lncRNA subcellular localization prediction method, device and medium based on graph convolutional network

Abstract

The present invention discloses a method, device, and medium for predicting the subcellular localization of lncRNA based on a graph convolutional network. The method comprises: (1) converting a lncRNA sequence to be tested into a k-mer sequence, constructing a weighted directed graph using each k-mer unit in the converted sequence as a node, and normalizing the edge weights of the graph to obtain a directed graph with normalized edge weights; (2) extracting distributed word vectors of each node in the directed graph based on word2vec technology and embedding them into the directed graph with normalized edge weights; and (3) using a pre-trained prediction model based on a graph convolutional network to output the subcellular localization of the lncRNA sequence to be tested based on the directed graph embedded with the distributed word vectors. The present invention has better prediction performance than traditional machine learning models and existing methods.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically relates to a lncRNA subcellular localization prediction method based on graph convolutional networks. Background Art

[0002] Long noncoding RNAs (lncRNAs), typically containing more than 200 nucleotides, are a crucial class of biological macromolecules. With the rapid development of high-throughput sequencing technologies, increasing evidence indicates that lncRNAs are involved in virtually every aspect of the cellular life cycle, including metabolic processes, epigenetic regulation, cell differentiation and apoptosis, chromosomal abnormalities, and organ or tissue development. For example, lncRNAs regulate gene expression by interacting with chromatin-modifying proteins or transcription factors and their specific protein-binding motifs. lncRNAs can also directly bind to complementary DNA sequences, forming RNA-DNA triplet structures that can block transcription. Furthermore, many human diseases are closely associated with mutations or dysregulation of lncRNAs, including breast cancer, prostate cancer, hepatocellular carcinoma, colon cancer, bladder cancer, thyroid cancer, lung cancer, ovarian cancer, Alzheimer's disease, diabetes, and AIDS. Consequently, research on the functions of lncRNAs in biology has intensified in recent years.

[0003] Research reports indicate that the subcellular localization of lncRNAs varies, and the mechanisms underlying their subcellular localization are also diverse. Understanding the subcellular localization of lncRNAs can provide valuable insights into their functions. For example, the nuclear lncRNA PVT1 interferes with MYC phosphorylation at Thr58 in the nucleus, thereby increasing MYC stability and leading to elevated MYC levels in cancer. The cytoplasmic lncRNA linc-MD1 can inhibit miR-133, thereby affecting the effects of transcription factors on muscle-specific gene expression. Furthermore, lncRNAs located in exosomes are believed to mediate intercellular communication through RNA transport. Therefore, identifying the subcellular localization of lncRNAs is crucial for understanding their biological functions.

[0004] Single-molecule fluorescence in situ hybridization (smFISH) is a biochemical experimental method for determining the subcellular localization of RNA. Although such image data is ideal for determining the localization range of lncRNAs, the technology is expensive, time-consuming, and technically difficult. Given these shortcomings, developing accurate and reliable computational methods to predict the subcellular localization of lncRNAs is extremely valuable to biologists. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a lncRNA subcellular localization prediction method based on graph convolutional network with good prediction performance, in order to address the shortcomings of existing lncRNA subcellular localization prediction methods in that the k-mer frequency features cannot effectively extract high-level sequence features.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A lncRNA subcellular localization prediction method based on graph convolutional networks, comprising:

[0008] (1) Convert the lncRNA sequence to be tested into a k-mer sequence, construct a weighted directed graph with each k-mer unit in the converted sequence as a node, and normalize the edge weights of the graph to obtain a directed graph with normalized edge weights;

[0009] (2) Based on word2vec technology, the distributed word vectors of each node in the directed graph are extracted and embedded into the directed graph with normalized edge weights;

[0010] (3) Using a pre-trained graph convolutional network-based prediction model, the subcellular localization of the lncRNA sequence to be tested is output according to the directed graph embedded with distributed word vectors.

[0011] Furthermore, the specific process of step (1) is as follows:

[0012] The lncRNA sequence to be tested is converted into a k-mer sequence, and each k-mer unit is assigned to a node in sequence, and adjacent k-mer unit nodes are connected using directed edges to obtain an unweighted directed graph;

[0013] The nodes with the same k-mer unit in the unweighted directed graph are glued together, and a weight is assigned to each directed edge based on the frequency of the (k+1)-mer unit formed by the adjacent nodes to obtain a weighted directed graph.

[0014] The edge weights in the weighted directed graph are normalized according to the following calculation formula to obtain a directed graph with normalized edge weights:

[0015]

[0016] Where, e ji represents the original assigned weight of the directed edge from node j to node i, e jk represents the original assigned weight of the directed edge from node j to node k, e ki represents the original assigned weight of the directed edge from node k to node i, N(i) represents the set of all neighbor nodes of node i, and w jirepresents the weight of the directed edge from node j to node i after normalization.

[0017] Furthermore, the specific process of step (2) is as follows:

[0018] All lncRNA sequences in the benchmark dataset were used as a corpus;

[0019] Each lncRNA sequence in the corpus is converted into a k-mer sequence;

[0020] Each k-mer sequence is treated as a sentence, with each k-mer unit as a word. The Skip-gram model in word2vec is used to learn the contextual information of nucleotides in lncRNA. Finally, the distributed word vector of each k-mer unit is obtained to obtain the word vector library.

[0021] The distributed word vectors of each k-mer unit in the lncRNA sequence to be tested are obtained from the word vector library and embedded into each node of the directed graph with normalized edge weights.

[0022] Furthermore, the prediction model based on graph convolutional network is constructed by stacking multiple layers of GCN and fully connected layers with activation functions.

[0023] Furthermore, k=4.

[0024] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor implements any of the above-mentioned lncRNA subcellular localization prediction methods based on graph convolutional networks.

[0025] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned lncRNA subcellular localization prediction methods based on graph convolutional networks.

[0026] Beneficial effects

[0027] Current existing technical methods often rely on k-mer frequency features to encode lncRNA sequences. In machine / deep learning models, how to encode the original lncRNA sequence into discriminant features is one of the most important issues. However, there are some defects in using k-mer frequency features to encode lncRNA sequences. (1) It can only reflect frequency information while ignoring sequence order information; (2) When k is fixed, it cannot capture motif fragments of different lengths. (3) Moreover, when k is small, the encoding method cannot obtain sufficient feature information or cannot capture useful features, resulting in underfitting of the prediction model. When k is large, the dimension of the encoding vector grows exponentially, making the encoding vector sparse, wasting computing resources, and may cause overfitting problems.

[0028] Unlike previous studies that used k-mer frequency features to encode lncRNA sequences, the present invention converts lncRNA sequences into de Bruijn graphs, which can provide more comprehensive information. In the de Bruijn graph, the nodes of the graph are k-mer units, and the direction of the edges is determined by the sequence order. Then, the pre-trained k-mer word2vec embedding vectors are used as node features, and the edges are weighted. Next, a graph convolutional network is used to learn implicit representations and extract high-level features from the de Bruijn graph. Finally, a fully connected layer is used to perform the prediction task. Therefore, the advantages of converting sequences into graphs in the present invention are as follows:

[0029] (1) The local ordered information of lncRNA sequences can be preserved in the form of a directed graph.

[0030] (2) By connecting multiple nodes in the graph to form a path, motifs of different lengths in lncRNA sequences can be automatically captured.

[0031] (3) Through the aggregation operation, multiple adjacent nodes can be aggregated to form communities and subgraphs, thereby capturing the global and high-level features of the entire lncRNA sequence.

[0032] (4) Data from other types can be integrated as node features to provide more comprehensive feature encoding for lncRNA sequences.

[0033] Therefore, the present invention converts lncRNA sequences into graphs with more distinguishable features and is more robust, thereby more accurately predicting the subcellular localization of lncRNA. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a diagram of the overall architecture of the method described in the embodiments of the present application; (a) graph construction, (b) word vector extraction, (c) feature extraction based on GCN, and (d) full connection and activation classification;

[0035] Figure 2 This is a t-SNE visualization of the graph vector and 4-mer frequency features used in the experimental validation of this application. Each point represents a sample, and its color represents its true class: (a) using 4-mer frequency features, (b) using graph vectors;

[0036] Figure 3 This is the performance of GraphLncLoc on the original and mutated datasets under different mutation rates in the experimental verification of this application, where (a) the mutation rate is 0.0001, (b) the mutation rate is 0.001, and (c) the mutation rate is 0.01. DETAILED DESCRIPTION

[0037] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present invention.

[0038] This example provides a lncRNA subcellular localization prediction method based on graph convolutional network. Figure 1 As shown, the following steps are included:

[0039] 1. Graph construction: converting lncRNA sequences into directed graphs

[0040] Formally, for a lncRNA sequence:

[0041] lncRNA=N1,N2,N3,...,N L-1 ,N L

[0042] Where L represents the length of lncRNA, N j It is one of the four nucleotide bases (A, C, G and U) at position j in the lncRNA sequence. Its k-mer component set (here taking 4-mer as an example) is {N1N2N3N4,N2N3N4N5,N3N4N5N6,...,N L-3 N L-2 N L-1 N L These 4-mers are then assigned to nodes, following the order of the k-mer set (from left to right), adding one 4-mer at a time, and using these 4-mers to reconstruct the lncRNA sequence. After the reconstruction process is complete, nodes with the same label are glued together to form a de Bruijn graph.

[0043] Then, each directed edge is assigned a weight, which is the frequency of the (k+1)-mer consisting of the two nodes that make up this edge.

[0044] In order to reduce the impact of the absolute difference between the edge weight frequency values, the weights of the edges in the graph are further standardized, that is, the frequency weights are normalized:

[0045]

[0046] Where, e ji represents the frequency weight of the edge from node j to node i, and N(i) represents the set of neighbor nodes of node i.

[0047] 2. Word Vector Embedding

[0048] This example uses a continuous distributed vector representation of k-mers as node features. The k-mer units in lncRNA sequences are similar to words in a text, so using a continuous distributed word vector representation of k-mers can naturally represent the contextual information of nucleotides in lncRNAs.

[0049] Specifically, this embodiment uses all lncRNA sequences in the benchmark data set as a corpus, and applies word2vec technology to obtain the encoding vector of each 4-mer unit in the lncRNA sequence corpus, that is, the distributed word vector, as the node feature vector of the graph. The Skip-gram model in word2vec technology is used to predict the surrounding words of a given central word, that is, the context information, as the distributed word vector of the k-mer unit shown by the central word. Following the idea of the Skip-Gram model, word2vec technology aims to maximize the possibility of co-occurrence between the target 4-mer and its context 4-mer. By using the continuous distributed vector representation of k-mer as a node feature, the semantic information of the constructed de Bruijn graph is enriched.

[0050] After extracting the distributed word vector of each k-mer unit in the benchmark database as described above, the distributed word vector of each k-mer unit in the lncRNA sequence to be tested can be obtained from the word vector library and embedded into each node of the directed graph with normalized edge weights.

[0051] 3. Positioning Prediction

[0052] After constructing a de Bruijn graph of the lncRNA sequence to be tested and obtaining the distributed word vectors of each node, a GCN model is used to extract high-level features. The prediction model based on the graph convolutional network used in this embodiment is constructed by stacking multiple layers of GCN and fully connected layers with activation functions.

[0053] GCN can learn the deep topological structure and node features of the graph by performing convolution operations on the graph. In GCN, the propagation rule can be expressed by the following formula:

[0054]

[0055] in is the adjacency matrix of the graph with self-connections added. N is the identity matrix, yes The degree matrix, W (l) represents the weight of layer (l), H (l) represents the activation matrix of layer (l), and σ represents the nonlinear activation function.

[0056] The main idea of the GCN layer is to learn a transformation function that generates a new embedding matrix H for node i by aggregating the features of node i itself and its neighbors in the directed graph G and considering the normalized edge weights. (l+1) By stacking multiple layers of GCN, information transfer between nodes can be achieved and high-level features of the graph can be captured. Specifically, GCN aggregates the embedding matrices of all nodes or edges and takes their average as the final graph encoding vector. The average aggregation formula of its node features is as follows:

[0057]

[0058] where h G is the encoding vector of the directed graph G, V is the set of all nodes in the directed graph G, h v It is the embedding vector obtained after node v is learned by stacking multiple layers of GCN.

[0059] Finally, we will use the high-level features extracted from the de Bruijn graph using GCN, i.e., the encoding vector h G , fed into the fully connected layer to perform the classification task.

[0060] IV. Experimental Verification

[0061] In order to verify the effectiveness of the present invention [hereinafter referred to as GraphLncLoc] for predicting lncRNA subcellular localization and its performance superiority compared to other methods, this section evaluates the performance of GraphLncLoc through extensive experiments. Comparison with traditional machine learning classifiers using different k-mer frequency features shows the benefits of converting sequences into graphs. Comparison with existing prediction methods shows the effectiveness of GrahLncLoc in predicting lncRNA subcellular localization. In addition, some analyses were performed. The analysis showed that GrahLncLoc can produce more distinguishable features than k-mer frequency features. GrahLncLoc is also more stable than methods using k-mer frequency features.

[0062] First, ACC, MacroPrecision, MacroRecall, MacroF1-score, and AUC are used as evaluation indicators. The calculation formulas for MacroPrecision, MacroRecall, and MacroF1-score are as follows:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] Among them TP (i) , FP (i) 、FN (i) Indicates the number of true positive cases, false positive cases, and false negative cases in the i-th category. (i) and recall (i) represents the precision and recall of the i-th category, and n is the number of sample categories.

[0069] Table 1 Performance comparison of GraphLncLoc and different machine learning models using different kmer frequency features

[0070]

[0071] Table 2 Performance comparison of GraphLncLoc and existing prediction methods on the test set

[0072]

[0073] Table 1 first focuses on the results of the machine learning models. Looking at the MacroF1-score, SVM, RF, LR, and NN achieve the highest MacroF1-score values when k = 3, k = 4, k = 6, and k = 5, respectively. These results indicate that different machine learning classifiers have preferred k values for optimal performance. Secondly, GraphLncLoc achieves higher performance across all evaluation metrics than other machine learning classifiers using k-mer frequency features. The best machine learning classifier is the RF model with k = 3, which achieves an ACC of 0.572, a Macro F1-score of 0.391, a Macro Precision of 0.511, and a Macro Recall of 0.380. GraphLncLoc outperforms the 3-mer RF model in terms of ACC (0.612), Macro F1-score (0.506), Macro Precision (0.691), and Macro Recall (0.475). In summary, the results show that GraphLncLoc performs better than these traditional machine learning classifiers using different kmer frequency features, which reflects the advantage of using graph vectors.

[0074] As shown in Table 2, GraphLncLoc outperformed other prediction methods in all evaluation metrics. In terms of ACC, GraphLncLoc reached 0.579, significantly higher than lncLocator (0.421), iLoc-lncRNA (0.509), Locate-R (0.368), DeepLn-cLoc (0.561), and iLoc-lncRNA2.0 (0.404). Other evaluation metrics (MacroPrecision, MacroRecall, Macro F1-score, AUC) showed similar results. These results indicate that GraphLncLoc is an effective tool for predicting lncRNA subcellular localization.

[0075] To show the difference between graph vectors and 4-mer frequency features, the embedding space of the two features was projected onto two dimensions using t-distributed stochastic neighbor embedding (t-SNE) for visualization. Different subcellular localization categories are marked with different colors. Figure 2 As shown in Figure 3, the use of graph vectors can better separate samples between different categories of cytoplasm, nucleus, ribosomes, and exosomes compared to using 4-mer frequency features. These results demonstrate the benefits of using graph vectors, meaning that the learned graph vectors can clearly distinguish different cellular localizations.

[0076] To further demonstrate the advantages of converting sequences into graphs, we conducted several experiments to test the robustness of GraphLncLoc and k-mer frequency features. Ideally, a good feature representation should remain largely unchanged if a lncRNA sequence undergoes minor changes. In other words, a good feature representation should be robust to the small changes that actually occur during sequencing. To quantitatively measure this, we designed the following experiments.

[0077] First, a "mutation" dataset is generated from the original dataset by introducing three mutation behaviors, including insertion, deletion, and mutation. Specifically, the generation process of the "mutation" dataset is as follows:

[0078] 1. Set a point mutation rate M.

[0079] 2. For each nucleotide in the lncRNA sequence, a probability is randomly generated. If the probability is greater than the point mutation rate M, the nucleotide remains unchanged; if the probability is less than or equal to the point mutation rate M, one of three actions is randomly performed to change the nucleotide.

[0080] Action 1 (insertion): randomly insert a nucleotide (A, U, C, G) before the nucleotide.

[0081] Action 2 (deletion): Delete the nucleotide in the sequence.

[0082] Action 3 (mutation): Randomly change the nucleotide to three other nucleotides.

[0083] 3. Repeat step 2 for all lncRNA sequences in the benchmark dataset until all sequences have been "mutated".

[0084] It is worth noting that the labels of lncRNA sequences did not change during the generation process. After the generation process, a "mutation" dataset was obtained from the original dataset. Then, GraphLncLoc and k-mer frequency features were used to encode the sequences in the "mutation" dataset and compare the differences with the original dataset. Since GraphLncLoc uses 4-mer as a node and the RF model achieved the best performance among traditional machine learning models (see Table 1), 4-mer+RF was used as a baseline for comparison. By querying some databases, it was found that the mutation rate of the human genome was estimated to be approximately 1×10 -8 However, a mutation rate that is too small has little effect on the machine learning model. Therefore, the point mutation rate is set from 0.01 to 0.0001 (0.01, 0.001, 0.0001) to observe the difference between the original dataset and the "mutated" dataset. The results are shown in Figure 3 middle.

[0085] from Figure 3 It can be observed that when M is 0.0001, the results are basically unchanged. When M is 0.001, in terms of MacroF1-score, 4-mer+RF drops from 0.377 to 0.325 (about 13.8%), while GraphLncLoc only drops from 0.506 to 0.493 (about 2.6%). When M is 0.01, in terms of MacroF1-score, 4-mer+RF drops from 0.377 to 0.300 (about 20.4%), while GraphLncLoc only drops from 0.506 to 0.450 (about 11.1%). It can be seen that GraphLncLoc is more robust than 4-mer+RF, which means that the robustness of using graph vectors is better than using k-mer frequency features. Other evaluation indicators (MacroPrecision, MacroRecall) show similar results. Therefore, the evaluation of GraphLncLoc on the "Mutation" dataset confirms its robustness.

[0086] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A lncRNA subcellular localization prediction method based on graph convolutional network, characterized in that: include: (1) Convert the lncRNA sequence to be tested k -mer format sequence, with each k -mer units are used as nodes to construct a weighted directed graph, and the edge weights of the graph are normalized to obtain a directed graph with normalized edge weights; The specific process of step (1) is as follows: Convert the lncRNA sequence to k -mer form sequence, each k -mer units are assigned to nodes and adjacent k -mer unit nodes are connected using directed edges, and k The nodes with the same -mer unit are glued into one, and an unweighted directed graph is obtained, which is the de Bruijn graph of the lncRNA sequence to be tested; Based on adjacent nodes ( k +1)-mer unit frequency, assign weights to each directed edge in the unweighted directed graph to obtain a weighted directed graph; The edge weights in the weighted directed graph are normalized according to the following calculation formula to obtain a directed graph with normalized edge weights: ； Where, Represents a slave node Pointing to a node The original assigned weights of the directed edges, Represents a slave node Pointing to a node The original assigned weights of the directed edges, Represents a slave node Pointing to a node The original assigned weights of the directed edges, Representation node The set of all neighbor nodes of Represents a slave node Pointing to a node The weight of the directed edge after normalization; (2) Based on word2vec technology, the distributed word vectors of each node in the directed graph are extracted and embedded into the directed graph with normalized edge weights; (3) Using a pre-trained graph convolutional network-based prediction model, the subcellular localization of the lncRNA sequence to be tested is output according to the directed graph embedded with distributed word vectors.

2. The lncRNA subcellular localization prediction method based on graph convolutional network according to claim 1, characterized in that: The specific process of step (2) is as follows: All lncRNA sequences in the benchmark dataset were used as a corpus; Each lncRNA sequence in the corpus is converted into k -mer-form sequences; Each k -mer form as a statement, where each k -mer units are used as words, and the Skip-gram model in word2vec is used to learn the context information of nucleotides in lncRNA. Finally, each k -The distributed word vector of the mer unit is used to obtain the word vector library; Get the words in the lncRNA sequence to be tested from the word vector library k The distributed word vectors of the -mer units are embedded into each node of the directed graph with normalized edge weights.

3. The lncRNA subcellular localization prediction method based on graph convolutional network according to claim 1, characterized in that: The prediction model based on graph convolutional network is constructed by stacking multiple layers of GCN and fully connected layers with activation functions.

4. The lncRNA subcellular localization prediction method based on graph convolutional network according to claim 1, characterized in that: k =4 。 5. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the computer program is executed by the processor, the processor is caused to implement the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images