Quantum and region-aware protein methylation site prediction method

Abstract

This invention provides a method for predicting protein methylation sites that integrates quantum and region-aware technologies, comprising the following steps: Step 1, obtaining protein sequences as data sources and constructing training and independent test sets respectively; Step 2, constructing multimodal features for each protein sequence using a three-way nested scattering network, fusing the multimodal features to obtain an optimized fused feature tensor; Step 3, inputting the optimized fused feature tensor into a RaQMeNet network model to perform methylation site prediction. This invention not only significantly outperforms existing technologies in terms of performance indicators but also possesses stronger adaptability, stability, and interpretability. It can be widely applied in multiple bioinformatics and biomedical fields such as protein functional annotation, disease mechanism research, and drug target discovery, demonstrating promising application prospects and commercial value.

Description

Technical Field

[0001] This invention belongs to the field of computer bioinformatics technology, and in particular relates to a method for predicting protein methylation sites that integrates quantum and region-aware technologies. Background Technology

[0002] Protein methylation is a key post-translational modification (PTM) that mainly occurs on residues such as lysine, arginine, and histidine, and is involved in regulating gene expression, protein function, and cell signaling pathways. Traditional experimental methods (such as mass spectrometry and antibody enrichment) have high accuracy in identifying sites, but they generally suffer from problems such as complex experimental procedures, low throughput, and high cost, making it difficult to meet the needs of large-scale proteomics analysis.

[0003] To improve the efficiency of protein methylation site prediction, researchers have proposed various computational methods. Early prediction models were mainly based on traditional machine learning methods such as Support Vector Machines (SVM) and Random Forests (RF). For example, the MethK predictor combines amino acid composition and physicochemical properties as input features to train an SVM classifier for prediction. However, these methods heavily rely on manually constructed features, making it difficult to effectively capture the complex nonlinear relationships and long-range dependencies in protein sequences, and their generalization ability is limited.

[0004] In recent years, with the rise of deep learning, models based on structures such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely used in the prediction of protein modification sites. For example, DeepRMethylSite uses deep CNNs to extract local features for predicting arginine methylation sites; PRMxAI combines interpretable AI methods and amino acid spatial distribution modeling to improve the interpretability and accuracy of the model. However, these models still have the following technical bottlenecks:

[0005] 1. Limited features and insufficient modality fusion: Most existing models only use basic encoding of amino acid sequences (such as One-Hot, PSSM, etc.), failing to fully utilize the contextual semantic features (such as pre-trained language model embeddings) and structural information in protein sequences, resulting in limited model expressive power.

[0006] 2. Insufficient structural design and lack of long-distance dependency modeling ability: Models such as DeepPRMS mostly use shallow convolutional structures, which make it difficult to model long-distance interactions and contextual dependencies in protein sequences, affecting prediction accuracy.

[0007] 3. Difficulty in handling class imbalance: Methylation sites account for a very small percentage of proteins, resulting in a severe imbalance between positive and negative samples. Conventional cross-entropy loss functions show a significant bias towards the majority class (non-methylated sites) during training, making it difficult for the model to accurately identify the minority class.

[0008] Furthermore, different deep learning modules often lack effective collaboration. For example, convolution, attention mechanisms, and graph neural networks operate independently, failing to achieve full integration and transmission of information, which further limits the performance of the model.

[0009] In summary, current methods for predicting protein methylation sites have certain limitations in terms of feature expression capabilities, structural design, and handling of class imbalances. Summary of the Invention

[0010] Purpose of the invention: The technical problem to be solved by the present invention is to address the shortcomings of the existing technology by providing a protein methylation site prediction method that integrates quantum and region awareness. By introducing pre-trained language model embedding, graph attention mechanism, multi-scale convolution structure and improved loss function, the invention aims to improve the model's comprehensive modeling ability of local patterns and global dependencies of protein sequences, and effectively improve the accuracy, stability and generalization ability of prediction.

[0011] The method of the present invention includes the following steps:

[0012] Step 1: Obtain protein sequences as data sources and construct training sets and independent test sets respectively;

[0013] Step 2: A three-way nested scattering network is used to construct multimodal features for each protein sequence. The multimodal features are then fused to obtain an optimized fused feature tensor.

[0014] Step 3: Input the optimized fusion feature tensor into the RaQMeNet network model to perform the methylation site prediction task.

[0015] Step 1 includes: using protein sequences from the UniProt Universal Protein Resource Knowledgebase (UniProtKB) and the Protein Lysine Methylation Database 3.0 (PLMD 3.0) as data sources, the collected protein sequences are processed for redundancy removal using the Cluster Database at High Identity with Tolerance (CD-HIT) tool and the cross-database redundancy comparison tool (CD-HIT-2D) to control the similarity between sequences to not exceed a threshold. Then, the protein sequences are uniformly pruned or padded to a length of 731 amino acid residues. The length of 731 is obtained by statistically analyzing the 95th percentile of the sequence lengths in the training and testing sets. This method can maximize the preservation of the original protein information while ensuring a balance between memory usage and computational efficiency during the training phase, thereby improving the stability and scalability of the model in processing ultra-long sequences.

[0016] Step 1 also includes: constructing a scattering feature space through a three-way scattering network consisting of sequence scattering, structure scattering, and coding scattering; and introducing a clustering method based on density peaks and shared nearest neighbors into the scattering feature space: first, the initial high-dimensional scattering feature tensor corresponding to each protein sequence is used as the local density ρ of the sample. i Similarity s with shared nearest neighbors ij This achieves automatic clustering without requiring manual specification of cluster numbers, and divides sequences into feature subgroups with similar expression distributions based on the similarity of scattering features. This is used for local fine-tuning and differential expression modeling in the region-aware modeling stage, resulting in a density peak clustering structure.

[0017]

[0018] Where, δ p ρ represents the minimum Euclidean distance between the p-th protein sequence and all other sequences q with a local density higher than its own; p ρ represents the local density of the p-th protein sequence in the scattering feature space; q The δq value represents the local density of the q-th protein sequence in the scattering feature space; dist(p,q) represents the Euclidean distance between the p-th and q-th protein sequences; δq is calculated for each sequence. pThe clustering method identifies sequences with high local density and distance from other high-density points as cluster centers (i.e., cluster foci), enabling automatic clustering without manual setting of cluster numbers. It further constructs sub-regions with high feature expression consistency, serving as the basis for parameter fine-tuning and differential expression learning in subsequent region-aware models. Through clustering methods based on density peaks and shared nearest neighbors, two or more sub-clusters are obtained, ultimately resulting in a set R of protein sequence clusters.

[0019] Step 1 further includes: based on the density peak clustering structure introduced in the scattering feature space, all deredundant protein sequences are allocated according to the balance and representativeness of each subgroup to construct training sets, validation sets, and independent test sets; wherein, the training set is used for parameter training of the RaQMeNet network model; the validation set is used for model parameter tuning and setting early stopping criteria; the independent test set does not participate in the training process, but is only used for final performance evaluation to ensure the generalization ability and stability of the prediction model. In this invention, the following sets were constructed: 5016 training sequences, 1254 validation sequences, and 2819 independent test sequences.

[0020] Step 2 includes:

[0021] Step 2-1, Constructing Sequence Scattering Features: The 731-amino acid residue length of the protein backbone obtained in Step 1 is used as the protein primary structure sequence (derived from the UniProtKB and PLMD 3.0 databases) and mapped to a physicochemical signal s(t). A sequence scattering network constructed using Morlet wavelets is used to generate multi-scale locally stable features:

[0022]

[0023] Where s(t) represents the physicochemical signal of the t-th residue, The center frequency is λ m Morlet wavelet kernel, S represents the low-pass filter function, |*| represents the complex modulus operation, and * represents the one-dimensional convolution operation; S0 represents the 0th-order scattering coefficient, i.e., the input signal s(t); S m S represents the m-th order scattering coefficient; m+1 (t) represents the scattering coefficient S of the m-th order. m The value of the (m+1)th order scattering coefficient at time point t after wavelet transform, modulus taking, and low-pass filtering;

[0024] This network captures local variation patterns at different frequency scales layer by layer, enabling stable representation of sequences at different biophysical granularities.

[0025] Multiscale local stability features S0, S1, ..., S mTogether, they constitute the sequence scattering feature tensor. The complex modulus value of the multi-scale scattering feature tensor is integrated through a low-pass filter φ to obtain local homology information, which represents the consistency between the local fidelity and co-occurrence pattern of amino acid residues. This information reflects the synchronous changes in the physicochemical properties of the protein sequence within adjacent residue windows, and is used to enhance the model's ability to identify local conserved regions and functional motifs.

[0026] Step 2-2, Constructing structural scattering features: Based on the AlphaFold2 structure prediction software (released by DeepMind in 2021), the primary structure sequence of the protein is inferred to obtain three-dimensional structural coordinate information; the three-dimensional structural coordinate information is the prediction result of an existing publicly available model, which is directly called as existing technology in this invention and has high reproducibility;

[0027] A spatial contact map is constructed based on the Euclidean distance between Cα atoms of each amino acid residue in the three-dimensional structural coordinates. The spatial proximity of residues is determined to serve as the input map structure for generating structural scattering features. Simultaneously, a spectral domain scattering transformation is performed on the graph Laplacian matrix L to generate rotationally equivalent structural scattering features. The initial structural feature tensor U0 = X, where X represents the initial feature vector corresponding to each amino acid residue in the spatial contact map. This initial feature vector is composed of node degree, spatial adjacency information between Cα atoms, and node type (amino acid category). The generation process of the spectral domain scattering features is a layer-by-layer recursive process.

[0028]

[0029] Where e represents the natural constant; Indicates the kth spec The diffusion timescale of the spectral transform; k spec The value is a natural number; This represents the diffusion operation of the graph signal in the spectral domain; Indicates the kth spec The characteristic tensor obtained by +1st order spectral domain scattering transformation;

[0030] Characteristic tensors obtained from spectral domain scattering transformations of each order Together, they constitute the rotationally equivariant structural scattering characteristics;

[0031] This process constructs a multiscale graph representation with rotational isovariability, capturing local-global hierarchical topological patterns in protein structures;

[0032] Steps 2-3: Construct coding scattering features: Obtain the coding region DNA sequence corresponding to each protein sequence. The DNA sequence is derived from the cross-referenced CDS entry in the UniProtKB database.

[0033] The DNA sequence is broken down into continuous triplet nucleotide-coding fragments, which are triplet tokens. The triplet nucleotide-coding fragments are then input into the pre-trained language model CodonBERT (an open-source nucleic acid language model based on the BERT structure) to extract the contextual semantic representation of each triplet.

[0034] Then, using the same isomorphic nested scattering network (i.e., the same convolution-nonlinear transformation-pooling cascade framework) as the sequence scattering network described in step 2-1, the embedding tensor formed by the context semantic representation of each triplet is processed to generate a multi-scale stable semantic representation tensor.

[0035] The semantic representation tensor is defined as encoding scattering features, which is essentially the expression of contextual features related to translation regulation mechanisms in DNA sequences at different scales, constituting the translation regulation layer representation part in multimodal feature fusion;

[0036] Steps 2-4: The sequence scattering features, structure scattering features, and coded scattering features are concatenated to form the initial high-dimensional feature tensor. in It is a real number space; D represents the total number of feature channels corresponding to each residue, specifically the sum of the number of feature channels of each mode: 128 dimensions (wavelet-based sequence scattering network, 3-layer nested structure) + 128 dimensions (3rd order of map diffusion) + 256 dimensions (dimension reduction after CodonBERT nested scattering) = 512;

[0037] Introducing a quantum core alignment feature selection strategy, firstly through parameterizing the quantum circuit U(θ) PQC Map each sequence feature to a quantum state |ψ i >, Define the quantum kernel:

[0038]

[0039] Among them, F p and F q Let ψ represent the initial high-dimensional feature tensor of the p-th protein sequence and the initial high-dimensional feature tensor of the q-th protein sequence, respectively; p This indicates that the parameterized quantum circuit U(θ) is used to represent PQC ) F p The quantum state vector obtained by mapping; ψ q This indicates that the parameterized quantum circuit U(θ) is used to representPQC ) F q The quantum state vector obtained by mapping;

[0040] <ψ p ∣ψ q > represents ψ p and ψ q The inner product between them (i.e., the amplitude overlap of the superposition states); |·|2 represents the modulus square operation, used to measure the similarity between two quantum states;

[0041] It is a kernel function that reflects the global embedding similarity between sequences in the quantum feature space and is an important foundation for the subsequent quantum alignment objective function.

[0042] by The feature kernel is K, which is constructed by combining the label kernel K from the label vector Y. Y (Constructed using Gaussian radial basis function (RBF) kernels or linear kernels), which are used together to calculate the Hilbert–Schmidt Independence Criterion (HSIC), expressed as:

[0043]

[0044] Where J(θ) PQC ) represents the parameter θ PQC A correlation evaluation index between the feature subspace generated by the controlled quantum circuit and the label distribution; express and κ Y The degree of statistical dependence between them;

[0045] By maximizing J(θ) PQC The top 512 feature channels were selected (the top 512 feature channels with the highest discriminative power were selected from all channel ranking results; this dimension represents the optimal number of channels chosen through experimental optimization and cross-validation in this invention. In actual experiments, the 512-dimensional feature channels achieved a good balance between prediction performance and computational efficiency. Compared to the 256-dimensional and 1024-dimensional settings, it achieved the highest performance in metrics such as F1-score and AUC, and significantly reduced training memory consumption and computational latency. This channel number setting has a clear technical effect: while maintaining optimal prediction performance, it significantly reduces redundant feature interference and computational resource consumption). Finally, an optimized feature subset was constructed, and each protein sequence obtained an optimized fusion feature tensor with a dimension of 731×512.

[0046] Step 3 includes: The RaQMeNet network model first models local features of the protein sequence using a quantum region-aware multi-scale convolution module. This module employs three sets of one-dimensional convolutional kernels with receptive fields of 3, 5, and 7 to encode the optimized fusion feature tensor F′ at multiple scales. Each set of convolutional channels is used to extract residue co-occurrence patterns within different length ranges. Specifically, the one-dimensional convolutional kernels with receptive fields of 3, 5, and 7 are used to capture co-occurrence feature patterns within windows of 3, 5, and 7 consecutive residues, respectively. For example, a convolutional kernel with a receptive field of 3 can identify short-range interaction signals, such as the conserved tripeptide motif [methionine–glutamic acid–lysine] (MEK); a convolutional kernel with a receptive field of 7 can model mid-range dependencies across regions, such as upstream and downstream regulatory structures of enzyme active regions or modification sites. Through multi-scale encoding, cooperative patterns of local residues can be extracted at different sequence granularities, thereby enhancing the model's ability to identify functional motifs and conserved domains.

[0047] For the cluster subgroups obtained in step 1, the local sequence regions in each protein sequence that correspond to the features of the cluster subgroup to which the sequence belongs are identified; each cluster subgroup represents a set of samples that are similarly distributed and have consistent feature expressions in the scattering feature space, and the corresponding local sequence regions are the regional context ranges that the model needs to model separately, serving as the target regions for subsequent regional fine-tuning and quantum gating regulation;

[0048] For the local sequence regions corresponding to cluster subgroups, a region-adjustable parameter fine-tuning mechanism is introduced. For any cluster subgroup r∈R, the k-th... conv The actual weights of the group convolution kernel on the local sequence region r Shared main core With region-specific fine-tuning items Composed of:

[0049]

[0050] Construct a quantum-inspired gating network (QGN) to generate scale-wise gating vectors based on the embedding representation embed(r) of the local sequence region corresponding to each cluster subgroup r. In this context, the local sequence region refers to the clustered subgroups automatically partitioned in step 1 using a clustering method based on density peaks and shared nearest neighbors. The fragment intervals located within these clustered subgroups in the protein sequence serve as the basic units for quantum regulation and parameter fine-tuning in RaQMeNet. Each region embedding (embed(r)) is generated by feature average pooling of all residues within the local sequence region corresponding to each clustered subgroup, reflecting the comprehensive expression mode of the local sequence region corresponding to each clustered subgroup under multi-scale representation. The gating function used in the quantum-gated network adopts the form of an activation function simulating the probability amplitude of quantum superposition measurement.

[0051] σ q (z)=sin 2 (θ QGN z),

[0052] Where, σ q (z) represents the nonlinear adjustment result of the quantum gating function on the input tensor z; z is the response intensity of the convolutional feature channel in the local sequence region r; θ QGN z is a trainable scaling factor used to control the magnitude of the nonlinear transformation; sin 2 The · symbol represents the squared form of the probability amplitude during simulated qubit measurement, derived from the probabilistic calculation mechanism of state projection in quantum measurement. This activation function can endow the gated output with higher nonlinear expressive power, enhance the model's response sensitivity to weak local features, and is particularly suitable for low-signal / high-sparseness biometric modeling tasks.

[0053] The contribution strength of each convolutional channel to the final output is controlled in a non-linear manner, and the convolutional output is multiplied element-wise with the gate vector:

[0054]

[0055] in, Indicates the kth conv Convolutional output features at scale and local sequence region r; This indicates that the k-th element is applied to the input tensor X. conv The output tensor after forming a one-dimensional convolutional kernel, and the input tensor X is the optimized fusion feature tensor F′; With Same dimension; the symbol ⊙ represents the Hadamard product, which is the element-wise multiplication of two tensors of the same dimension; this structure can realize personalized adjustment of feature channels in each region and at each scale, enhancing the model's ability to adapt to region-specific expression patterns.

[0056] In the convolution output fusion stage, the quantum region-aware multi-scale convolution module concatenates the outputs of each scale channel and maps them back to a unified channel dimension through linear projection. Simultaneously, it performs residual connections with the input sequence features to form the final multi-scale region-aware representation.

[0057]

[0058] in (For example: Z in the formula) (3,r) ) represents the receptive field as k conv The output tensor of the convolutional channels over the local sequence region r; the symbol || denotes the tensor concatenation operation;

[0059] Subsequently, the number of channels of the concatenated tensor is mapped back to the same dimension as the optimized fusion feature tensor F′ using the linear mapping operator Proj(·), and then element-wise addition (i.e., residual connection) is performed with the original input tensor X to form the final multi-scale representation result on the local sequence region r. This structure helps to integrate local features at different scales while preserving the backbone information of the input features, thereby improving feature stability and expression depth.

[0060] In the modeling stage based on protein 3D structure information, a quantum projection graph attention network module is established. This module is based on the protein structure adjacency graph predicted by AlphaFold2, constructing a structure dependency graph G = (V, E) between residues, where V is the set of nodes, the i-th node vi ∈ V corresponds to the i-th amino acid residue in the protein sequence, and E is the set of edges, with each edge e... ij ∈E indicates that there is a spatial contact relationship between the i-th residue and the j-th residue; when the Euclidean distance between the Cα atoms of the two residues is less than a set threshold (e.g. When ), a connecting edge is defined between the Cα atoms of two residues. This graph structure is used to model the spatial dependencies between residues in the three-dimensional folded structure of proteins and is the input basis for the quantum graph attention mechanism; for the i-th node feature h i Through the variable quantum coding circuit U φ Mapped to quantum states:

[0061] |Ψ i >=U φ (h i )|0>,

[0062] Among them |Ψ i > indicates that the feature h of the i-th node i After passing through the variable quantum coding circuit U φ The quantum state vector obtained after applying the initial state |0>;

[0063] The attention weights of any pair of adjacent nodes (i,j) are determined by the quantum state overlap:

[0064]

[0065] Where, α ij Ψ represents the quantum attention weight between the i-th residue node and the j-th residue node; i and Ψ j Don't use node v i and v j After the variable quantum coding circuit The quantum state vector obtained by mapping; N(i) represents the i-th node v i The set of all adjacent nodes; ∈ is a numerical stability constant used to avoid zero denominator. This weight is used to measure the similarity of adjacent nodes in the quantum representation space and is the core scoring index in the quantum graph attention mechanism;

[0066] Where the denominator For all adjacent nodes in |ψ i The sum of the projected intensities in space;

[0067] Ultimately, the output of each node is represented as an aggregation of features from its neighboring nodes under the quantum attention weights:

[0068]

[0069] Where, h′ i Represents the i-th node v i The updated feature representation; h j Indicates adjacent node v j The original feature vector; Describes a parameter θ that can be learned. map The control feature mapping function consists of a linear transformation and an activation function. This update method enables nonlinear integration of neighbor information in the structure space, enhancing the node's ability to represent the local and global topological environment.

[0070] Multi-head quantum attention mechanisms utilize two or more quantum encoders Uφ (h) To achieve parallel representation, the outputs of all heads are concatenated into a graph attention feature representation H′.

[0071] Subsequently, the output features of the quantum region-aware multi-scale convolutional module and the quantum projection graph attention network module are concatenated and fed into a bidirectional Transformer module with a region-aware mechanism to capture long-range dependencies in the sequence. The bidirectional Transformer module employs a strategy combining parameter sharing and region-specific fine-tuning to construct a region-aware feedforward network. For samples divided by cluster subgroup r, the feedforward network (FFN) in its Transformer sublayer is defined under the region-aware mechanism as follows:

[0072]

[0073] Among them, FFN (r) (x) represents the region-aware feedforward output corresponding to the cluster subgroup r; x is the input feature tensor; W1 (s) and These are the weight matrices corresponding to two different linear layers in the feedforward network (FFN), and are shared by all regions; W1 (s) The weight matrix is ​​shared by the first layer of the feedforward network FFN. ΔW1 is the weight matrix shared by the second layer of the feedforward network FFN. (r) , For the local sequence region r, the weight fine-tuning term is used. For shared bias terms, The term represents the region-specific bias fine-tuning term; ReLU is a common linear rectified activation function. This structure combines region sharing with region-specific modeling, enabling the model to simultaneously learn global feature representations and model local differences.

[0074] The output of the bidirectional Transformer module is then passed through a multi-head self-attention layer to achieve non-linear interaction modeling between features;

[0075] The Quantum Region-Aware Multi-Scale Convolutional Module (QMSC), the Quantum Projection Map Attention Network Module (QP-GAT), and the Bidirectional Transformer Module are connected in tensor structure via standard operations. Specifically, this includes the region features output by the Quantum Region-Aware Multi-Scale Convolutional Module. The graph attention feature representation H′ output by the quantum projection graph attention network module is concatenated with the multimodal feature tensor in the channel dimension. The concatenated multimodal feature tensor is used as the input tensor of the bidirectional Transformer module, keeping the original sequence length (731) unchanged, so as to ensure that the final model output is still the predicted representation of each amino acid residue site.

[0076] In step 3, a classifier module is connected as the output layer in the output stage of the RaQMeNet network model to predict the methylation probability of each amino acid residue position.

[0077] The classifier module consists of a three-layer feed-forward neural network, which uses the sigmoid function as the output activation function. It maps the output features of the bidirectional Transformer module into a residue probability vector of length 731, where each value represents the predicted probability that the corresponding residue is a methylation modification site.

[0078] During the training of the RaQMeNet network model, an improved Region-aware Weighted Binary Cross-Entropy (RW-BCE) loss function is used. Introduce independent weighting parameters w for positive and negative samples in different cluster subgroups r. r + With w r - To address the issues of uneven sample distribution and imbalanced positive-negative class ratios in different regions of the training set, the calculation formula is as follows:

[0079]

[0080] Where R represents the set of all cluster subgroups; y i p is the true label of the i-th residue site in the cluster subgroup r. i w is the predicted probability of the i-th residue site in the cluster subgroup r; r + w r - These are the weights of the positive and negative classes within region r, respectively, dynamically calculated based on the ratio of positive to negative samples in region r of the training data. This mechanism can automatically adjust the penalty intensity according to the difficulty of samples within the region, improving training stability and prediction accuracy, especially when dealing with extremely imbalanced or structurally complex regions;

[0081] A dynamic adjustment mechanism based on the difficulty of samples within clustering regions is introduced to automatically adjust the class weights: The difficulty η of samples within region r is defined. r The average prediction error of all training samples in region r (i.e., |y|) i -p i |Average):

[0082]

[0083] Subsequently, based on η rDynamically update the positive class weight w r + and negative class weight w r - This mechanism assigns higher weights to difficult-to-learn regions and moderately relaxes them to achieve adaptive adjustment of class weights. This mechanism works in conjunction with the RW-BCE loss function, automatically updating η at each training epoch. r The corresponding weights enable continuous perception and dynamic response to regional difficulty during training, which is one of the key strategies of this invention to improve the robustness of the model.

[0084] Step 3 also includes: introducing a masking mechanism to avoid the impact of padding sites on the loss function; selecting Adam as the optimizer, setting the initial learning rate to 1e-3, the number of training epochs to 100, and the batch size to 32; re-executing the quantum feature subset selection at regular intervals during training to adaptively track the dynamic changes of features and the model, and using five-fold cross-validation to select the optimal model parameters. During the RaQMeNet model training process, this invention sets up an automatic re-execution of the quantum kernel alignment feature selection process every T = 10 training epochs to dynamically update and optimize the channel selection results of the fused feature tensor F′. This strategy can adapt to changes in the model's representational ability at different training stages, avoiding the limitation of the initially selected feature channels on the later generalization ability. The quantum kernel is recalculated in each re-selection. With tag kernel K Y Maximize the Hilbert–Schmidt independence criterion J(θ) PQC The optimal 512 feature channels for the current stage are selected for training in the next cycle. This adaptive subset update mechanism has a clear cycle control strategy, manageable computational overhead, and significantly improves the model's robustness and generalization performance on the test set in experiments.

[0085] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0086] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0087] Beneficial effects: This invention not only significantly outperforms existing technologies in terms of performance indicators, but also has stronger adaptability, stability and interpretability. It can be widely applied in multiple bioinformatics and biomedical fields such as protein functional annotation, disease mechanism research, and drug target discovery, and has good application prospects and commercial value. Attached Figure Description

[0088] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0089] Figure 1 This is a partial module diagram of QRA-MSC.

[0090] Figure 2 This is a partial module diagram of QP-GAT.

[0091] Figure 3 This is a local module diagram of the region-aware bidirectional Transformer.

[0092] Figure 4 This is a diagram of the overall structure of RaQMeNet. Detailed Implementation

[0093] This invention provides a method for predicting protein methylation sites that integrates quantum and region-aware technologies. First, protein sequences from the UniProtKB and PLMD 3.0 databases are used as data sources to construct training and independent test sets, respectively. The collected protein sequences are deredundant using CD-HIT and CD-HIT-2D tools, controlling the similarity between sequences to no more than 40%. Subsequently, the sequences are uniformly pruned or padded to a length of 731 amino acid residues. Based on this, to further improve the rationality of data partitioning, this method introduces a clustering method based on density peaks and shared nearest neighbors in the scattering feature space.

[0094] First, calculate the local density ρ of the sample. i Similarity s with shared nearest neighbors ij Then, through the density-distance index:

[0095]

[0096] Multiple high-density cluster centers were identified to achieve automatic cluster number partitioning and feature-consistent subgroup construction, thereby constructing more balanced and representative training, validation, and independent test sets. The final result was a training set of 5,016 records, a validation set of 1,254 records, and an independent test set of 2,819 records.

[0097] In the feature extraction stage, this method uses a three-way nested scattering network to construct multimodal features for each protein sequence, namely: (1) Sequence scattering features, which map the main chain amino acid sequence to a physicochemical signal s(t), and use a scattering network constructed with Morlet wavelets to generate multi-scale local stable features:

[0098]

[0099] Output local coherence information;

[0100] (2) Structural scattering features: A contact map is constructed based on the three-dimensional coordinates predicted by AlphaFold2, and a spectral domain scattering transformation is performed on the graph Laplacian matrix L to generate rotationally equivalent structural features:

[0101]

[0102] (3) Encoding scattering features: The DNA sequence of the protein-coding region is decomposed into triplet tokens, which are then input into the CodonBERT model to obtain contextual embeddings. The isomorphic scattering framework is then used to extract translation regulatory layer features. The above three types of features are concatenated to form an initial high-dimensional feature tensor. To further compress the feature dimension and improve the representation quality, this method introduces a quantum kernel-aligned feature selection strategy. This strategy first maps each sequence feature to a quantum state |ψ| using a parameterized quantum circuit U(θ). i >, Define the quantum kernel:

[0103] κ θ (F i ,F j )=|<ψ i |ψ j >| 2 ,

[0104] And calculate the Hilbert–Schmidt independence criterion (HSIC) between it and the tag kernel:

[0105]

[0106] By maximizing The first 512 feature channels were selected to construct an optimized feature subset with strong information expression and high discriminative power. Finally, each protein sequence obtained an optimized fusion feature tensor with a dimension of 731×512, which was fed into the downstream prediction model for methylation site prediction.

[0107] The fused features are input into the RaQMeNet network model proposed in this invention. The RaQMeNet network model first models the local features of the protein sequence using the Quantum Region-Aware Multi-Scale Convolution (QRA-MSC) module designed in this invention. This module systematically reconstructs the traditional multi-scale convolutional network structure (MSCANet), innovatively integrating a region-aware fine-tuning mechanism with a quantum-heuristic gating adjustment method, achieving collaborative modeling of the local motivation, multi-scale contextual relationships, and regional heterogeneity patterns of the protein sequence. Specifically, the QRA-MSC module uses three sets of one-dimensional convolutional kernels with receptive fields of 3, 5, and 7 to encode the input sequence at multiple scales. Each set of convolutional channels is used to extract residue co-occurrence patterns within different length ranges. Unlike the uniform parameter setting of convolutional kernels in traditional networks, this invention introduces a region-adjustable parameter fine-tuning mechanism for the sequence regions defined by the clustered subgroups. For the k-th convolutional kernel, its actual weight in region r is shared by the master kernel. With region-specific fine-tuning items Combining, that is:

[0108]

[0109] This "shared + differential" convolutional architecture enables the model to uniformly extract features from the backbone sequence while adapting to subtle differences in convolutional representations across different regions. To further enhance the information modulation capability between convolutional channels, this invention designs a feature modulation method based on a quantum-inspired gating mechanism. This invention constructs a Quantum-inspired Gating Network (QGN) to generate scale-wise gating vectors based on the embedding representation embed(r) of each region. The gating function adopts the activation function form of the simulated quantum superposition measurement probability amplitude, such as σ. q (z)=sin 2 (θz) controls the contribution of each convolutional channel to the final output in a non-linear manner. The convolutional output is then element-wise multiplied with the gate vector:

[0110]

[0111] Where ⊙ represents the Hadamard product. By guiding the expression preferences of different regions at different scales through a quantum gating mechanism, the model can dynamically reconstruct local information channels in the feature space, thereby improving its ability to model regional feature changes, weak signal patterns, and non-stationary distributions. In the convolutional output fusion stage, the QRA-MSC module concatenates the outputs of channels at each scale and maps them back to a unified channel dimension through linear projection. Simultaneously, it performs residual connections with the input sequence features to form the final multi-scale region-aware representation.

[0112]

[0113] This design not only integrates information from different receptive field convolutional channels, but also preserves the principal components of the input, thereby improving the stability and discriminativeness of feature representation.

[0114] To further explore the spatial structural relationships between protein residues, this invention innovatively proposes a Quantum-Projected Graph Attention Network (QP-GAT) module in the structural modeling stage. This module is based on the protein structure adjacency graph predicted by AlphaFold2, constructing a structural dependency graph G = (V, E) between residues, where each node represents an amino acid residue, and the edge set is determined by the spatial contact relationships in the three-dimensional structure. Unlike traditional graph attention networks (GAT), the QP-GAT module constructs quantum state representations and utilizes their orthogonal projection relationships to calculate attention weights, thereby making the propagation of structural information between residues more physically plausible and possessing greater nonlinear expressive power. Specifically, for each node feature h... i The variable quantum coding circuit Uφ is used to map the quantum state:

[0115] |Ψ i >=U φ (h i )|0>,

[0116] The attention weights of any pair of adjacent nodes (i,j) are determined by their quantum state overlap:

[0117]

[0118] The denominator is the sum of all adjacent nodes in |ψ i The sum of projection intensities in the quantum space, ∈, is a numerically stable term. This attention mechanism, based on the interference amplitude between quantum states, can identify higher-order dependencies and structural commonalities among residues in the folded space. Ultimately, the output of each node is represented as an aggregation of the features of its neighboring nodes under this quantum attention weight:

[0119]

[0120] in This represents a combination of a linear mapping and an activation function. The multi-head quantum attention mechanism utilizes multiple quantum encoders Uφ. (h) Parallel expression is achieved, and the outputs of all heads are concatenated into a graph attention feature representation H′. This module, through a quantum physics-driven attention mechanism, better models the structural semantic propagation relationships between protein residues while maintaining topological consistency, providing a feature foundation rich in structural cognition for subsequent BiTransformer layers.

[0121] Subsequently, the output features of QRA-MSC and QP-GAT are concatenated and fed into a bidirectional Transformer module with a region-aware mechanism to capture long-range dependencies in the sequence. This Transformer module employs a combination of parameter sharing and region-specific fine-tuning to effectively balance the model's generalization performance with regional feature differences. This module uses a strategy combining parameter sharing and region-specific fine-tuning to construct a region-aware feedforward network. Specifically, for samples partitioned by cluster subgroup r, the feedforward part of its Transformer sublayer is defined as follows:

[0122]

[0123] Among them, FFN (r) (x) represents the region-aware feedforward output corresponding to the cluster subgroup r; x is the input feature tensor; W1 (s) and These are the weight matrices corresponding to two different linear layers in the feedforward network (FFN), and are shared by all regions; W1 (s) The weight matrix is ​​shared by the first layer of the feedforward network FFN. ΔW1 is the weight matrix shared by the second layer of the feedforward network FFN. (r) , For the local sequence region r, the weight fine-tuning term is used. For shared bias terms, The term represents the region-specific bias fine-tuning term; ReLU is a common linear rectified activation function. This structure combines region sharing with region-specific modeling, enabling the model to simultaneously learn global feature representations and model local differences.

[0124] The output of the Transformer module is then passed through a multi-head self-attention layer to achieve more complex nonlinear interaction modeling between features.

[0125] The modules described above are connected using standard tensors to maintain the sequence dimension and ensure that the final output corresponds to the prediction result for each amino acid residue position. The classifier is a three-layer feedforward neural network with a sigmoid activation function, outputting the probability value of each site being a methylation site. During model training, an improved region-aware weighted binary cross-entropy loss function is used, and a dynamic adjustment mechanism based on the difficulty of samples in clustered regions is introduced to automatically adjust class weights to alleviate the sample imbalance problem. Simultaneously, a masking mechanism is introduced to avoid the influence of padding sites on the loss function. The optimizer is Adam, with an initial learning rate of 1e-3, 100 training epochs, and a batch size of 32. During training, quantum feature subset selection is re-executed at regular intervals to adaptively track the dynamic changes of features and the model, and five-fold cross-validation is used to select the optimal model parameters.

[0126] To verify the effectiveness of the model of this invention, the performance of RaQMeNet was compared with that of existing mainstream methods (including DeepPRMS, MeSEP, PRMxAI, etc.) on the same dataset in the embodiments. Experimental results show that on the UniProtKB test set, the RaQMeNet model achieves an F1 score of 93.607%, an MCC value of 93.549%, and an AUC of 98.916%, all of which are significantly better than existing methods (such as DeepPRMS, whose F1 score is only 84.32% and MCC is 78.35%). On the PLMD3.0 test set, the F1 score is 90.022% and the AUC reaches 99.050%, maintaining excellent performance while demonstrating extremely strong generalization ability.

[0127] Furthermore, the hardware configuration used in the experimental operation of this invention included: an Intel(R) Core(TM) i9-14900K 3.20GHz CPU, an NVIDIA GeForce RTX 4090 24GB GPU, 64GB of RAM, a 1TB SSD hard drive, and Windows 10 Enterprise LTSC operating system. The CUDA version was 11.2.2, and the CUDNN version was 8.2.0. The development tools were PyCharm 2024.3.1.1 and the Python interpreter version was 3.8.20. The main third-party libraries relied upon included: TensorFlow 2.10.0, PyTorch 2.4.1, NumPy 1.24.3, SciPy 1.10.1, Scikit-learn 1.3.2, Bio 1.6.2, Pandas 2.0.3, and Matplotlib 3.7.3.

[0128] As can be seen from the above implementation methods, the RaQMeNet model proposed in this invention systematically improves upon several advanced deep learning modules in terms of structural design, including Quantum Region-Aware Multi-Scale Convolution (QRA-MSC), Quantum Projection Graph Attention Network (QP-GAT), and a bidirectional Transformer with a partition-aware mechanism, comprehensively modeling the local residue co-occurrence, multi-scale contextual dependence, and complex structural relationships of protein sequences. In terms of feature construction, it innovatively introduces a multimodal feature extraction framework based on a three-way scattering network, integrating sequence physicochemical signals, structural Laplacian spectral domain information, and semantic representation of the coding layer; simultaneously, it combines a feature selection strategy driven by quantum kernel alignment and the HSIC criterion, significantly improving the representation quality and discriminative ability of the high-dimensional feature space. In the model training strategy, a region-aware weighted binary cross-entropy loss function is further designed, coupled with a masking and dynamic feature refresh mechanism, to alleviate the problems of class imbalance and feature redundancy from the source. The overall model exhibits extremely high accuracy, stability, and cross-dataset generalization ability in the protein methylation site prediction task, making it suitable for accurate mining and high-throughput annotation of large-scale biological omics data.

[0129] In terms of data preparation, this invention uses protein sequences from the UniProtKB and PLMD 3.0 databases as a basis. The original sequences are filtered for redundancy using CD-HIT and CD-HIT-2D tools, strictly controlling the similarity between sequences to no more than 40%, and standardizing the sequence length to 731 amino acid residues. Short sequences are padded using padding. Furthermore, an automatic clustering method based on density peaks and shared nearest neighbor similarity is introduced to construct subgroup consistency regions in the scattering feature space, achieving a more reasonable, balanced, and representative division between the training set, validation set, and independent test set. The final result is 5,016 sequences in the training set, 1,254 sequences in the validation set, and 2,819 sequences in the independent test set.

[0130] In terms of feature construction, this invention innovatively designs a three-way nested scattering network to extract multi-source heterogeneous features: (1) After the main chain sequence is physically mapped, a Morlet wavelet scattering network is constructed to extract local scale-invariant features; (2) AlphaFold2 predicts the structure to generate a contact map, and performs structural scattering transformation based on the graph Laplacian spectral domain to extract rotationally equivalent high-dimensional structural features; (3) After the DNA coding region is embedded by CodonBERT, isomorphic scattering is performed to capture the contextual expression of the translation regulatory layer. The three types of features are spliced ​​together to form a high-dimensional tensor. Quantum states are constructed through quantum circuit mapping, the overlap between states is defined to form the kernel, and feature selection is performed using the HSIC objective function. Finally, the 512 feature channels with the strongest expressive power are retained to generate an optimized fusion feature tensor.

[0131] In terms of model architecture, RaQMeNet consists of three main structural units: (1) The QRA-MSC module uses three-scale convolution (receptive fields of 3, 5, and 7), combined with a region fine-tuning mechanism and a quantum-gated network, to guide the dynamic adjustment of convolution channels in different regions; (2) The QP-GAT module maps residue nodes to quantum states, and uses the orthogonal projection relationship between states to construct an attention distribution, thereby achieving structure-aware nonlinear aggregation expression; (3) The Region-Aware BiTransformer module is used for clustering region division to achieve adaptive modeling that combines parameter sharing and region feedforward fine-tuning, capturing long-distance dependencies across residues. After the outputs of convolution, graph modeling, and sequence modeling are spliced ​​together, the methylation probability of each residue is finally output through a three-layer feedforward neural network.

[0132] In the output prediction stage, Sigmoid activation is used to perform binary classification for each site. To effectively alleviate the class imbalance problem caused by the extreme scarcity (1.43%) of methylated samples, this invention designs a weighted BCE loss function based on the proportion of region subgroups, introduces a masking mechanism to filter the influence of padding sites, and periodically refreshes the feature subset during training to adaptively update the model representation. The training configuration uses the Adam optimizer with a learning rate of 1e-3, a batch size of 32, a total of 100 rounds, and five-fold cross-validation to select the optimal model.

[0133] The technical solution of this invention has the following outstanding innovations:

[0134] (1) A multimodal feature representation framework based on scattering mechanism was first proposed, which integrates three types of hierarchical information: physicochemical sequence, structural topology and genetic regulation, to improve the full coverage of representation;

[0135] (2) Introduce a kernel alignment method based on quantum state mapping to achieve highly expressive screening of feature subsets, taking into account both quantum information relevance and label dependence;

[0136] (3) Constructing region partitioning in the feature space based on shared nearest neighbor density to support the region perception mechanism of subsequent structure and loss function, and enhance the model’s adaptability to heterogeneous regions;

[0137] (4) Construct a multi-scale convolutional module that integrates regional fine-tuning and quantum gating mechanisms to dynamically model motivational patterns and regional changes in sequences;

[0138] (5) A quantum projection-driven graph attention mechanism is proposed to model the spatial semantic propagation path in folded structures in a physically interpretable manner;

[0139] (6) Enhance the training stability and generalization performance of the model in the context of extreme class imbalance by using dynamic quantum feature update and mask loss mechanism.

[0140] The RaQMeNet framework proposed in this invention achieves significant performance improvements over existing models on two independent test sets, UniProtKB and PLMD 3.0, with an F1-score of up to 93.607% and an AUC of 99.050%, verifying the effectiveness, scalability, and industrial transformation potential of this method in protein methylation site prediction.

[0141] The method of this invention demonstrates superior performance and broad applicability in the protein methylation site prediction task. Compared with existing methods such as DeepPRMS (F1-score: 84.32%), PRMxAI (F1-score: 88.84%), and MeSEP (F1-score: 81.00%), this invention achieves an F1-score of 93.607%, a precision of 96.153%, and a recall of 91.193% on the UniProtKB independent test set, with an MCC value of 93.549% and an AUC of 98.916%. It significantly outperforms existing models in all key performance indicators, with an average improvement of more than 10 percentage points. Its performance on the PLMD 3.0 independent dataset is also stable, with an F1-score of 90.022%, an MCC of 89.893%, and an AUC of 99.050%, demonstrating that this model has good generalization ability and robustness.

[0142] This invention provides a protein methylation site prediction method that integrates quantum and region-aware technologies. Without departing from the fundamental principles, various modifications and extensions of the related technologies can be made, all of which should be considered within the scope of this invention. Sub-modules not described in detail herein can be implemented using existing mature technologies.

Claims

1. A method for predicting protein methylation sites by fusing quantum and regional perception, characterized in that, Includes the following steps: Step 1: Obtain protein sequences as data sources and construct training sets and independent test sets respectively; Step 2: A three-way nested scattering network is used to construct multimodal features for each protein sequence, and the multimodal features are fused to obtain an optimized fused feature tensor; the three-way nested scattering network is used to construct sequence scattering features, structure scattering features, and coding scattering features. Step 3: Input the optimized fusion feature tensor into the RaQMeNet network model to perform the methylation site prediction task; Step 1 includes: using protein sequences from the UniProt Protein Resource Knowledge Base and the Protein Lysine Methylation Database version 3.0 as data sources, the collected protein sequences are processed for redundancy removal using a clustered fast homology redundancy filtering tool and a cross-database redundancy comparison tool, respectively, to control the similarity between sequences to not exceed the threshold, and then the protein sequences are uniformly pruned or padded to a length of 731 amino acid residues.

2. The method of claim 1, wherein, Step 1 also includes: constructing a scattering feature space through a three-way scattering network consisting of sequence scattering, structure scattering, and coding scattering; and introducing a clustering method based on density peaks and shared nearest neighbors into the scattering feature space: firstly, the initial high-dimensional scattering feature tensor corresponding to each protein sequence is used as the local density of the sample. Similarity with shared nearest neighbors This method enables automatic clustering and partitions sequences into feature subgroups with similar expression distributions based on the similarity of scattering features. This is used for local fine-tuning and differential expression modeling in the region-aware modeling stage, resulting in a density peak clustering structure. ， in, Let represent the minimum Euclidean distance between the p-th protein sequence and all other sequences q with a local density higher than itself; This represents the local density of the p-th protein sequence in the scattering feature space; This represents the local density of the q-th protein sequence in the scattering feature space; This represents the Euclidean distance between the p-th and q-th protein sequences; it is calculated by... The value is used to achieve automatic clustering and partitioning. By using a clustering method based on density peaks and shared nearest neighbors, two or more cluster subgroups are obtained, and finally a set R of protein sequence cluster subgroups is obtained.

3. The method of claim 2, wherein, Step 1 also includes: based on the density peak clustering structure introduced in the scattering feature space, all the deredundant protein sequences are allocated according to the balance and representativeness of each subgroup to construct training set, validation set and independent test set; wherein, the training set is used for parameter training of RaQMeNet network model; the validation set is used for model parameter tuning and setting early stopping criteria; the independent test set does not participate in the training process, but is only used for final performance evaluation.

4. The method according to claim 3, characterized in that, Step 2 includes the following steps: Step 2-1, Constructing Sequence Scattering Features: The 731-amino acid residue length of the protein backbone obtained in Step 1 is used as the primary structure sequence of the protein and mapped to physicochemical signals. ,use Wavelet-constructed sequence scattering networks generate multi-scale locally stable features: ， in This represents the physicochemical signal of the t-th residue. The center frequency is Morlet wavelet kernel, Denotes the low-pass filter function, | | indicates the modular arithmetic operation. This represents a one-dimensional convolution operation; This represents the 0th-order scattering coefficient, i.e., the input signal s(t); Represents the m-th order scattering coefficient; This indicates that by adjusting the m-th order scattering coefficient The value of the (m+1)th order scattering coefficient at time point t after wavelet transform, modulus taking, and low-pass filtering; Multiscale local stability features , , ..., Together, they constitute the sequence scattering feature tensor. The complex modulus values ​​of the multi-scale scattering feature tensor are then passed through a low-pass filter. Integrating the data yields local cohomology information representing the consistency between the local fidelity of amino acid residues and their co-occurrence patterns. Step 2-2, Constructing structural scattering features: Based on the AlphaFold2 structure prediction software, the primary structure sequence of the protein is inferred to obtain three-dimensional structural coordinate information; Based on the amino acid residues in the three-dimensional structural coordinates The Euclidean distance between atoms is used to construct a spatial contact map to determine whether residues have spatial proximity relationships, serving as the input map structure for generating structural scattering features. Simultaneously, a spectral domain scattering transformation is performed on the graph Laplacian matrix L to generate rotationally equivalent structural scattering features. Let the initial structural feature tensor U0 = X, where X represents the initial feature vector corresponding to each amino acid residue in the spatial contact map. This initial feature vector is composed of node degree, spatial adjacency information between Cα atoms, and node type. The generation process of the spectral domain scattering features is a layer-by-layer recursive process. ， Where e represents the natural constant; Indicates the first The diffusion timescale of the order spectrum transformation; The value is a natural number; This represents the diffusion operation of the graph signal in the spectral domain; Indicates the first The characteristic tensor obtained by spectral domain scattering transformation; characteristic tensors obtained by the spectral domain scattering transform of each order together constitute a rotationally isotropic structural scattering signature Steps 2-3: Construct coding scattering features: Obtain the coding region DNA sequence corresponding to each protein sequence, wherein the DNA sequence is derived from the cross-annotation field of the UniProtKB database; The DNA sequence is broken down into continuous triplet nucleotide-coding fragments, which are triplet tokens. The triplet nucleotide-coding fragments are then input into the pre-trained language model CodonBERT to extract the contextual semantic representation of each triplet. Then, using the same isomorphic nested scattering network structure as the sequence scattering network described in step 2-1, the embedding tensor formed by the context semantic representation of each triplet is processed to generate a multi-scale stable semantic representation tensor; Steps 2-4: The sequence scattering features, structure scattering features, and coded scattering features are concatenated to form the initial high-dimensional feature tensor. ,in It is a real number space; D represents the total number of feature channels corresponding to each residue; Introducing a quantum core alignment feature selection strategy, firstly through parameterized quantum circuits Map each sequence feature to a quantum state Define the quantum kernel: ， in, and Let represent the initial high-dimensional feature tensor of the p-th protein sequence and the initial high-dimensional feature tensor of the q-th protein sequence, respectively. This indicates the use of parameterized quantum circuits. Will The quantum state vector obtained by mapping; This indicates the use of parameterized quantum circuits. Will The quantum state vector obtained by mapping; express and The inner product between; This represents the modulo-square operation; is a kernel function; by The feature kernel is combined with the label kernel constructed from the label vector Y. Together, they are used to calculate the Hilbert–Schmidt independence criterion (HSIC), which is expressed as: ， in Indicates parameters A correlation evaluation index between the feature subspace generated by the controlled quantum circuit and the label distribution; express and The degree of statistical dependence between them; By maximizing The first 512 feature channels were selected, and an optimized feature subset was finally constructed. The dimension of each protein sequence was [dimension value missing]. Optimized fusion feature tensor .

5. The method according to claim 4, characterized in that, Step 3 includes: The RaQMeNet network model first models local features of the protein sequence through a quantum region-aware multi-scale convolutional module. The quantum region-aware multi-scale convolutional module uses three sets of one-dimensional convolutional kernels with receptive fields of 3, 5, and 7 to optimize and fuse the feature tensor. Multi-scale encoding is performed, with each set of convolutional channels used to extract residue co-occurrence patterns within different length ranges; one-dimensional convolutional kernels with receptive fields of 3, 5, and 7 are used to capture co-occurrence feature patterns in 3, 5, and 7 consecutive residue windows, respectively. For the cluster subgroups obtained in step 1, the local sequence regions in each protein sequence that correspond to the features of the cluster subgroup to which the sequence belongs are identified; each cluster subgroup represents a set of samples that are similarly distributed and have consistent feature expressions in the scattering feature space, and the corresponding local sequence regions are the regional context ranges that the model needs to model separately, serving as the target regions for subsequent regional fine-tuning and quantum gating regulation; For the local sequence regions corresponding to the cluster subgroups, a region-adjustable parameter fine-tuning mechanism is introduced. For any cluster subgroup... , No. The actual weights of the group convolution kernel on the local sequence region r Shared main core With region-specific fine-tuning items Composed of: ， Construct a quantum-gated network to generate scale-wise gate vectors based on the embedding representation embed(r) of the local sequence region corresponding to each cluster subgroup r. In this context, the local sequence region refers to the clustered subgroups automatically partitioned in step 1 using a clustering method based on density peaks and shared nearest neighbors. The fragment intervals located within these clustered subgroups in the protein sequence serve as the basic units for quantum regulation and parameter fine-tuning in RaQMeNet. Each region embedding (embed(r)) is generated by the feature average pooling of all residues within the local sequence region corresponding to each clustered subgroup, reflecting the comprehensive expression mode of the local sequence region corresponding to each clustered subgroup under multi-scale representation. The gating function used by the quantum-gated network adopts the form of an activation function simulating the probability amplitude of quantum superposition measurement. ， in, This represents the nonlinear adjustment result of the quantum gating function on the input tensor z; z is the response intensity of the convolutional feature channel in the local sequence region r; is a trainable scaling factor used to control the amplitude of the nonlinear transformation; sin²(·) represents the square form of the probability amplitude when simulating qubit measurement; The contribution strength of each convolutional channel to the final output is controlled in a non-linear manner, and the convolutional output is multiplied element-wise with the gate vector: ， in, Indicates the first Convolutional output features at scale and local sequence region r; This indicates that the first step is to apply the first step to the input tensor X. The output tensor after forming a one-dimensional convolution kernel, and the input tensor X is the optimized fused feature tensor. ; With Same dimension; the symbol ⊙ represents the Hadamard product; In the convolution output fusion stage, the quantum region-aware multi-scale convolution module concatenates the outputs of each scale channel and maps them back to a unified channel dimension through linear projection. Simultaneously, it performs residual connections with the input sequence features to form the final multi-scale region-aware representation. ， in Indicates the sensory field as The output tensor of the convolutional channels over the local sequence region r; symbol This indicates a tensor concatenation operation; Subsequently, using the linear mapping operator The number of channels in the concatenated tensor is mapped back to the same dimension as the optimized fusion feature tensor F′, and then element-wise addition is performed with the original input tensor X to form the final multi-scale representation result on the local sequence region r. ; In the modeling stage based on protein 3D structure information, a quantum projection graph attention network module is established. This module is based on the protein structure adjacency graph predicted by AlphaFold2, constructing a structure dependency graph G = (V, E) between residues, where V is the set of nodes, and the i-th node vᵢ∈V corresponds to the i-th amino acid residue in the protein sequence; E is the set of edges, and each edge... ∈E indicates that there is a spatial contact relationship between the i-th residue and the j-th residue; when the Euclidean distance between the Cα atoms of the two residues is less than a set threshold, a connecting edge is defined between the Cα atoms of the two residues; for the i-th node feature Through variable quantum coding circuit Mapped to quantum states: ， in Represents the features of the i-th node After the variable quantum coding circuit Acting on the initial state The resulting quantum state vector; The attention weights of any pair of adjacent nodes (i, j) are determined by the quantum state overlap: ， in, This represents the quantum attention weight between the i-th residue node and the j-th residue node; and Don't be a node and After the variable quantum coding circuit The quantum state vector obtained by mapping; N(i) represents the i-th node. The set of all adjacent nodes; It is the numerical stability constant; where the denominator is the sum of the projection intensities of all adjacent nodes in space. Ultimately, the output of each node is represented as an aggregation of features from its neighboring nodes under the quantum attention weights: ， in, Represents the i-th node The updated feature representation; Indicates adjacent nodes The original feature vector; Represents a set of learnable parameters The feature mapping function of control; Multi-head quantum attention mechanisms utilize two or more quantum encoders. To achieve parallel representation, the outputs of all heads are concatenated into a graph attention feature representation. ; Subsequently, the output features of the quantum region-aware multi-scale convolutional module and the quantum projection graph attention network module are concatenated and fed into a bidirectional Transformer module with a region-aware mechanism to capture long-range dependencies in the sequence. The bidirectional Transformer module employs a strategy combining parameter sharing and region-specific fine-tuning to construct a region-aware feedforward network. For samples divided by cluster subgroup r, the feedforward network in the Transformer sublayer is defined under the region-aware mechanism as follows: ， in, denoted by r, representing the region-aware feedforward output corresponding to the cluster subgroup r; x is the input feature tensor; and These are the weight matrices corresponding to two different linear layers in the feedforward network, and are shared by all regions. The weight matrix is ​​shared by the first layer of the feedforward network. The weight matrix is ​​shared by the second layer of the feedforward network FFN. For the local sequence region r, the weight fine-tuning term is used. For shared bias terms, For region-specific bias fine-tuning terms; ReLU is the linear rectified activation function; The output of the bidirectional Transformer module is then passed through a multi-head self-attention layer to achieve non-linear interaction modeling between features; The quantum region-aware multi-scale convolutional module, the quantum projection map attention network module, and the bidirectional Transformer module are connected in the tensor structure via standard operations. Specifically, this includes the region features output by the quantum region-aware multi-scale convolutional module. Graph attention feature representation output by the quantum projection graph attention network module The sequence is spliced ​​along the channel dimension, and the spliced ​​multimodal feature tensor is used as the input tensor of the bidirectional Transformer module, keeping the original sequence length unchanged to ensure that the final model output is still a predicted representation of each amino acid residue site.

6. The method of claim 5, wherein, In step 3, a classifier module is connected as the output layer in the output stage of the RaQMeNet network model to predict the methylation probability of each amino acid residue position. The classifier module consists of a three-layer feedforward neural network, which uses the Sigmoid function as the output activation function and maps the output features of the bidirectional Transformer module into a residue probability vector of length 731. The value of each bit represents the predicted probability that the corresponding residue is a methylation modification site. During the training of the RaQMeNet network model, an improved region-aware weighted binary cross-entropy loss function is used. Independent weighting parameters are introduced for positive and negative samples in different cluster subgroups r. and The calculation formula is: ， Where R represents the set of all cluster subgroups; The true label for the i-th residue site in the cluster subgroup r. is the predicted probability of the i-th residue site in the cluster subgroup r; These are the weights of the positive class and the weights of the negative class within region r, respectively. A dynamic adjustment mechanism based on the difficulty of samples within clustering regions is introduced to automatically adjust class weights: The difficulty of samples within region r is defined. The average prediction error for all training samples in region r is: ， Subsequently based on Dynamically update positive class weight and negative class weights Automatic updates for each training epoch With corresponding weights.

7. The method of claim 6, wherein, Step 3 also includes: introducing a masking mechanism, selecting Adam as the optimizer, setting the initial learning rate to 1e-3, training epochs to 100, and batch size to 32; re-performing quantum feature subset selection every certain number of epochs T during training, and using five-fold cross-validation to select the optimal model parameters; and recalculating the quantum kernel in each re-selection. With tag kernel Maximizing the Hilbert–Schmidt independence criterion The 512 best feature channels for the current stage are selected for training in the next cycle.

8. An electronic device, comprising: It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A storage medium, characterized by It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 7.

Images