Adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agent

Abstract

The invention discloses a self-adaptive multi-channel fusion lncRNA subcellular localization prediction method based on a reinforcement learning agent, and relates to a long-chain non-coding RNA subcellular localization prediction method. The method aims at solving the problems that sample heterogeneity is ignored and RNA-protein interaction characteristics are ignored due to the fact that a fixed feature fusion strategy is adopted in an existing method. According to the method, firstly, multi-modal feature extraction is carried out, original feature vectors are obtained through splicing, the original feature vectors are projected to a reinforcement learning state, actions are obtained through an intelligent agent obtained through PPO training based on near-end strategy optimization, and the actions comprise feature selection mask codes, fusion weight guidance and dynamic hyper-parameter configuration. A personalized fusion strategy is obtained through action vectors generated by the reinforcement learning agent, the fusion prediction network fuses the features in the multi-modal feature set based on the personalized fusion strategy, and finally the positioning probability of the lncRNA at each subcell position is obtained.

Description

Technical Field

[0001] This invention falls within the fields of bioinformatics and artificial intelligence, and relates to a method for predicting the subcellular localization of long non-coding RNAs (lncRNAs). Background Technology

[0002] Long non-coding RNAs (lncRNAs) are an important class of regulatory molecules that play a crucial role in various life processes, including gene expression, cell differentiation, and disease development. The biological function of lncRNAs is closely related to their spatial distribution within cells, i.e., their subcellular localization. Therefore, accurately predicting the subcellular localization of lncRNAs is of paramount importance for elucidating their functional mechanisms and screening for disease biomarkers.

[0003] To support costly, low-throughput experimental methods, various computational prediction models have been developed. In their early stages, these methods primarily relied on traditional machine learning frameworks. For example, models like iLoc-LncRNA and LncFinder typically extract artificially designed features from RNA sequences, such as k-mer frequencies (e.g., trinucleotide and tetranucleotide combinations) and sequence physicochemical properties. These features are then fed into classifiers like Support Vector Machines (SVM) and Random Forests for training and prediction. The performance of these methods is highly dependent on the quality of feature engineering and struggles to capture deep sequence patterns. With the rise of deep learning, prediction models have entered a new phase, capable of automatically learning high-level representations from raw sequences. Models based on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have become mainstream. For instance, DeepLncLoc utilizes CNNs to capture conserved motifs in sequences and combines them with Long Short-Term Memory Networks (LSTMs) to handle long-range dependencies. In recent years, drawing on the successes of natural language processing, models based on the Transformer architecture have demonstrated superior performance. Models such as LncLocFormer utilize the self-attention mechanism, which can capture contextual dependencies across the entire lncRNA sequence more effectively than RNNs. They have broken performance records in multiple benchmark tests and represent the current state-of-the-art level in the field.

[0004] However, existing state-of-the-art methods all face a common configuration bottleneck at their underlying logic: they almost all employ fixed, pre-designed feature fusion strategies. For example, they might simply concatenate features from different sources or use a fixed-structure attention network for weighting. This fusion approach is constant for all input lncRNA samples, leading to several insurmountable limitations: First, it ignores sample heterogeneity; the localization signals of different lncRNAs may be dominated by different modalities. Fixed fusion models cannot capture such sample-specific decision preferences, resulting in poor performance on complex samples. Second, it introduces noise and redundancy; under a fixed strategy, all pre-defined features are input indiscriminately, potentially introducing irrelevant noise and impairing the model's generalization ability. Finally, it lacks integration of biological mechanisms; most models are limited to sequence and secondary structure information, ignoring biological evidence such as RNA-protein interactions (RPI) that directly reflect localization and regulatory mechanisms. Therefore, how to break through the "static fusion" paradigm and design an intelligent fusion framework capable of adaptive decision-making based on the characteristics of each sample is a pressing technical challenge in the field. Summary of the Invention

[0005] This invention addresses the problems of existing methods that use fixed feature fusion strategies, which ignore sample heterogeneity and features such as RNA-protein interactions.

[0006] An adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents includes:

[0007] S1. Multimodal Feature Extraction: From the input lncRNA sequence, RNA sequence features, RNA secondary structure features, and RNA-protein interaction features are extracted in parallel and then concatenated to obtain a multimodal feature set. The number of features in the multimodal feature set is denoted as N. m ;

[0008] S2. Concatenate the features of the multimodal feature set to obtain the original feature vector. Projecting the state into the reinforcement learning state Actions are obtained by optimizing the agent trained using PPO based on proximal policy. The policy network in reinforcement learning contains a mechanism for processing input states. The shared backbone network, and three independent output branches branching off in parallel from this backbone network; the three output branches output feature selection masks. Integration weight guidance Dynamic hyperparameter configuration , thus obtaining the action vector ; It is an Nm A binary vector of dimension 1, where 0 or 1 in the i-th bit represents disabling or enabling the i-th feature channel, respectively; It is an N m A weight vector of dimension , where the i-th element represents the weight vector of the i-th feature channel; It is a parameter value vector with a dimension equal to the total number of parameter options of the fusion prediction network, which serves as the hyperparameter configuration of the fusion prediction network;

[0009] S3, Action vectors generated by the reinforcement learning agent. A personalized fusion strategy is obtained. The fusion prediction network fuses features in the multimodal feature set based on the personalized fusion strategy, and finally obtains the localization probability of lncRNA at each subcellular location.

[0010] Furthermore, the RNA sequence features include one or more of the following: local semantic features, context-aware features, deep biological representation features, and visual pattern features; wherein,

[0011] (a) The process of local semantic feature extraction includes:

[0012] (1) Perform k-merization on the sequence to generate a sequence composed of overlapping k-mers. ;

[0013] (2) Each k-mer Mapped to dense vectors Transform the sequence K into a vector matrix V;

[0014] (3) Divide V evenly into m consecutive, non-overlapping sub-segments along the dimension representing the k-mer permutation order. For each sub-segment Apply average pooling to all vectors within the segment to obtain a representative vector for that segment. ; to vector By concatenating them sequentially, local semantic features are obtained. ;

[0015] (b) The process of context-aware feature extraction includes:

[0016] The lncRNA sequence was segmented, and a special classifier was added at the beginning of the sequence. The processed token sequence is input into a Transformer-based model to obtain context-aware features. ;

[0017] (c) The process of deep biological characterization feature extraction includes:

[0018] The lncRNA sequence S was fed into the RNA-FM model to obtain the sequence consisting of each nucleotide. Corresponding embedding vector The embedded matrix Deep biological characterization features are obtained by applying average pooling along the sequence length dimension. ;

[0019] (d) The process of visual pattern feature extraction includes:

[0020] A one-dimensional sequence S is mapped to a two-dimensional image using chaotic game theory, and visual pattern features are obtained based on a neural network model. .

[0021] Furthermore, the extraction process of the RNA secondary structure features includes:

[0022] Secondary structure prediction is performed on the input lncRNA sequence S. Based on the secondary structure information of lncRNA, multiple graph data structures that can be processed by graph neural networks are obtained. For each graph data structure, one secondary structure feature is obtained based on the graph neural network to obtain the set of secondary structure features corresponding to multiple graph data structures.

[0023] Furthermore, the process of transforming the secondary structure information of lncRNA into a graph data structure that can be processed by a graph neural network includes:

[0024] The L nucleotides in the input lncRNA sequence S are used as L nodes V in the graph. Based on the secondary structure prediction results, edges E are established between paired nucleotide nodes to construct a graph topology G=(V,E) shared by all views.

[0025] Based on the secondary structure prediction results, we obtain pairing information P representing the pairing state of the nucleotide, binding probability B representing the type of structural element to which the nucleotide belongs, structural context information C encoded by structural context information, and position encoding E representing the absolute position of the nucleotide in the sequence; based on Different combinations yield multiple graphical views.

[0026] Furthermore, for each graph data structure, the process of obtaining one secondary structure feature based on the graph neural network includes:

[0027] The node feature matrix of the v-th graph view The input to the GNN encoder yields one secondary structure feature. The processing steps of the GNN encoder include:

[0028] (1) GIN processing for graph isomorphic networks: The representation of the central node is updated by aggregating the features of neighboring nodes. The update rule is as follows:

[0029]

[0030] in, It is a node In the The layer is represented as In the equation, N(c) represents the row vector corresponding to node c; N(c) is the set of neighbors of node c. It is a learnable parameter;

[0031] (2) Graph Attention Network (GAT) Processing: Based on the GIN output, a self-attention mechanism is used to dynamically assign different weights to neighboring nodes, enabling the model to focus on more important neighborhood information; attention coefficients The calculation is as follows:

[0032]

[0033]

[0034] in, For activation functions; and These are the feature vectors of nodes x and y after GIN encoding, respectively; W is a learnable linear transformation weight matrix applied to all nodes. It is a learnable single-layer feedforward network weight vector; This indicates vector concatenation; It is the set of neighbors of node x;

[0035] The new representation of node x is a weighted sum:

[0036]

[0037] in, Indicates the activation function;

[0038] Global average pooling is applied to the final node representation matrix of the view to aggregate the information of all nodes into a graph-level representation vector.

[0039] Furthermore, the extraction process of the RNA-protein interaction features includes:

[0040] First, for the target species, a comprehensive list of RBPs is retrieved from the protein information database. Based on the RBP list, a set of candidate RBPs is obtained, where each RBP is associated with one or more standardized subcellular localization categories. Further screening is performed based on the subcellular regions in the RNA dataset to ensure that the subcellular labels of the RBPs are consistent with the subcellular labels of the RNA. Finally, a fixed panel containing N RBPs that are identical to the subcellular regions in the current RNA dataset is determined.

[0041] RPI Score Prediction: For an input lncRNA sequence S, multiple RPI prediction tools are used to predict its interaction scores with N RBPs in the panel, resulting in multiple independent score vectors. ;

[0042] For each RPI prediction tool, the following processing is performed:

[0043] For RBP identity embedding matrix The p-th row vector of the matrix , representing the vectorized representation of the p-th RBP in the RBP panel; based on the RBP identity embedding matrix and positioning information matrix To construct the query matrix Q; use the RBP identity embedding matrix. As the bond matrix K; and Combining matrices to generate value matrices It employs a standard scaled dot product attention mechanism, utilizing a query matrix Q to query and weight RPI information. Then, global average pooling is performed on the output matrix of the attention network to obtain RNA-protein interaction features.

[0044] Furthermore, the process by which the fusion prediction network fuses features from the multimodal feature set based on a personalized fusion strategy, and ultimately obtains the localization probability of lncRNA at each subcellular location, includes:

[0045] S31, According to the mask Select the enabled feature subset from the multimodal feature set. ; mask With weight guidance Element-wise multiplication yields the effective weight vector. Based on effective weight vector Obtain the modulation weight vector ;Will The feature matrix is ​​obtained by stacking the feature vectors. After two independent linear transformation layers, the original bond moments are generated respectively. Sum matrix Construct a diagonal matrix Its diagonal elements are The corresponding modulation weight values ​​are then used to obtain the key matrix. Sum matrix ;

[0046] S32, Stack-based fusion block processing: For the first A merge block that receives the bottleneck token from the previous merge block. As input, output the updated bottleneck token. Internally, it performs the following two core operations:

[0047] (1) Cross-attention information aggregation: bottleneck token of the previous layer's output After layer normalization, it is used as query Q, combining the feature key matrix and value matrix. These are used as keys and values, respectively; through a multi-head cross-attention mechanism (MHCA), feature information is aggregated onto the token and residually joined with the original token to obtain an intermediate token with injected information. ;

[0048] (2) Self-interaction and integration between tokens: for intermediate tokens Perform multi-head self-attention operations and make residual connections to obtain tokens that have undergone internal information exchange. ,Will Feed into feedforward network Perform a nonlinear transformation and then perform residual connection again to obtain the first... The final output of each fusion block ;

[0049] The final bottleneck token is obtained by stacking L such fusion blocks. ;

[0050] S33, will The fused representation vector is obtained by applying average pooling along the token dimension. ,based on The localization probabilities of lncRNAs at predefined subcellular locations are obtained.

[0051] Furthermore, based on the effective weight vector The obtained modulation weight vector :

[0052]

[0053] The temperature parameter τ is used to adjust the smoothness of the weight distribution. It is an effective weight vector The One element; index and The traversal is of the feature subset selected by the strategy. .

[0054] Furthermore, in the process of optimizing the PPO training agent based on the proximal strategy, the following is adopted: Calculate rewards; It is a hyperparameter. Prediction refers to actions generated by an agent based on training samples. The final subcellular localization prediction probability vector is obtained through fusion prediction; label is the real and known subcellular localization label corresponding to the training sample; , It is the first feature selection mask. Each element.

[0055] Furthermore, in the prediction process, the initial bottleneck token in step S32 is obtained at the end of the end-to-end training process of the entire model, specifically including:

[0056] During end-to-end training of the entire model, including reinforcement learning and the fusion prediction network, the initial fusion bottleneck token is set as a learnable parameterized vector. ,Will Used for processing fusion blocks and optimized along with other parameters; at the end of training, the initial fusion bottleneck token corresponding to the best-performing model on the validation set is fixed as the initial bottleneck token in the prediction process.

[0057] Beneficial effects:

[0058] Compared with existing technologies, the adaptive multi-channel fusion lncRNA subcellular localization prediction system and method based on proximal strategy optimization agent provided by this invention has the following significant advantages:

[0059] (1) It realizes the personalization and precision of prediction strategies, and greatly improves prediction accuracy:

[0060] Existing methods generally employ fixed feature fusion and prediction models, which cannot adapt to the significant heterogeneity among lncRNA molecules. This invention introduces an adaptive fusion strategy generation module based on reinforcement learning, enabling personalized configuration for each input lncRNA sample and dynamically generating an optimal set of feature selection, fusion weights, and learning parameters. This strategy allows the model to accurately capture key biological signals that determine the specific lncRNA, effectively suppressing interference from irrelevant or noisy features, thus fundamentally overcoming the limitations of static models and significantly improving prediction accuracy and robustness.

[0061] (2) It innovatively introduces prior biological knowledge to guide feature extraction, thereby improving the interpretability of prediction results:

[0062] In particular, the RPI modality extraction submodule of this invention features an innovatively designed location-guided attention integration unit. This unit utilizes known subcellularly localized RNA-binding proteins (RBPs) as biological anchors, enabling the model to actively learn and focus on interaction signals most relevant to specific localization functions through an attention mechanism. This is not only an efficient feature extraction method, but more importantly, it directly injects explicit biological prior knowledge into the model, giving the final RPI features stronger biological significance and interpretability, providing valuable clues for subsequent biological validation.

[0063] (3) An efficient and flexible dynamic fusion network architecture is introduced, which balances performance and computational efficiency:

[0064] To address the challenge of dynamically selecting features by intelligent agents, this invention introduces a High-Bottlehole Cross-Attention Fusion Network (H-CAFN). This network uses a small set of fixed bottleneck tokens to query and aggregate a variable number of input features. Its computational complexity is linearly related to the number of input features, rather than quadratically. This design not only perfectly meets the requirements of adaptive feature selection but also significantly reduces computational overhead while ensuring deep information interaction, enabling the entire system to maintain high efficiency and scalability when processing large-scale datasets. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the prediction process of the present invention;

[0066] Figure 2 This is a graph showing the results of five-fold cross-validation of the present invention on two public datasets;

[0067] Figure 3 This is a performance comparison chart between the present invention and the ablation model of key components;

[0068] Figure 4 This is a performance comparison chart between the present invention and the prior art (SOTA model). Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. The drawings in this specification are schematic diagrams and are used only to explain the invention, and are not intended to limit the actual application scenarios of the invention. For ease of description and reference, a specific embodiment of the method proposed in this invention is referred to as LncAPNet (LncRNA Adaptive Policy Network) in the experimental section below.

[0070] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0071] The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents described in this embodiment includes the following steps:

[0072] S1. Multimodal Feature Extraction. Information is extracted in parallel and from multiple dimensions from the input lncRNA sequence to construct a comprehensive feature set, providing a solid foundation for subsequent adaptive fusion decisions. Specifically, this step generates 15 independent feature vectors from three biological modalities: RNA sequence (modality A), RNA secondary structure (modality B), and RNA-protein interactions (modality C).

[0073] S11. RNA Sequence Modality Extraction. This step directly analyzes the primary sequence of lncRNA, i.e., the sequence of nucleotides (A, U, G, C), and performs deep encoding from four complementary perspectives to generate feature vectors. The specific process includes:

[0074] S111, Local Semantic Feature Extraction. This aims to capture local nucleotide combination patterns (k-mers) and their distribution information within a sequence. The specific steps are:

[0075] (1) Sequence k-merization: The input lncRNA sequence of length L is k-merized. ,in, The sequence is segmented using a sliding window with a step size of 1, generating a sequence composed of overlapping k-mers. In this embodiment, k is preferably 3. ,in, .

[0076] (2) k-mer vectorization: Using a Word2Vec model pre-trained on a large-scale unlabeled lncRNA corpus, each k-mer is vectorized. Mapped to a Dense vectors of dimension In this embodiment, Therefore, the k-mer sequence K is transformed into a vector matrix V.

[0077] (3) Segmented pooling and concatenation: In order to preserve the positional distribution information of k-mer in the sequence and generate a fixed-length feature vector, the following operations are performed:

[0078] a. will be by A matrix V consisting of k-mer vectors is uniformly divided into m consecutive, non-overlapping segments along the dimension representing the k-mer order. In this embodiment, m=64.

[0079] b. For each sub-segment Apply average pooling to all vectors within the segment to obtain a representative vector for that segment. :

[0080]

[0081] in, It is a sub-segment The number of vectors in the middle; for In the dense vector, j represents the segment index.

[0082] c. Calculate all m pooled vectors The components are concatenated sequentially to form the final local semantic feature vector. .

[0083]

[0084] In this embodiment, The dimension is 64×64=4096.

[0085] S112. Context-Aware Feature Extraction. This step utilizes a large-scale pre-trained nucleic acid language model to capture long-range dependencies and global contextual information. The specific method is as follows:

[0086] (1) Model selection: A pre-trained model based on the Transformer architecture, such as RNABERT-2, was selected. This model has been pre-trained on a large number of nucleic acid sequences and is able to understand complex grammatical and semantic rules.

[0087] (2) Sequence encoding: The input lncRNA sequence S is tokenized according to the requirements of RNABERT-2, and a special classifier is added at the beginning of the sequence. The processed token sequence is then input into the RNABERT-2 model.

[0088] (3) Feature Extraction: After processing through multiple layers of self-attention mechanisms, the model outputs a context-dependent representation of each token. The feature extraction process involves extracting the context-dependent representation from the output of the last Transformer encoder layer. The hidden state vector corresponding to the token is denoted as . In practice with BERT-like models (such as RNABERT-2 used in this embodiment), the final hidden state of the [CLS] token is trained to capture global summary information of the entire input sequence and is therefore used as a sequence-level feature representation. This vector is designed as an aggregated representation of the entire sequence:

[0089]

[0090] In this embodiment, the dimension of RNABERT-2 output .

[0091] S113. Deep Biological Characterization Extraction. Another RNA Foundation Model, the RNA-FM model, is employed. This model is a publicly published, large-scale language model pre-trained on massive amounts of RNA sequences, capable of generating biologically meaningful representations for RNA sequences. The specific method is as follows:

[0092] (1) Model selection: The RNA-FM model was selected, which was pre-trained on a database containing more than 23 million real RNA sequences.

[0093] (2) Nucleotide-by-nucleotide embedding: The input lncRNA sequence S is fed into the RNA-FM model. The model is for each nucleotide in the sequence. Generate a context-sensitive embedding vector To form an embedding matrix In this embodiment, .

[0094] (3) Global pooling: Apply average pooling to the embedding matrix E along the sequence length dimension to aggregate the representations of all positions into a single, fixed-length sequence-level feature vector. :

[0095]

[0096] S114. Visual pattern feature extraction: The one-dimensional RNA sequence is converted into a two-dimensional image, and the visual and spatial patterns contained therein are extracted using a convolutional neural network (CNN). The specific method is as follows:

[0097] (1) A one-dimensional sequence S is mapped to a two-dimensional image using the chaotic game representation (CGR) method:

[0098] First, define the four vertices of a square, each corresponding to one of the four types of nucleotides, for example... Then, starting from the center point of the square... Begin by analyzing the nucleotides in the sequence. Iteratively calculate the position of the next point. :

[0099]

[0100] in, It is a nucleotide The corresponding vertex coordinates. This is achieved by drawing all generated points. This yields a CGR image with fractal characteristics, whose texture and pattern reflect the compositional regularity of the sequence.

[0101] (2) CNN Feature Extraction: The generated CGR image is input into a pre-defined convolutional neural network. In this embodiment, a Centralized Feature Pyramid (CFP) network is used, which is good at capturing and fusing multi-scale visual features.

[0102] (3) Feature vectorization: The feature map output by the CFP network is subjected to global average pooling or flattening to obtain a fixed-length visual pattern feature vector. In this embodiment, .

[0103] S12. RNA Secondary Structure Modal Feature Extraction. This invention considers that the function of RNA molecules is closely related to their three-dimensional spatial structure formed by folding, and that secondary structure is its key framework. Therefore, RNA secondary structure is modeled as a graph, and a graph neural network (GNN) is used to learn structural features from it to generate feature vectors. to The specific process includes:

[0104] S121. Secondary Structure Prediction: Using the mature computational tool RNAfold, the secondary structure of the input lncRNA sequence S is predicted. The output mainly includes:

[0105] (1) Dotted bracket notation: a string that describes the base pairing relationship.

[0106] (2) Base pairing probability matrix: an L×L matrix, where elements This represents the probability that the a-th and b-th nucleotides pair up.

[0107] (3) Structural context: The type of structural element in which each nucleotide is located.

[0108] S122, Graph Construction and Multi-View Node Feature Definition: Based on the prediction results obtained in step S121, this step transforms the secondary structure information of lncRNA into a graph data structure that can be processed by graph neural networks. This process includes two core parts: unified graph topology construction and diverse node feature definitions.

[0109] (1) Construction of unified graph topology: First, construct a graph topology G=(V,E) shared by all views.

[0110] Node (V): The L nucleotides in the input lncRNA sequence S are used as the L nodes of the graph.

[0111] Edge (E): Based on the dotted bracket notation or base pairing probability matrix output by S121, establish edges between nucleotide nodes where pairing occurs.

[0112] This forms the basic framework of the diagram, reflecting the folding pattern of RNA.

[0113] (2) Definition of features for multi-view nodes:

[0114] Pairing information (P): A scalar or vector representing the pairing status of the nucleotide, which can be obtained from the dotted bracket notation.

[0115] Binding probability (B): A scalar representing the total probability that the nucleotide will participate in pairing, directly derived from the base pairing probability matrix output by S121.

[0116] Structural context (C): A 4-dimensional one-hot vector representing the type of structural element to which the nucleotide belongs (such as stem, loop, etc.), derived from the structural context information output by S121.

[0117] Position code (E): A fixed coding vector generated by sine and cosine functions, representing the absolute position of the nucleotide in the sequence.

[0118] (3) Generation of Seven Graph Views: By combining the four basic features mentioned above, seven different sets of node features are provided for a unified graph topology G, thereby generating seven graph views. Each view shares the same nodes and edges, but the initial feature representations of its nodes are different. The specific feature combinations are as follows: For example, for a view {P,B,C}, its node features include three types of features: pairing information (P), combination probability (B), and structural context (C).

[0119] S123, Two-Layer GNN Encoding: Each graph view is equipped with an independent two-layer GNN encoder. This encoder consists of a Graph Isomorphism Network (GIN) layer and a Graph Attention Network (GAT) layer stacked together. For the v-th graph view, its node feature matrix is ​​denoted as... The input to the GNN encoder is the node representation of layer 0. That is The row vector corresponding to node c.

[0120] (1) GIN layer: The representation of the central node is updated by aggregating the features of neighboring nodes. The update rule is as follows:

[0121]

[0122] in, It is a node In the The layer representation, where N(c) is the set of neighbors of node c. It is a learnable parameter. GIN layers are able to effectively capture the topology of a graph.

[0123] (2) GAT Layer: Based on the output of the GIN layer, a self-attention mechanism is used to dynamically assign different weights to neighboring nodes, enabling the model to focus on more important neighborhood information. Attention coefficients The calculation is as follows:

[0124]

[0125]

[0126] in, This is the activation function. and are the feature vectors of nodes x and y after being encoded by the GIN layer (GIN as a whole), respectively. W is a learnable linear transformation weight matrix applied to all nodes. It is a learnable single-layer feedforward network weight vector. This indicates vector concatenation. It is the set of neighbors of node x.

[0127] The new representation of node x is a weighted sum:

[0128]

[0129] in, This represents the activation function.

[0130] S124. Graph Representation Generation: After two-layer GNN encoding, global average pooling is applied to the final node representation matrix of each view, aggregating the information of all nodes into a graph-level representation vector. Thus, seven independent secondary structure feature vectors are obtained from the seven graph views. .

[0131] S13. RNA-protein interaction mode feature extraction. This invention considers that lncRNAs exert their biological functions through interactions with specific RNA-binding proteins (RBPs). Therefore, by predicting the interaction profiles of lncRNAs with a set of functionally known RBPs and utilizing prior knowledge of RBP location for information extraction, feature vectors are generated. The specific process includes:

[0132] S131, RBP Panel Selection: To correlate RBP interaction information with subcellular localization knowledge, this step aims to construct a high-quality reference panel containing N RBPs. The panel's construction follows the principles of species specificity and explicit localization information, ensuring its biological relevance and information reliability. The specific construction process is as follows:

[0133] First, for the target species (e.g., human Homo sapiens), a comprehensive list of RBPs is retrieved from authoritative, publicly available protein information databases. In this embodiment, the UniProt (Universal Protein Resource) database is used. Advanced search functions are employed, using keyword combinations for filtering, with three specific limitations:

[0134] (1) Functional annotation (Gene Ontology, GO): Screening for proteins containing terms such as “RNA binding” (GO:0003723).

[0135] (2) Species (Organism): limited to the target species, such as "Homo sapiens (human)

[9606] ".

[0136] (3) Reviewed status: Only items that have been manually reviewed by experts (i.e. UniProtKB / Swiss-Prot section) are selected to ensure the high quality and accuracy of protein information.

[0137] Next, for each RBP in the RBP set obtained through the above filtering operations, its known subcellular location information is extracted. The "Subcellular location" annotation section under each protein entry in the UniProt database is used again. To ensure the reliability and consistency of the information, the following operations are performed:

[0138] (1) Evidence level screening: Prioritize retaining location information supported by experimental evidence and filter out locations based solely on computational predictions.

[0139] (2) Location Category Normalization: Detailed location descriptions in the database (such as "Nucleus>nucleolus", "Cytoplasm>cytoskeleton") are normalized to predefined, more macroscopic subcellular compartment categories, such as: nucleus, cytoplasm, ribosome, etc. A single RBP may be labeled as belonging to multiple compartments.

[0140] After the above screening, a candidate RBP set is obtained, where each RBP is associated with one or more standardized subcellular localization categories. Further screening is performed based on the subcellular regions in the RNA dataset, ensuring that the subcellular labels of the RBPs are consistent with the subcellular labels of the RNA. For example, if the subcellular labels of the RNA dataset are chromatin, cytoplasm, insoluble-cytoplasm, and nucleus, then the subcellular labels of the RBPs should also be these four labels. Finally, a fixed panel containing N RBPs that are identical to the subcellular regions in the current RNA dataset is determined.

[0141] This panel and its associated location information matrix (If RBP has a certain subcellular label, the corresponding value in this matrix is ​​1; otherwise, it is 0. Where C is the total number of subcellular labels.) This serves as the prior knowledge basis for subsequent steps S132 and S133.

[0142] S132, RPI Score Prediction: Four RPI (RNA-Protein Interaction) prediction tools with different principles were selected, such as RPISeq-SVM and RPISeq-RF based on machine learning, and LPI-KCGCN and RPI-GGCN based on graph neural networks. For the input lncRNA sequence S, these four tools were used to predict its interaction score with N RBPs in the panel, resulting in four independent score vectors. ,in, .

[0143] S133. Location-Guided Attention Mechanism: The core task of this step is to construct an attention module that can learn to evaluate the importance of each RBP in the RBP panel for predicting the location of any subcellular compartment. Unlike simple information weighting, the guided attention mechanism of this invention utilizes both static embedding information learned from the RBPs themselves and dynamic RPI scores calculated from the current lncRNA, thereby achieving a context-aware, highly adaptive feature aggregation. The detailed execution flow of this step is as follows:

[0144] (1) To implement the attention mechanism, this invention predefines an RBP identity embedding matrix. One dimension is The matrix. The vector of the p-th row of this matrix. , represents a vectorized representation of the inherent, abstract biological function of the p-th RBP in the RBP panel, where For the embedded dimension.

[0145] (2) Constructing the Query: Based on the RBP identity embedding matrix and positioning information matrix To construct the query matrix Q, for example Therefore, the dimension of the resulting query matrix Q is The c-th row vector in the Q matrix is ​​formed by aggregating the identity embeddings of all known RBPs present in subcellular compartment c. Therefore, this row vector can be understood as a high-dimensional, learnable vectorized representation of the biological environmental characteristics of compartment c itself. It will serve as a guiding signal to query the RPI information most relevant to that compartment.

[0146] (3) Constructing Keys and Values: Directly using the RBP identity embedding matrix As the bond matrix K. K= The dimension of K is The q-th row represents the identity information of the q-th RBP, which is used to match the query Q.

[0147] Then, dimensional score spectrum vector and Vie Combining matrices to generate value matrices Preferably, this is achieved through a linear transformation layer (parameterized by learnable weights) or by... Each score The corresponding RBP embedding vector is scaled as a modulation factor. Finally, a dimension is generated. Value matrix , It is the dimension of the value vector. The q-th row vector contains both the inherent information of RBPq and reflects the strength of the interaction between the current lncRNA and the RBP through its magnitude.

[0148] (4) Attention Calculation: The standard scaled dot product attention mechanism is adopted, using the query matrix Q to query and weight RPI information. ):

[0149]

[0150] The output of this operation is The matrix, where each row represents the integrated RPI features obtained under the guidance of a specific subcellular localization.

[0151] S134, Feature Aggregation: Perform global average pooling on the output matrix of the attention network to compress it into a single, information-condensed feature vector.

[0152] Perform this operation on the outputs of the four RPI prediction tools respectively, and you will get four RPI feature vectors. .

[0153] At this point, step S1 is complete, generating a total of 4 + 7 + 4 = 15 multimodal feature vectors, forming the feature set. .

[0154] It should be noted that, based on the research and experiments of this invention, these 15 multimodal feature vectors have good performance. However, in actual processing, the 15 multimodal feature vectors can be filtered according to the computational requirements, or additional features can be added on this basis. The subsequent processing is the same, only the feature sets corresponding to the 15 features need to be adjusted accordingly.

[0155] S2. Adaptive Fusion Strategy Generation. This step is one of the core innovations of LncAPNet, aiming to overcome the limitations of the "one-size-fits-all" static fusion strategy in traditional methods. At its core is a reinforcement learning (RL)-based agent that dynamically and adaptively generates an optimal fusion strategy based on the specific multimodal feature set generated for each lncRNA sample in step S1. This strategy guides the subsequent S3 step in selecting, weighting, and processing these features to achieve optimal predictive performance. The implementation of this step can be modeled as a Markov Decision Process (MDP), with its key elements defined as follows:

[0156] S21. State Representation. The state is the sole basis for the agent's decision-making, and it must comprehensively and compactly represent the characteristic information of the current lncRNA sample.

[0157] S211, Feature Aggregation. All 15 feature vectors generated in step S1 are aggregated. Concatenate along the feature dimension to form a high-dimensional original feature vector. .

[0158]

[0159] in, It is the dimension of the f-th feature vector.

[0160] S212, Status Coding. Because... The high dimensionality and redundant information necessitate a state encoder network to project the data into a low-dimensional, information-dense representation space. In this embodiment, the state encoder is a multi-layer perceptron (MLP), denoted as... .

[0161]

[0162] in, These are the learnable parameters of the MLP, and ReLU is the modified linear unit activation function. It is the dimension of the final state vector.

[0163] This state vector The data captured from 15 feature channels will be used as input to the agent's policy network.

[0164] S22. Action Space Design. An action is a decision output by the agent, i.e., a fusion strategy tailored to the current sample. LncAPNet's action space design is multidimensional and structured, comprising three cooperating components that together form a complete action vector. ,in, Choose a mask for the feature. To integrate weight guidance, For dynamic hyperparameter configuration, to achieve efficient decision-making in the aforementioned heterogeneous action space (including discrete selection, continuous weights, and classification configuration), the policy network of the reinforcement learning agent used in this invention has an output layer designed as a multi-branch output structure. Specifically, this policy network includes a layer for processing the input state. The system consists of a shared backbone network and three independent output branches that branch off in parallel from this backbone network. Each output branch is specifically responsible for generating action vectors. This is one of the components. The three branches are: feature selection branch, fusion weight branch, and hyperparameter configuration branch. The generation mechanism of each branch will be described in detail below.

[0165] S221, Feature selection mask ( This section determines which of the 15 feature channels should be used for subsequent fusion.

[0166] (1) Form: A 15-dimensional binary vector , where 0 or 1 in the i-th bit represents disabling or enabling the i-th feature channel, respectively.

[0167] (2) Generation Mechanism: The feature selection branch of the policy network receives state representations from the shared backbone network. This branch contains a linear layer with an output dimension of 15. By applying the Sigmoid activation function to the output of this linear layer, an independent probability value in the range [0,1] is generated for each of the 15 feature channels. During decision-making, a threshold (e.g., 0.5) can be set, and the mask elements corresponding to probability values ​​greater than the threshold are set to 1, while the rest are set to 0, thereby generating a binary mask vector. .

[0168] S222, Guidance on Fusion Weights ( This section provides a continuous value guide for the relative importance of all 15 feature channels (whether or not they are selected).

[0169] (1) Form: A 15-dimensional continuous value vector .

[0170] (2) Generation Mechanism: The fusion weight branch of the policy network also receives state representations from the shared backbone network. This branch directly outputs 15 consecutive real values ​​through a linear layer, forming the original weight vector. Optionally, to give the weights a clear relative meaning, these 15 values ​​can be normalized using the Softmax function to obtain the final weight vector. .

[0171] S223, Dynamic Fusion Network Parameter Configuration ( This part is a unique design of LncAPNet that allows the agent to dynamically adjust the learning or behavioral parameters of the downstream fusion network for each batch of samples.

[0172] (1) Form: A set of discrete parameter values ​​(fusion network parameters). Taking the learning rate of a single-step update of the fusion network as an example, the action space can be a predefined set, such as... These parameters include: learning rate, fusion network depth, modality gating temperature coefficient, feedforward network expansion ratio, and attention dropout rate in the fusion block.

[0173] (2) Generation Mechanism: The parameter configuration branch of the fusion network in the policy network is responsible for making decisions about the parameters of the fusion network. The output dimension of the linear layer of this branch is equal to the total number of all available parameter options. By applying the Softmax function to this output, a probability distribution is obtained, and the agent selects the option with the highest probability as the parameter configuration of the fusion network for the current step. .

[0174] S23. Reward Function Design. The reward function is used to evaluate the actions performed by the agent. The quality of the reward function serves as a supervisory signal driving policy learning. A well-designed reward function is crucial for successful training. The reward function of LncAPNet... It consists of two parts, designed to balance prediction accuracy and model efficiency. The reward function is used only during the training phase of the reinforcement learning agent to guide the parameter optimization of the policy network. After training is complete, when entering the actual prediction (inference) phase, the agent directly outputs the optimal action without calculating the reward.

[0175] S231, Accuracy Bonus ( This part of the reward aims to encourage the agent to choose actions that lead to high prediction accuracy. The agent performs these actions during the training phase. Afterward, the system completes the S3 fusion and prediction steps, obtaining a subcellular prediction result. This reward value is determined by calculating the macro-average F1 score between this prediction result and the ground truth of the training samples, using the following formula:

[0176]

[0177] Among them, prediction refers to the actions generated by the agent on the training samples. The final subcellular localization prediction probability vector is obtained through the S3 dynamic fusion prediction network module; label is the real and known subcellular localization label corresponding to the training sample. This function calculates the macro-average F1 score. It takes the model's predictions and the true labels as input and outputs a scalar value between 0 and 1.

[0178] S232, Sparsity penalty ( This section aims to encourage agents to complete tasks using as few features as possible, thereby improving the model's computational efficiency and generalization ability. The penalty is proportional to the number of features selected.

[0179]

[0180] in, It is the first feature selection mask. Each element.

[0181] S233, Total Reward, Final Reward It is a weighted sum of accuracy reward and sparsity penalty:

[0182]

[0183] in, It is a hyperparameter used to balance the importance of prediction accuracy and model simplicity. The larger the value, the more likely the agent is to select fewer features.

[0184] S24. Policy Optimization Algorithm: This step uses the Proximal Policy Optimization (PPO) algorithm to train the agent. PPO is an advanced policy gradient algorithm known for its stable training, simple implementation, and high sample efficiency.

[0185] S241. The intelligent agent consists of two neural networks:

[0186] (1) Policy network, with parameters It is parameterized. It receives the state. As input, and output action The probability distribution.

[0187] (2) Value network, based on parameters It is parameterized. It receives the state. As input, output a scalar value. This is used to estimate the expected value of the total reward that may be obtained in the future under this state.

[0188] S242. Advantage Function: To reduce the variance of policy gradient estimation, PPO uses an advantage function. To evaluate how good an action is relative to the average action in that state.

[0189]

[0190] in, It is a discount factor.

[0191] S243, PPO Clipped Objective Function: The core of PPO is to limit the step size of each policy update by using a clipped objective function, thus avoiding performance crashes caused by excessively large updates. Policy Network The optimization objective is:

[0192]

[0193] in, Actions taken under both old and new strategies The probability ratio, Indicates the old strategy; It is an estimate of the dominance function; It is a small hyperparameter (e.g., 0.2) that defines the clipping range; The function restricts the value of x to 1 / 2. Within the range.

[0194] This objective function pessimistically limits the magnitude of policy updates by taking the minimum value, thereby ensuring the stability of training.

[0195] S244. The training process is iterative. In each iteration, the agent interacts with the environment (i.e., the entire prediction process) using the current policy, collecting a batch of experience trajectories, i.e., (State, Action, Reward) tuples. Then, the advantage function is calculated using these data, and the parameters of the actor network and critic network are updated according to the PPO objective function. This process is repeated until the agent's policy converges, enabling the generation of efficient and accurate fusion strategies for different lncRNA samples.

[0196] At this point, the S2 training phase is complete. It produces a well-trained reinforcement learning agent with fixed parameters. This agent will play a key decision-making role in the subsequent model prediction phase, and its specific workflow is described in step S25.

[0197] S25. Workflow of the Model Prediction Phase. After the agent has been fully trained in step S24, and its policy network parameters have been finalized and frozen, the entire LncAPNet model can be used to predict new, unknown lncRNA samples. This phase's workflow is a deterministic, feedback-free, unidirectional process, with the following specific steps:

[0198] S251, Initial State Construction: Upon receiving a new lncRNA sample, the system first performs multimodal feature extraction in S1, obtaining 15 initial feature vectors for the lncRNA. Subsequently, the system performs the state representation process in S21, concatenating these feature vectors into... And through the state encoder network 𝑀𝐿𝑃𝑠𝑡𝑎𝑡𝑒, it is mapped to a unique, information-dense state vector for that sample, with the initial state as... .

[0199] S252, Optimal Action Generation and execution: The initial state constructed above. The input is fed into a pre-trained policy network with fixed parameters. The policy network performs a forward propagation calculation on the state and directly outputs the current optimal action. (Including feature mask) fusion weight and network configuration This action. The results are then passed to the S3 dynamic fusion prediction network module, which uses this strategy to re-fuse and calculate the features, and outputs the subcellular localization prediction results for the lncRNA sample.

[0200] S3. Dynamic Fusion Prediction Network Module. This step is the execution terminal of the LncAPNet method. Its core task is to accurately and unconditionally execute the personalized fusion strategy generated by the reinforcement learning agent in step S2. To this end, LncAPNet designs a High-bottleneck Cross-attention Fusion Network (H-CAFN) specifically for handling dynamically variable inputs. This module receives the agent's action instructions and the selected feature vector, and through deep information interaction guided by the strategy, generates a highly condensed discriminative representation, and finally outputs the localization probability of lncRNAs at various subcellular locations.

[0201] S31, Policy-Guided Modulation. This sub-step is the key bridge connecting decision-making (S2) and execution (S3). It is responsible for modifying the abstract action vector output by the agent. This is transformed into a specific modulation signal that can be directly injected into the computation process of the fusion network. The specific process includes:

[0202] S311, Policy Resolution and Weight Generation. This step consists of the following parts:

[0203] (1) Receive the 15-dimensional feature selection mask output from step S2. 15-dimensional fusion weight guidance .

[0204] (2) First, according to the mask Select the enabled feature subset from the complete set F containing 15 feature vectors. .

[0205] (3) Mask With weight guidance Perform element-wise Hadamard product to obtain an effective weight vector that has non-zero values ​​only on the selected channels. :

[0206]

[0207] (4) To enhance numerical stability and highlight the relative importance among features, the effective weight vector is... All non-zero element applications include temperature parameters. The modulation weight vector is obtained by normalizing the Softmax function. :

[0208]

[0209] The temperature parameter τ (set to 1.0 in this embodiment) is used to adjust the smoothness of the weight distribution. It is an effective weight vector The , Element, index and The traversal is of the feature subset selected by the strategy. Instead of all 15 original features; for The elements in.

[0210] S312, Modulation Signal Injection. To transmit the agent's policy intent losslessly to the downstream cross-attention network, LncAPNet employs an energy-conserving scaling mechanism. The normalized modulation weights are then injected... Used to scale the key (K) and value (V) matrices in the cross-attention mechanism. Specifically:

[0211] (1) In The eigenvectors are stacked to form a feature matrix. Where d is the feature dimension, This is the number of feature vectors currently selected for feature fusion.

[0212] (2) The characteristic matrix After two independent linear transformation layers, the original bond moments are generated respectively. Sum matrix .

[0213] (3) Construct a diagonal matrix Its diagonal elements are The corresponding modulation weight value.

[0214] (4) Calculate the modulated key matrix Sum matrix :

[0215]

[0216] In this way, the weights assigned by the agent will simultaneously affect the calculation of the attention score (through...). ) and weighted aggregation of subsequent information (through This enables precise, dual guidance of the fusion process.

[0217] S32, Core Fusion Unit (H-CAFN Network). This unit is specifically designed to address the challenge of agents dynamically selecting features. It can efficiently fuse a variable number of feature vectors from heterogeneous sources. Its core component is a stackable fusion block.

[0218] (a) Inputs and core component definitions of the fusion block:

[0219] Dynamic input: Feature subsets filtered and modulated through step S31 and its corresponding key matrix Sum matrix Output: Initial fusion bottleneck token.

[0220] Initial Fusion Bottleneck Tokens: A set of learnable parameterized vectors with a fixed number (e.g., B=4). These tokens are learnable parameter matrices that are used throughout the end-to-end training of the entire model, including the module in step S3. It is a parameterized vector that is optimized along with other network parameters. After training, the values ​​of these initial tokens are fixed (actually, they are the initial fusion bottleneck tokens used in the last training process at the end of training) as fixed information digests. In actual use (inference), they are directly used as initial fusion bottleneck tokens to participate in the processing of fusion blocks, in order to extract key information from diverse input features.

[0221] (b) Internal workflow of the Fusion Block: LncAPNet implements the core fusion unit by stacking L fusion blocks. For the first... Each fusion block receives the fusion bottleneck token output by the previous fusion block. As input, output the updated bottleneck token. The model uses an initial fusion bottleneck token in the initial phase. Input the first fusion block. Internally, it performs the following two core operations:

[0222] (1) Cross-attention information aggregation: This step realizes the one-way aggregation of information from the external feature vector to the internal bottleneck token.

[0223] Role allocation: Bottleneck tokens for output in the previous layer After layer normalization, it becomes the query (Q), while the feature key matrix and value matrix generated by S31 and modulated by the policy are... These serve as the key and value, respectively.

[0224] Calculation process: Through the Multi-Head Cross-Attention (MHCA) mechanism, feature information is aggregated onto the token, and residual connections are made with the original token to obtain the intermediate token after information injection. :

[0225]

[0226]

[0227] The above formula performs a cross-attention operation. For this cross-attention operation, the required query is a tensor. supply, It is a layer normalization operation. The essence of this process is that each bottleneck token queries all selected features and, based on its attention score, weights and aggregates the most important information into itself, completing an efficient information extraction.

[0228] (2) Token self-interaction and integration: This step allows bottleneck tokens to communicate with each other to integrate, coordinate, and deeply process the information they draw from the feature set. This step adopts a standard Transformer encoder block structure, which includes two sub-modules: Multi-Head Self-Attention (MHSA) and Feed-Forward Network (FFN).

[0229] Calculation process:

[0230] First, for the intermediate token Perform multi-head self-attention operations and make residual connections to obtain tokens that have undergone internal information exchange. :

[0231]

[0232] Then, the results after self-attention processing Feed into feedforward network Perform a nonlinear transformation and then perform residual connection again to obtain the first... The final output of each fusion block :

[0233]

[0234] By stacking L such fusion blocks, information undergoes multiple rounds of deep interaction and refinement between features and tokens, and between tokens themselves, gradually generating a highly abstract and refined representation.

[0235] S33 represents generation and classification. This sub-step is responsible for transforming the deeply fused information into the final prediction result. This sub-step can be further divided into two smaller steps:

[0236] S331. The final bottleneck token is obtained after the fusion representation is generated and processed by L fusion blocks. The essence of all selected features has been highly condensed. To generate a single representation of the entire lncRNA sample, mean pooling is applied along the token dimension to these B token vectors:

[0237]

[0238] Received It is a fixed-dimensional, highly condensed fusion representation vector.

[0239] S332, Final classification prediction, which will fuse the representation vectors. The data is fed into a final fully connected classifier. Since lncRNA subcellular localization is a multi-label classification task, a sigmoid activation function is appended to the output layer of the classifier to independently calculate the localization probability of the lncRNA at each of the C predefined subcellular locations.

[0240]

[0241] in and These are the learnable weights and biases of the classification layer, and P is the probability vector of the final output, where the 6th... The element indicates that the lncRNA is located at the [number]th [element]. The probability of a subcellular location.

[0242] At this point, step S3 is complete. The entire process begins with receiving precise instructions from the agent, proceeds through policy-guided modulation and efficient cross-attention deep fusion, and finally generates localization probabilities, forming a complete, closed-loop, and highly adaptive prediction process.

[0243] The performance of LncAPNet on five-fold cross-validation on two public datasets is shown in Table 1 and Table 2. Figure 2 As shown in Table 2, the performance comparison of LncAPNet with existing state-of-the-art methods on two public datasets is presented in Table 2. Figure 4 As shown in Figures 3-4, LncAPNet's performance is compared with that of the key component ablation model. Figure 3 As shown.

[0244] Table 1. Performance of LncAPNet on five-fold cross-validation on two public datasets.

[0245]

[0246]

[0247] Table 2. Performance comparison of LncAPNet with existing state-of-the-art methods on two public datasets.

[0248]

[0249] Table 3. Performance Comparison of LncAPNet and Key Component Ablation Model on the Human Genome Dataset

[0250]

[0251] Table 4. Performance comparison of LncAPNet and key component ablation model on the Human and mouse gene mixed dataset.

[0252]

[0253] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents, characterized in that, include: S1. Multimodal feature extraction; From the input lncRNA sequence, RNA sequence features, RNA secondary structure features, and RNA-protein interaction features are extracted in parallel and then concatenated to obtain a multimodal feature set. The number of features in the multimodal feature set is denoted as N. m ; S2. Concatenate the features of the multimodal feature set to obtain the original feature vector. Projecting the state into the reinforcement learning state Actions are obtained by optimizing the agent trained using PPO based on proximal policy. ; The policy network in reinforcement learning contains a function for processing the input state. The shared backbone network, and three independent output branches branching off in parallel from this backbone network; the three output branches output feature selection masks. Integration weight guidance Dynamic hyperparameter configuration , thus obtaining the action vector ; It is an N m A two-dimensional vector, where 0 or 1 in the i-th bit represents disabling or enabling the i-th feature channel, respectively; The image src='' class="img-anchor" img-id="QLYQS_9" / > is an N m A weight vector of dimension , where the i-th element represents the weight vector of the i-th feature channel; It is a parameter value vector with a dimension equal to the total number of parameter options of the fusion prediction network, which serves as the hyperparameter configuration of the fusion prediction network; S3, Action vectors generated by the reinforcement learning agent. A personalized fusion strategy is obtained. The fusion prediction network fuses features in the multimodal feature set based on the personalized fusion strategy, and finally obtains the localization probability of lncRNA at each subcellular location.

2. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 1, characterized in that, The RNA sequence features include one or more of the following: local semantic features, context-aware features, deep biological representation features, and visual pattern features; wherein... (a) The process of local semantic feature extraction includes: (1) Perform k-merization on the sequence to generate a sequence composed of overlapping k-mers. ; (2) Each k-mer Mapped to dense vectors Transform the sequence K into a vector matrix V; (3) Divide V evenly into m consecutive, non-overlapping sub-segments along the dimension representing the k-mer permutation order. For each sub-segment Apply average pooling to all vectors within the segment to obtain a representative vector for that segment. ; to vector By concatenating them sequentially, local semantic features are obtained. ; (b) The process of context-aware feature extraction includes: The lncRNA sequence was segmented, and a special classifier was added at the beginning of the sequence. The processed token sequence is input into a Transformer-based model to obtain context-aware features. ; (c) The process of deep biological characterization feature extraction includes: The lncRNA sequence S was fed into the RNA-FM model to obtain the sequence of each nucleotide. Corresponding embedding vector The embedded matrix Deep biological characterization features are obtained by applying average pooling along the sequence length dimension. ; (d) The process of visual pattern feature extraction includes: A one-dimensional sequence S is mapped to a two-dimensional image using chaotic game theory, and visual pattern features are obtained based on a neural network model. .

3. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 1, characterized in that, The extraction process of the RNA secondary structure features includes: Secondary structure prediction is performed on the input lncRNA sequence S. Based on the secondary structure information of lncRNA, multiple graph data structures that can be processed by graph neural networks are obtained. For each graph data structure, one secondary structure feature is obtained based on the graph neural network to obtain the set of secondary structure features corresponding to multiple graph data structures.

4. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 3, characterized in that, The process of transforming the secondary structure information of lncRNA into a graph data structure that can be processed by a graph neural network includes: The L nucleotides in the input lncRNA sequence S are used as L nodes V in the graph. Based on the secondary structure prediction results, edges E are established between paired nucleotide nodes to construct a graph topology G=(V,E) shared by all views. Based on the secondary structure prediction results, we obtain pairing information P representing the pairing state of the nucleotide, binding probability B representing the type of structural element to which the nucleotide belongs, structural context information C encoded by structural context information, and position encoding E representing the absolute position of the nucleotide in the sequence; based on Different combinations yield multiple graphical views.

5. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 4, characterized in that, For each graph data structure, the process of obtaining one secondary structure feature based on a graph neural network includes: The node feature matrix of the v-th graph view The input to the GNN encoder yields one secondary structure feature. The processing steps of the GNN encoder include: (1) GIN processing for graph isomorphic networks: The representation of the central node is updated by aggregating the features of neighboring nodes. The update rule is as follows: in, It is a node In the The layer is represented as In the equation, N(c) represents the row vector corresponding to node c; N(c) is the set of neighbors of node c. It is a learnable parameter; (2) Graph Attention Network (GAT) Processing: Based on the GIN output, a self-attention mechanism is used to dynamically assign different weights to neighboring nodes, enabling the model to focus on more important neighborhood information; attention coefficients The calculation is as follows: in, For activation functions; and These are the feature vectors of nodes x and y after GIN encoding, respectively; W is a learnable linear transformation weight matrix applied to all nodes. It is a learnable single-layer feedforward network weight vector; This indicates vector concatenation; It is the set of neighbors of node x; The new representation of node x is a weighted sum: in, Indicates the activation function; Global average pooling is applied to the final node representation matrix of the view to aggregate the information of all nodes into a graph-level representation vector.

6. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 1, characterized in that, The extraction process of the RNA-protein interaction features includes: First, for the target species, a comprehensive list of RBPs is retrieved from the protein information database. Based on the RBP list, a set of candidate RBPs is obtained, where each RBP is associated with one or more standardized subcellular localization categories. Further screening is performed based on the subcellular regions in the RNA dataset to ensure that the subcellular labels of the RBPs are consistent with the subcellular labels of the RNA. Finally, a fixed panel containing N RBPs that are identical to the subcellular regions in the current RNA dataset is determined. RPI Score Prediction: For an input lncRNA sequence S, multiple RPI prediction tools are used to predict its interaction scores with N RBPs in the panel, resulting in multiple independent score vectors. ; For each RPI prediction tool, the following processing is performed: For RBP identity embedding matrix The p-th row vector of the matrix , representing the vectorized representation of the p-th RBP in the RBP panel; based on the RBP identity embedding matrix and positioning information matrix To construct the query matrix Q; use the RBP identity embedding matrix. As the bond matrix K; and Combining matrices to generate value matrices It employs a standard scaled dot product attention mechanism, utilizing a query matrix Q to query and weight RPI information. Then, global average pooling is performed on the output matrix of the attention network to obtain RNA-protein interaction features.

7. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to any one of claims 1 to 6, characterized in that, The process by which the fusion prediction network fuses features from a multimodal feature set based on a personalized fusion strategy, and ultimately obtains the localization probability of lncRNAs at various subcellular locations, includes: S31, According to the mask Select the enabled feature subset from the multimodal feature set. ; mask With weight guidance Element-wise multiplication yields the effective weight vector. Based on effective weight vector Obtain the modulation weight vector ;Will The feature matrix is ​​obtained by stacking the feature vectors. After two independent linear transformation layers, the original bond moments are generated respectively. Sum matrix Construct a diagonal matrix Its diagonal elements are The corresponding modulation weight values ​​are then used to obtain the key matrix. Sum matrix ; S32, Stack-based fusion block processing: For the first A merge block that receives the bottleneck token from the previous merge block. As input, output the updated bottleneck token. Internally, it performs the following two core operations: (1) Cross-attention information aggregation: bottleneck token of the previous layer's output After layer normalization, it is used as query Q, combining the feature key matrix and value matrix. These are used as keys and values, respectively; through a multi-head cross-attention mechanism (MHCA), feature information is aggregated onto the token and residually joined with the original token to obtain an intermediate token with injected information. ; (2) Self-interaction and integration between tokens: for intermediate tokens Perform multi-head self-attention operations and make residual connections to obtain tokens that have undergone internal information exchange. ,Will Feed into feedforward network Perform a nonlinear transformation and then perform residual connection again to obtain the first... The final output of each fusion block ; The final bottleneck token is obtained by stacking L such fusion blocks. ; S33, will The fused representation vector is obtained by applying average pooling along the token dimension. ,based on The localization probabilities of lncRNAs at predefined subcellular locations are obtained.

8. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 7, characterized in that, Based on effective weight vector The obtained modulation weight vector : The temperature parameter τ is used to adjust the smoothness of the weight distribution. It is an effective weight vector The One element; index and The traversal is of the feature subset selected by the strategy. .

9. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 7, characterized in that, In the process of optimizing the PPO training agent based on the proximal policy, the following methods are adopted: Calculate rewards; It is a hyperparameter. Prediction refers to actions generated by an agent based on training samples. The final subcellular localization prediction probability vector is obtained through fusion prediction; label is the real and known subcellular localization label corresponding to the training sample; , It is the first feature selection mask. Each element.

10. The adaptive multi-channel fusion lncRNA subcellular localization prediction method based on reinforcement learning agents according to claim 9, characterized in that, During the prediction process, the initial bottleneck token in step S32 is obtained at the end of the entire end-to-end training process of the model. The specific process includes: During end-to-end training of the entire model, including reinforcement learning and the fusion prediction network, the initial fusion bottleneck token is set as a learnable parameterized vector. ,Will Used for processing fusion blocks and optimized along with other parameters; at the end of training, the initial fusion bottleneck token corresponding to the best-performing model on the validation set is fixed as the initial bottleneck token in the prediction process.

Images