A dna-rna interaction pair prediction system based on a pre-trained nucleic acid sequence model

Abstract

The present application relates to the technical field of computational genomics, and particularly relates to a DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model, comprising: extracting full amounts of to-be-predicted DNA sequences and to-be-predicted RNA sequences from cells to be detected; taking window-level representations of the to-be-predicted DNA sequences to construct a DNA representation set, and taking window-level representations of the to-be-predicted RNA sequences to construct an RNA representation set; pairing to obtain paired features, predicting interaction probabilities of each paired feature, and screening to obtain candidate DNA-RNA interaction pairs. In view of the problems of relatively difficult DNA-RNA interaction pair screening and relatively long pairing time in the prior art, by constructing window-level representations containing local context and functional patterns for DNA sequences and RNA sequences, and pairing the DNA representation set and the RNA representation set to predict the probabilities corresponding to the interaction pairs, the automatic screening process of the DNA-RNA interaction pairs is realized, and effective RNA-chromatin one-to-one interaction pairs can be screened in the whole genome and specific chromatin sites.

Description

Technical Field

[0001] This invention relates to the field of computational genomics, specifically to a DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model. Background Technology

[0002] DNA-RNA interactions are crucial molecular events in fundamental life processes such as gene expression regulation, chromatin structure maintenance, transcriptional activation or repression, recruitment of epigenetic modifiers, and bridging of distant regulatory elements. In particular, chromatin-associated RNAs (caRNAs) can form direct or indirect interactions with specific genomic regions, thereby participating in the regulation of promoters, enhancers, heterochromatin regions, and other functional genomic elements. Accurate identification of DNA-RNA interaction pairs helps to elucidate intracellular transcriptional regulatory networks, discover physiological / disease-related regulatory RNAs, locate potential drug targets, and provide high-confidence candidates for subsequent experimental validation.

[0003] Current research on DNA-RNA interactions primarily relies on high-throughput experimental techniques or detection techniques for specific structural categories. While these methods can capture a large number of interaction signals, they typically suffer from high experimental costs, high requirements for cell materials, limited resolution, and difficulty in directly resolving one-to-one interaction relationships from complex population signals. Some methods can only reflect the correlation between RNA and chromatin states, making it difficult to directly determine the one-to-one pairing relationship between specific RNA sequences and specific DNA regions; other methods focus on specific conformations such as RNA:DNA hybridization, making it difficult to cover the broader forms of RNA-chromatin interactions.

[0004] In terms of computational prediction, existing models mostly focus on predicting the function of a single DNA sequence, a single RNA sequence, or a specific molecular structure. There is a lack of a universal technical framework that can simultaneously utilize DNA and RNA sequence information and fuse them at the pairing level to output interaction probabilities. Especially when whole-genome regions and cell-type-specific RNA sets are involved in prediction, pan-combinatorial pairing screening results in a significant number of pairs. If the number of candidate DNA regions is N_DNA and the number of candidate RNAs is N_RNA, the number of candidate interaction pairs to be evaluated is N_DNA × N_RNA. For example, when there are 10,000 candidate DNA regions and 10,000 candidate RNAs, the total number of combinations reaches 100,000,000. If the DNA regions and RNA sequences are further divided into multiple windows, the actual number of window-level pairs to be evaluated will continue to increase. Such a large number of candidate combinations significantly increases computational resource consumption, inference time, and the burden of subsequent experimental validation, making direct exhaustive experimental validation or manual screening difficult to achieve. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, a DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model is provided.

[0006] The specific technical solution is as follows: A DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model includes: A sequence generation module, which acquires cells to be tested and extracts full amounts of DNA and RNA sequences to be predicted from the cells to be tested; A window representation module, wherein the window representation module is connected to the sequence generation module; The window representation module extracts a window-level representation of the DNA sequence to be predicted to construct a DNA representation set, and extracts a window-level representation of the RNA sequence to be predicted to construct an RNA representation set; A pairing prediction module is used to pair the DNA representation set and the RNA representation set to obtain pairing features, and then predict the interaction probability of each pairing feature. A screening module, which filters the pairing features based on the interaction probability to obtain candidate DNA-RNA interaction pairs.

[0007] On the other hand, the window representation module includes: A window partitioning module, wherein the window partitioning module receives the DNA sequence to be predicted and the RNA sequence to be predicted, respectively; The window segmentation module performs a sliding window operation on the DNA sequence to be predicted and marks the position information to obtain a DNA sequence window, and performs a sliding window operation on the RNA sequence to be predicted and marks the position information to obtain an RNA sequence window; The location information includes the source sequence identifier, the window start position, and the window end position; A sequence representation module that converts the DNA sequence window into a DNA window-level representation based on a pre-trained DNA sequence model, and converts the RNA sequence window into an RNA window-level representation based on a pre-trained RNA sequence model; The DNA sequence model is used to extract a sequence representation containing local context and functional patterns from the DNA nucleotide sequence in the DNA sequence window as the DNA window-level representation; The RNA sequence model is used to extract a sequence representation containing local context and functional patterns from the RNA nucleotide sequence in the RNA sequence window as the RNA window-level representation; The set addition module adds the location information of the DNA window-level representation annotation to the DNA representation set, and adds the location information of the RNA window-level representation annotation to the RNA representation set.

[0008] On the other hand, the window representation module also includes: A projection module is connected to both the sequence representation module and the set addition module. The projection module outputs the DNA window-level representation and the RNA window-level representation after dimensional alignment.

[0009] On the other hand, the DNA sequence model and the RNA sequence model are based on any one of the following implementations: nucleic acid language model, Transformer model, BERT-like model, masked language model, autoregressive language model, and contrastive learning model.

[0010] On the other hand, the pairing prediction module includes: The feature pairing module selects a specific splicing method according to the feature type selection instruction input by the external input, and extracts the DNA window-level representation and the RNA window-level representation to splice them to obtain paired features; An interaction prediction model is used to predict the interaction probability by analyzing the paired features.

[0011] On the other hand, the splicing method is one of the following: direct splicing of the DNA window-level representation and the RNA window-level representation, splicing after projection alignment, difference features after projection alignment, element-wise product features after projection alignment, similarity features, bilinear scoring features, and attention fusion features.

[0012] On the other hand, the interaction prediction model is based on one of the following implementations: multilayer perceptron, linear classifier, attention classifier, and bilinear scorer.

[0013] On the other hand, when the interaction prediction model outputs a logical value, a mapping layer is also connected to the subsequent stage of the interaction prediction model. The mapping layer converts the logical value into the interaction probability based on the Sigmoid function.

[0014] On the other hand, the filtering module includes: An extraction module, wherein the extraction module selects the DNA window-level representation or the RNA window-level representation to be output respectively; A probability output module, which obtains the associated interaction probability for the DNA window-level representation or the RNA window-level representation, and extracts associated RNA or associated DNA based on the interaction probability; The output module generates corresponding candidate DNA-RNA interaction pairs based on the DNA window-level representation or the RNA window-level representation and the screened associated RNA or the associated DNA.

[0015] On the other hand, it also includes a drug target screening module; The drug target screening module acquires RNA targets obtained from external experiments, and then screens the candidate DNA-RNA interaction pairs based on the RNA targets to obtain matching DNA sites and outputs them.

[0016] The above technical solution has the following advantages or beneficial effects: To address the challenges of screening DNA-RNA interaction pairs in existing technologies, which are relatively difficult and time-consuming, this paper proposes an automated screening process for DNA-RNA interaction pairs. This is achieved by constructing window-level representations of DNA and RNA sequences that include local context and functional patterns, pairing DNA and RNA representation sets, and predicting the probability of each pair corresponding to an interaction pair. This process enables the screening of effective one-to-one RNA-chromatin interaction pairs across the entire genome and at specific chromatin sites. Attached Figure Description

[0017] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0018] Figure 1 This is an overall schematic diagram of an embodiment of the present invention; Figure 2 This is a schematic diagram of the window representation module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the projection module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the pairing prediction module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the screening module in an embodiment of the present invention; Figure 6A This is a schematic diagram of the ROC AUC test results of the model in an embodiment of the present invention; Figure 6B This is a schematic diagram of the Precision-Recall AP test results of the model in an embodiment of the present invention; Figure 7AThis is a schematic diagram of the first result of the HEK293T-related RNA-chromatin screening interaction pair in an embodiment of the present invention; Figure 7B This is a schematic diagram of the second result of the HL-60 related RNA-chromatin screening interaction pair in an embodiment of the present invention; Figure 7C This is a schematic diagram of the third result of the HCCLM3-related RNA-chromatin screening interaction pair in an embodiment of the present invention; Figure 8A This is a schematic diagram of the PCR test results for HEK293T in an embodiment of the present invention; Figure 8B This is a schematic diagram of the PCR test results for HL-60 in an embodiment of the present invention; Figure 8C This is a schematic diagram of the PCR test results for HCCLM3 in an embodiment of the present invention; Figure 9A This is a schematic diagram of the fluorescence in situ hybridization results of HEK293T in an embodiment of the present invention; Figure 9B This is a schematic diagram of the fluorescence in situ hybridization results of HL-60 in an embodiment of the present invention; Figure 9C This is a schematic diagram of the fluorescence in situ hybridization results of HCCLM3 in an embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the screening results of chromosome 21 interaction pairs in HEK293T cells in an embodiment of the present invention; Figure 11 This is a schematic diagram of the PCR test results for chromosome 21 interaction pairs in HEK293T cells in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0022] This invention includes: A DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model, such as Figure 1 As shown, it includes: Sequence generation module 1 acquires the cell to be tested and extracts the full amount of DNA and RNA sequences to be predicted from the cell to be tested. Window representation module 2, window representation module 2 connects to sequence generation module 1; Window representation module 2 extracts window-level representations of the DNA sequence to be predicted to construct a DNA representation set, and extracts window-level representations of the RNA sequence to be predicted to construct an RNA representation set; Pairing prediction module 3: Pairing prediction module 3 pairs DNA representation set and RNA representation set to obtain pairing features, and then predicts the interaction probability of each pairing feature; Screening module 4 filters pairing features based on interaction probability to obtain candidate DNA-RNA interaction pairs.

[0023] Specifically, addressing the issues of relatively difficult screening of DNA-RNA pairs and long pairing time in existing technologies, this paper proposes an automated screening process for DNA-RNA pairs by constructing window-level representations of DNA and RNA sequences, pairing DNA and RNA representation sets, and predicting the probability of them corresponding to pairs. This enables the screening of effective one-to-one RNA-chromatin pairs across the entire genome and at specific chromatin sites.

[0024] Specifically, in practical implementation, the aforementioned DNA-RNA interaction pair prediction system is primarily configured as part of a large software platform within a software implementation. It is used to assist in predicting DNA-RNA interactions and identifying effective DNA loci during non-diagnostic medical research processes, such as analyzing intracellular transcriptional regulatory networks, discovering physiological / disease-related regulatory RNAs, and locating potential drug targets. In some embodiments, candidate DNA-RNA interaction pairs can also be added to the whole-genome sequencing results of a cell to be predicted for labeling.

[0025] DNA-RNA interaction refers to the direct or indirect binding, localization, enrichment, or functional association between DNA sequence regions and RNA molecules. In the context of RNA-chromatin interaction, the DNA sequence region can be a genomic fragment, promoter, enhancer, open chromatin region, region related to specific epigenetic modifications, or other candidate chromatin regions; the RNA molecule can be coding RNA, non-coding RNA, chromatin-associated RNA, candidate regulatory RNA, or other unpredicted RNA molecules.

[0026] In practical biomedical research for non-diagnostic purposes, gene sequencing of the cells to be tested is required to obtain DNA and RNA sequence listings. DNA sequences can be extracted from a reference genome and candidate region coordinates, or provided by FASTA, BED, or other genomic region files. RNA sequences can be provided from transcript sequences, candidate RNA sequences, cell type-specific expressed RNA lists, or experimentally captured RNA lists.

[0027] The raw data is cleaned, including removing duplicate pairs, duplicate or invalid sequences, records lacking matching DNA or RNA entries, and standardizing sequence identifiers. DNA and RNA sequences can be standardized according to the input vocabulary, character set, or tokenization rules of the pre-trained nucleic acid sequence model used. For example, in one implementation, DNA sequences can be converted to uppercase, and characters not supported by the model can be replaced with preset unknown base symbols or filtered; RNA sequences can be converted to uppercase, and whether to replace T with U can be selected based on the input requirements of the RNA sequence model used. For pre-trained models that support extended base symbols, modified base symbols, or smaller character sets, character mapping, retention, or filtering rules that match the input requirements of the model can be used.

[0028] After obtaining the processed DNA and RNA sequences to be predicted, a window-level representation is established according to the aforementioned process, and potential DNA-RNA interaction pairs are obtained through interaction probability screening. These pairs are then added to the genome-wide DNA-RNA interaction pair set for further research.

[0029] For example, when applied to predict potential drug targets, the process of verifying the authenticity of interactions and verifying the functional mechanism can be followed to clarify that the corresponding DNA / RNA segments have the function of drug targets.

[0030] For example, DNA-RNA immunoprecipitation (DRIP) combined with qPCR is first used to verify the interaction relationships predicted by the model. This technique can specifically enrich naturally occurring DNA-RNA hybrid chains within cells and verify endogenous DNA-RNA interactions. The procedure specifically includes preparing cell samples and performing lysis and nucleic acid fragmentation, followed by specific immunoenrichment to capture DNA-RNA interaction complexes, and setting up a negative control group to exclude non-specific interference. If the signal in the experimental group is significantly higher than that in the negative control group, the existence of the interaction can be confirmed.

[0031] Then, the enriched DNA and RNA were purified, and specific primers were designed for the DNA-RNA interaction pairs predicted by the model. The enrichment fold of candidate DNA segments and RNA molecules was quantitatively detected by qPCR. If the enrichment level of the target interaction pair in the disease model cells is significantly higher than that in normal cells and is consistent with the model prediction, it can be confirmed that the predicted DNA-RNA interaction truly exists and is disease-specific.

[0032] Furthermore, to verify the predicted biological functions of the DNA-RNA interaction pair and its availability as a drug target, we observed changes in cell phenotype by regulating the expression of upstream and downstream molecules involved in the interaction, thus clarifying its disease regulatory function.

[0033] For example, a blank control group, a negative interference group, and a targeted interference group can be set up. By knocking down candidate RNA molecules with siRNA and silencing the corresponding DNA regulatory segments with CRISPR interference (CRISPRi), the target DNA-RNA interaction system can be specifically disrupted.

[0034] Cell proliferation was then assessed using the CCK-8 assay and colony formation assay, apoptosis rate was detected by flow cytometry, and cell migration and invasion were assessed by Transwell assay. Simultaneously, gene and protein expression levels in key disease pathways were detected by qPCR and Western blot. Disruption of this DNA-RNA interaction significantly suppressed the malignant phenotype of diseased cells and downregulated the expression of disease-related pathways, demonstrating that this DNA-RNA interaction is a key functional unit regulating disease development and progression, and that the corresponding segment possesses the core functional attributes of a drug target.

[0035] In one embodiment, such as Figure 2 As shown, window representation module 2 includes: Window partitioning module 21 receives the DNA sequence to be predicted and the RNA sequence to be predicted, respectively. The window partitioning module 21 performs a sliding window operation on the DNA sequence to be predicted and marks the position information to obtain a DNA sequence window, and performs a sliding window operation on the RNA sequence to be predicted and marks the position information to obtain an RNA sequence window; Location information includes the source sequence identifier, the start position of the window, and the end position of the window; Sequence representation module 22 converts DNA sequence windows into DNA window-level representations based on a pre-trained DNA sequence model, and converts RNA sequence windows into RNA window-level representations based on a pre-trained RNA sequence model. DNA sequence models are used to extract sequence representations containing local context and functional patterns from DNA nucleotide sequences within a DNA sequence window as a DNA window-level representation. RNA sequence models are used to extract sequence representations containing local context and functional patterns from RNA nucleotide sequences within an RNA sequence window as RNA window-level representations. Set addition module 23 adds DNA window-level representation annotation location information to the DNA representation set and adds RNA window-level representation annotation location information to the RNA representation set.

[0036] Specifically, in response to the problem that simple truncation of sequences in the prior art can cause the loss of local regulatory information and make it difficult to map the prediction results back to the original genome coordinates or RNA fragment positions, this embodiment first uses a window partitioning module to slide the DNA sequence to be predicted and mark the position information to obtain a DNA sequence window, and then slides the RNA sequence to be predicted and marks the position information to obtain an RNA sequence window.

[0037] The sliding window processing involves segmenting the window based on a preset window length and step size, resulting in multiple DNA sequence windows and RNA sequence windows, including: Given the original sequence Window length and step length The original sequence is divided into an ordered set of windows: ; During the segmentation process, the source sequence identifier, start position, and end position of each window are saved as location information. For sequences whose length does not exceed the window length, they can be used as a single window input to the model. This process avoids information loss caused by simple truncation and allows the prediction results to be traced back to the original DNA coordinates and RNA fragment positions.

[0038] Subsequently, the sequence representation module 22 converts the DNA sequence window into a DNA window-level representation based on a pre-trained DNA sequence model, and converts the RNA sequence window into an RNA window-level representation based on a pre-trained RNA sequence model. In one embodiment, the DNA sequence model and the RNA sequence model are based on any one of the following implementations: nucleic acid language model, Transformer model, BERT-like model, masked language model, autoregressive language model, and contrastive learning model.

[0039] Specifically, DNA sequence models and RNA sequence models refer to deep learning models obtained through large-scale DNA or RNA sequence pre-training. They can receive nucleotide sequences or nucleotide tokens as input and output token-level and window-level representations.

[0040] Sequence representation refers to the vector obtained by pooling, projection, or aggregation after output from pre-trained DNA sequence models and RNA sequence models.

[0041] By constructing the DNA sequence model and RNA sequence model described above, it is possible to convert the sequences in the corresponding DNA window or RNA window into their corresponding sequence representations for output.

[0042] When the DNA sequence model and RNA sequence model output a token-level representation, a window-level representation can be obtained through CLS token extraction, average pooling, max pooling, attention pooling, or other aggregation methods.

[0043] In one embodiment, such as Figure 3 As shown, window representation module 2 also includes: Projection module 24 is connected to sequence representation module 22 and set addition module 23 respectively; The projection module 24 outputs the DNA window-level representation and the RNA window-level representation after dimensional alignment.

[0044] Specifically, when the dimensions of the DNA window-level representation and the RNA window-level representation are inconsistent, the projection module 24 can be used to align the DNA window-level representation and the RNA window-level representation respectively, thereby... and Mapped to the same or preset dimension:

[0045] in, and It can be a linear layer, a multilayer perceptron, or a projection module with normalization. This is a window-level representation of DNA before projection. This is a window-level representation of the projected DNA. This is a window-level representation of the RNA before projection. This is a window-level representation of the projected RNA.

[0046] After obtaining the DNA window-level representation and the RNA window-level representation, they can be associated and stored with the window source sequence identifier, start position, end position and model version information, respectively, to form the DNA representation table and the RNA representation table. The set addition module 23 adds the DNA window-level representation annotation position information to the DNA representation set and adds the RNA window-level representation annotation position information to the RNA representation set.

[0047] As an optional implementation, for the same batch of candidate sequences, subsequent pairwise inference can directly call the cached window-level representation, without repeatedly running the pre-trained DNA sequence model and the pre-trained RNA sequence model for each DNA-RNA combination. Thus, in the context of... DNA window and In the candidate set of RNA windows, the number of encodings of the pre-trained sequence model can be determined by the number of encodings in combination repetition. The level was reduced to level.

[0048] In one embodiment, such as Figure 4 As shown, the pairing prediction module 3 includes: Feature matching module 31 selects a specific splicing method according to the feature type selection instruction input by the external input, and extracts DNA window-level representation and RNA window-level representation to splice to obtain paired features; Interaction prediction model 32 predicts the interaction probability by matching features.

[0049] Specifically, after obtaining the DNA and RNA representations, DNA-RNA pairing features can be constructed. These pairing features can be selected from one or more of the following: direct splicing of DNA and RNA representations, splicing after projection alignment, difference features after projection alignment, element-wise product features after projection alignment, similarity features, bilinear scoring features, and attention fusion features. These features are then fed into the interaction prediction model 32 to predict the interaction probability. The interaction probability refers to the numerical value output by the interaction prediction model that characterizes the likelihood of interaction between the DNA sequence window and the RNA sequence window.

[0050] In one embodiment, the splicing method is one of the following: direct splicing of DNA window-level representation and RNA window-level representation, splicing after projection alignment, difference features after projection alignment, element-wise product features after projection alignment, similarity features, bilinear scoring features, and attention fusion features.

[0051] The set of candidate pairing features that can be generated is represented as follows:

[0052] in, Represents the set of candidate pairing features. Indicates splicing characteristics, Indicates the difference characteristics, This represents the element-wise product characteristic. Represents similarity features. This represents the bilinear scoring feature. This represents attention fusion features. The above features can be represented in the following example forms:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] in, This represents the learnable parameter matrix. It can be cosine similarity, dot product similarity, or other similarity functions. It can be a unidirectional attention, cross attention, or multi-head attention mechanism.

[0059] This is a window-level representation of DNA before projection. This is a window-level representation of the projected DNA. This is a window-level representation of the RNA before projection. This represents the projected RNA at the window level. for and The DNA window-level representation and the RNA window-level representation need to have the same dimension, or be pre-mapped to the same dimension via a projection layer. The above formulas are all example implementations. The pairing features of the actual input interaction prediction model can be obtained by combining one or more features from the candidate pairing feature set. That is, it can use only spliced ​​features, or it can use a combination of multiple features.

[0060] For example, during processing, the DNA window-level representation and the RNA window-level representation can be directly concatenated to obtain paired features:

[0061] in, This represents a window-level representation of DNA. This indicates a window-level representation. This method represents vector concatenation. It does not require the DNA window-level representation and the RNA window-level representation to have the same dimension, and it can preserve the information of both the DNA sequence and the RNA sequence, allowing the subsequent interaction prediction model to learn the pairing relationship between them.

[0062] In another implementation, if the DNA window-level representation and the RNA window-level representation have the same dimensions, or have been mapped to aligned representations of the same dimensions by the projection module 24 respectively. and Then the pairing features can include:

[0063] in, This represents vector concatenation. This represents the element-wise absolute difference between the aligned DNA window-level representation and the RNA window-level representation. This represents element-wise product.

[0064] The aforementioned features are used to preserve the intrinsic information of the DNA sequence, the intrinsic information of the RNA sequence, the differences between the two, and the matching information between them in each representation dimension. This formula is only an example of dimensional consistency or projection alignment, and does not imply that the four types of features must be used simultaneously, nor does it limit the present invention to using only this combination of features.

[0065] For cases where the dimensions of DNA and RNA window-level representations are inconsistent and element-wise alignment has not been performed, paired features can be constructed using direct splicing, splicing after separate projection, cross-attention, bilinear mapping, or other fusion methods capable of accepting inputs of different dimensions. These paired features can then be further standardized, linearly projected, or nonlinearly transformed before being input into the interaction prediction model.32

[0066] In one embodiment, the interaction prediction model is implemented through pre-training based on one of the following: multilayer perceptron, linear classifier, attention classifier, or bilinear scorer. Essentially, it is a kind of supervised learning classifier that can be used to predict the probability of DNA-RNA interaction pairs.

[0067] In the training process, a training set is first constructed based on the training samples.

[0068] Training samples include positive DNA-RNA interaction pairs and negative DNA-RNA pairings. Positive samples can come from public databases, literature reports, experimental capture data, or manually compiled data. For scenarios lacking high-confidence non-interacting pairs, negative samples can be constructed by randomly recombining DNA and RNA identifiers, ensuring that the generated negative samples do not overlap with known positive samples.

[0069] To reduce the risk of data leakage, the training and validation sets can be randomly partitioned or grouped according to DNA identifiers, RNA identifiers, interaction pairing identifiers, or cell type identifiers. Before being fed into the classification model, the pairing features can be standardized using training set statistics. ; in, and These represent the mean and standard deviation of the corresponding feature dimension in the training set, respectively.

[0070] The selected pedestal model is then trained, and the model can be optimized using binary cross-entropy loss: ; in, Indicates the true label, This represents the model's output logical value. Model parameters can be updated using AdamW or other optimizers, and the best model can be saved based on validation set performance. After training, the pre-trained model selection, projection layer parameters, interaction prediction model parameters, feature normalization statistics, and prediction thresholds are fixed for subsequent inference.

[0071] In one embodiment, when the interaction prediction model outputs a logical value, a mapping layer is also connected to the subsequent level of the interaction prediction model. The mapping layer uses the Sigmoid function to transform logical values ​​into interaction probabilities.

[0072] Specifically, depending on the chosen base model, the output logic value may be... Then, it is converted into interaction probability using the Sigmoid function: .

[0073] In one embodiment, such as Figure 5 As shown, the filtering module 4 includes: Extraction module 41 selects the DNA window-level representation or RNA window-level representation to be output respectively; The probability output module 42 obtains the associated interaction probability for the DNA window-level representation or the RNA window-level representation, and extracts the associated RNA or associated DNA based on the interaction probability. Output module 43 generates corresponding candidate DNA-RNA interaction pairs based on the DNA window-level representation or RNA window-level representation and the screened associated RNA or associated DNA.

[0074] Specifically, to achieve better screening results, this embodiment can score pre-specified candidate pairs, or perform exhaustive prediction on all combinations of DNA and RNA window sets. When predicting for a specified DNA or RNA window representation, the DNA and RNA candidates are scored respectively to obtain the associated interaction probability, and the top-ranked ones are retained. DNA candidates and Each RNA candidate is designated as an associated RNA or associated DNA. Output module 43 then generates corresponding candidate DNA-RNA interaction pairs based on the DNA window-level representation or RNA window-level representation and the selected associated RNA or associated DNA. The output results include DNA source identifier, DNA window start position, DNA window end position, RNA source identifier, RNA window position, interaction probability, and binary prediction label. For scenarios requiring sequence-level results, the maximum, average, weighted average, or Top-K average probabilities of multiple window-level probabilities for the same DNA sequence and the same RNA sequence can be aggregated.

[0075] In one embodiment, a drug target screening module is also included; The drug target screening module acquires RNA targets obtained from external experiments. Based on the RNA targets, the drug target screening module screens candidate DNA-RNA interaction pairs to obtain matching DNA sites and outputs them.

[0076] Example 1: Constructing the MiRCIC-AI model using DNABERT-2 and RNA-FM In this embodiment, the pre-trained DNA sequence model specifically uses DNABERT-2, and the pre-trained RNA sequence model specifically uses RNA-FM. The DNA sequence is first segmented and tensorized, then input into DNABERT-2 to extract information from the last layer of representation and aggregated into a fixed-length DNA sequence embedding h_d. The RNA sequence, after standardization, is input into RNA-FM, where information is extracted from the last layer of representation and aggregated into an RNA sequence embedding h_r.

[0077] For sequences whose length exceeds the encoder input range, this embodiment adopts a sliding window segmentation strategy with a window length of 1000 nucleotides and a step size of 500 nucleotides. The generated DNA and RNA windows are then expanded together with paired annotations to form window-level training data.

[0078] In this embodiment, the positive sample DNA-RNA pairings are primarily sourced from the NPInter database. Due to the lack of widely accepted high-confidence non-interacting datasets, this embodiment constructs negative samples by randomly recombining DNA and RNA identifiers, ensuring that the negative samples do not overlap with the known set of positive samples. The negative sampling ratio is set to 1:1. The samples are randomly divided into training and validation sets at a 9:1 ratio.

[0079] The constructed MiRCIC-AI model employs a dual-tower representation and pairwise fusion architecture. In this embodiment, the paired features are directly concatenated from the DNA sequence embedding h_d and the RNA sequence embedding h_r, i.e., x_(d,r)=[h_d,h_r], without using differential features, element-wise product features, or attention fusion features. The concatenated paired features are input into a multilayer perceptron, which learns the nonlinear pairing relationship between the DNA and RNA representations and outputs the interaction probability for a given DNA-RNA sequence window pair. In the validation data, such as... Figure 6A As shown, the model's ROC AUC reaches 0.877, as... Figure 6B As shown, the Precision-Recall AP reached 0.870. This result indicates that using DNA and RNA pre-trained sequence models to encode and then perform pairwise fusion can effectively identify high-confidence DNA-RNA interaction pairs.

[0080] Example 2: Screening one-to-one RNA-chromatin interaction pairs from MiRCIC detection data In this embodiment, the MiRCIC-AI model is applied to screen for one-to-one H3K27ac-related RNA-chromatin interactions.

[0081] Under physiological conditions, specific RNA-chromatin interactions are crucial for maintaining cell identity, coordinating developmental processes, and responding to stress, enabling precise temporal and spatial regulation of gene expression. However, under pathological conditions, this intricate regulatory network can be disrupted. Aberrant RNA-chromatin interactions can form pathological interaction complexes in the promoter / enhancer regions of specific oncogenes or immune regulatory genes. By misrecruiting activating or repressive complexes, these complexes lead to persistent aberrant activation or silencing of genes, thereby driving tumor progression or suppressing anti-tumor immunity. These aberrant interaction structures themselves have become a highly promising new class of therapeutic targets. Obtaining one-to-one RNA-chromatin interaction relationships is essential for elucidating the function of RNA-chromatin interactions under physiological and pathological conditions.

[0082] Based on this, DNA and RNA fragments captured from three tumor cell lines (HEK293T, HL-60, and HCCLM3 cells) were analyzed using this model to score candidate DNA regions and candidate RNA sequences in pairs, and high-probability one-to-one interactions were selected. This validated the system's ability to detect one-to-one H3K27ac-related RNA-chromatin interactions in a tumor disease model.

[0083] The model scores candidate DNA regions and candidate RNA sequences in pairs and filters out high-probability one-to-one interaction pairs. Genome-wide distribution analysis is performed on the top-ranked interaction pairs, and the results are as follows: Figure 7A , Figure 7B and Figure 7C As shown in the figure. The results show that the screened interaction pairs are enriched in the promoter region near the transcription start site (TSS), especially within ±1 kb of the TSS. These results demonstrate that this invention can not only output interaction probabilities but also resolve candidate pairings between specific DNA regions and specific RNA molecules from complex RNA-chromatin interaction signals, providing candidates for subsequent experimental verification.

[0084] Furthermore, in HEK293T, HL-60, and HCCLM3 cells, we used a specific RNA-targeting oligonucleotide-guided MiRCIC-PCR method to actually detect high-scoring one-to-one interaction pairs reported by the MiRCIC-AI model. The results are as follows: Figure 8A , Figure 8B and Figure 8C As shown in the figure. Simultaneously, its authenticity was verified by using specific DNA / RNA probes for fluorescence in situ hybridization (FISH), and the results are as follows. Figure 9A , Figure 9B , Figure 9C As shown above, the laboratory results fully demonstrate the effectiveness of the MiRCIC-AI model's predictions.

[0085] Example 3: Screening of RNAs with Clustered RNA-Chromatin Interaction Structures In this embodiment, the MiRCIC-AI model was used to screen candidate regions related to chromosome 21 in HEK293T cells to predict RNA molecules that may be responsible for forming clustered RNA-chromatin interaction structures. Figure 10 As shown, the model predictions indicate that RNAs such as BNIP3P4, CAV1, and MET have high interaction probabilities, suggesting that the clustered interaction structures in this region may be formed by a specific small number of RNA molecules. Furthermore, we used the MiRCIC-PCR method to actually detect the high-scoring clustered one-to-one interaction pairs reported by the MiRCIC-AI model, demonstrating their existence. The results are as follows: Figure 11 As shown.

[0086] Example 5: Large-scale candidate pool inference In this embodiment, for an input set containing 10^4 DNA candidates and 10^4 RNA candidates, full combinatorial prediction would require evaluating 10^8 DNA-RNA combinations. By enabling the single-sequence Top-K pre-screening module, high-scoring DNA and RNA candidates can be retained separately, and then pairwise predictions can be performed on the retained sets, thereby significantly reducing computational load. This approach is suitable for large-scale searches of whole-genome candidate regions, cell type-specific expressed RNA sets, or epigenetic modification-related regions.

[0087] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A DNA-RNA interaction pair prediction system based on a pre-trained nucleic acid sequence model, characterized in that, include: A sequence generation module, which acquires cells to be tested and extracts full amounts of DNA and RNA sequences to be predicted from the cells to be tested; A window representation module, wherein the window representation module is connected to the sequence generation module; The window representation module extracts a window-level representation of the DNA sequence to be predicted to construct a DNA representation set, and extracts a window-level representation of the RNA sequence to be predicted to construct an RNA representation set; A pairing prediction module is used to pair the DNA representation set and the RNA representation set to obtain pairing features, and then predict the interaction probability of each pairing feature. A screening module, which filters the pairing features based on the interaction probability to obtain candidate DNA-RNA interaction pairs.

2. The DNA-RNA interaction pair prediction method according to claim 1, characterized in that, The window representation module includes: A window partitioning module, wherein the window partitioning module receives the DNA sequence to be predicted and the RNA sequence to be predicted, respectively; The window segmentation module performs a sliding window operation on the DNA sequence to be predicted and marks the position information to obtain a DNA sequence window, and performs a sliding window operation on the RNA sequence to be predicted and marks the position information to obtain an RNA sequence window; The location information includes the source sequence identifier, the window start position, and the window end position; A sequence representation module that converts the DNA sequence window into a DNA window-level representation based on a pre-trained DNA sequence model, and converts the RNA sequence window into an RNA window-level representation based on a pre-trained RNA sequence model; The DNA sequence model is used to extract a sequence representation containing local context and functional patterns from the DNA nucleotide sequence in the DNA sequence window as the DNA window-level representation; The RNA sequence model is used to extract a sequence representation containing local context and functional patterns from the RNA nucleotide sequence in the RNA sequence window as the RNA window-level representation; The set addition module adds the location information of the DNA window-level representation annotation to the DNA representation set, and adds the location information of the RNA window-level representation annotation to the RNA representation set.

3. The DNA-RNA interaction pair prediction method according to claim 2, characterized in that, The window representation module also includes: A projection module is connected to both the sequence representation module and the set addition module. The projection module outputs the DNA window-level representation and the RNA window-level representation after dimensional alignment.

4. The DNA-RNA interaction pair prediction method according to claim 2, characterized in that, The DNA sequence model and the RNA sequence model are based on any one of the following implementations: nucleic acid language model, Transformer model, BERT-like model, masked language model, autoregressive language model, and contrastive learning model.

5. The DNA-RNA interaction pair prediction method according to claim 2, characterized in that, The pairing prediction module includes: The feature pairing module selects a specific splicing method according to the feature type selection instruction input by the external input, and extracts the DNA window-level representation and the RNA window-level representation to splice them to obtain paired features; An interaction prediction model is used to predict the interaction probability by analyzing the paired features.

6. The DNA-RNA interaction pair prediction method according to claim 5, characterized in that, The splicing method is one of the following: direct splicing of the DNA window-level representation and the RNA window-level representation, splicing after projection alignment, difference features after projection alignment, element-wise product features after projection alignment, similarity features, bilinear scoring features, and attention fusion features.

7. The DNA-RNA interaction pair prediction method according to claim 5, characterized in that, The interaction prediction model is based on one of the following implementations: multilayer perceptron, linear classifier, attention classifier, and bilinear scorer.

8. The DNA-RNA interaction pair prediction method according to claim 7, characterized in that, When the interaction prediction model outputs a logical value, a mapping layer is also connected to the subsequent level of the interaction prediction model. The mapping layer converts the logical value into the interaction probability based on the Sigmoid function.

9. The method for predicting DNA-RNA interaction pairs according to claim 2, characterized in that, The filtering module includes: An extraction module, wherein the extraction module selects the DNA window-level representation or the RNA window-level representation to be output respectively; A probability output module, which obtains the associated interaction probability for the DNA window-level representation or the RNA window-level representation, and extracts associated RNA or associated DNA based on the interaction probability; The output module generates corresponding candidate DNA-RNA interaction pairs based on the DNA window-level representation or the RNA window-level representation and the screened associated RNA or the associated DNA.

10. The method for predicting DNA-RNA interaction pairs according to claim 1, characterized in that, It also includes a drug target screening module; The drug target screening module acquires RNA targets obtained from external experiments, and then screens the candidate DNA-RNA interaction pairs based on the RNA targets to obtain matching DNA sites and outputs them.

Images