RNA (Ribonucleic Acid) far homologous detection method and system based on deep learning

Abstract

The invention discloses an RNA (Ribonucleic Acid) far homology detection method and system based on deep learning, and belongs to the field of bioinformatics, an RNA sequence data set is divided into a plurality of sets and then a plurality of training data sets containing far homology pairs are constructed, so that an RNA far homology detection model is trained on the basis of each training data set, a plurality of model variants are obtained, and the RNA far homology detection model is obtained. Performance evaluation is carried out based on the independent test set to obtain a model with optimal performance; in the application stage, an RNA sequence data set is converted into vector representation based on the model, a vector database is constructed, after a to-be-detected sequence is received, the to-be-detected sequence is converted into vector representation based on the same mode, then the vector database is retrieved, similar RNA sequences including far homologous sequences are obtained, and an end-to-end deep learning framework is achieved. And the RNA sequence can be directly mapped to the structural similarity without depending on the structural information of RNA, so that the semantic gap between the sequence and the structure is effectively bridged.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, and more specifically, relates to a method and system for detecting distant homology of RNA based on deep learning. Background Technology

[0002] Homology detection is one of the core challenges in bioinformatics, crucial for understanding the function, evolution, and structure of proteins and nucleic acids. Sequence similarity-based methods have been widely applied in areas such as structural design, functional annotation, and interaction prediction. However, the performance of these traditional methods degrades significantly when sequence similarity is low (distant homology).

[0003] In recent years, pre-trained language models have brought revolutionary progress to this field. In the protein domain, models such as TM-Vec, DeepBLAST, and PLMsearch have greatly improved the sensitivity of distant homology detection by mining deep semantic features of sequences. However, transferring this successful paradigm to RNA research faces unique challenges: RNA consists of only four bases, and its information density is far lower than that of proteins, which are composed of 20 amino acids; at the same time, the available high-quality RNA sequence data is relatively limited, which restricts the model's ability to learn complex structural information.

[0004] When the three-dimensional structure of RNA is known, structure-based alignment tools (such as USalign and RMalign) can provide the most reliable homology determination. However, the practical application of this strategy is limited by two major bottlenecks: first, the structure of most RNAs is not experimentally resolved; second, relying solely on single sequences for high-precision three-dimensional structure prediction is still computationally expensive and difficult to support rapid retrieval of large-scale genome databases.

[0005] To reduce reliance on known or predicted structures, researchers have developed various homology search pipelines, such as the RNAcmap series and its improved versions. These methods, by integrating tools like BLAST and Infernal and iteratively searching large databases to construct multiple sequence alignments, have improved detection sensitivity to some extent. However, their computational cost is extremely high, and their performance is heavily dependent on the accuracy of the predicted secondary structure, thus limiting their large-scale applications. Summary of the Invention

[0006] To address the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for RNA distant homology detection based on deep learning. This solves the technical problems of existing RNA distant homology detection methods, which heavily rely on computationally expensive structural alignment or homology search, have low computational efficiency, cannot support rapid retrieval of large-scale databases, and exhibit significantly reduced detection performance under conditions of low sequence similarity.

[0007] To achieve the above objectives, according to a first aspect of the present invention, a deep learning-based method for detecting distant homology of RNA is provided, comprising: An RNA sequence dataset is obtained. Based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence dataset, the RNA sequence dataset is divided into multiple sets with different biological meanings. Multiple training datasets containing distantly homologous RNA pairs are constructed based on the multiple sets with different biological meanings. The RNA distant homology detection model was trained on each training dataset to obtain multiple variants of the RNA distant homology detection model. The performance of each variant was evaluated on an independent test set. Based on the performance evaluation results of each variant, the RNA distant homology detection application model was selected. The RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences. Based on the RNA distant homology detection application model, each RNA sequence in the RNA sequence dataset is converted into a vector representation to construct a vector database. After receiving the RNA sequence to be detected, the RNA sequence to be detected is converted into a vector representation based on the RNA distant homology detection application model. Then, the vector database is searched based on the vector of the RNA sequence to be detected to obtain the similar RNA sequence corresponding to the RNA sequence to be detected, including distant homology sequences.

[0008] Based on the aforementioned deep learning-based RNA distant homology detection method, the RNA sequence dataset is divided into multiple sets with different biological significance based on the sequence identity and structural similarity between each pair of RNA sequences. Multiple training datasets and independent test sets containing distant homologous RNA pairs are then constructed based on these sets, specifically including: Based on sequence identity thresholds and structural similarity thresholds, RNA sequence pairs formed by each RNA sequence in the RNA sequence dataset are divided into four sets with different biological meanings. Specifically, RNA sequence pairs with both sequence identity and structural similarity greater than the sequence identity threshold are divided into the first set, RNA sequence pairs with both sequence identity and structural similarity less than the sequence identity threshold are divided into the second set, RNA sequence pairs with both sequence identity and structural similarity less than the sequence identity threshold are divided into the third set, and RNA sequence pairs with both sequence identity and structural similarity greater than the sequence identity threshold are divided into the fourth set. A first training dataset is constructed based on the complete second set, a second training dataset is constructed based on the complete second set and partial RNA sequence pairs in the third set, and a third training dataset is constructed based on the complete first set, the second set, the fourth set, and partial RNA sequence pairs in the third set.

[0009] According to the above-mentioned deep learning-based RNA distant homology detection method, the RNA distant homology detection model includes an input representation layer, a feature encoder layer, and a similarity calculation layer connected in sequence. The input representation layer is constructed based on the RNA language model and is used to extract the embeddings of the input RNA sequence pairs to obtain the RNA embedding representations of each RNA sequence in the RNA sequence pair. The feature encoder layer is built based on the Transformer encoder layer and is used to extract sequence features from the RNA embedding representation of each RNA sequence in the RNA sequence pair to obtain the RNA sequence feature vector of each RNA sequence in the RNA sequence pair. The similarity calculation layer is used to calculate the similarity between the RNA sequence feature vectors of each RNA sequence in the RNA sequence pair, which is used as the structural similarity of the RNA sequence pair.

[0010] According to the above-mentioned deep learning-based RNA distant homology detection method, the feature encoder layer includes two stacked Transformer encoder layers, an average pooling layer, a dropout layer, and a fully connected layer.

[0011] Based on the aforementioned deep learning-based RNA distant homology detection method, the RNA distant homology detection model is trained using any training dataset, specifically including: Each RNA sequence pair in any of the training datasets is sequentially input into the RNA distant homology detection model to obtain the structural similarity of the corresponding RNA sequence pairs predicted by the RNA distant homology detection model. Based on the difference between the structural similarity of each RNA sequence pair predicted by the RNA distant homology detection model and the actual structural similarity of the corresponding RNA sequence pair, the model loss is calculated, and the parameters of the feature encoder layer in the RNA distant homology detection model are adjusted based on the model loss.

[0012] Based on the aforementioned deep learning-based RNA distant homology detection method, the performance evaluation of various RNA distant homology detection model variants using independent test sets specifically includes: The RNA sequence pairs in the independent test set are input into each RNA distant homology detection model variant to obtain the structural similarity of each RNA sequence pair predicted by each RNA distant homology detection model variant. Based on the structural similarity of each RNA sequence pair predicted by each RNA distant homology detection model variant and the actual structural similarity of the corresponding RNA sequence pair, the performance index value of each RNA distant homology detection model variant is calculated; the performance index value includes at least one of Pearson correlation coefficient, Spearman correlation coefficient, precision, recall, F1-score, area under the receiver operating characteristic curve and area under the precision-recall curve.

[0013] According to the aforementioned deep learning-based RNA distant homology detection method, obtaining the RNA sequence dataset specifically includes: Multiple RNA structure data were obtained from a protein data bank, and the RNA structure data were subjected to length filtering and redundancy removal to obtain a non-redundant RNA sequence dataset.

[0014] According to a second aspect of the present invention, a deep learning-based RNA distant homology detection system is provided, comprising: The dataset construction unit is used to acquire RNA sequence datasets, divide the RNA sequence datasets into multiple sets with different biological meanings based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence datasets, and construct multiple training datasets containing distantly homologous RNA pairs based on the multiple sets with different biological meanings. The model training and evaluation unit is used to train the RNA distant homology detection model on each training dataset to obtain multiple RNA distant homology detection model variants, and to evaluate the performance of each RNA distant homology detection model variant on an independent test set. Based on the performance evaluation results of each RNA distant homology detection model variant, the application model for RNA distant homology detection is selected. The RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences. The database construction unit is used to convert each RNA sequence in the RNA sequence dataset into a vector representation based on the RNA distant homology detection application model, and construct a vector database. The distant homology detection unit is used to, after receiving the RNA sequence to be detected, convert the RNA sequence to be detected into a vector representation based on the RNA distant homology detection application model, and then search the vector database based on the vector of the RNA sequence to be detected to obtain similar RNA sequences corresponding to the RNA sequence to be detected, including distant homology sequences.

[0015] According to a third aspect of the present invention, an electronic device is provided, comprising: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0016] According to a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the first aspect.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: By acquiring an RNA sequence dataset, and dividing the RNA sequences into multiple sets with different biological meanings based on the sequence identity and structural similarity between each pair of RNA sequences, multiple training datasets containing distantly homologous RNA pairs are constructed based on these sets. The RNA distant homology detection model is then trained on each training dataset, resulting in multiple variants of the RNA distant homology detection model. The performance of each variant is evaluated on an independent test set to obtain the best-performing RNA distant homology detection application model. In the application phase, based on this RNA distant homology detection application model, each RNA sequence in the RNA sequence dataset is converted into a vector representation, and a vector database is constructed. Upon receiving a target RNA sequence, the target RNA sequence is converted into a vector representation in the same way, and the vector database is retrieved to obtain similar RNA sequences, including distantly homologous sequences, corresponding to the target RNA sequence. This achieves an end-to-end deep learning framework, enabling direct mapping from RNA sequences to structural similarity without relying on RNA structural information, effectively bridging the semantic gap between sequence and structure.

[0018] Furthermore, in this embodiment of the invention, a pre-trained RNA language model is combined with a Transformer encoder for RNA distant homology detection, which can effectively capture the deep structural features of RNA sequences and maintain excellent detection performance even under low sequence similarity conditions. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the deep learning-based RNA distant homology detection method provided in this embodiment of the invention; Figure 2 A scatter plot of sequence identity-structural similarity provided in an embodiment of the present invention; Figure 3 This is a model architecture diagram of the RNA distant homology detection model provided in an embodiment of the present invention; Figure 4 This is a graph showing the performance evaluation results of far-homogeneous detection on an independent test set provided in this embodiment of the invention; Figure 5 Scatter plot of Pearson correlation coefficients between structural similarity and TM scores predicted by the model provided in this embodiment of the invention; Figure 6 A prediction accuracy analysis diagram of different sequence identity intervals provided in an embodiment of the present invention; Figure 7 A flowchart illustrating the construction and retrieval process of a large-scale RNA vector database provided in this embodiment of the invention; Figure 8 This is an analysis diagram of a distant homology detection example provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] This invention provides a deep learning-based method for detecting distant homology of RNA, such as... Figure 1 As shown, it includes: Step 110: Obtain the RNA sequence dataset. Based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence dataset, divide the RNA sequence dataset into multiple sets with different biological meanings, and construct multiple training datasets containing distantly homologous RNA pairs based on the multiple sets with different biological meanings. Step 120: Train the RNA distant homology detection model on each training dataset to obtain multiple RNA distant homology detection model variants, and evaluate the performance of each RNA distant homology detection model variant on an independent test set. Select the RNA distant homology detection application model based on the performance evaluation results of each RNA distant homology detection model variant; wherein, the RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences; Step 130: Based on the RNA distant homology detection application model, convert each RNA sequence in the RNA sequence dataset into a vector representation and construct a vector database; Step 140: After receiving the RNA sequence to be detected, the RNA sequence to be detected is converted into a vector representation based on the RNA distant homology detection application model. Then, the vector database is searched based on the vector of the RNA sequence to be detected to obtain similar RNA sequences corresponding to the RNA sequence to be detected, including distant homology sequences.

[0022] Specifically, RNA sequences can be systematically screened and downloaded from the Protein Data Bank, containing ≥1 RNA instance or strand with a resolution ≤4.5 Å, to construct an RNA sequence dataset. In some embodiments, to ensure model processing efficiency and reduce sequence redundancy, after acquiring multiple RNA structure data from the Protein Data Bank, these RNA structure data can be length-filtered and redundancy-removed to obtain a non-redundant RNA sequence dataset. Specifically, RNA sequences with a length not exceeding 1024 nucleotides can be selected, and redundancy removal can be performed using the cd-hit-est tool (with a sequence identity threshold set to 0.8), ultimately yielding an RNA sequence dataset containing non-redundant standard nucleotide sequences.

[0023] Subsequently, RNA sequences in the RNA sequence dataset were paired to generate all possible RNA sequence pairs. Sequence identity of the RNA sequence pairs was calculated based on global sequence alignment. The ratio of the number of matching residues between two RNA sequences to their total length can be used as the sequence identity of the RNA sequence pair. Additionally, the USalign tool was used to calculate the TM score (TM-score) of the RNA sequence pairs as their structural similarity. A scatter plot of sequence identity versus structural similarity was then plotted based on the sequence identity and structural similarity of each RNA sequence pair. Figure 2 As shown in the figure, the horizontal axis of the scatter plot represents sequence identity, and the vertical axis represents structural similarity (TM-score). Dividing the scatter plot yields multiple biologically significant regions, thus forming multiple datasets with different biological meanings. These datasets can then be used to construct multiple training datasets containing distantly homologous RNA pairs. The RNA distant homology detection model can then be trained on each of these datasets, resulting in multiple variants of the RNA distant homology detection model.

[0024] In some embodiments, sequence identity thresholds and structural similarity thresholds can be set to divide RNA sequence pairs in the RNA sequence dataset into four sets with different biological meanings. For example, the sequence identity threshold can be set to 0.6 (below this value indicates a weakened association between sequence and structural conservation), and the structural similarity threshold can be set to 0.45 (this cutoff value serves as the threshold between RNA structural similarity and dissimilarity). RNA sequence pairs with both sequence identity and structural similarity greater than the sequence identity threshold are assigned to the first set; those with less than the sequence identity threshold but greater than the structural similarity threshold are assigned to the second set; those with less than the sequence identity threshold and less than the structural similarity threshold are assigned to the third set; and those with both are assigned to the fourth set. The biological meanings corresponding to these four sets are shown in Table 1.

[0025]

[0026] As shown in Table 1, the first set contains 1254 RNA sequence pairs, which show high values ​​for both sequence similarity and structural similarity. The second set contains 5621 RNA sequence pairs with low sequence similarity but high structural similarity, i.e., distant homologous pairs, which are the focus of this embodiment of the invention. The third set contains 739442 RNA sequence pairs, in which each pair of RNA sequences is randomly arranged and has different structures and sequences. The fourth set contains 935 RNA sequence pairs with similar sequences but large structural differences.

[0027] To comprehensively and progressively evaluate the model performance of different model variants and select the best-performing model for RNA sequence distant homology detection, a first training dataset can be constructed based on the complete second set, a second training dataset can be constructed based on partial RNA sequence pairs from the complete second and third sets, and a third training dataset can be constructed based on partial RNA sequence pairs from the complete first, second, and fourth sets and the third set.

[0028] The first training dataset, containing all long-range homology RNA sequence pairs from the second set, is designed to train the model to identify core long-range homology relationships. Based on this, it can be randomly divided into training, validation, and test sets in a 9:1:1 ratio to train the RNA long-range homology detection model. Thus, the first training dataset is used to train the model to identify the core feature of long-range homology relationships in RNA, namely, structural conservation despite high sequence differentiation. Therefore, the model variant trained using the first training dataset focuses on learning the essential features of long-range homology relationships, establishing a discriminative basis for the model based on structural conservation.

[0029] The second training dataset introduces negative samples on top of the first training dataset to enhance the model's discriminative ability. It contains all RNA sequence pairs from the second set and randomly selects multiple RNA sequence pairs (e.g., 3000 pairs) from the third set. Similarly, it can be randomly divided into training, validation, and test sets in a 9:1:1 ratio to train the RNA distant homology detection model. As can be seen, the second training dataset introduces negative samples on top of the first training dataset to improve the model's discriminative ability. The model variant trained using the second training dataset can establish clear decision boundaries. By introducing negative samples with well-defined features, the model is forced to learn to distinguish between "structurally related" and "structurally unrelated" sequences, adding a discriminative ability for correlation to simple similarity learning.

[0030] The third training dataset encompasses representative samples from all four sets. It includes all data from the first, second, and fourth sets, and randomly selects multiple RNA sequence pairs (e.g., 3000 pairs) from the third set. Subsequently, it is divided into training, validation, and test sets in equal proportions. As can be seen, the third training dataset covers representative samples from all four sets, enabling the model to adapt to various complex relationships present in real-world biological data. In particular, by introducing samples from the fourth set, the model is forced to learn to weigh sequence and structural evidence, and to suppress deceptive sequence similarity signals when the two conflict, thus significantly improving the model's robustness in real-world complex scenarios.

[0031] By adopting a progressive training strategy using the first training dataset, the second training dataset, and the third training dataset, the model can learn core biometric features, establish discrimination boundaries, and ultimately make robust judgments in complex scenarios. This effectively solves the problem of misjudgment caused by sequence similarity in RNA distant homology detection, and significantly improves the detection accuracy.

[0032] It should be noted that the aforementioned RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between them. The core idea of ​​the model is to learn the mapping relationship from RNA sequences to structural similarity, where structural similarity can be supervised by the TM-score. The TM-score can be calculated using the USalign tool, with a value ranging from 0 to 1. Values ​​below the structural similarity threshold (e.g., 0.45) indicate structural irrelevance, values ​​above the threshold indicate structural similarity, and 1 indicates a perfect match. In some embodiments, the model architecture of the RNA distant homology detection model includes three core components: an input representation layer, a feature encoder layer, and a similarity calculation layer, such as... Figure 3 As shown.

[0033] The input representation layer is constructed based on the RNA language model and is used to extract the embeddings from input RNA sequence pairs, obtaining the RNA embedding representations of each RNA sequence in the RNA sequence pair. In some embodiments, the RNA-FM language model can be used to extract the embeddings from the input RNA sequence pairs, and each RNA sequence is transformed into an (L, 640) RNA embedding representation through the RNA-FM model. Here, L represents the length of the RNA sequence, and the output dimension is 640. This embedding representation can effectively capture the evolutionary conservation and structural features of RNA sequences.

[0034] The feature encoder layer, built upon the Transformer encoder layer, extracts sequence features from the RNA embedding representations of each RNA sequence in the RNA sequence pair, yielding RNA sequence feature vectors for each RNA sequence in the RNA sequence pair. In some embodiments, the feature encoder layer includes two stacked Transformer encoder layers, an average pooling layer, a dropout layer (dropout rate set to 0.2), and a fully connected layer to effectively capture global features of the RNA sequence. The number of Transformer encoder layers is specified by the `num_layers` parameter, which is set to 2. Each Transformer encoder layer includes a multi-head attention layer and a feedforward network layer. This design enables the feature encoder layer to capture long-range dependencies and global features in the RNA sequence, and the stacking of Transformer encoder layers allows the feature encoder layer to progressively extract higher-level sequence features. The multi-head attention layer captures dependencies between different positions in the input RNA sequence; its number of heads is specified by the `nhead` parameter, which can be set to 4, determining the feature encoder layer's ability to compute the attention mechanism in parallel. The dimension of the feedforward network layer is specified by the `dim_feedforward` parameter, which can be set to 2048 to enhance representational capabilities through non-linear transformations.

[0035] The similarity calculation layer is used to calculate the similarity between the RNA sequence feature vectors of each RNA sequence in an RNA sequence pair, which is used as the structural similarity of the RNA sequence pair.

[0036] As can be seen, compared with traditional RNA distant homology detection methods, the RNA distant homology detection model provided in this embodiment of the invention only requires RNA sequence information to calculate structural similarity and achieve RNA distant homology detection. Specifically, the RNA distant homology detection model can effectively capture the hidden structural information and evolutionary conservation between RNA sequences by using an architecture of an input representation layer and a feature encoder layer. The addition of the feature encoder layer enables context-based comparison, which can explicitly model global dependencies and capture long-range interactions between RNAs and dependencies between different positions.

[0037] When training an RNA distant homology detection model based on any training dataset, each RNA sequence pair in the training dataset can be sequentially input into the model to obtain the structural similarity predicted by the model for each RNA sequence pair. Based on the difference between the predicted structural similarity and the actual structural similarity of the corresponding RNA sequence pairs, the model loss is calculated. This model loss is then used to adjust the parameters of the feature encoder layer in the RNA distant homology detection model. The Adam optimizer can be used for training, with a learning rate set to 1×10⁻⁶. -4 The entire RNA distant homology detection model can be trained on a single GPU.

[0038] The training objective of this RNA distant homology detection model is to minimize the difference between the predicted structural similarity of RNA sequence pairs and the actual structural similarity (i.e., TM-score), which is to minimize the L-distance of all samples.

[0039] Where N is the number of RNA sequence pairs, h (1) and h (2) This represents the RNA sequence feature vector of the two RNA sequences in an RNA sequence pair. similarity () is used to measure the directional similarity between two RNA sequences:

[0040] The three RNA distant homology detection model variants (Model-1, Model-2, and Model-3) trained on three training datasets exhibit different convergence patterns due to the varying complexity of the datasets used for training. Model-1, trained on the first training dataset, converged the fastest; Model-2, trained on the second training dataset incorporating the third set of data, required a longer period to fully converge; and Model-3, trained on the third training dataset, required the longest period to fully converge (this final model will be referred to as Model-3). 2 It is worth noting that during Model-3 training, checkpoints where its performance is excellent can be obtained. These checkpoints (denoted as Model-3) are the ones where the Model-3 performs exceptionally well during training. 1 It exhibits the highest Pearson correlation coefficient on the validation set, and therefore can also be included in subsequent performance evaluation.

[0041] After training multiple RNA distant homology detection model variants, their performance can be evaluated using an independent test set (constructed from the authoritative independent database RR3DD, ensuring that all RNA sequences in this set do not appear in any of the training datasets). Based on the performance evaluation results, the variant with the best performance is selected as the application model for RNA distant homology detection. During performance evaluation, RNA sequence pairs from the independent test set are input into each RNA distant homology detection model variant to obtain the predicted structural similarity of each RNA sequence pair. The performance index of each RNA distant homology detection model variant is then calculated based on the predicted structural similarity of each RNA sequence pair and the actual structural similarity of the corresponding RNA sequence pairs.

[0042] It should be noted that this embodiment of the invention adopts a progressive training strategy, systematically training three model variants—Model-1, Model-2, and Model-3—by constructing a first training dataset, a second training dataset, and a third training dataset. The core advantage of this training method lies in constructing a complete capability ladder from "core feature perception" to "discrimination boundary" and then to "anti-interference decision-making," tackling the core challenges of distant homology detection in stages. By accurately capturing feature information, it ultimately outputs an RNA distant homology detection application model with strong generalization capabilities. Specifically, Model-1 is trained on clean distant homology data, focusing on the fundamental ability to perceive structural signals from low sequence similarity, verifying the feasibility of the model architecture; Model-2 introduces negative samples, focusing on establishing a clear boundary for distinguishing homology from non-homology; Model-3 faces the most difficult-to-distinguish confusing samples, refining its ability to suppress sequence misleading and make robust decisions based on structural information. Crucially, the Model-3 model captured during training... 1 Its optimal detection performance on the independent test set reveals the best balance of the model's generalization ability when learning to resist sequence misleading. Its integrated perception, discrimination and anti-interference capabilities enable it to exhibit the highest overall performance and reliability on the final test set, thus providing a solid guarantee for building large-scale databases and RNA similarity search.

[0043] To comprehensively evaluate model performance, embodiments of this invention can use various evaluation metrics, including the Pearson correlation coefficient (r). p Spearman correlation coefficient (r) s Precision, Recall, F1 score, Area Under the Receiver Operational Characteristic Curve (AUC-ROC), and Area Under the Precision-Recall Curve (AUC-PR) are detailed below:

[0044]

[0045]

[0046]

[0047]

[0048] in, The structural similarity score (similarity score of the i-th RNA sequence) is the structural similarity score of the i-th predicted RNA sequence pair. This represents the average structural similarity of the predicted RNA sequence pairs. Let be the i-th actual structural similarity (the i-th TM-score value). This represents the average of the actual structural similarity, where n is the number of samples, TP stands for True Positive (the number of positive samples correctly predicted as positive), FN for False Negative (the number of positive samples incorrectly predicted as negative), TN for True Negative (the number of negative samples correctly predicted as negative), and FP for False Positive (the number of negative samples incorrectly predicted as positive). Here, "positive" means the input RNA sequence pair is homologous, and "negative" means the input RNA sequence pair is not homologous.

[0049] The AUC-ROC curve, with the false positive rate on the x-axis and the true positive rate on the y-axis, measures the model's ability to distinguish between positive and negative samples at different thresholds. The AUC-PR curve, with the true positive rate on the x-axis and precision on the y-axis, measures the model's accuracy in predicting positive samples at different thresholds.

[0050] The classification performance of each model on Test-Data2400 is as follows: Figure 4 As shown in Table 2, the specific values ​​are shown in Table 3, and the correlation analysis results between the structural similarity predicted by each model and the actual structural similarity are shown in Table 3.

[0051]

[0052]

[0053] To systematically evaluate the predictive ability of RNA distant homology detection models, comprehensive tests were conducted on three independent test sets for each variant of the RNA distant homology detection model, such as... Figure 5 As shown in Table 4.

[0054]

[0055] like Figure 5As shown, this figure displays a Pearson correlation scatter plot of structural similarity and TM scores for each model variant across various independent test sets. The horizontal axis represents the structural similarity predicted by the corresponding model variant, and the vertical axis represents the TM score calculated by USalign. The blue scatter points represent the original data points, and the red line represents the linear regression line; the closer the data points are to the red line, the stronger the linear relationship.

[0056] like Figure 6 As shown in the figure, this plot displays the prediction accuracy of each model variant across different sequence identity intervals. The sequence identity range is divided into intervals of [0, 0.3] (representing the lowest sequence similarity), (0.4, 0.5], (0.5, 0.6], and (0.6, 1.0] (representing the highest sequence similarity; only Test-Data541 includes this interval).

[0057] from Figure 5 , Figure 6 As shown in Table 4, the RNA distant homology detection model variants provided in this embodiment of the invention can effectively capture deep structural features beyond sequence similarity, and can directly learn structural similarity features from RNA sequences. All model variants exhibit a highly significant correlation between the predicted structural similarity and the TM score calculated by USalign, verifying the effectiveness of the method proposed in this embodiment. Especially under challenging low sequence similarity conditions, all model variants maintain high prediction accuracy, demonstrating their ability to capture RNA structural conservation and providing a new computational paradigm for RNA distant homology detection.

[0058] After evaluating the performance of various RNA distant homology detection model variants and selecting the best-performing variant as the application model for RNA distant homology detection, a large-scale RNA vector database based on RNAcentral can be constructed. Among them, such as... Figure 7 As shown, the database construction phase (solid arrows) includes: obtaining RNA sequences from the RNA sequence dataset and inputting them into an RNA distant homology detection application model (such as...). Figure 7 The trained RSSearch model shown in the diagram generates corresponding high-dimensional vectors using its feature encoder layer, ultimately constructing a large-scale RNA vector database. In the query retrieval phase (dashed arrow), 5500 sequences from the RR3DD dataset are used as query sequences (i.e., the RNA sequences to be detected). Following the same feature extraction and vectorization process, the RNA distant homology detection application model converts the RNA sequences to be detected into vector representations. By searching the vector database using an approximate nearest neighbor search method, the top-k most similar RNA sequences, including distant homology sequences, can be obtained.

[0059] In a single-CPU environment, a query on a database containing nearly 2 million sequences takes an average of only about 6 seconds, demonstrating that the method proposed in this embodiment can meet the real-time search requirements of large-scale databases. During distant homology detection using 5500 RNA sequences from the RR3DD dataset as the query set, the RNA sequences in the k-nearest neighbor search results were analyzed: from Model-3... 1 An instance was randomly selected from the search results for in-depth analysis, such as... Figure 8 As shown, taking the typical tRNA sequence 1ASY_R (yeast aspartic tRNA) as an example, the sequence similarity between the query sequence 1ASY_R and the hit sequence URS000220A056_1913988 is only 0.41, which is below the distant homology threshold, but Model-3 1 The predicted structural similarity between the two RNAs reached 0.98; calculated using USalign, their true TM score was 0.67 (>0.45), confirming the reliability of this distant homology relationship. This case demonstrates that the method proposed in this invention can effectively capture deep structural features beyond sequence similarity and accurately identify distantly homologous RNAs, maintaining excellent performance even under low sequence similarity conditions. This provides a powerful tool for RNA evolutionary relationship analysis and functional annotation.

[0060] In summary, the method provided by this invention acquires an RNA sequence dataset, divides it into multiple sets with different biological meanings based on the sequence identity and structural similarity between each pair of RNA sequences in the dataset, constructs multiple training datasets containing distantly homologous RNA pairs based on these sets, and trains an RNA distant homology detection model on each training dataset to obtain multiple RNA distant homology detection model variants. The performance of each RNA distant homology detection model variant is evaluated based on an independent test set to obtain the best-performing RNA distant homology detection application model. In the application stage, based on this RNA distant homology detection application model, each RNA sequence in the RNA sequence dataset is converted into a vector representation to construct a vector database. Upon receiving a RNA sequence to be detected, the RNA sequence to be detected is converted into a vector representation in the same way, and the vector database is retrieved to obtain similar RNA sequences, including distantly homologous sequences, corresponding to the RNA sequence to be detected. This achieves an end-to-end deep learning framework, enabling direct mapping from RNA sequences to structural similarity without relying on RNA structural information, effectively bridging the semantic gap between sequence and structure.

[0061] Furthermore, in this embodiment of the invention, a pre-trained RNA language model is combined with a Transformer encoder for RNA distant homology detection, which can effectively capture the deep structural features of RNA sequences and maintain excellent detection performance even under low sequence similarity conditions.

[0062] The deep learning-based RNA distant homology detection system provided by this invention is described below. The deep learning-based RNA distant homology detection system described below can be referred to in correspondence with the deep learning-based RNA distant homology detection method described above.

[0063] This invention provides a deep learning-based RNA distant homology detection system, comprising: The dataset construction unit is used to acquire RNA sequence datasets, divide the RNA sequence datasets into multiple sets with different biological meanings based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence datasets, and construct multiple training datasets containing distantly homologous RNA pairs based on the multiple sets with different biological meanings. The model training and evaluation unit is used to train the RNA distant homology detection model on each training dataset to obtain multiple RNA distant homology detection model variants, and to evaluate the performance of each RNA distant homology detection model variant on an independent test set. Based on the performance evaluation results of each RNA distant homology detection model variant, the application model for RNA distant homology detection is selected. The RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences. The database construction unit is used to convert each RNA sequence in the RNA sequence dataset into a vector representation based on the RNA distant homology detection application model, and construct a vector database. The distant homology detection unit is used to, after receiving the RNA sequence to be detected, convert the RNA sequence to be detected into a vector representation based on the RNA distant homology detection application model, and then search the vector database based on the vector of the RNA sequence to be detected to obtain similar RNA sequences corresponding to the RNA sequence to be detected, including distant homology sequences.

[0064] This invention provides an electronic device, including: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any of the above embodiments.

[0065] This invention provides a computer-readable storage medium storing computer instructions that cause a processor to perform the method described in any of the above embodiments.

[0066] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the method described in any of the above embodiments.

[0067] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting distant homology of RNA based on deep learning, characterized in that, include: An RNA sequence dataset is obtained. Based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence dataset, the RNA sequence dataset is divided into multiple sets with different biological meanings. Multiple training datasets containing distantly homologous RNA pairs are constructed based on the multiple sets with different biological meanings. The RNA distant homology detection model was trained on each training dataset to obtain multiple variants of the RNA distant homology detection model. The performance of each variant was evaluated on an independent test set. Based on the performance evaluation results of each variant, the RNA distant homology detection application model was selected. The RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences. Based on the RNA distant homology detection application model, each RNA sequence in the RNA sequence dataset is converted into a vector representation to construct a vector database. After receiving the RNA sequence to be detected, the RNA sequence to be detected is converted into a vector representation based on the RNA distant homology detection application model. Then, the vector database is searched based on the vector of the RNA sequence to be detected to obtain the similar RNA sequence corresponding to the RNA sequence to be detected, including distant homology sequences.

2. The RNA distant homology detection method based on deep learning as described in claim 1, characterized in that, Based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence dataset, the RNA sequence dataset is divided into multiple sets with different biological significance. Based on these multiple sets with different biological significance, multiple training datasets and independent test sets containing distantly homologous RNA pairs are constructed, specifically including: Based on sequence identity thresholds and structural similarity thresholds, RNA sequence pairs formed by each RNA sequence in the RNA sequence dataset are divided into four sets with different biological meanings. Specifically, RNA sequence pairs with both sequence identity and structural similarity greater than the sequence identity threshold are divided into the first set, RNA sequence pairs with both sequence identity and structural similarity less than the sequence identity threshold are divided into the second set, RNA sequence pairs with both sequence identity and structural similarity less than the sequence identity threshold are divided into the third set, and RNA sequence pairs with both sequence identity and structural similarity greater than the sequence identity threshold are divided into the fourth set. A first training dataset is constructed based on the complete second set, a second training dataset is constructed based on the complete second set and partial RNA sequence pairs in the third set, and a third training dataset is constructed based on the complete first set, the second set, the fourth set, and partial RNA sequence pairs in the third set.

3. The RNA distant homology detection method based on deep learning as described in claim 1, characterized in that, The RNA distant homology detection model comprises an input representation layer, a feature encoder layer, and a similarity calculation layer connected in sequence. The input representation layer is constructed based on the RNA language model and is used to extract the embeddings of the input RNA sequence pairs to obtain the RNA embedding representations of each RNA sequence in the RNA sequence pair. The feature encoder layer is built based on the Transformer encoder layer and is used to extract sequence features from the RNA embedding representation of each RNA sequence in the RNA sequence pair to obtain the RNA sequence feature vector of each RNA sequence in the RNA sequence pair. The similarity calculation layer is used to calculate the similarity between the RNA sequence feature vectors of each RNA sequence in the RNA sequence pair, which is used as the structural similarity of the RNA sequence pair.

4. The RNA distant homology detection method based on deep learning as described in claim 3, characterized in that, The feature encoder layer comprises two stacked Transformer encoder layers, an average pooling layer, a dropout layer, and a fully connected layer.

5. The deep learning-based RNA distant homology detection method as described in claim 3, characterized in that, The RNA distant homology detection model is trained based on any training dataset, specifically including: Each RNA sequence pair in any of the training datasets is sequentially input into the RNA distant homology detection model to obtain the structural similarity of the corresponding RNA sequence pairs predicted by the RNA distant homology detection model. Based on the difference between the structural similarity of each RNA sequence pair predicted by the RNA distant homology detection model and the actual structural similarity of the corresponding RNA sequence pair, the model loss is calculated, and the parameters of the feature encoder layer in the RNA distant homology detection model are adjusted based on the model loss.

6. The RNA distant homology detection method based on deep learning as described in claim 1, characterized in that, The performance evaluation of various RNA distant homology detection model variants based on independent test sets specifically includes: The RNA sequence pairs in the independent test set are input into each RNA distant homology detection model variant to obtain the structural similarity of each RNA sequence pair predicted by each RNA distant homology detection model variant. Based on the structural similarity of each RNA sequence pair predicted by each RNA distant homology detection model variant and the actual structural similarity of the corresponding RNA sequence pair, the performance index value of each RNA distant homology detection model variant is calculated; the performance index value includes at least one of Pearson correlation coefficient, Spearman correlation coefficient, precision, recall, F1-score, area under the receiver operating characteristic curve and area under the precision-recall curve.

7. The RNA distant homology detection method based on deep learning as described in claim 1, characterized in that, The acquisition of the RNA sequence dataset specifically includes: Multiple RNA structure data were obtained from a protein data bank, and the RNA structure data were subjected to length filtering and redundancy removal to obtain a non-redundant RNA sequence dataset.

8. A deep learning-based RNA distant homology detection system, characterized in that, include: The dataset construction unit is used to acquire RNA sequence datasets, divide the RNA sequence datasets into multiple sets with different biological meanings based on the sequence identity and structural similarity between each pair of RNA sequences in the RNA sequence datasets, and construct multiple training datasets containing distantly homologous RNA pairs based on the multiple sets with different biological meanings. The model training and evaluation unit is used to train the RNA distant homology detection model on each training dataset to obtain multiple RNA distant homology detection model variants, and to evaluate the performance of each RNA distant homology detection model variant on an independent test set. Based on the performance evaluation results of each RNA distant homology detection model variant, the application model for RNA distant homology detection is selected. The RNA distant homology detection model is used to vectorize two input RNA sequences and measure the structural similarity between the two RNA sequences. The database construction unit is used to convert each RNA sequence in the RNA sequence dataset into a vector representation based on the RNA distant homology detection application model, and construct a vector database. The distant homology detection unit is used to, after receiving the RNA sequence to be detected, convert the RNA sequence to be detected into a vector representation based on the RNA distant homology detection application model, and then search the vector database based on the vector of the RNA sequence to be detected to obtain similar RNA sequences corresponding to the RNA sequence to be detected, including distant homology sequences.

9. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method as described in any one of claims 1-7.

Images