A TF-DNA binding recognition method based on multi-feature fusion

Abstract

The present invention discloses a TF-DNA binding recognition method based on multi-feature fusion, comprising: S1: obtaining five types of raw data related to transcription factor binding of common human tissues; S2: preprocessing multiple data; S3: performing data encoding processing on DNA sequence data in the preprocessed raw data; S4: normalizing other data in the preprocessed raw data; S5: using a self-attention mechanism of multi-feature fusion and a convolutional neural network to perform global dependency extraction, feature extraction and feature fusion on the encoded DNA sequence data and the normalized data to obtain an overall feature mapping combination; S6: training a multi-feature fusion TF-DNA binding recognition model according to the overall feature mapping combination to obtain a trained multi-feature fusion TF-DNA binding recognition model; S7: using the trained multi-feature fusion TF-DNA binding recognition model to identify the data to be identified.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a TF-DNA binding recognition method based on multi-feature fusion. Background Art

[0002] Precision medicine, as a next-generation diagnostic and treatment technology, is an emerging approach to disease prevention and treatment that takes into account individual differences in genes, environment, and lifestyle. It offers significant technical advantages over traditional diagnostic and treatment methods, placing greater emphasis on the in-depth characteristics of the disease. By precisely identifying the causes and treatment targets of diseases and accurately classifying the disease's progression, it ultimately achieves personalized precision treatment for specific diseases and patients, improving the effectiveness of disease prevention, diagnosis, and treatment. Key to this is identifying various gene transcription regulatory elements and analyzing their relationships with various chromatin features, thereby understanding human life at the molecular level.

[0003] Transcription factors (TFs) are specialized proteins that control transcription by specifically binding to DNA sequences, thereby regulating gene expression. Numerous studies have confirmed that TFs, due to their broad tissue-specific binding, play a crucial role in human physiology. They can regulate differential gene expression within tissues, influence the development and progression of human diseases, and guide cellular activity within tissues. Therefore, studying TF-DNA binding and exploring their tissue-specific binding mechanisms is crucial for understanding how TFs participate in transcriptional regulation, probing gene function, and understanding cellular activity in different tissues.

[0004] With the widespread adoption of high-throughput sequencing technology, the vast amount of experimental data it has accumulated has laid the foundation for computational approaches to identifying gene transcripts, eliminating the time-consuming and labor-intensive nature of traditional biological experiments. Currently, the experimental approach of first predicting gene transcripts through computational methods and then verifying them through biological experiments has gradually become mainstream. DeepBind, utilizing convolutional neural networks, pioneered deep learning to predict transcription factor binding sites. However, because many TFs do not bind specifically to DNA sequences, sequence-based methods result in a large number of false positives. Since then, numerous improved methods have been proposed, among which gene transcript identification methods based on multi-feature fusion have become a mainstream research direction in this field.

[0005] The integrated analysis of multiple features can provide a more comprehensive understanding of biological processes, leading to better predictions of TF-DNA binding. Studies have shown that the binding of transcription factors to DNA sequences is not only related to DNA sequence, but also to multiple features such as DNA shape, histone modifications, and chromatin accessibility.

[0006] Most existing studies focus solely on DNA sequences, without systematically analyzing multiple features. Furthermore, current research on transcription factors at the tissue level is very limited. Building high-performance, robust gene transcription recognition models based on multiple features, effectively addressing the small sample size problem, and resolving the high-dimensional heterogeneity between multi-omics data, as well as rationally characterizing features, remain current challenges. Furthermore, addressing the "black box" problem of models and developing effective interpretable methods for models, thereby identifying key regions that influence model decisions, are crucial for understanding biological processes. Summary of the Invention

[0007] The purpose of the present invention is to provide a TF-DNA binding recognition method based on multi-feature fusion, which can effectively solve the problem of high-dimensional heterogeneity between small samples and multi-omics data while being able to reasonably characterize the features.

[0008] The technical solution of the present invention to solve the above technical problems is as follows:

[0009] The present invention provides a TF-DNA binding recognition method based on multi-feature fusion, the TF-DNA binding recognition method based on multi-feature fusion comprising:

[0010] S1: Obtained original data related to transcription factor binding in five common human tissues;

[0011] S2: Preprocessing the multiple data to obtain preprocessed raw data;

[0012] S3: performing data encoding processing on the DNA sequence data in the preprocessed raw data to obtain encoded DNA sequence data;

[0013] S4: performing normalization processing on other data in the preprocessed original data to obtain normalized data;

[0014] S5: performing global dependency extraction, feature extraction, and feature fusion on the encoded DNA sequence data and the normalized data using a multi-feature fusion attention mechanism and a convolutional neural network to obtain an overall feature mapping combination;

[0015] S6: training the multi-feature fusion TF-DNA binding recognition model using a transfer learning method according to the overall feature mapping combination to obtain a trained multi-feature fusion TF-DNA binding recognition model;

[0016] S7: Using the trained multi-feature fusion TF-DNA binding recognition model to identify the data to be identified, to obtain a recognition result.

[0017] Optionally, in step S2, the pre-processed raw data includes: DNA sequence data and its corresponding DNA shape data, chromatin accessibility data, histone modification data and conservation data.

[0018] Optionally, step S2 includes:

[0019] S21: extracting DNA sequence data from the raw data using a Chip-seq dataset;

[0020] S22: Processing the DNA sequence data using GKM-SVM to obtain positive samples and negative samples;

[0021] S23: Acquire three types of DNA shape data of the positive sample and the negative sample;

[0022] S24: extracting chromatin accessibility data and histone modification data from the DNA sequence data;

[0023] S25: generating conservation data of positive samples and negative samples using phastCons100way according to the gene coordinates of the DNA sequence data;

[0024] S26: Outputting the DNA sequence data, the DNA shape data, the chromatin accessibility data, the histone modification data and the conservation data as the preprocessed original data, and the chromatin accessibility data, the histone modification data and the conservation data are the other data.

[0025] Optionally, step S22 includes:

[0026] S221: For each data set, based on the gene coordinates and centered on the peak, a sequence of 101 bp is expanded as the positive sample.

[0027] S222: Select a region in the whole genome with a similar GC content to the positive sample as the negative sample.

[0028] Optionally, the three types of DNA shape data include:

[0029] A: inter-bp: HelT, Rise, Roll, Shift, Slide and Tilt;

[0030] B: intra-bp: Buckle, Opening, ProT, Shear, Stagger and Stretch;

[0031] C:MGW and EP.

[0032] Optionally, step S3 includes:

[0033] S31: dividing the positive sample and the negative sample into a plurality of k-mer base segments of equal sequence length;

[0034] S32: Encoding the plurality of k-mer base segments using hot-single encoding to obtain a plurality of encoded k-mer base segments;

[0035] S33: According to the word2vec strategy, the high-dimensional sparse single-hot encoding is converted into distributed encoding;

[0036] S34: Characterizing the encoded k-mer base segments using the distributed encoding to obtain encoded DNA sequence data.

[0037] Optionally, in step S4, the normalized data is:

[0038]

[0039] in, S m 、S h 、S d 、S c Represents the sequence S The corresponding feature matrix of DNA shape, chromatin accessibility, histone modification and conservation score, 、 、 、 They represent the DNA shape, chromatin accessibility, histone modification and conservation scores corresponding to the i-th base in the DNA sequence, respectively.

[0040] Optionally, step S5 includes: the multi-feature fusion convolutional neural network includes a multi-head self-attention mechanism module, a first feature fusion module, a convolutional feedforward neural network module, a second feature fusion module, a convolutional layer 1, a maximum pooling layer, a convolutional layer 2, a global pooling layer, a fully connected layer and a ReLU activation layer, which are arranged in sequence.

[0041] The multi-head self-attention mechanism module is used to obtain the encoded DNA sequence data / normalized data and capture the long-range and short-range dependencies between different features in the encoded DNA sequence data / normalized data;

[0042] The first feature fusion module is used to fuse the encoded DNA sequence data / normalized data with the long-distance and short-distance dependencies to obtain a first fusion result;

[0043] The convolutional feedforward neural network module is used to perform feature enhancement on the first fusion result to obtain an enhanced result;

[0044] The second feature fusion module is used to fuse the first fusion result and the enhanced result to obtain a global dependency extraction result;

[0045] The convolutional layer 1 is used to extract initial features from the global dependency extraction result and transmit the initial features to the maximum pooling layer;

[0046] The maximum pooling layer is used to sample the feature map of the initial features;

[0047] The convolution layer 2 is used to perform convolution calculation on the feature map to obtain a calculated feature map;

[0048] The global pooling layer is used to sample the calculated feature map to obtain a feature extraction result;

[0049] The fully connected layer and the ReLU activation layer are used to fuse the global dependency extraction result and the feature extraction result to obtain an overall feature mapping combination.

[0050] Optionally, in step S6, the transfer learning method includes a weak transfer learning method and a strong transfer learning method, and the weak transfer learning method includes:

[0051] A1: For the transcription factor AT in tissue T, use the remaining datasets for pre-training until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases;

[0052] A2: Fine-tune the multi-feature fusion TF-DNA binding recognition model using the AT dataset until the loss no longer decreases;

[0053] A3: Outputting the multi-feature fusion TF-DNA binding recognition model;

[0054] The strong transfer learning method includes:

[0055] B1: For a transcription factor A, pre-train using the datasets of all tissues of A until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases;

[0056] B2: Freeze all network structures except the fully connected layer;

[0057] B3: Fine-tune the multi-feature fusion TF-DNA binding recognition model using target tissue data until the loss no longer decreases;

[0058] B4: Output the multi-feature fusion TF-DNA binding recognition model.

[0059] The present invention also provides a TF-DNA binding recognition system, which includes a processor, a memory, and a computer program stored in the memory, wherein the computer program executes the above-mentioned TF-DNA binding recognition method based on multi-feature fusion on the processor.

[0060] The present invention has the following beneficial effects:

[0061] 1) This invention uses wor2vec and feature matrix methods to represent different data, effectively solve the high-dimensional heterogeneity between different data, and achieve data fusion;

[0062] 2) This paper uses a self-attention mechanism and convolution to set up two branches to learn DNA sequence and other feature information in parallel, avoiding confusion between the two features, thereby fully extracting the potential features that enhance TF-DNA binding and significantly improving model performance;

[0063] 3) The proposed method can analyze the importance of various features for TF-DNA binding, thereby better understanding tissue-specific TF-DNA binding;

[0064] 4) This method can identify key regions in the genome that affect TF-DNA binding. Analyzing these key regions can explore the mechanism of gene transcription.

[0065] 5) This study has made recent progress in the prediction of transcription factor binding sites (TFBSs), addressing the shortcomings of previous studies that failed to fully consider both local and global dependencies, and fully accounting for global information.

[0066] 6) The model pre-training method based on transfer learning proposed in this paper has a certain degree of portability and can be applied to problems with less data to expand the data set;

[0067] 7) The present invention constructs a neural network learning based on known gene sequence data to predict whether unverified data has an interaction relationship to guide biological experiments, effectively reducing experimental time and financial losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of the TF-DNA binding recognition method based on multi-feature fusion of the present invention;

[0069] Figure 2 Data flow diagram encoding the DNA sequence and other features proposed by the present invention;

[0070] Figure 3 Schematic diagram of the neural network structure for TF-DNA binding recognition based on multi-feature fusion constructed in an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram of the structure of the C-Transformer proposed in the present invention;

[0072] Figure 5 This is a diagram of experimental results in Example 2 of the present invention;

[0073] Figure 6 Schematic diagram of attention score extraction in implementation case 3 of the present invention;

[0074] Figure 7 This is the attention visualization result in implementation case 4 of the present invention. DETAILED DESCRIPTION

[0075] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0076] The present invention provides a TF-DNA binding recognition method based on multi-feature fusion, referring to Figure 1 As shown, the TF-DNA binding recognition method based on multi-feature fusion includes:

[0077] S1: Obtained original data related to transcription factor binding in five common human tissues;

[0078] In this study, the TF-DNA binding sequence, histone modification, and chromatin accessibility data were obtained from the Human DNA Encyclopedia (ENCODE). DNA shape data were generated using DNAshapeR (a DNA shape generation tool), and conservation data were generated using phastCons100way. The original TF-DNA binding sequence was stored in the bed file format, with its gene coordinates stored. Gene sequences were generated from these gene coordinates using GKMSVM. Histone modification and chromatin accessibility data were stored in bigwig (a binary file format), and features of the corresponding sequences were extracted using deepTools. Finally, the five data types were stored in fasta (a biological sequence storage file format) in the format "chromosome name:sequence position," [specific feature, label (0 or 1)].

[0079] S2: Preprocessing the multiple data to obtain preprocessed raw data;

[0080] Optionally, step S2 includes:

[0081] S21: extracting DNA sequence data from the raw data using a chromatin immunoprecipitation sequencing (Chip-seq) dataset;

[0082] S22: Processing the DNA sequence data using GKM-SVM (a k-mer-based DNA sequence processing tool) to obtain positive samples and negative samples;

[0083] Optionally, step S22 includes:

[0084] S221: For each data set, based on the gene coordinates and centered on the peak, a sequence of 101 bp is expanded as the positive sample.

[0085] Here, the present invention expands each Chip-seq dataset to a 101bp sequence centered on the peak, based on the gene coordinates. If the length L is less than 101bp, the sequence is padded with a length of (101 − L) / 2 at the front and back, respectively. If the length N of the front end of the sequence is less than (101 − L) / 2, the front end is padded with a length of N and the back end is padded with a length of 101 − L– N. If the sequence length L exceeds 101bp, only the 101bp sequence centered on the peak is retained.

[0086] S222: Selecting a region in the whole genome that has a similar guanine-cytosine (GC) content to the positive sample as the negative sample.

[0087] Because the number of positive and negative samples in TF-DNA binding is imbalanced, which can easily affect model performance, this paper adopts a sample data balancing method to ensure that the number of positive and negative samples is approximately equal.

[0088] 1) Oversampling of positive samples

[0089] i. Extract the features of both the positive and negative strands of DNA simultaneously, expanding the dataset of m positive samples to 2m.

[0090] 2) Downsampling of negative samples

[0091] i. Select n regions in the whole genome with similar GC content to the positive sample as negative samples. Randomly select 2m sequences from the n negative samples;

[0092] ii. Make the number of positive and negative samples equal.

[0093] S23: Obtain three types of DNA shape data of the positive sample and the negative sample, and normalize them to [0, 1];

[0094] The three types of DNA shape data include:

[0095] A: Inter-base (inter-bp): Helix Twist (HelT), Rise, Roll, Shift, Slide and Tilt;

[0096] B: Intra-bp: Buckle, Opening, Propeller Twist (ProT), Shear, Stagger and Stretch;

[0097] C: Minor Groove Width (MGW) and Electrostatic Potential (EP).

[0098] S24: extracting chromatin accessibility data and histone modification data from the DNA sequence data;

[0099] S25: generating conservation data of positive samples and negative samples using phastCons100way (a conservation score calculation tool) according to the gene coordinates of the DNA sequence data;

[0100] S26: Outputting the DNA sequence data, the DNA shape data, the chromatin accessibility data, the histone modification data and the conservation data as the preprocessed original data, and the chromatin accessibility data, the histone modification data and the conservation data are the other data.

[0101] S3: performing data encoding processing on the DNA sequence data in the preprocessed raw data to obtain encoded DNA sequence data;

[0102] Here, different data types are represented in different ways. Specifically, for DNA sequences, the word2vec (word embedding) strategy of Skip-Gram (a word2vec model) is used to represent k-mer base segments in a distributed manner, while the remaining data is encoded using a feature matrix.

[0103] Optionally, step S3 includes:

[0104] S31: dividing the positive sample and the negative sample into a plurality of k-mer base segments of equal sequence length;

[0105] refer to Figure 2As shown in the figure, the positive and negative sample sequences are divided into a large number of k-mer base segments, one-hot encoding is used based on these base segments, and 'N' is used before and after the sequence to ensure the consistency of sequence length.

[0106] DNA sequences are encoded using k-mer-based one-hot encoding. The DNA sequence "ATCG" is divided into base segments (NTC, ATC, TCG, CGN) using 3-mer decomposition.

[0107] S32: Encoding the plurality of k-mer base segments using hot-single encoding to obtain a plurality of encoded k-mer base segments;

[0108] Each base segment is represented using a one-hot encoding. Since there are five possible nucleotide occurrences at each position (A, G, C, T, and N), the dimension of the one-hot encoding is .

[0109] S33: According to the word2vec strategy, the high-dimensional sparse single-hot encoding is converted into distributed encoding;

[0110] To solve the problem of high-dimensional sparsity in the encoding method, the word2vec method of Skip-Gram is used to learn the low-dimensional representation of base segments and design the objective function L. The formula of the objective function L is as follows:

[0111]

[0112] in, p Represents the probability, that is, we hope that given the input word w In the case of w Surrounding words The higher the probability, express w The surrounding words, w Represents any word in a sentence. For example, a sentence " there is an apple on the table ”, the expectation is that given a word w (For example, apple ), can predict w Surrounding words context(w) have an, on, the, table And other words.

[0113] The Skip-Gram word2vec strategy is used to learn the distributed representation of the one-hot encoded k-mer base segment to convert the one-dimensional DNA sequence into a (101, 16) two-dimensional vector.

[0114] 1) The Skip-gram strategy predicts the probability of surrounding base segments by inputting a DNA base segment, thereby obtaining high-order correlations.

[0115] 2) Randomly sample 5,000 sequences from each positive dataset to form a training set of approximately 430,000 sequences.

[0116] 3) The loss function is optimized using mini-batch gradient descent with a batch size of 256, Adam as the optimizer, and a learning rate of 0.001. After approximately five epochs, the loss no longer decreases, so training is terminated. The weight matrix of the hidden layer can be defined as a piecewise distributed representation.

[0117] S34: Characterizing the encoded k-mer base segments using the distributed encoding to obtain encoded DNA sequence data.

[0118] Other features were encoded using a feature matrix. DNA shape was converted into a two-dimensional vector of (101, 14), eight histone modifications were converted into a two-dimensional vector of (101, 8), chromatin accessibility was converted into a two-dimensional vector of (101, 1), and conservation score was converted into a two-dimensional vector of (101, 1).

[0119] The encoded DNA sequence data is:

[0120]

[0121] in, Indicates the i Characteristic encoding of base segments.

[0122] S4: performing normalization processing on other data in the preprocessed original data to obtain normalized data;

[0123] The normalized data is:

[0124]

[0125] in, S m 、S h 、S d 、S c Represents the sequence S The corresponding feature matrix of DNA shape, chromatin accessibility, histone modification and conservation score, 、 、 、 They represent the DNA shape, chromatin accessibility, histone modification and conservation scores corresponding to the i-th base in the DNA sequence, respectively.

[0126] S5: performing global dependency extraction, feature extraction, and feature fusion on the encoded DNA sequence data and the normalized data using a multi-feature fusion convolutional neural network to obtain an overall feature mapping combination;

[0127] refer to Figure 3 and Figure 4 As shown, the multi-feature fusion convolutional neural network includes a multi-head self-attention mechanism module, a first feature fusion module, a convolutional feedforward neural network module, a second feature fusion module, a convolutional layer 1, a maximum pooling layer, a convolutional layer 2, a global pooling layer, a fully connected layer and a ReLU (linear rectification function) activation layer, which are arranged in sequence.

[0128] The multi-head self-attention mechanism module is used to obtain the encoded DNA sequence data / normalized data and capture the long-range and short-range dependencies between different features in the encoded DNA sequence data / normalized data;

[0129] In order to capture the long and short distance dependencies in different features, a Transformer-based self-attention mechanism is used, and the number of stacking n of Transformer modules is 2.

[0130] The first feature fusion module is used to fuse the encoded DNA sequence data / normalized data with the long-distance and short-distance dependencies to obtain a first fusion result;

[0131] The convolutional feedforward neural network module is used to perform feature enhancement on the first fusion result to obtain an enhanced result;

[0132] The second feature fusion module is used to fuse the first fusion result and the enhanced result to obtain a global dependency extraction result;

[0133] The dimension of the final DNA sequence feature map is 101×16, and the dimension of the other transcription factor binding feature maps is 101×24.

[0134] The convolutional layer 1 is used to extract initial features from the global dependency extraction result and transmit the initial features to the maximum pooling layer;

[0135] The number of filters is set to 128, the size of the convolution kernel is 1×8, and the stride of the convolution window is 1. This layer extracts 128 features from the input data, and the dimension of the output DNA sequence feature map and other features is 94×128.

[0136] The maximum pooling layer is used to sample the feature maps in the initial features; the size of the maximum pooling window is set to 1×2, the step size is 2, and the dimension of the output DNA sequence features and other feature maps is 47×64.

[0137] The convolutional layer 2 is used to perform convolution calculation on the feature map to obtain the calculated feature map; the number of filters of the convolutional layer 2 of the present invention is 256, the convolution kernel size is 1×8, the step size is 2, the activation function is ReLU, and the dimension of the output DNA sequence features and other feature maps is 40×256.

[0138] The global pooling layer is used to sample the calculated feature map to obtain feature extraction results. The sliding window size of the global pooling layer is the same as that of the entire feature map, and each W×H×C input feature map is converted into a 1×1×C output. Using GlobalMaxPooling2D in pytorch, the dimension of the output DNA sequence feature map is 20×256, and the dimension of other feature maps is 1×256.

[0139] The fully connected layer and the ReLU activation layer are used to fuse the global dependency extraction result and the feature extraction result to obtain an overall feature mapping combination.

[0140] Each feature map of the DNA sequence and other data after the pooling layer is added together for fusion. During the fusion process, each corresponding position data of the DNA sequence data and other TF combined data is added together, and the dimension after addition is ;

[0141] The fused feature map is passed to the regularization layer for regularization, using the dropout function. In this embodiment, the dropout probability is set to 0.2. The regularized feature map is passed to the fully connected layer. The dimension of the feature map output by the hidden layer is 1×64 and activated using the ReLU activation function. The final output layer has a dimension of 1×2 and is activated using the softmax function.

[0142] When given an input feature matrix X, the multi-head attention mechanism can be expressed as:

[0143]

[0144]

[0145] represents the feature matrix containing position information, A represents the output of multi-head self-attention, P Represents position information, which can be obtained through training and learning. represents the multi-head attention mechanism, Represents the Layernormalization operation. It can be specifically expressed as:

[0146]

[0147] in,

[0148] Represents the output of the i-th attention head, there are 2 dmodels in total, dmodel represents the dimension of the position vector, is the weight matrix of the model. Specifically, Indicates the i Attention heads H i The corresponding weight matrix. Concat(·) Represents a splicing operation, For the scaling factor, use Scaling to prevent The inner product is too large.

[0149] S6: training the multi-feature fusion TF-DNA binding recognition model using a transfer learning method according to the overall feature mapping combination to obtain a trained multi-feature fusion TF-DNA binding recognition model;

[0150] A1: For the transcription factor AT in tissue T, use the remaining datasets for pre-training until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases;

[0151] A2: Fine-tune the multi-feature fusion TF-DNA binding recognition model using the AT dataset until the loss no longer decreases;

[0152] A3: Outputting the multi-feature fusion TF-DNA binding recognition model;

[0153] The strong transfer learning method includes:

[0154] B1: For a transcription factor A, pre-train using the datasets of all tissues of A until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases;

[0155] B2: Freeze all network structures except the fully connected layer;

[0156] B3: Fine-tune the multi-feature fusion TF-DNA binding recognition model using target tissue data until the loss no longer decreases;

[0157] B4: Output the multi-feature fusion TF-DNA binding recognition model.

[0158] S7: Using the trained multi-feature fusion TF-DNA binding recognition model to identify the data to be identified, to obtain a recognition result.

[0159] Example 2

[0160] The beneficial effects of the present invention are verified by comparative experiments below.

[0161] This experiment used data extracted from a public database containing five types of TF-DNA binding data from 34 tissues. Prediction performance was compared using DeepBind (method 1), DeepSEA (method 2), CNN_Zeng (method 3), DLBSS (method 4), FactorNet (method 5), and our proposed method. Four proposed methods were used: GHTNet (pre-trained without transfer learning), GHTNet-DNA (pre-trained without transfer learning and trained solely on DNA sequence data), GHTNet-transOne (pre-trained with a weak transfer learning strategy), and GHTNet-transTwo (pre-trained with a strong transfer learning strategy). The proposed method was comprehensively compared with other methods using the ACC, AUROC, and AUPRC metrics.

[0162] Table 1 Comparison results of TF-DNA binding recognition

[0163]

[0164] Overall, the method of the present invention can accurately predict whether a sequence contains transcription factor binding sites, with an average accuracy of 92.54% across 86 datasets, significantly surpassing the 84.77% of method two, 84.86% of method three, 84.50% of method one, 84.30% of method four, and 91.93% of method five. Similarly, the method of the present invention achieved the best average performance in terms of AUC and PRC compared to other methods.

[0165] Furthermore, it was observed that the combined CNN and RNN models performed better than models based solely on CNN. On average, Method 3 achieved an AUROC score 0.0041 (p=2.30e-5) and 0.0035 (p=3.10e-4) higher than Methods 1 and 2, respectively. This suggests that the combined CNN and RNN models can more accurately identify gene transcripts. When the method of the present invention predicts TFBSs using only DNA sequence data, its performance surpasses that of sequence-based methods 1, 2, and 3 across all evaluation metrics. This demonstrates that the proposed method outperforms existing RNN+CNN-based methods. Incorporating transfer learning methods significantly improved the model's performance, demonstrating that the same transcription factor shares many common features across different tissues.

[0166] It can be concluded that compared with existing TF-DNA binding prediction methods, the method of the present invention has higher prediction accuracy.

[0167] Example 3

[0168] In the third embodiment of the present invention, to analyze the importance of various features to TF-DNA, the present invention uses a leave-one-feature-out method to analyze the importance of each feature. The model construction and training process is consistent with the implementation example 1.

[0169] First, the present invention analyzed the effects of 13 DNA shape and EP features on gene transcription recognition. Experiments showed that relying solely on these 14 features could predict transcription factor binding sites, but the performance was lower than that of models relying on DNA sequence ( Figure 5 b). On average, the AUROC decreased by 0.055 (p=3.06e-19). This suggests that, while DNA shape is less important than DNA sequence in identifying gene transcription binding, it may be helpful in identifying tissue-specific binding of transcription factors.

[0170] Next, the present invention fused DNA sequence and DNA shape data. After incorporating DNA shape, model performance improved, with an AUROC increase of 0.0027 (p=5.32e-3) compared to a model relying solely on DNA sequence (Figure 5b). This demonstrates that incorporating DNA shape can effectively identify tissue-specific binding of transcription factors. To further investigate the impact of DNA shape on transcriptional binding, the present invention categorized DNA shape into three separate categories: intra-base, inter-base, and MWG. A leave-one-out approach was used to analyze the importance of each DNA shape within each of these three categories. Overall, the three DNA shape categories had different average contributions to gene transcription recognition, with intra-base, inter-base, and MWG contributing 36.4%, 37.9%, and 25.6%, respectively. Furthermore, the present invention analyzed each DNA shape within the inter-base and intra-base categories and found that each DNA shape contributed differently. Roll and Buckle are the most important features of the two types of DNA shapes, contributing 25.56% and 37.54% respectively (Figure 5a). In particular, the two DNA shapes of HelT and Rise have a negative contribution to the recognition of gene transcription binding. After removing these two shapes, the area under the receiver operating characteristic curve (AUROC) of the model increased by 0.0021 and 0.0017. The present invention analyzes the importance of DNA sequence and DNA shape in different tissues. Take CTCF (a transcription factor) and POLR2A (a transcription factor) as examples. The present invention found that there are significant differences in the importance of DNA sequence and DNA shape in different tissues. For CTCF and POLR2A, the importance of DNA sequence is lowest in vagina and Peyer's patch tissues, respectively; the importance of DNA shape is lowest in spleen and thyroid gland tissues, respectively ( Figure 5 c). This indicates that both DNA shape and DNA sequence are important factors affecting the tissue-specific binding of transcription factors.

[0171] The present inventors have discovered that epigenomic signatures can predict gene transcriptional binding using only eight histone modification and DNase data, but their performance is somewhat inferior to models that rely on DNA sequence. In general, DNase I hypersensitive sites (DNase) are more important than histone modifications for predicting TF-DNA binding. Furthermore, the present inventors have discovered that the importance of these signatures exhibits strong tissue specificity, with the same signature for the same TF exhibiting significant differences in importance across different tissues.

[0172] It can be concluded that analyzing the importance of different features for TF-DNA binding by the method of the present invention helps to explain the differential expression of genes in different human tissues.

[0173] Example 4

[0174] In the fourth embodiment of the present invention, in order to solve the "black box" problem of the model, based on the trained model in the first implementation case, the attention mechanism is visualized to extract the areas that affect the model decision, and the convolution kernel is used to perform motif mining, thereby analyzing the tissue-specific binding of transcription factors. The following steps are included:

[0175] X1: Model training. Use the model trained in Implementation Case 1 to conduct the following experiments.

[0176] X2: Motif detector selection. Potential motif detectors are selected in the convolution kernel through max pooling.

[0177] X3: Extract TF-DNA binding regions. The maximum activation value of the motif detector is I. For each positive sequence, the position with the maximum activation value selected by the current motif detector is selected. Positions are considered potential binding sites only if their activation value exceeds a manually selected threshold (0.7I).

[0178] X4: Motif similarity comparison: The motifs extracted by the present invention are compared and visualized with the motifs verified in the database using the TOMTOM tool.

[0179] X5: Attention visualization mechanism. Use self-attention to evaluate and visualize the importance of each position in the sequence ( Figure 6 ).

[0180] Specifically, the detailed process of the above step 2 is as follows:

[0181] X21: For each sequence, calculate the maximum activation value of each convolution kernel.

[0182] X22: Find the convolution kernel with the maximum value among these maximum activation values.

[0183] X23: Perform the two-step operations X21 and X22 on the entire positive sample data set, and count the number of times the maximum value of each convolution kernel appears.

[0184] X24: The convolution kernel whose maximum value occurs more than one-fifth of the number of samples is used as a motif detector.

[0185] The data of this experiment is the data in Example 1. DeepBind (method 1), MEME-ChIP (method 2), gkm-SVM (method 3) and the method of the present invention were used to perform motif similarity comparison.

[0186] The prediction method provided by the present invention can be implemented in a software manner to realize the automatic operation of the process. The device for running the process should also be within the scope of protection of the present invention.

[0187] The beneficial effects of the present invention are verified by the following experimental results.

[0188] Table 2 Motif similarity comparison results

[0189]

[0190] As can be seen from Table 2, the method of the present invention can capture transcription factor binding motifs more effectively than other methods. It is significantly better than other methods in three evaluation indicators. In addition, Figure 7 The attention scores and corresponding attention maps for four ChIP-Seq-validated CTCF binding sites in pancreatic tissue are displayed. This demonstrates that the attention mechanism employed in this paper can prioritize TF-DNA binding regions, helping us better understand the mechanisms of TF-DNA binding. This also indirectly supports the ability of this method to better identify TF-DNA binding.

[0191] In summary, the present invention has designed a method for identifying TF-DNA binding based on multi-feature fusion, which can effectively improve the prediction performance of TF-DNA binding. The research results of the present invention can be applied in the biomedical field. Researchers can use the present method to predict TF-DNA binding and screen potential binding regions from massive amounts of data, effectively overcoming the time-consuming and labor-intensive shortcomings of traditional experiments. The present method can also analyze features that influence the specific binding of transcription factors, helping to explain differential gene expression in different human tissues and providing clues for disease diagnosis, development of therapeutic targets, and elucidation of disease etiology. In addition, the present invention can identify important regions in the genome, and analysis of these important regions can also provide important biological insights.

[0192] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A TF-DNA binding recognition method based on multi-feature fusion, characterized in that: The TF-DNA binding recognition method based on multi-feature fusion includes: S1: Obtained original data related to transcription factor binding in five common human tissues; S2: Preprocessing the raw data to obtain preprocessed raw data; the preprocessed raw data includes: DNA sequence data and its corresponding DNA shape data, chromatin accessibility data, histone modification data and conservation data; S3: performing data encoding processing on the DNA sequence data in the pre-processed raw data to obtain encoded DNA sequence data; S4: normalizing the DNA shape data, chromatin accessibility data, histone modification data, and conservation data in the pre-processed raw data to obtain normalized data; S5: using a multi-feature fusion attention mechanism and a convolutional neural network to perform global dependency extraction, feature extraction, and feature fusion on the encoded DNA sequence data and the normalized data to obtain an overall feature mapping combination; the multi-feature fusion convolutional neural network includes a multi-head self-attention mechanism module, a first feature fusion module, a convolutional feedforward neural network module, a second feature fusion module, a convolutional layer 1, a maximum pooling layer, a convolutional layer 2, a global pooling layer, a fully connected layer, and a ReLU activation layer, which are arranged in sequence; The multi-head self-attention mechanism module is used to obtain the encoded DNA sequence data / normalized data and capture the long-range and short-range dependencies between different features in the encoded DNA sequence data / normalized data; The first feature fusion module is used to fuse the encoded DNA sequence data / normalized data with the long-distance and short-distance dependencies to obtain a first fusion result; The convolutional feedforward neural network module is used to perform feature enhancement on the first fusion result to obtain an enhanced result; The second feature fusion module is used to fuse the first fusion result and the enhanced result to obtain a global dependency extraction result; The convolutional layer 1 is used to extract initial features from the global dependency extraction result and transmit the initial features to the maximum pooling layer; The maximum pooling layer is used to sample the feature map of the initial features; The convolution layer 2 is used to perform convolution calculation on the feature map to obtain a calculated feature map; The global pooling layer is used to sample the calculated feature map to obtain a feature extraction result; The fully connected layer and the ReLU activation layer are used to fuse the global dependency extraction result and the feature extraction result to obtain an overall feature mapping combination; S6: training the multi-feature fusion TF-DNA binding recognition model using a transfer learning method according to the overall feature mapping combination to obtain a trained multi-feature fusion TF-DNA binding recognition model; Transfer learning methods include weak transfer learning methods and strong transfer learning methods. Weak transfer learning methods include: A1: For the transcription factor AT in tissue T, use the remaining datasets for pre-training until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases; A2: Fine-tune the multi-feature fusion TF-DNA binding recognition model using the AT dataset until the loss no longer decreases; A3: Outputting the multi-feature fusion TF-DNA binding recognition model; Strong transfer learning methods include: B1: For a transcription factor A, pre-train using the datasets of all tissues of A until the loss of the multi-feature fusion TF-DNA binding recognition model no longer decreases; B2: Freeze all network structures except the fully connected layer; B3: Fine-tune the multi-feature fusion TF-DNA binding recognition model using target tissue data until the loss no longer decreases; B4: Outputting the multi-feature fusion TF-DNA binding recognition model; S7: Using the trained multi-feature fusion TF-DNA binding recognition model to identify the data to be identified, to obtain a recognition result.

2. The TF-DNA binding recognition method based on multi-feature fusion according to claim 1, characterized in that: The step S2 comprises: S21: extracting DNA sequence data from the raw data using a chromatin immunoprecipitation sequencing dataset; S22: Processing the DNA sequence data using GKM-SVM to obtain positive samples and negative samples; S23: Acquire three types of DNA shape data of the positive sample and the negative sample; the three types of DNA shape data include: A: inter-bp: HelT, Rise, Roll, Shift, Slide and Tilt; B: intra-bp: Buckle, Opening, ProT, Shear, Stagger and Stretch; C: MGW and EP; S24: extracting chromatin accessibility data and histone modification data from the DNA sequence data; S25: generating conservation data of positive samples and negative samples using phastCons100way according to the gene coordinates of the DNA sequence data; S26: Outputting the DNA sequence data, the DNA shape data, the chromatin accessibility data, the histone modification data and the conservation data as the pre-processed raw data.

3. The TF-DNA binding recognition method based on multi-feature fusion according to claim 2, characterized in that: The step S22 includes: S221: For each data set, based on the gene coordinates, the sequence centered at the peak is expanded to 101 bp as the positive sample; S222: Select a region in the whole genome with a similar GC content to the positive sample as the negative sample.

4. The TF-DNA binding recognition method based on multi-feature fusion according to claim 2, characterized in that: The step S3 comprises: S31: dividing the positive sample and the negative sample into a plurality of k-mer base segments of equal sequence length; S32: Encoding the plurality of k-mer base segments using hot-single encoding to obtain a plurality of encoded k-mer base segments; S33: According to the word2vec strategy, the high-dimensional sparse single-hot encoding is converted into distributed encoding; S34: Characterizing the encoded k-mer base segments using the distributed encoding to obtain encoded DNA sequence data.

5. The TF-DNA binding recognition method based on multi-feature fusion according to claim 2, characterized in that: In step S4, the normalized data is: in, S m 、S h 、S d 、S c Represents the sequence S The corresponding feature matrix of DNA shape, chromatin accessibility, histone modification and conservation score, 、 、 、 They represent the DNA shape, chromatin accessibility, histone modification and conservation scores corresponding to the i-th base in the DNA sequence, respectively.

6. A TF-DNA binding recognition system, characterized in that: The TF-DNA binding recognition system includes a processor, a memory, and a computer program stored in the memory, wherein the computer program executes the TF-DNA binding recognition method based on multi-feature fusion according to any one of claims 1 to 5 on the processor.

Images