synthetic biology practice, the relative strength of ribosome stagnation is often determined by the synergistic effect of multiple regulatory elements, such as the combined effects of promoter strength, RBS efficiency, and codon bias. By simultaneously sliding multiple K-mer windows, the model can capture the relative positional relationships and combination patterns among these regulatory elements, identifying sequence context effects that influence the relative strength of ribosome stagnation. This parallel scanning simulates the working mechanism of multi-factor synergistic regulation in organisms, ensuring a comprehensive exploration of the complex regulatory logic hidden in DNA sequences and providing a rich data foundation for understanding the complex relationship between sequence design and the relative strength of ribosome stagnation.

[0017] 1-3. Building upon the K-mer sequence fragments obtained in 1-2, the model intelligently expands the feature space using One-Hot encoding technology. Each K-mer fragment in the DNA sequence represents a specific biological function, such as an amino acid codon, a transcription factor recognition sequence, or a regulatory protein binding site. One-Hot encoding maps each K-mer to a high-dimensional sparse vector, where each dimension corresponds to a possible K-mer combination, and the activated dimension indicates the specific composition of the current sequence fragment. This encoding method not only preserves the category information of K-mers but, more importantly, converts discrete sequence symbols into continuous numerical representations, enabling subsequent neural networks to learn the complex mapping relationship between sequence features and the relative intensity of ribosome stagnation through linear algebraic operations.

[0018] 1-4. Based on the encoding in step 1-3, non-standard sequence patterns are processed by special character mapping; rare K-mer combinations or sequencing noise in the sequence are uniformly mapped to all--1 vectors for identification by the model; a composite encoding matrix of [sequence_length, 64] is constructed, in which the row dimension of the matrix maintains the spatial order of the original sequence, and the column dimension encodes the K-mer composition features, forming a structured representation that contains both positional and compositional information.

[0019] 1-5. Transpose the composite matrix from 1-4 into a feature tensor of [64, sequence_length].

[0020] Step 2 specifically involves:

[0021] 2-1. Based on the hierarchical structure characteristics of DNA sequences, this system adopts a parallel multi-scale CNN modular branching structure. The branching structure contains three parallel branches of convolutional layers. Each branch is configured with one-dimensional convolutional kernels of sizes 3, 5, and 7, respectively. Each branch includes: convolutional operations for sliding window scanning of the encoded DNA sequence; normalization and activation operations for batch normalization after each convolutional layer to reduce distribution shift; and regularization to ensure consistent overfitting.

[0022] 2-2. Convolutional kernels of different sizes target and capture multi-level biological features in DNA sequences:

[0023] 3-mer convolution kernel: Adapted to standard codon length, specifically capturing the characteristic patterns of amino acid coding units and identifying coding regions in open reading frames;

[0024] 5-mer convolutional kernel: covers the core region of typical transcription factor binding sites (TFBS) and recognizes specific sequence patterns for regulatory protein binding, such as TATA boxes and GC boxes;

[0025] 7-mer convolution kernels: capture a broader ensemble of regulatory elements, including characteristic patterns of ribosome binding site (RBS) box promoter regions;

[0026] Each convolutional kernel scans through a sliding window to extract locally conserved sequence patterns in the DNA sequence. These patterns are directly related to the relative strength of ribosome arrest: identifying Kozak sequence variants that affect translation initiation efficiency, detecting rare codon clusters that affect ribosome arrest, and discovering local sequence features that are related to the qualitative analysis of mRNA secondary structure.

[0027] S2-3: Based on the branches of S2-1, the feature matrix is ​​spliced ​​along the channel dimension to form a feature matrix that integrates multi-scale information. The spliced ​​feature matrix has a dimension of 192, which not only preserves the specific information of each scale, but also provides rich hierarchical control signal input for the subsequent MambaPlus / Transformer module.

[0028] Specifically, step S3 is as follows:

[0029] S3-1. Based on the multi-scale DNA sequence features extracted in S2, this step uses a bidirectional MambaPlus branch architecture to achieve bidirectional modeling of the sequence. The forward module extracts local context from the input features and uses the SSM state-space model to simulate the transcription process in the 5'→3' direction, capturing the temporal signals generated by RNA polymerase along the template strand. The backward module performs a reversal operation on the input features and then processes them through a MambaPlus module that is isomorphic to the forward module. After reversing, the sequence direction is restored, simulating the regulatory signal transmission of the reverse complementary strand. The gating mechanism enables the selective transmission of features through sigmoid-activated gating branches of each MambaPlus block, dynamically balancing the weights of convolutional local features and long-range dependencies of SSM.

[0030] S3-2. Based on the model processing method described in step S3-1, its biological significance is to restore the bidirectional regulatory characteristics of DNA sequences; the forward simulation matches the direction in which RNA polymerase moves from the promoter to the terminator during transcription, focusing on capturing the codon sequencing of coding regions, downstream regulatory signals of the transcription start site (TSS), and the temporal dynamics of mRNA synthesis; the backward simulation corresponds to the regulatory elements on the reverse complementary strand, covering natural regulatory modes such as the reverse interaction between enhancers and genes and the reverse complementary pairing of CRISPR repeat sequences; the bidirectional synergy realizes the superposition of positive and negative strand signals to restore the common phenomenon of poor positive and negative strand performance in the genome.

[0031] The discretization parameter Δ of the S3-3 Mamba module adapts to the temporal scale differences in transcription / translation in the following ways:

[0032] Δ is dynamically calculated from the input features, and its value is positively correlated with the action time of the sequence element. In terms of scale matching, for rapid transcription, Δ is adaptively increased to integrate long-range dependencies and solve the problem that a single time scale cannot cover multiple biological processes. The dynamic range of Δ is controlled by the degree kernel channel expansion coefficient to ensure sufficient modeling resolution for biological processes at different rates.

[0033] Specifically, step S4 is as follows:

[0034] S4-1. This step achieves dynamic fusion of MambaPlus features and Transformer features. The core is to dynamically adjust the weight ratio of the two types of features through learnable parameters. Specifically, when using a weighted fusion strategy, the model initializes two learnable parameters: the MambaPlus branch weights. With Transformer branch weights The weights are constrained to the (0,1) interval by sigmoid activation, and then... / ( + Normalization ensures that the sum of the weights is 1, enabling the model to adaptively allocate weights based on the characteristics of the input sequence.

[0035] S4-2. To highlight the contribution of key DNA fragments, the fused features are weighted and aggregated using an attention pooling module. This module consists of a fully connected layer and a Softmax layer: the former maps the feature at each location to a 1-dimensional attention score, while the latter normalizes the score to a probability distribution and performs a weighted summation. From a biological perspective, the attention weights automatically tilt towards regions with high information density, allowing the model to focus on functional elements when aggregating features and reducing the interference of redundant sequences on prediction.

[0036] S4-3. Compared with traditional multimodal fusion methods, this design solves the problem of coarse fusion through a dual mechanism: On the one hand, dynamic weights capture the complementarity of MambaPlus and Transformer features through learnable parameters. When processing sequences containing inverse regulatory elements, the model will increase the weight ratio of MambaPlus backward module features. On the other hand, attention pooling filters low-contribution segments and strengthens the feature signals of functional elements through spatial dimension weight allocation.

[0037] S4-4. The features resulting from dynamic fusion and attention pooling are not directly used for prediction, but are further integrated with biophysical features, achieving deep interaction of multi-dimensional features through fully connected layers. Specifically, the pooled sequence features and biophysical features are concatenated along the channel dimension, and after two layers of linear transformation, GELU activation, and layer normalization, the final prediction result is output. This design organically combines sequence-level regulatory signals with molecular-level physicochemical properties, enabling the model to simultaneously capture the structural and functional features of DNA sequences, further enhancing its ability to model complex regulatory mechanisms.

[0038] Specifically, step S5 is as follows:

[0039] S5-1. To ensure that biophysical features and sequence depth features have consistent scale and distribution during fusion, the model performs rigorous standardization on biophysical features. In the dataset, StandardScaler is used to standardize the 7 types of biophysical features, converting the feature values ​​into a 0-1 distribution to prevent any type of feature from dominating the fusion process due to excessively large numerical ranges.

[0040] S5-2. The dynamically weighted fused sequence features and standardized biophysical features are fused by concatenating them along the channel dimension. The concatenated features retain the original structure of both types of features, avoiding information loss caused by feature compression.

[0041] S5-3. The fused features after splicing achieve deep interaction through multi-layer fully connected layers, combined with batch normalization, GELU activation, and Dropout regularization. The first-layer linear transformation maps the spliced ​​features to a higher dimension, capturing the synergistic relationship between sequence features and biophysical features through nonlinear activation; the second-layer transformation further refines key interaction patterns, ultimately outputting the prediction result. This design breaks through the coarse mode of simple splicing and fusion, and mines the implicit correlations between multi-source features through nonlinear transformation.

[0042] S5-4. This fusion strategy accurately captures the natural synergistic mechanism between intracellular DNA sequences and biophysical properties: on the one hand, sequence features provide the structural basis for biophysical processes; on the other hand, biophysical features influence the function of sequences. For example, when sequence features show the presence of a strong Shine-Dalgarno sequence, and the utrCdsStructureMFE value in the biophysical features is low and negative, the model, through interactive computation in fully connected layers, strengthens the prediction weights for high translation efficiency. This is consistent with the mechanism in real cells where "sequence motifs + structural flexibility" jointly determine translation efficiency.

[0043] The ribosome stagnation relative intensity prediction method based on the MambaPlus and Transformer hybrid architecture provided in the above technical solution has the following advantages compared with the prior art:

[0044] 1. This invention utilizes a parallel multi-branch convolutional structure, using convolutional kernels of different sizes to simultaneously scan DNA sequences. It can simultaneously identify fine base combination features and extensive sequence context, more comprehensively covering multi-level biological information from nucleotides to regulatory modules, significantly improving the decoding ability of complex regulatory logic, and solving the problem that existing single-scale sequence encoding is difficult to capture short-range local patterns and long-range regulatory relationships in DNA sequences.

[0045] 2. This invention employs a bidirectional MambaPlus architecture to simulate the transcription process and potential reverse regulatory mechanisms in the 5'→3' direction, which more closely reflects phenomena such as promoter directionality and antisense transcription in actual organisms. The discretization parameter Δ of the state-space model adaptively simulates the timescale differences of different regulatory stages, enabling the model to distinguish between rapid transcription and slow translation elongation processes, and improving the accuracy of predicting the relative intensity of temporal dynamic ribosome arrest.

[0046] 3. Most existing methods process sequence features and biophysical features independently, neglecting their interaction. This invention organically integrates sequence logical features and biophysical constraints through dynamic weighted fusion and attention pooling mechanisms, simulating the complete path of sequence information influencing the relative intensity of ribosome stagnation through structural mediation in the actual cellular environment, thus overcoming the limitations of single-modal prediction.

[0047] 4. Compared to black-box prediction models, the modular design of this invention has a clear biological correspondence: parallel multi-scale CNNs correspond to the identification of local regulatory elements, MambaPlus simulates temporal dependencies, Transformer captures global interactions, and feature fusion reflects the synergy of multiple factors. This hierarchical design allows prediction results to be traced back to specific sequence regions or biophysical properties, providing feasible guidance for the rational design of genes in synthetic biology.

[0048] 5. Traditional methods often process feature extraction and regression prediction in stages, leading to suboptimal solutions. This invention uses an end-to-end training strategy to coordinate and optimize the feature extraction, fusion, and prediction modules, adaptively learning the complex mapping relationship between DNA sequences and the relative intensity of ribosome stagnation. It fully utilizes the complementarity of multi-source information, achieving higher prediction accuracy and generalization ability while maintaining biological rationality.

[0049] 6. By using the correct state-space model discretization formula and parameter definition, the problem of numerical indeterminate form in long sequence training is solved, which improves the model convergence speed and prediction consistency. Attached Figure Description

[0050] Figure 1 This is a general framework diagram of the present invention;

[0051] Figure 2 This is a framework diagram showing the correlation between the multi-level biophysical characteristics and the relative intensity of ribosome stagnation in this invention;

[0052] Figure 3 This is a diagram of the deep neural network model architecture of the present invention;

[0053] Figure 4 This is a schematic diagram of the K-mer One-Hot composite encoding process of the present invention;

[0054] Figure 5 This is a macroscopic branch structure diagram of the MambaPlus of the present invention;

[0055] Figure 6 This is a flowchart of the internal structure of the BiMambaPlusEncoder and MambaPlus block of the present invention.

[0056] Figure 7 This is a macroscopic branch structure diagram of the Transformer of this invention;

[0057] Figure 8 This is a structural diagram of the SequenceTransformer block in the Transformer branch of this invention;

[0058] Figure 9 This is a visualization of some ablation experimental indicators of the present invention. Figure 10 This is a schematic diagram of the complete network topology of the present invention. Detailed Implementation

[0059] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0060] This embodiment uses model data utilization efficiency and ablation experiments as examples, combined with the attached... Figures 1-9 The present invention will be further analyzed and explained.

[0061] A method for predicting the relative intensity of ribosomal stagnation based on a parallel architecture of MambaPlus and Transformer includes the following steps:

[0062] Example 1: Validating model performance using different sample sizes:

[0063] 1.1 The E. colis fGFP expression dataset published in the reference paper "Accuracy and data efficiency in deep learning models of protein expression" (Nature Communications, 2022, 13: 7755) was used. It contains 227,024 DNA sequences of 96 nucleotides in length and corresponding standardized relative strength measurements of ribosome stagnation. The measurements were logarithmically transformed and Z-score standardized to eliminate technical variation and batch effects.

[0064] 1.2 Model Training Configuration: Learning rate 1e-3; L2 regularization with coefficient 1e-5; batch size 64; gradient accumulation steps 1; maximum training epochs 80; early stopping patience value 8; termination occurs if validation set performance shows no improvement for 8 consecutive epochs; mean squared error (MSE) loss function; AdamW optimizer (where momentum parameter...). The learning rate tuner used was ReduceLROnPlateau with a patience value of 3 and a decay factor of 0.5; the fusion method used was weighted fusion with the learnable parameter initialized to 0.5; the evaluation metric used was the coefficient of determination (COP). ), root mean square error (RMSE) and mean absolute error (MAE);

[0065] 1.3 The publicly available dataset from the referenced paper in 1.1 was divided into training sets with 5K, 23K, 46K, 92K, 160K, and 184K samples respectively, for use in subsequent model performance experiments.

[0066] 1.4 The determination coefficients of the model under different training data volumes were obtained by inputting different training samples into the model for training, as shown in Table 1:

[0067] Number of training set samples R² Approximately 5K 0.4896 Approximately 23K 0.6550 Approximately 46K 0.7089 Approximately 92K 0.7861 Approximately 160K 0.8281 Approximately 184K 0.8467

[0068] Table 1

[0069] 1.5 Experimental data analysis results show that: this model uses 80% of the DNA sequence and biophysical features simultaneously for multimodal training, and performs well on the test set. The value reached 0.8467; the table data results show that when the amount of data was reduced by nearly half, The model remains at a high predictive level, indicating that it uses data efficiently.

[0070] Example 2: Evaluation of the contribution of each part of the model to the model's prediction accuracy through ablation experiments:

[0071] 2.1 This example demonstrates ablation of various model structures or techniques, such as multi-scale CNN, bidirectional MambaPlus architecture, Transformer architecture, using average fusion instead of dynamic fusion, using only MambaPlus branches, using only Transformer branches, and removing biophysical features, by sequentially eliminating each part of the model or replacing it with an identity module. The performance of each ablated model is then compared with the performance of the complete model.

[0072] 2.2 Multi-scale CNN ablation experiments Figure 1The convolutional layers were replaced with simple fully connected layers to adapt to the input dimensions of the subsequent MambaPlus and Transformer branches. Information processed by the encoding layer then flows directly into the two branches via the fully connected layers. Bidirectional MambaPlus architecture ablation experiments and Transformer architecture ablation experiments respectively transformed the two branches of the complete model (i.e.,...) Figure 5 MambaPlus macro branches and Figure 7 The macroscopic branches of the Transformer were removed, and implicit feature fusion was used; the branch dynamic fusion ablation experiment was conducted using average fusion and using only one branch respectively; the biophysical feature ablation experiment used only DNA sequence features for model training and canceled the multimodal function.

[0073] 2.3 The model performance indicators obtained through ablation experiments, including the coefficient of determination, root mean square error, mean absolute error, mean residual, standard deviation residual, and Pearson correlation coefficient, are summarized in Table 2:

[0074] Model variants R² RMSE MAE Mean Residual Std Residual Pearson Complete model 0.8493 0.4078 0.2763 0.0453 0.4063 0.9216 Single-kernel CNN (kernel size 3) 0.7813 0.4270 0.3102 0.0341 0.4200 0.8839 Single-kernel CNN (kernel size 5) 0.7865 0.4288 0.3121 0.0400 0.4102 0.8868 Single-kernel CNN (kernel size 7) 0.8060 0.4270 0.3102 0.0325 0.4256 0.8978 Remove bidirectional MambaPlus 0.7257 0.5237 0.3807 0.0400 0.5222 0.8519 Remove Transformer 0.8079 0.4383 0.3024 0.0185 0.4379 0.8988 Remove dynamic fusion (average fusion) 0.8014 0.4457 0.3154 0.0354 0.4442 0.8952 Remove dynamic fusion (Mamba only) 0.8178 0.4136 0.2858 0.0067 0.4135 0.9043 Remove dynamic fusion (transformer only) 0.4523 0.8230 0.6949 -0.0277 0.8225 0.6725 Remove biophysical features 0.7603 0.4896 0.3466 -0.0882 0.4816 0.8720

[0075] Table 2

[0076] Visualization of the metrics generated by each ablation experiment code, as shown below Figure 9 As shown.

[0077] 2.4 The ablation experiment results shown in Table 2 verify the necessity of the multimodal deep learning architecture involved in this model through systematic ablation experiments. Experimental data show that the complete model achieves a coefficient of determination of 0.8467 on the task of predicting the relative intensity of ribosome stagnation. (The complete model is then used in the text.) Based on this benchmark, the bidirectional MambaPlus module, based on a state-space model, demonstrates its core value in modeling long-range dependencies in DNA sequences. Its removal significantly reduces performance by 14.55%, indicating that the module effectively captures interactions of distal regulatory elements in the sequence through selective scanning and bidirectional state transfer. The correctness of the SSM discretization process has a crucial impact on prediction accuracy. The multi-scale convolutional neural network extracts multi-granularity local sequence patterns through parallel convolutional kernels. Using kernel size 3 alone causes an 8.01% performance loss, kernel size 5 alone causes a 7.39% performance loss, and kernel size 7 alone causes a 5.10% performance loss. This demonstrates that multi-scale feature extraction is crucial for identifying local functional elements such as promoters and codon preferences. The dynamic feature fusion mechanism adaptively adjusts the contributions of the MambaPlus and Transformer branches through learnable weight parameters, improving performance by approximately 5.98% compared to the average fusion strategy. However, using the Transformer branch alone results in a sharp performance drop of 46.75%, while using the MambaPlus branch alone only results in a 3.71% drop. This fully reveals the structural advantages of the state-space model in DNA sequence modeling, and its selective scanning mechanism is more suitable for capturing long-range dependencies in biological sequences. The biophysical feature module integrates physicochemical properties such as codon fitness index and free energy; its removal leads to a 10.48% performance decrease, indicating the complementary value of multimodal feature fusion for predicting the relative intensity of ribosome stagnation. These experimental data, from a computational biology perspective, validate the technological innovation of the proposed parallel architecture of MambaPlus and Transformer for predicting the relative intensity of ribosome stagnation in biological sequence analysis tasks.

[0078] In summary, the method of this invention effectively solves the key bottlenecks in existing methods for the relative intensity of ribosome stagnation: it overcomes the limitation of single-scale feature extraction in insufficient identification of complex DNA regulatory elements by using a multi-scale convolutional neural network; it overcomes the defect of traditional unidirectional modeling in failing to reflect the directional regulatory characteristics of DNA sequences by using a bidirectional state-space model; it avoids the information dilution problem caused by the simple splicing of DNA sequence features and biophysical features through a dynamic weighted fusion mechanism; and it balances the contradiction between computational efficiency and accuracy in long-range dependency modeling with a collaborative architecture of MambaPlus and Transformer.

[0079] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for predicting the relative intensity of ribosome stagnation using a parallel architecture of MambaPlus and Transformer, characterized in that, The method is specifically designed to predict the relative strength of ribosome stagnation based on DNA sequence and biophysical characteristics, and includes the following steps: S1. A serial composite encoding method combining K-mer sliding window with One-Hot encoding based on the codon scale is used to convert the DNA sequence into a high-dimensional numerical feature representation. S2. Input the feature matrix described in S1 into a parallel multi-scale convolutional neural network, and extract local sequence patterns of different sizes in the DNA sequence simultaneously through parallel convolutional branches; S3. Convert the feature tensors output by the parallel convolutional layers. Input the bidirectional MambaPlus encoder branch and the Transformer encoder branch respectively to model the DNA sequence state space information and global context information respectively; S4. Adaptively and dynamically weightedly fuse the temporal features output by the MambaPlus branch with the global attention features output by the Transformer branch; S5. The obtained DNA deep features and biological features are fused in a multimodal manner and then fed into the fully connected layer to predict the relative intensity of ribosome stagnation.

2. The method according to claim 1, characterized in that: The composite encoding performs one-hot encoding on the DNA sequence information after K-mer encoding. K-mer encoding is used to capture local pattern information of the DNA sequence and capture the synergistic effect of adjacent bases. One-hot encoding is then used to encode each extracted short base sequence to make up for the lack of numerical feature representation in K-mer encoding, while retaining K-mer context information and providing numerical high-dimensional feature representation.

3. The method according to claim 1, characterized in that: The parallel multi-scale convolutional neural network uses parallel convolutional branches. Each convolutional kernel branch contains a convolutional layer, a batch normalization layer, an activation function layer, and a dropout layer. The features output by each branch are concatenated along the channel dimension to obtain a feature tensor that fuses multi-scale local information. ； in, Let be the feature tensor of the DNA sequence after compound encoding, with dimension . 'batch' refers to the batch size. , The length of consecutive bases used to extract local patterns of DNA sequences. The length of the DNA sequence; , This represents the features output by the parallel convolutional layer; Local information extracted for the i-th convolutional branch; The kernel size parameter represents the kernel size of a one-dimensional convolutional layer, and the sliding window length of the kernel over a DNA sequence of length degrees. This represents the kernel size of a 1D convolutional layer; This is the parameter name for the number of output channels of a one-dimensional convolutional layer, used to represent the channel dimension of the final output feature of this convolutional layer; This represents the number of output channels of a one-dimensional convolutional layer; It is the core parameter that controls the grouping method of convolution kernels; For input channels; To achieve batch normalization, the feature data within a single training batch is standardized to avoid activation value saturation. Let it be the activation function; It is a regularization technique that masks neurons to prevent the model from overfitting; It is a concatenation function used to concatenate the local feature information extracted from the three convolutional branches.

4. The method according to claim 1, characterized in that: The bidirectional MambaPlus encoder consists of forward MambaPlus blocks and backward MambaPlus blocks. Each MambaPlus block contains one-dimensional depthwise separable convolutions for local feature enhancement of the input features, a selective state-space model to capture long sequence dependencies, a gated fusion mechanism using SiLU activation and sigmoid-gated dynamic weighted fusion of convolutional features and SSM output, and residual connections and normalized layers for stabilizing training and gradient flow. The MambaPlus block is computed as follows: ； ； ； ； ； ； in, For feature tensors that fuse multi-scale local information; For one-dimensional depthwise separable convolution, the number of parameters is reduced by grouping convolutions; It is an activation function used to enhance the expressive power of convolutional features and gating vectors, and to adapt to the nonlinear response of local features of DNA sequences; For layer normalization; These are the local feature vectors after convolution enhancement; It is a selective state-space model used to model the temporal dynamics of DNA sequences; Output features for the state-space model; It is the sigmoid activation function; and These are the weights and biases of the gated linear layer, respectively. For gated vectors, the tensor dimension is increased through a linear expansion operation; The main feature dimension is The dimension; The dimensions are the expanded dimensions within the MambaPlus module; This indicates that the gate vector Repeat the copying process in the last dimension until the dimension matches. Consistency, allowing the gated vector to be consistent with and Perform element-wise operations; This is the expanded gate vector; This is a Hadamard product, implemented with element-wise weighting. The gated fusion feature vector is formed by fusing local convolutional features and long-range state features. This is the weight matrix for the output projection layer; This is the final output of the MambaPlus module. The bidirectional MambaPlus encoder simulates the reverse complementarity of DNA double strands and the biological function of bidirectional promoters. The forward MambaPlus block corresponds to the transcription process where the fifth carbon atom of the pentose sugar points to the third carbon atom of the pentose sugar (5'→3'), capturing the cis-regulatory logic from promoter to terminator. The backward MambaPlus block corresponds to the transcription process where the third carbon atom of the pentose sugar points to the fifth carbon atom of the pentose sugar (3'→5'), simulating the regulatory role of the reverse complementary strand and the antisense transcription mechanism. The discretization parameter Δ in the state-space model corresponds to the temporal scale differences in different stages of transcription and translation, including the temporal heterogeneity of promoter binding, mRNA stability, and ribosome binding. The Transformer module models global context relationships through a multi-head attention mechanism. The calculation process is as follows: ； ； ； ； ； in, For feature tensors that fuse multi-scale local information; To create a learnable positional encoding matrix, model the 5'→3' orientation of the DNA sequence, with dimension [missing information]. ,in DNA sequence length, For feature dimensions (and) (The number of channels is the same). Features after adding learnable location encoding information; , and These are the projection matrices for query, key, and value, used to map sequence features to different semantic subspaces. Each attention head learns to focus on different types of biological relationships in the nucleotide sequence. For the Query matrix, by through Obtained by projection; The key matrix is ​​formed by... through Obtained by projection; The Value matrix is ​​formed by... through Obtained by projection; It is the output after attention weighting; This is a scaling factor to prevent the dot product result from becoming too large. Gradient vanishing; For activation functions; Weighted output for attention; For layer normalization; This is the result of normalizing the feature dimensions of each sample after adding the attention output to the original input; It is a convolutional feedforward network; This is the final output of the Transformer encoder; Global feature vectors are generated using attention pooling: ； ； in, These are learnable parameters; This is the feature matrix after fusing features from MambaPlus and Transformer, with dimensions of [dimensional value missing]. ; For activation functions; This is the attention weight matrix; The length of the DNA sequence; The attention weights are the values ​​corresponding to the i-th base position in the DNA sequence. The model learns these weights through training. Sequence segments that play a key role in the relative intensity of ribosome stagnation through autofocus; The fusion feature at the i-th base position is a matrix. The i-th row; The global feature vector generated by attention pooling; The Transformer encoder models the global spatial conformation of the DNA sequence and the long-range interactions between different regions through a multi-head attention mechanism, simulating the influence of the 5' UTR and coding region structure of messenger ribonucleic acid (mRNA) on ribosome scanning efficiency; different attention heads focus on different levels of biological characteristics such as codon usage frequency, regulatory protein binding site recognition, and mRNA secondary structure stability.

5. The method according to claim 1, characterized in that: The MambaPlus encoder performs forward processing (inputting the feature tensor into a forward MambaPlus block for processing) and backward processing (reversing the feature tensor along the sequence dimension and inputting it into a backward MambaPlus block for processing, then reversing the processing result again). The outputs of the forward and backward processing steps are then added together to obtain the bidirectional fused features. ； ； ； in, For feature tensors that fuse multi-scale local information; For the MambaPlus module in the model; This indicates a reversal along the DNA sequence dimension, simulating the biological significance of inverted complementary sequences; This represents the forward processing output of DNA sequence features on the MambaPlus branch, used to simulate biological processes in the 5'→3' direction; This represents the reverse processing output of DNA sequence features on the Mamba branch, used to simulate the regulatory mechanism in the 3'→5' direction; This indicates the addition and fusion of bidirectional outputs, where the feature vectors at corresponding positions are added element by element to simulate the synergistic regulatory effect of DNA double strands. The bidirectional MambaPlus encoder also includes a feedforward network module: ； ； in, Used for fusing bidirectional information Perform linear projection processing; The activation function for the Gaussian error linear unit; For layer normalization operation; The feature vector output by the feedforward network represents the advanced modulation features after nonlinear enhancement. These are complete sequence features that have undergone bidirectional processing and feature enhancement.

6. The method according to claim 1, characterized in that: The state space information extracted from the MambaPlus branch and the global context information extracted from the Transformer branch are fused using an adaptive dynamic weighting method: ； ； in, and The weights are learnable parameters (all initialized to 0.5). The weights are mapped to the [0,1] interval using sigmoid and normalized to ensure that the sum of the weights is 1. and These represent the state space information and global context information extracted from the MambaPlus branch and the Transformer branch, respectively. The weights correspond to the weights of the time-series dynamic features of the MambaPlus branches; The weights correspond to the weights of the global interaction features in the Transformer branch; The adaptive dynamic weighted fusion strategy simulates the synergistic effects of different regulatory levels in DNA sequences. The weights correspond to the importance of the kinetic processes of ribosome transcription rate and scanning efficiency; The weights correspond to the importance of steady-state properties of promoter strength and codon preference; the attention pooling mechanism automatically identifies regulatory hotspots in the DNA sequence that have a decisive influence on the relative strength of ribosome arrest, including core promoters, translation start sites, and rare codon clusters.

7. The method according to claim 1, characterized in that: Sequence depth features are concatenated with seven biophysical features, and regression prediction is achieved through a fully connected network. ； ； in, Main feature dimension; Biophysical characteristics dimension; The dimension extracted from the fused features by attention pooling is The global sequence feature vector represents the integrated functional signal of important regulatory elements in the DNA sequence; The standardized biophysical feature vector is obtained by Z-score standardization, which makes the features follow a normal distribution with a mean of 0 and a variance of 1. This is a multimodal fusion feature vector.

Images