A method for predicting gene expression levels based on transcription factor combination patterns

Abstract

This invention provides a method for predicting gene expression levels based on transcription factor combination patterns. It addresses the shortcomings of existing models, such as insufficient interpretability. The method includes data acquisition and preprocessing; transcription factor sequence encoding; feature engineering and label generation; dataset construction; deep learning model construction; model evaluation; visualization analysis; model interpretability analysis; and regulatory pattern discovery. This invention offers advantages such as high automation, good pre-compression effect, and strong versatility.

Description

Technical Field

[0001] This invention belongs to the field of healthcare informatics technology, specifically relating to a method for predicting gene expression levels based on transcription factor combination patterns. Background Technology

[0002] Gene expression levels or gene activity are influenced by inherent genomic properties such as DNA sequence and epigenetic modifications, and can be modeled and predicted using machine learning or deep learning methods. Machine learning methods (such as linear regression, support vector machines, random forests, and XGBoost) and deep learning methods (convolutional neural networks, Transformers) have been widely used for gene expression prediction. Commonly used models include Basset, Basenji, Enformer, Puffin, and GET. In the study of gene expression levels, most existing methods are based on indirect regulatory factors such as DNA sequence, histone modifications, and chromatin structure. While these methods map DNA sequence context and chromatin state to gene expression levels to some extent, they cannot reflect the contribution of transcription factor combination patterns that gene transcription directly depends on. Therefore, existing models have shortcomings in interpretability.

[0003] This approach proposes a method for predicting gene expression levels based on transcription factor combinatorial patterns, aiming to reveal gene expression regulation mechanisms using novel research methods. First, the binding maps of transcription factors in gene promoter regions are encoded with novel sequences. Then, a deep learning model is constructed to learn the combinatorial semantics of the transcription factor coding sequences, improving the accuracy of gene expression level prediction while identifying important transcription factor combinatorial patterns, thus leading to a better understanding of gene expression regulation mechanisms. Summary of the Invention

[0004] The purpose of this invention is to address the above-mentioned problems by providing a well-designed and interpretable method for predicting gene expression levels based on transcription factor combination patterns.

[0005] To achieve the above objectives, the present invention employs the following technical solution: a method for predicting gene expression levels based on transcription factor combinatorial patterns, comprising the following steps:

[0006] S1: Data acquisition and preprocessing;

[0007] S2: Encodes a transcription factor sequence;

[0008] S3: Feature Engineering and Tag Generation;

[0009] S4: Dataset Construction;

[0010] S5: Deep learning model building;

[0011] S6: Model Evaluation;

[0012] S7: Visual Analysis;

[0013] S8: Model interpretability analysis;

[0014] S9: Regulation pattern discovery.

[0015] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S1 includes the following steps:

[0016] S11: Data download, obtain transcription factor ChIP-seq data, transcription factor motif files, genome annotation files, and gene expression data TPM respectively;

[0017] S12: Process gene annotation, select protein-coding genes, and extract transcription start sites (TSS);

[0018] S13: Process the expression data, select the gene expression data TPM of the corresponding cell line, obtain the gene expression data of 3 replicate samples and take the average value to obtain the gene expression level data of the corresponding gene.

[0019] S14: Calculate the transcription factor association strength (TFAS). The formula for calculating the transcription factor association strength (TFAS) is as follows:

[0020] ;

[0021] in, It is the TFAS score between transcription factor j and gene i. It is the distance from the center of the k-th peak of the j-th transcription factor to the transcription start site (TSS) of the target gene i. It is the intensity of the k-th peak. It is a constant whose value is set to 2kb based on the distribution of peaks in the region;

[0022] S15: Correlation analysis, calculating the Spearman correlation between TFAS and TPM.

[0023] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S2 includes the following steps:

[0024] S21: Define the 5kb upstream and downstream of the transcription start site as the promoter region;

[0025] S22: The identified transcription factors and TSS are letter-mapped to participate in the encoding;

[0026] S23: The transcription factor binding map is converted into transcription factor sequence encoding. In the promoter region, the corresponding transcription factor representative letters are arranged in sequence according to the coordinates of the center sites of the binding peaks of each transcription factor with the genome to obtain the sequence encoding of the transcription factor.

[0027] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S3 includes the following steps:

[0028] S31: All sequence samples from both cell lines are uniformly encoded using one-hot encoding, and the binding strength of transcription factors is also extracted as a one-dimensional feature for model learning.

[0029] S32: Genes are categorized into high, medium, and low expression levels based on the upper and lower quartiles of 10lg(TPM+1).

[0030] In the above method for predicting gene expression levels based on transcription factor combination patterns, step S4 divides the dataset by dividing high-expression and low-expression samples into a training set and an independent test set in an 8:2 ratio.

[0031] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S5 involves constructing a convolutional neural network model, DeepTF, to classify gene expression levels. This model consists of three convolutional blocks, a fully connected layer, and an output layer, and includes the following steps:

[0032] S51: Extracting local features of transcription factor sequences through convolutional layers;

[0033] S52: Reduce feature dimensionality using pooling layers;

[0034] S53: Feature integration and classification are performed through fully connected layers;

[0035] S54: Finally, the predicted value of gene expression level is output through the sigmoid activation function;

[0036] S55: The Adam optimizer is used to dynamically adjust the learning rate based on the gradient of the parameters to generate the final predicted classification result.

[0037] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S6 includes performance metrics for model evaluation such as accuracy, loss, and area under the ROC curve (AUC).

[0038] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S7 plots loss curves and ROC curves to demonstrate the model's accuracy and generalization ability. The ROC curve uses the relationship between TPR and FPR to evaluate the classification model's performance, showing the classification ability at different thresholds. The AUC value quantifies the overall performance of the model. The AUC value is between 0 and 1. The closer the value is to 1, the better the model's performance. A value of 0.5 indicates that the model's predictive ability is comparable to random guessing.

[0039] In the above-mentioned method for predicting gene expression levels based on transcription factor combinatorial patterns, step S8 includes the following steps:

[0040] S81: Background data selection: 100 samples are randomly selected from the test set as the background dataset;

[0041] S82: SHAP value calculation, using the DeepExplainer interpreter to calculate the contribution of each feature to the model output;

[0042] S83: Traverse each transcription factor, extract its SHAP value at all positions, and calculate the average of the absolute values ​​of these SHAP values;

[0043] S84: Based on the obtained average values, draw a ranking diagram of the importance of transcription factors to the model.

[0044] In the above-mentioned method for predicting gene expression levels based on transcription factor combination patterns, step S9 uses MEME and Weblogo tools to discover transcription factor combination patterns.

[0045] Compared with existing technologies, the advantages of this invention are as follows: Encoding the binding position information of transcription factors in promoter regions into transcription factor sequences for studying gene expression regulation, compared with research methods based on DNA sequences and histone modifications, focuses only on transcription factor binding events directly involved in transcription regulation, making it easier to discover the transcriptional regulatory code; Utilizing the powerful feature extraction capabilities of deep learning, especially when processing complex biological data, it can capture more nonlinear features and patterns, thereby improving the predictive performance of the model; Based on transcription factor sequence encoding and deep learning, it discovers transcription factor combination patterns closely related to gene expression, providing new insights into the study of transcriptional regulatory language. Attached Figure Description

[0046] Figure 1 This is a flowchart of the process of this invention;

[0047] Figure 2 This is a Spearman correlation diagram of transcription factor association strength and gene expression level according to the present invention;

[0048] Figure 3 This is a box plot of gene expression levels from the present invention;

[0049] Figure 4 This is a diagram of the DeepTF model architecture of the present invention;

[0050] Figure 5 This is a loss graph of DeepTF of the present invention on the training set of GM12878 and K562 cell lines;

[0051] Figure 6 This is the ROC curve of DeepTF in the GM12878 and K562 cell lines on independent test sets;

[0052] Figure 7 This is a ranking diagram of the importance of transcription factors in the GM12878 cell line of the present invention;

[0053] Figure 8 This is a ranking diagram of the importance of transcription factors in the K562 cell line of this invention;

[0054] Figure 9 This is a diagram showing the results of the MEME tool of this invention in discovering transcription factor combinatorial patterns;

[0055] Figure 10 This is a diagram showing the results of the Weblogo tool of this invention discovering transcription factor combination patterns. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1-10 As shown, a method for predicting gene expression levels based on transcription factor combinatorial patterns includes the following steps:

[0058] S1: Data acquisition and preprocessing;

[0059] S2: Encodes a transcription factor sequence;

[0060] S3: Feature Engineering and Tag Generation;

[0061] S4: Dataset Construction;

[0062] S5: Deep learning model building;

[0063] S6: Model Evaluation;

[0064] S7: Visual Analysis;

[0065] S8: Model interpretability analysis;

[0066] S9: Regulation pattern discovery.

[0067] Specifically, in step S1, we first downloaded the narrowpeak format files of ChIP-seq data for 13 transcription factors shared by the GM12878 and K562 cell lines and closely related to transcription from the ENCODE database; secondly, we downloaded transfac format transcription factor motif files from the JASPAR database; then, we selected the genome annotation file with the gene annotation version number Feb.2009 (hg19 / CRCh37) and downloaded it from UCSGCenomeBrowser. Based on the gene annotation, we selected protein-coding genes and obtained the corresponding transcription start sites (TSS); finally, we selected the gene expression data (TPM) of the corresponding cell lines. For each cell line, we obtained gene expression data from three replicates. The average of the three samples was used to obtain the gene expression level data of the corresponding gene. To quantify the strength of the association between transcription factors and genes, we calculated the transcription factor association strength (TFAS). The formula for calculating TFAS is:

[0068] ;

[0069] in, It is the TFAS score between transcription factor j and gene i. It is the distance from the center of the k-th peak of the j-th transcription factor to the transcription start site (TSS) of the target gene i. It is the intensity of the k-th peak. It is a constant whose value is set to 2kb based on the distribution of peaks in the region;

[0070] like Figure 2 As shown, in order to quantify the relationship between transcription factor association strength and gene expression level, the Spearman correlation between TFAS and TPM was calculated.

[0071] In-depth, after identifying genes and transcription factors in step S2, transcription factor sequence encoding is performed, with the specific steps as follows:

[0072] The promoter region is defined as 5 kb upstream and downstream of the transcription start site.

[0073] The 13 identified transcription factors and TSS were assigned letter mappings for coding (TBP: "T", SP1: "S", YY1: "Y", BHLHE40: "B", ETS1: "E", NFYB: "F", POLAR2A: "P", CTCF: "C", NRF1: "N", YBX1: "X", CEBPZ: "Z", NFYA: "A", NR2C2: "R", TSS: "O").

[0074] The transcription factor binding map is converted into transcription factor sequence encoding. Specifically, in the promoter region, the corresponding letters of the transcription factor are arranged sequentially according to the coordinates of the center sites of the binding peaks of each transcription factor with the genome to obtain the sequence encoding of the transcription factor. To conform to the input of the model, the letter "O" in the already encoded sequence is padded with equal-length "I" letters on both sides.

[0075] like Figure 3 As shown, all sequence samples from both cell lines were uniformly encoded using one-hot encoding. The binding strength of transcription factors was also extracted and used as a one-dimensional feature for model learning. Genes were categorized into high, medium, and low expression levels based on the upper and lower quartiles of 10lg(TPM+1): data below Q1 (first quartile) were defined as low expression, equivalent to the 25th percentile of all data in the sample arranged from smallest to largest; data above Q3 (upper quartile) were defined as high expression, equivalent to the 75th percentile of all data in the sample arranged from smallest to largest; and data between Q1 and Q3 were defined as medium expression.

[0076] Furthermore, step S4, in order to study the contribution of transcription factors to gene expression regulation, involves a binary classification problem of highly expressed genes and low-expressed (including non-expressed) genes. The dataset is divided into a training set and an independent test set, with the high-expression and low-expression samples split in an 8:2 ratio.

[0077] Step S5 constructs the DeepTF convolutional neural network model to classify gene expression levels. The model framework diagram is shown below. Figure 4 As shown, this model consists of three convolutional blocks, fully connected layers, and an output layer, used to classify gene expression levels. Local features of transcription factor sequences are extracted through convolutional layers, feature dimensionality is reduced using pooling layers, features are integrated and classified through fully connected layers, and finally, the predicted gene expression level is output through a sigmoid activation function. During training, the model is optimized using a binary cross-entropy loss function and the Adam optimizer, and the learning rate is adjusted using a learning rate scheduler to generate the final predicted classification result. The specific steps are as follows:

[0078] Convolutional Blocks: Each convolutional block contains two convolutional layers used to extract local features from the input data. The convolutional layers slide across the input data using a 3×3 kernel to perform convolution operations, thereby extracting spatially correlated features. Each convolutional layer is followed by a ReLU activation function, introducing non-linearity into the model and enabling it to learn more complex feature representations. Each convolutional block is connected to a max-pooling layer to reduce the spatial size of the feature map, decrease the number of parameters and computational cost, and improve the model's abstraction ability, allowing it to capture higher-level features. Each convolutional block also contains a dropout layer to randomly discard a portion of neurons, preventing overfitting and improving the model's generalization ability.

[0079] Fully Connected Layers: Fully connected layers further integrate and transform the features extracted by convolutional and pooling layers to learn complex relationships between features. The first fully connected layer maps the output features of the convolutional block to an intermediate dimension, providing richer feature representations for subsequent classification tasks. The first fully connected layer is followed by a ReLU activation function and a dropout layer to further introduce non-linearity and prevent overfitting. The second fully connected layer maps the intermediate features to the final output dimension for binary classification tasks, i.e., predicting whether gene expression levels are low or high.

[0080] Loss function: Binary Cross-Entropy Loss (BCEWithLogitsLoss), commonly used for binary classification tasks. Its input is the predicted probability after activation by the Sigmoid function (usually represented using one-hot encoding). The formula for calculating the loss function is:

[0081] ;

[0082] in It is the Sigmoid function, where N is the number of samples. It is the probability that the model predicts the i-th sample to be of the positive class. It is the true label of the i-th sample.

[0083] Optimizer: The Adam (Adaptive Moment Estimation) optimizer is used. It is a widely used adaptive learning rate optimization algorithm that combines the advantages of momentum and RMSProp optimizers. It can dynamically adjust the learning rate according to the gradient of the parameters, thereby accelerating the convergence speed of the model and improving the stability of training.

[0084] Meanwhile, step S6 evaluates the model's performance metrics, including accuracy, loss, and area under the ROC curve (AUC).

[0085] Accuracy refers to the proportion of samples correctly predicted by the model out of the total number of samples. The calculation formula is as follows:

[0086] ;

[0087] Where TP is the number of samples correctly predicted as positive, TN is the number of samples correctly predicted as negative, FP is the number of samples incorrectly predicted as positive, and FN is the number of samples incorrectly predicted as negative.

[0088] A loss function measures the difference between the model's predicted values ​​and the true values. In binary classification tasks, the commonly used loss function is the binary cross-entropy loss.

[0089] AUC (Area Under the ROC Curve) is the area under the ROC curve, used to measure a model's ability to distinguish between positive and negative samples. A higher AUC value indicates better model performance. The calculation formula is as follows:

[0090] ;

[0091] Where TPR is the true positive rate and FPR is the false positive rate.

[0092] As can be seen, step S7 plots the loss curve and ROC curve to show the model's accuracy and generalization ability; the ROC curve uses the relationship between TPR and FPR to evaluate the classification model's performance and shows the classification ability at different thresholds; the AUC value quantifies the overall performance of the model. The AUC value is between 0 and 1. The closer the value is to 1, the better the model's performance. A value of 0.5 indicates that the model's predictive ability is comparable to random guessing.

[0093] Step S8 uses the SHAP interpretation method to calculate the contribution of transcription factors to the model, as follows: Figure 7-8As shown in the diagram, firstly, the top 100 samples from the test set were selected as background data. The DeepExplainer interpreter was chosen to calculate the contribution of each feature to the model output, i.e., the SHAP value. Secondly, each transcription factor was traversed, and its SHAP value at all positions was extracted. The average of the absolute values ​​of these SHAP values ​​was then calculated. Finally, the transcription factors were sorted according to the obtained average values, and a ranking diagram of the importance of the transcription factors to the model was drawn. It can be seen that, based on the model SHAP values, the transcription factors TBP, YY1, SP1, and RNA polymerase POLAR2A made the greatest contributions to the prediction, which is consistent with the statistical results based on the transcription factor association strength (TFAS). TBP is the core subunit of the transcription factor TFIID. It binds to the TATA-box, anchors the promoter region, and helps POLAR2A to locate and form the transcription initiation complex PIC. The transcription factor SP1 regulates gene expression by binding to the GC-box and helps TFIID bind to DNA in the TATA-less promoter. YY1 binds downstream of TSS and may help POLAR2A overcome promoter-proximal pause, promoting transcription to enter the elongation stage.

[0094] Preferably, step S9 uses MEME and Weblogo tools to discover transcription factor combinatorial patterns, and the specific steps are as follows:

[0095] MEME Tools: MEMESuite is an online platform that aggregates numerous tools for predicting and annotating motifs. Here, we use MEME to discover transcription factor combinatorial patterns de novo. First, we define a custom alphabet containing 14 letters corresponding to transcription factor names and colors. Second, we divide the sequence samples from each cell line into positive and negative sets based on the upper and lower quartiles of 10lg(TPM+1). Sequences smaller than Q1 are classified as negative, and those larger than Q3 as positive. Coding sequences shorter than 3 bytes are deleted, and the remaining sequences are written in FASTA format. Finally, we use the positive (or negative) sets as the sequences to search in MEME. The results are as follows: Figure 9 As shown, both cell lines exhibit a preference for the "STOY" combination pattern in highly expressed genes, while the "CYO" pattern frequently appears in low-expressed genes. The discovery of these transcription factor combination patterns is significant for decoding gene regulatory syntax.

[0096] in Figure 9 Part (a) shows that the GM12878 cell line used the negative set as the control sequence and the positive set as the sequence to be searched, and the resulting combination of transcription factors was "SSTOYP" and "SPPTOY". Figure 9Part (b) shows that the GM12878 cell line used the positive set as the control sequence and the negative set as the sequence to be searched, and the resulting combination of transcription factors was "CYO" and "FSFO". Figure 9 Part (c) shows that the K562 cell line used the negative set as the control sequence and the positive set as the sequence to be searched, and the resulting combination of transcription factors was "SSCTOYP" and "CTOYPP". Figure 9 Part (d) shows that the K562 cell line used the positive set as the control sequence and the negative set as the sequence to be searched, and the resulting combination of transcription factors was "NCTY" and "SNNO".

[0097] Weblogo Tool: WebLogo is a web-based sequence logo generator. Specifically, it divides the processed equal-length sequence sample data into high-expression and low-expression data according to the upper and lower quartiles of 10lg(TPM+1), with values ​​less than Q1 indicating low expression and values ​​greater than Q3 indicating high expression. WebLogo then loads the high-expression and low-expression data respectively to generate sequence logo diagrams. The results are as follows... Figure 10 As shown, POLAR2A has the highest probability of occurrence at all loci in highly expressed genes, which is consistent with the dynamic process of POLAR2A sliding from the 5' end to the 3' end of the gene during transcription. Furthermore, the sequence logo further confirms the influence of patterns such as "STOY" and "CYO" on gene expression.

[0098] In summary, this embodiment proposes an encoding method based on the binding position and sequence of transcription factors. This method transforms the physical location information of transcription factors into sequence information and uses numerical vectorization to make it suitable as input for deep learning models, providing a novel data dimension for predicting gene expression levels. The method utilizes CNNs combined with transcription factor sequences to accurately predict gene expression levels. Based on transcription factor sequence identification, transcription factor combination patterns closely related to gene expression, such as "STOY" and "CYO," are obtained.

[0099] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0100] Although this paper uses terms such as transcription factors frequently, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A method for predicting gene expression levels based on transcription factor combinatorial patterns, characterized in that, Includes the following steps: S1: Data acquisition and preprocessing, downloading data and processing gene annotations, selecting and processing gene expression data, quantifying the strength of the association between transcription factors and genes and calculating their correlation with gene expression levels; S2: Transcription factor sequence encoding, defining the promoter region, performing letter mapping to participate in encoding, and combining the map to obtain the transcription factor sequence encoding; S21: Define the 5kb upstream and downstream of the transcription start site as the promoter region; S22: The identified transcription factors and TSS are letter-mapped to participate in the encoding; S23: The transcription factor binding map is converted into transcription factor sequence coding. In the promoter region, the corresponding transcription factor representative letters are arranged in sequence according to the coordinates of the center sites of the binding peaks of each transcription factor with the genome to obtain the sequence coding of the transcription factor. S3: Feature engineering and tag generation, which encodes samples and tags genes; S4: Dataset construction, splitting the training set and independent test set; S5: Deep learning model construction, building a convolutional neural network model DeepTF and performing prediction and classification; S6: Model Evaluation, set the performance metrics for model evaluation and perform calculations; S7: Visual analysis, plotting loss curves and ROC curves; S8: Model interpretability analysis, using the SHAP interpretation method to calculate the contribution of transcription factors to the model; S9: Regulation pattern discovery.

2. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 1, characterized in that, Step S1 includes the following steps: S11: Data download, obtain transcription factor ChIP-seq data, transcription factor motif files, genome annotation files, and gene expression data TPM respectively; S12: Process gene annotation, select protein-coding genes, and extract transcription start sites (TSS); S13: Process the expression data, select the gene expression data TPM of the corresponding cell line, obtain the gene expression data of 3 replicate samples and take the average value to obtain the gene expression level data of the corresponding gene. S14: Calculate the transcription factor association strength (TFAS). The formula for calculating the transcription factor association strength (TFAS) is as follows: ； in, It is the TFAS score between transcription factor j and gene i. It is the distance from the center of the k-th peak of the j-th transcription factor to the transcription start site (TSS) of the target gene i. It is the intensity of the k-th peak. It is a constant whose value is set to 2kb based on the distribution of the peaks; S15: Correlation analysis, calculating the Spearman correlation between TFAS and TPM.

3. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 1, characterized in that, Step S3 includes the following steps: S31: All sequence samples from both cell lines are uniformly encoded using one-hot encoding, and the binding strength of transcription factors is also extracted as a one-dimensional feature for model learning. S32: Genes are categorized into high, medium, and low expression levels based on the upper and lower quartiles of 10lg(TPM+1).

4. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 3, characterized in that, Step S4 involves dividing the dataset by splitting the high-expression and low-expression samples into a training set and an independent test set in an 8:2 ratio.

5. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 4, characterized in that, In step S5, the DeepTF convolutional neural network model is constructed to classify gene expression levels. This model consists of three convolutional blocks, a fully connected layer, and an output layer, and includes the following steps: S51: Extracting local features of transcription factor sequences through convolutional layers; S52: Reduce feature dimensionality using pooling layers; S53: Feature integration and classification are performed through fully connected layers; S54: Finally, the predicted value of gene expression level is output through the sigmoid activation function; S55: The Adam optimizer is used to dynamically adjust the learning rate based on the gradient of the parameters to generate the final predicted classification result.

6. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 3, characterized in that, The performance metrics for model evaluation in step S6 include accuracy, loss, and area under the ROC curve (AUC).

7. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 6, characterized in that, Step S7 involves plotting loss curves and ROC curves to demonstrate the model's accuracy and generalization ability. The ROC curve uses the relationship between TPR and FPR to evaluate the classification model's performance, showing the classification ability at different thresholds. The AUC value quantifies the overall performance of the model. The AUC value is between 0 and 1. The closer the value is to 1, the better the model's performance. A value of 0.5 indicates that the model's predictive ability is comparable to random guessing.

8. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 6, characterized in that, Step S8 includes the following steps: S81: Background data selection: 100 samples are randomly selected from the test set as the background dataset; S82: SHAP value calculation, using the DeepExplainer interpreter to calculate the contribution of each feature to the model output; S83: Traverse each transcription factor, extract its SHAP value at all positions, and calculate the average of the absolute values ​​of these SHAP values; S84: Based on the obtained average values, draw a ranking diagram of the importance of transcription factors to the model.

9. The method for predicting gene expression levels based on transcription factor combinatorial patterns according to claim 6, characterized in that, Step S9 uses MEME and Weblogo tools to discover transcription factor combination patterns.

Images