Protein palmitoyl transferase prediction method and system based on multi-branch deep convolutional neural network

Abstract

The application discloses a protein palmitoyl transferase prediction method and system based on a multi-branch deep convolutional neural network, and belongs to the technical field of bioinformatics and artificial intelligence. The method comprises the following steps: S1, obtaining a protein sequence to be detected; S2, inputting the protein sequence into an iPalmT model pre-trained; and S3, judging whether the target protein is a palmitoyl transferase according to the model output result. The iPalmT model comprises an encoding module, two parallel convolution branches, a feature fusion module and a classification module; an SE module is arranged after convolution layers of each convolution branch, and is used for channel weighting and feature recalibration. The model extracts multi-level sequence features through different scale convolution kernels, realizes high-precision prediction through feature fusion and a residual structure, can automatically learn multi-scale features from large-scale data, realizes end-to-end palmitoyl transferase identification, and has high accuracy and good universality.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, specifically to a method and system for predicting protein palmitoylation transferases based on a multi-branch deep convolutional neural network. Background Technology

[0002] Palmitoyltransferases are the key catalytic enzymes mediating protein palmitoylation. Palmitoylation is an important post-translational lipid modification, typically referring to the transfer of palmitic acid from palmitoyl-CoA to cysteine ​​residues in target proteins. The dynamic balance between protein S-acylation and deacylation has been shown to profoundly influence a variety of cellular processes, including innate immunity, inflammation, glucose and lipid metabolism, and tumorigenesis and progression. Therefore, accurate identification of palmitoyltransferases is crucial for understanding multiple cellular signaling pathways and developing therapeutic strategies for related diseases.

[0003] Traditional biological experimental methods (such as mass spectrometry and enzyme activity assays) for identifying palmitoyltransferases are typically costly, low-throughput, and time-consuming. Meanwhile, some computational biology methods attempt to make predictions through sequence alignment or machine learning.

[0004] Several machine learning-based computational methods (such as SwissPalm, CSS-Palm, and PalmPred) are available for predicting palmitoylation sites on proteins. However, these methods aim to identify modified residues, rather than determining whether a protein itself is a palmitoylate. Tools such as Pfam, InterProScan, and HMMER rely on statistical modeling and sequence alignment algorithms to infer potential palmitoylate activity by detecting known functional domains (such as DHHC motifs). However, these tools are inherently limited to predefined conserved domains in existing databases, often failing to cover proteins with high sequence diversity, significant evolutionary differentiation, or atypical / non-canonical DHHC structures, leading to frequent false negatives. Furthermore, their performance depends on manual collation and database updates, limiting their timely adaptation to emerging palmitoylate variants. Most importantly, these methods lack the ability to directly learn entirely new characteristic patterns from the original sequence, thus limiting their predictive power for functionally uncharacterized proteins.

[0005] In recent years, deep learning has demonstrated significant advantages in predicting protein function and post-translational modifications. Unlike traditional machine learning methods (such as support vector machines or random forests) that rely on manual feature engineering, deep learning can automatically extract hierarchical feature representations directly from protein sequences, thereby greatly improving prediction accuracy. Furthermore, deep convolutional neural networks (CNNs), recurrent neural networks, and attention-based Transformer architectures exhibit stronger generalization ability and robustness when processing large-scale sequence data. Deep learning also possesses the flexibility to integrate multimodal information (including structural, evolutionary, and functional annotations), making it more efficient and scalable than traditional methods.

[0006] Therefore, there is an urgent need in this field for a new technical solution that can automatically, accurately, and efficiently predict palmityltransferases from raw protein sequences. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks includes the following steps:

[0010] S1. Obtain the protein sequence of the protein to be predicted;

[0011] S2. Input the protein sequence into a pre-trained palmitoyltransferase prediction model;

[0012] S3. Based on the output of the palmitoyltransferase prediction model, determine whether the protein is a palmitoyltransferase.

[0013] The palmitoyltransferase prediction model includes:

[0014] The encoding module is used to encode the protein sequence into a one-hot encoding matrix with a length of 1000 and a width of 21, where 20 dimensions correspond to standard amino acids and 1 dimension is used for filling gaps or unknown symbols;

[0015] Two parallel convolutional branches are used to extract multi-scale local sequence features from the one-hot encoding matrix;

[0016] The feature fusion module is used to concatenate the feature vectors output by the two convolutional branches;

[0017] The classification module is used for palmitoyltransferase prediction based on the fused features.

[0018] Each convolutional branch includes several two-dimensional convolutional layers of different sizes, and after the convolutional layers are stacked, an SE module is connected to perform adaptive recalibration of the feature channels of the convolutional output.

[0019] Furthermore, the feature fusion module performs the following operations in sequence:

[0020] Global pooling and flattening are performed on the feature maps output by the two convolutional branches respectively to obtain two 64-dimensional feature vectors.

[0021] The two feature vectors are concatenated into a 128-dimensional vector;

[0022] The data is projected through a fully connected layer to 512-dimensional and 128-dimensional latent feature spaces for subsequent classification.

[0023] Furthermore, each convolutional branch contains three two-dimensional convolutional layers. The kernel sizes of the first branch are (3,21), (3,1), and (3,1) respectively; the kernel sizes of the second branch are (5,21), (5,1), and (5,1) respectively. After each convolutional layer, a batch normalization layer and a ReLU activation layer are connected in sequence, and an SE module is set at the end of the branch to weight and calibrate the feature channels.

[0024] Furthermore, the SE module performs the following operations in sequence:

[0025] Global average pooling is performed on the feature map to extract overall sequence information;

[0026] The weights for each channel are generated using two fully connected layers and a sigmoid function.

[0027] The generated weights are multiplied channel by channel with the original feature map to obtain a weighted output feature map.

[0028] Furthermore, the classification module includes at least two fully connected layers, and after the fully connected layers are sequentially connected a batch normalization layer, a ReLU activation function layer, and a random zeroing layer.

[0029] Furthermore, residual connections are set in the fully connected layer structure to enable cross-layer transfer of features from previous layers, thereby improving the training stability of the network and enhancing the feature propagation effect.

[0030] Furthermore, the palmitoyltransferase prediction model was obtained in the following manner:

[0031] 12,788 palmitoyltransferase sequences were obtained from the UniProt database as positive samples, and 38,364 sequences were proportionally extracted from 383,640 candidate non-palmitoyltransferase sequences to form a negative sample dataset.

[0032] The positive and negative sample datasets were divided into training, validation and test sets in a ratio of 8:1:1 using stratified sampling.

[0033] The palmitoyltransferase prediction model was trained and its hyperparameters optimized using the training and validation sets.

[0034] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that the processor, when executing the program, implements the above-described protein palmitoylation transferase prediction method based on a multi-branch deep convolutional neural network.

[0035] A computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the above-described method for predicting protein palmitoylation transferases based on a multi-branch deep convolutional neural network.

[0036] The beneficial effects of this invention: This application discloses a deep learning model, iPalmT, for predicting whether a protein possesses palmitoyltransferase activity. iPalmT is built upon a convolutional neural network architecture and incorporates a Squeeze-and-Excitation module, parallel convolutional branches, and residual connections to enhance feature extraction and representation capabilities. Through systematic optimization of the network structure and hyperparameter configuration, iPalmT achieves superior accuracy and generalization performance.

[0037] iPalmT can automatically learn multi-scale hierarchical sequence features from large-scale training data to build an end-to-end prediction framework, thereby achieving accurate classification of palmitoyltransferases. Importantly, the training dataset covers both eukaryotic and prokaryotic species, ensuring broad applicability. Attached Figure Description

[0038] Figure 1 This is the network structure of iPalmT.

[0039] Figure 2 This describes the specific architecture of the iPalmT network structure.

[0040] Figure 3Evaluation of iPalmT's recognition efficiency and comparison with other models. (A) Training and validation loss curves of iPalmT. The figure shows the changes in training loss (green) and validation loss (orange) with the number of epochs. Both gradually decrease during training, with the validation loss reaching its minimum at the 12th epoch, which is determined to be the optimal epoch. The area under the ROC curve (AUC) of iPalmT is also shown. (B) Accuracy, precision, recall, F1 score, and Matthews correlation coefficient (MCC) of iPalmT. (C) Comparison of iPalmT with Pfam (protein family database) / InterProScan (protein sequence functional classification tool) and HMMER (hidden Markov model-based sequence analysis tool).

[0041] Figure 4 These are the key domains identified during the iPalmT model's identification process. Amino acid fragments identified by the model that play a crucial role in palmitoyltransferase prediction were mapped to their corresponding functional domains. The top three domains were the DHHC domain (A), the Ankyrin repeat domain (B), and the SH3 domain (C).

[0042] Figure 5 This section presents the statistical results of the number of candidate palmitoyltransferases identified by the iPalmT model and marked as "unreviewed" in the UniProt database among human proteins. Sequence and structural characterization analyses were performed on two representative candidate molecules. (A) Venn diagram showing the intersection of proteins predicted as palmitoyltransferases by iPalmT, proteins annotated as palmitoyltransferases by UniProt, and unreviewed proteins. A0A0D9SEX5 and A0A1W2PRJ8 are located in the subset predicted as palmitoyltransferases by iPalmT and also marked as unreviewed. (B, C) Results of analysis of the transmembrane region, three-dimensional structure, DHHC and cysteine-rich domains, and their corresponding positions in the sequence of A0A0D9SEX5 and A0A1W2PRJ8 proteins.

[0043] Figure 6Results of palmitoylation sequencing analysis following overexpression of A0A0D9SEX5 or A0A1W2PRJ8 in Urotsa cells. (A) Experimental procedure for palmitoylation proteomics analysis. (B, F) Results of overexpression of A0A0D9SEX5 or A0A1W2PRJ8 in Urotsa cells and detection of its transcriptional level. (C, G) Total number of peptides with upregulated and downregulated palmitoylation levels in Urotsa cells after overexpression of A0A0D9SEX5 or A0A1W2PRJ8. (D, H) Distribution of peptides or proteins with altered palmitoylation levels inside / outside the membrane in Urotsa cells overexpressing A0A0D9SEX5 or A0A1W2PRJ8. (E, I) The top five peptide sequences with the most significant palmitoylation changes after overexpression of A0A0D9SEX5 or A0A1W2PRJ8, their fold changes relative to the control group, modification sites, corresponding protein names and related domains.

[0044] Figure 7 and Figure 8 They are respectively Figure 6 Enlarged schematic diagram of E and I in the figure. Detailed Implementation

[0045] 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 used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. 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.

[0046] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for predicting protein palmitoylation transferases based on a multi-branch deep convolutional neural network. The method includes the following steps:

[0047] S1. Obtain the protein sequence of the protein to be predicted;

[0048] S2. Input the protein sequence into a pre-trained palmitoyltransferase prediction model;

[0049] S3. Based on the output of the palmitoyltransferase prediction model, determine whether the protein is a palmitoyltransferase.

[0050] The palmitoyltransferase prediction model includes:

[0051] The encoding module is used to encode the protein sequence into a 21-dimensional one-hot encoding matrix;

[0052] Furthermore, the encoding module is used to standardize and vectorize the input protein sequences. To address the issue of inconsistent sequence lengths among different proteins, this embodiment uniformly sets the sequence length to 1000 amino acid residues. Preferably, when the input sequence length is less than 1000, it is padded with underscores ("_") at the end; when the sequence length exceeds 1000, it is not included in the dataset. This method ensures that all input sequences maintain dimensionality, thereby meeting the requirements for batch input to deep learning models and improving the stability of model training and inference.

[0053] Furthermore, this embodiment employs a 21-dimensional one-hot encoding rule to represent the protein sequence features. 20 dimensions correspond to 20 standard amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y), and the remaining dimension represents an unknown amino acid X or a void symbol "_". Each amino acid residue is encoded as a vector of length 21, with a value of 1 assigned to the dimension corresponding to the amino acid and 0 assigned to the remaining dimensions, forming a sparse one-hot representation matrix.

[0054] Preferably, for special or non-standard amino acid characters, a mapping strategy based on chemical similarity is used: for example, U is mapped to the same encoding vector as C; B is mapped to the average of D and N vectors; Z is mapped to the average of E and Q vectors; and J is mapped to the average of I and L vectors. For unrecognizable characters, they are uniformly mapped to the encoding vector of X. After processing by this encoding module, a one-hot encoding matrix with a dimension of 1000×21 can be generated as the input feature of the deep learning model for subsequent calculations in convolutional layers, attention layers, or fully connected layers.

[0055] Two parallel convolutional branches are used to extract multi-scale local sequence features from the one-hot encoding matrix. Parallelism means that the two branches process the input simultaneously, have the same structure but independent parameters. The convolutional layer configuration of the two parallel convolutional branches is as follows: the kernel size of the first branch is (3,21), (3,1), (3,1) in sequence; the kernel size of the second branch is (5,21), (5,1), (5,1) in sequence. After each convolutional layer, a batch normalization layer and a ReLU activation function layer are connected in sequence.

[0056] Furthermore, the two-dimensional convolutional kernels in the two parallel convolutional branches are designed in the form of "sequence length × number of channels" to extract local sequence features of different scales from the input 1000×21 one-hot encoding matrix. Specifically, the first dimension of the convolutional kernel corresponds to the amino acid sequence direction, reflecting the size of the model's receptive field on the sequence window; the second dimension corresponds to the encoding channel direction (21 amino acid categories), used to integrate features of all amino acid types in a single convolution. Through this two-dimensional convolutional structure, pattern capture along the sequence direction and cross-channel feature fusion can be achieved simultaneously.

[0057] Preferably, the three convolutional kernels of the first parallel branch have dimensions of (3,21), (3,1), and (3,1) respectively, and the three convolutional kernels of the second parallel branch have dimensions of (5,21), (5,1), and (5,1) respectively. The "×21" structure in the first convolutional kernel size is used to cover all channel dimensions, achieving full-channel feature aggregation of several adjacent residues to identify short-range amino acid combination patterns and functional motifs. Subsequent "×1" convolutional kernels slide only in the sequence direction, gradually expanding the receptive field on the feature map of the aggregated channels, thereby capturing longer contextual dependencies and continuous features between segments. Batch normalization layers and ReLU activation layers are sequentially set after each convolutional layer to improve training stability and enhance nonlinear expression capabilities.

[0058] Furthermore, the design intent of the convolutional kernel size is to achieve parallel extraction of multi-scale sequence patterns: the first branch uses three convolutional windows with sizes of (3,21), (3,1), and (3,1) respectively, focusing on capturing short segments and local motif features; the second branch uses three convolutional windows with sizes of (5,21), (5,1), and (5,1) respectively, used to identify relatively longer local segments and medium-scale sequence dependencies. The two branches are structurally consistent but parameter-independent, allowing for parallel extraction of complementary sequence features on the same input. These features are then concatenated and input to subsequent fully connected layers, achieving the fusion of information at different scales and enhancing discriminative capabilities.

[0059] Preferably, the convolution kernel employs a padding strategy to maintain the output sequence length unchanged, and the stride is set to 1 to ensure that the sequence position correspondence is completely preserved after multiple convolutions. Through the above configuration of two-dimensional convolution kernels of different sizes, this invention can significantly improve the joint modeling ability of short-range and medium-to-long-range patterns in protein sequences while maintaining the efficiency of model parameters, thereby improving the recognition accuracy and robustness of palmitoyltransferase candidate proteins.

[0060] The feature fusion module is used to concatenate the feature vectors output from the two convolutional branches. The feature fusion module performs the following operations in sequence:

[0061] Global pooling and flattening are performed on the feature maps output by the two convolutional branches respectively to obtain two 64-dimensional feature vectors.

[0062] The two feature vectors are concatenated into a 128-dimensional vector;

[0063] The data is projected through a fully connected layer to 512-dimensional and 128-dimensional latent feature spaces for subsequent classification.

[0064] Furthermore, the 128-dimensional latent space is defined based on a comprehensive optimization of feature compression and model performance. This dimension is used to reduce convolutional feature redundancy while preserving key discriminative information, thereby improving the model's generalization ability and computational efficiency.

[0065] Preferably, the setting of the 128-dimensional latent space is matched by the order of magnitude of the number of feature channels after the two convolutional branches are concatenated, which can achieve a balanced embedding mapping, avoiding information loss caused by excessive compression, and maintaining the linear separability of features in a lower dimension.

[0066] Furthermore, in a large number of experimental comparisons, when the latent space dimension is set to 128, the model converges quickly on the validation set, has high recognition accuracy, and a moderate number of parameters. Compared with other settings such as 64 dimensions or 256 dimensions, it has a comprehensive advantage in terms of accuracy and stability.

[0067] Therefore, by projecting the spliced ​​feature vectors into a 128-dimensional latent space, a balance can be achieved between model performance and computational complexity, allowing the sequence features of different types of palmitoyltransferases to form a clear and separable distribution in the latent space, thereby significantly improving the model's discriminative ability and prediction accuracy.

[0068] A classification module is used to predict palmitoyltransferase based on the fused features. The classification module includes at least two fully connected layers, and after the fully connected layers, a batch normalization layer, a ReLU activation function layer and a random zeroing layer are connected in sequence. A residual connection is provided between the output of the feature fusion module and the final output of the classification module.

[0069] The specific structure of the classification module is as follows: Figure 2 As shown, it includes two hidden layers connected in sequence. Following each hidden layer are a fully connected layer, a projected residual shortcut, feature channel compression, and a sigmoid activation function. The hidden layer includes a fully connected layer, followed by a batch normalization layer, a ReLU activation function layer, and a random zeroing layer.

[0070] The classification module receives a 128-dimensional feature vector from the feature fusion module as input, and processes it sequentially through the following hierarchical structure to finally output a prediction score:

[0071] First hidden layer:

[0072] Fully connected layer: projects the input 128-dimensional feature vector into a higher-dimensional space (e.g., 512-dimensional).

[0073] Batch normalization layer: Normalizes the output of the fully connected layer to accelerate the training process and improve model stability.

[0074] Modified linear unit: As an activation function, it introduces a nonlinear transformation to enhance the model's representational ability.

[0075] Randomly zeroed layers: This involves randomly setting the output of some neurons to zero with a given probability (e.g., 0.5) as an effective regularization method to prevent model overfitting.

[0076] Second hidden layer:

[0077] Fully connected layer: The output vector of the first hidden layer is linearly transformed again, usually projected to a lower dimension (e.g., 128-dimensional).

[0078] Batch normalization layer: Performs batch normalization on the output of this layer.

[0079] Modified linear unit: Nonlinearity is introduced again.

[0080] Randomly set to zero: Regularization is applied again by randomly discarding.

[0081] Output and residual connection:

[0082] Output layer: Consists of a fully connected layer that maps the output vector of the second hidden layer to the final output dimension (usually 1 dimension for binary classification tasks).

[0083] After stacking fully connected layers, residual connections are set up: when the input feature dimension and the output dimension are inconsistent, the dimensions are aligned by linear mapping and then added to the output of the main branch, thereby enhancing the convergence and stability of the model.

[0084] Feature fusion and activation: The calculation results of the output layer are added to the projection results of the residual shortcut to achieve the fusion of linear and residual information. The model output is summed to obtain the final prediction score. During the inference phase, the score is compressed to the (0,1) interval using a sigmoid activation function and used as the prediction probability that the protein is palmitoyltransferase.

[0085] Each convolutional branch contains several convolutional layers with two-dimensional convolutional kernels of different sizes. An SE module is connected after the convolutional layer sequence to recalibrate the feature channels of the convolutional output. The SE module includes a Squeeze module and an Excitation module. The Squeeze module performs global average pooling on the feature map, compressing the spatial information of the entire feature map into a channel descriptor. The Excitation module passes the channel descriptor through two fully connected layers and activates it with the Sigmoid function to generate weights (between 0 and 1) for each feature channel. Recalibration: The learned channel weights are multiplied with the original feature map channel by channel, thereby amplifying important features and suppressing minor features.

[0086] The SE module performs the following operations in sequence:

[0087] Compression involves performing global average pooling on the feature map to extract overall sequence information; global average pooling is also performed on the input original feature map. This operation compresses each two-dimensional feature channel into a single scalar statistic, thereby generating a channel descriptor that aggregates the global spatial information of the feature map.

[0088] The incentive is to generate the weights of each channel through two fully connected layers and a sigmoid function; the channel descriptors obtained in the previous step are input into a gating mechanism consisting of two fully connected layers to capture the non-linear dependencies between channels and generate the weights of each channel.

[0089] The first fully connected layer is a dimension reduction layer. In this embodiment, the number of channels is reduced by 1 / 16. This layer is followed by a modified linear unit activation function (ReLU).

[0090] The second fully connected layer is a dimension reduction layer that restores the number of channels from the reduced number to the original number of channels to ensure that it matches the number of channels in the original feature map. This layer is followed by a sigmoid activation function to obtain the weight coefficient of each channel.

[0091] Recalibration involves multiplying the generated weights channel-by-channel with the original feature map to obtain a weighted output feature map. This is achieved by multiplying the channel weights learned in the previous step with the original input feature map channel-by-channel. The final output feature map is the recalibrated feature map, whose dimensions are consistent with the input feature map.

[0092] By acquiring global context information through the "squeeze" operation, learning the importance weights of each channel through the "stimulation" operation, and finally enhancing informative features and suppressing redundant features adaptively according to the weights through the "recalibration" operation, the feature representation capability of the model is significantly improved.

[0093] Example

[0094] Sample preparation

[0095] Positive and negative samples were collected from UniProt (https: / / www.uniprot.org / , accessed June 11, 2025). For positive samples, the search criteria were set as follows: DHHCORpalmitoyltransferaseORZDHHCORZDHORS-acyltransferaseORPFAORPATORSWF (species and time not limited). Based on these criteria, 224,455 protein entries were initially obtained. Subsequently, proteins lacking distinctive characteristics were removed, and their identification as palmitoyltransferases was confirmed through literature review; proteins longer than 1,000 amino acids were also removed. After this screening, 12,788 positive samples and 182,630 negative samples remained. To further expand the negative samples, proteins annotated as not involving the palmitoylation pathway were searched in UniProt, and cross-validation with literature was performed to confirm their non-palmitoyltransferase status, resulting in 201,010 entries. This negative sample pool was merged with the previously obtained 182,630 negative samples to form a pool of 383,640 negative samples. One-tenth (38,364 samples) of this pool was randomly selected as the final negative sample set. The dataset containing both positive and negative samples was then divided into training, validation, and test sets in an 8:1:1 ratio. Stratified sampling was used during the partitioning process to ensure a balanced distribution of positive and negative samples in each subset.

[0096] In the human-specific protein analysis, all proteins were downloaded by querying organism_id:9606, yielding 205,205 results. Further screening for unreviewed proteins was performed using organism_id:9606ANDreviewed:false, resulting in 184,785 results. The iPalmT model was then used to screen these unreviewed human proteins for candidate palmitoyltransferases, thus providing functional annotations for proteins currently lacking human annotation.

[0097] To construct a comprehensive palmitoyltransferase database, a massive collection of protein sequences was downloaded from NCBI on September 22, 2025. The merged dataset covers 33,285 species-level groupings, totaling 147,847,003 protein sequences.

[0098] Construction process of palmitoyltransferase prediction model

[0099] iPalmT is developed based on the PyTorch deep learning framework and incorporates several structural improvements, including the introduction of a SE module, parallel convolutional branches, batch normalization, dropout, and residual connections. After building the basic modules, the convolutional neural network architecture was systematically optimized, combining 8 layer configurations with 5 hidden dimensions to obtain 40 structural variants. For each variant, the number of parameters and prediction accuracy were calculated, and the average accuracy was taken three times to select the optimal structure for subsequent analysis.

[0100] After determining the convolutional neural network architecture, we optimized the hyperparameters. Specifically, we combined 5 learning rates with 5 large and small batch sizes to create 25 parameter configurations, each of which was evaluated three times. The optimal learning rate and batch size were then determined based on the average accuracy. Furthermore, to further reduce the loss, we fine-tuned regularization-related parameters, including weight decay, early stopping patience, and dropout probability. The best performance corresponded to a weight decay of 10. -4 The early stopping patience value is 5, and the dropout value is 0.5. This yields the final iPalmT model.

[0101] iPalmT aims to identify palmitoyltransferases from primary amino acid sequences. Each protein sequence is normalized to a 1000×21 one-hot encoding matrix (sequences ≤1000 are padded to 1000 with underscores at the end; sequences >1000 are excluded from the dataset) and used as a single-channel input map. The model contains two parallel convolutional branches, each using two-dimensional convolutional kernels (e.g., 3×21 and 5×21), followed by smaller kernels (3×1 or 5×1) to simultaneously capture local motifs and long-range contextual features. Each convolutional layer is followed by batch normalization and ReLU activation to enhance training stability and non-linear representation capabilities. To strengthen channel-level feature representation, each branch integrates an SE module: global average pooling extracts overall sequence information, and channel weights are generated through two fully connected layers and a sigmoid function, adaptively emphasizing informative features and suppressing redundant signals. The outputs of the two branches are then compressed using adaptive max pooling and concatenated into a 128-dimensional feature vector. The vector is then passed through two fully connected layers (512 and 128 units), each followed by batch normalization, ReLU, and dropout (rate = 0.5) to reduce overfitting. A projection residual connection is set between the concatenated 128-dimensional features and the final output: when the dimensions are inconsistent, they are aligned by linear mapping before addition to improve training stability. The final prediction layer provides the binary classification result.

[0102] Model Attribution and Interpretability Framework Design

[0103] To clarify the criteria used by deep learning models in predicting palmitoyltransferases, this invention designs a comprehensive analysis workflow combining interpretable algorithms with structural annotation. First, the integral gradient (IG) method is used to evaluate the contribution of individual amino acids to the model's prediction. Specifically, the protein sequence is read from a FASTA file and a trained iPalmT model is loaded. Multi-step interpolation is performed between the zero-baseline tensor and the actual input, and the gradient is iteratively calculated and accumulated to obtain the attribution score for each residue. The program generates an attribution file for each sequence and outputs a summary file containing amino acid positions, residue types, and their importance scores. Subsequently, to locate potentially key functional regions in the sequence, quantile thresholds are set for the attribution scores, high-importance residues are extracted, merged into contiguous fragments, and the average importance of each fragment is calculated. This step highlights the amino acid regions most likely to play a crucial role in functional prediction. To achieve structure-function integration, InterProScan is further used to annotate the protein sequence with domains and functional regions. By mapping the top 15 most important residues identified by IG to annotated structural domains (such as DHHC domains or auxiliary domains), key residues can be associated with their corresponding structural regions, and the final results can be output.

[0104] Computing environment

[0105] All computational tasks in this invention were performed on a single CentOS Stream 9 (Linux kernel 5.14) node: an Intel® Core™ i9-14900K CPU (24 cores, 32 threads, up to 7.7GHz) and 125GB of memory. The Python environment was managed by Anaconda (conda24.5.0) using a separate virtual environment (iPalmT), Python 3.7.16, with major dependencies including NumPy 1.21.6, Pandas 1.3.5, Scikit-learn 1.0.2, SciPy 1.7.3, Matplotlib 3.5.3, Biopython 1.81, and PyTorch 1.13.1. CUDA libraries (cuDNN 8.5, CUDA 11.7) were installed to support deep learning model execution. This configuration ensured reproducibility and computational efficiency.

[0106] Protein structure prediction and analysis

[0107] The target protein sequence was downloaded from the UniProt database (https: / / www.uniprot.org / ), and its structure was predicted using the GaliLeo scientific computing platform (https: / / sidereus-ai.com / ). PyMOL was used for structural visualization. Domain annotations were obtained using InterPro / InterProScan. Amino acid-level scores were generated using the corresponding modules in iPalmT and visualized using Python packages (including pandas, logomaker, and matplotlib.pyplot). Transmembrane region prediction was performed using the TMHMM online analysis tool.

[0108] Overexpression of A0A0D9SEX5 and A0A1W2PRJ8 in Urotsa cells

[0109] An overexpression model was established in Urotsa cells using a lentiviral system. In short, the coding sequences of A0A0D9SEX5 and A0A1W2PRJ8 were cloned into the pLenti-EF1α-IRES-Puro vector with a FLAG or HA tag; lentivirus was prepared by co-transfection with packaging plasmids in HEK293T cells. Viral supernatants were collected and filtered, and Urotsa cells were infected in the presence of polycations; centrifugation could be used to promote efficiency. Stable overexpressing cells were screened with puromycin; an empty vector was used as a negative control. Overexpression efficiency was verified at the transcriptional level by qPCR.

[0110] Palmitylation sequencing experiments

[0111] Protein palmitoylation was analyzed using an ABE (Acyl-biotin exchange) coupled with data-independent acquisition (DIA) mass spectrometry strategy. The simplified procedure was as follows: Cells or protein samples were lysed with lysis buffer (4% SDS, 150 mM Tris-HCl, pH 8.0) (cell lysates were sonicated), centrifuged, and quantified using the BCA method. Equal volumes of protein (15 µg) were separated on 4%–20% SDS-PAGE gels and developed with Coomassie Brilliant Blue. Free thiol groups were blocked using a NEM, excess reagent was removed by TCA-acetone precipitation, and the samples were reconstituted in PBS containing 6 M urea and 2% SDS. Samples were separated into HA+ (hydroxylamine and biotin-HPDP treatment) and HA+. The control group was incubated at 25°C, concentrated, and then digested with trypsin. After desalting, biotin-labeled peptides were enriched with streptavidin agarose, eluted with TCEP, alkylated with iodoacetamide, desalted again, dried, and reconstituted in 0.1% formic acid containing iRT peptides. DIA was performed on Orbitrap™ Astral™ mass spectrometry coupled with Vanquish NeoLC: MS1 (380–980 m / z, resolution 240,000), MS2 using a 300-slot isolation window (2 m / z). Data processing was performed using Spectronaut: trypsin digestion, allowing ≤2 missed digests; fixed modification was carboxymethylation of cysteine; variable modification was methionine oxidation and N-terminal acetylation of proteins; FDR for peptide / protein identification was ≤1%. Sequencing results were visualized and analyzed using R. The workflow is described in [link to workflow]. Figure 6 A.

[0112] Statistical analysis

[0113] For two groups of continuous variables, a t-test is used if the parametric assumptions are met; otherwise, a Mann–Whitney U test is used. A two-sided significance level of 0.05 is used. Data are typically expressed as mean ± standard deviation (SD). Graphs are generated using R or GraphPadPrism.

[0114] result

[0115] Construction and Optimization of iPalmT Deep Neural Network

[0116] like Figure 1 and Figure 2 As shown, this invention discloses a deep convolutional neural network, iPalmT, for predicting whether a given protein is a palmityltransferase. First, the protein sequence is encoded using 21-dimensional one-hot encoding, and then input into two parallel convolutional branches; each branch contains three convolutional layers with different kernel sizes to achieve multi-scale local sequence feature extraction. Figure 1 and Figure 2 ).

[0117] To enhance feature representation and reduce redundancy, an SE module is placed after the convolutional stack of each branch for channel recalibration. After pooling and flattening ( Figure 1 and Figure 2 After that, the feature vectors from the two branches are concatenated and projected into a 128-dimensional latent space, and then the final classification is completed through a fully connected layer. Figure 1 and Figure 2During the model optimization phase, 40 different CNN architectures and hidden dimensions were evaluated. The optimal configuration was [[(3,21),(3,1),(3,1)],[(5,21),(5,1),(5,1)]], with a hidden dimension of 512. To further improve training efficiency and convergence, we tested 25 combinations of learning rate and batch size, and determined that a learning rate of 0.001 and a batch size of 16 performed best. Furthermore, we optimized hyperparameters including weight decay, early stopping patience, and dropout rate, with the best result achieved by a weight decay of 10. -4 The early stopping patience value is 5, and the dropout value is 0.5. Meanwhile, the model reaches its optimum in the 12th epoch. Figure 3 A).

[0118] The final iPalmT model achieved excellent performance on the test set: AUC 0.999, accuracy 0.989, precision 0.980, recall 0.977, F1 score 0.979, and MCC 0.971. Figure 3 B). These results demonstrate that iPalmT has excellent accuracy and robustness in predicting whether a protein is a palmitoyltransferase.

[0119] Based on iPalmT, we identified 10,365,644 predicted palmitoylation transferases from 147,847,003 protein sequences covering 33,285 species-level groups, which can serve as a candidate library for subsequent functional validation and application screening.

[0120] Comparison of iPalmT with existing forecasting tools

[0121] To date, no palmitoyltransferase prediction tools based on deep learning or machine learning have been reported. Existing software related to this task includes Pfam, InterProScan, and HMMER. Specifically, Pfam is a database of protein families and domains built based on Hidden Markov Models (HMMs); ​​InterProScan integrates multiple domain identification methods (such as PRINTS and PROSITE) for functional annotation; and HMMER uses an HMM-based sequence alignment algorithm to identify similarities with known protein families. These tools can only provide a rough assessment of potential palmitoyltransferase activity based on sequence similarity and conserved domain matching, and are not specifically designed for palmitoyltransferase prediction.

[0122] This study discloses iPalmT, the first deep learning model specifically designed for palmitoyltransferase prediction. Unlike traditional methods, iPalmT can automatically learn multi-scale hierarchical sequence features from large-scale training data to construct an end-to-end prediction framework, thereby achieving accurate classification of palmitoyltransferases. Importantly, the training dataset covers both eukaryotic and prokaryotic species, ensuring broad applicability. Figure 3 As shown in C, compared with Pfam, InterProScan, and HMMER, iPalmT has the following unique advantages: (1) end-to-end prediction; (2) identification of novel DHHC-like proteins; (3) consideration of auxiliary domains, including functional contributions beyond the DHHC domain; (4) prediction independent of database annotations; (5) adaptability to new species; (6) network independence; and (7) fast inference speed (0.039s per sequence), suitable for large-scale screening. In summary, iPalmT represents the first deep learning framework for palmitoyltransferase prediction, providing a powerful computational tool with high accuracy, broad applicability, and high efficiency.

[0123] iPalmT Predictive Attribution Explanation

[0124] To quantitatively interpret the prediction results of deep learning models and combine sequence-level attribution with structural / functional annotation to achieve mechanistic verification of the results, the Inductively Interpolated (IG) method was used to identify the amino acid region with the greatest impact on prediction. This method performs multi-step interpolation between the baseline input and the true input, iteratively calculates and accumulates gradients, and sums them to obtain the contribution of each residue to the model output. Subsequently, the top 15 most important residues are annotated and mapped to their corresponding structural domains, thus enabling a traceable explanation of the model's decision-making process.

[0125] Applying this analysis to all positive samples revealed that iPalmT primarily focuses on the DHHC domain when determining whether a protein is a palmityltransferase. Specifically, 10,758 sequences (approximately 84% of all positive samples) contain highly important residues within this domain. The DHHC domain is a classic conserved marker of palmityltransferases; the structure of a typical protein and its DHHC domain is shown below. Figure 4 As shown in A; despite the lack of prior knowledge during training, iPalmT can still spontaneously capture and identify the region, demonstrating its ability to learn key functional features directly from large-scale sequence data.

[0126] In addition to the DHHC domain, iPalmT also identified other common auxiliary functional domains, including Ankyrin repeat and SH3 domains. Figure 4(B–C). Notably, these domains have been independently confirmed in the literature to be associated with palmityltransferase activity. Overall, these findings confirm that iPalmT not only has high predictive accuracy but also demonstrates strong biological rationality and interpretability, providing confidence for its application in functional protein annotation.

[0127] Discovery of unannotated palmitoyltransferase

[0128] In the UniProt database, some proteins are marked as unreviewed, indicating a lack of manual annotation and functional description. Annotating these proteins and identifying potential palmitoyltransferase candidates is crucial for discovering novel protein functions. Using iPalmT, the inventors screened the entire human proteome, identifying 2,239 proteins predicted to have palmitoyltransferase activity. Of these, 23 have been annotated as palmitoyltransferases, 2,012 have other functional annotations but were additionally predicted to have palmitoyltransferase activity, and the remaining 204 are unreviewed entries but were identified by iPalmT as putative palmitoyltransferases. Figure 5 A). From these 204 candidates, A0A0D9SEX5 and A0A1W2PRJ8 were randomly selected for further analysis. Figure 5 A).

[0129] Previous studies have shown that palmitoyltransferases typically contain multiple transmembrane segments and a conserved DHHC domain. The inventors first analyzed the transmembrane structures of A0A0D9SEX5 and A0A1W2PRJ8. The results showed that both have more than three transmembrane regions, and the DHHC domain is located in the membrane intercalation region (…). Figure 5 B–C). Furthermore, both sequences contain cysteine-rich regions, which may contribute to protein conformation stability, facilitate acyl transfer reactions, and enhance substrate interactions. Figure 5 (B–C). These observations demonstrate that A0A0D9SEX5 and A0A1W2PRJ8 possess the classic structural features of palmitoyltransferases, thus supporting the reliability and biological rationale for iPalmT's predictions of unreviewed proteins.

[0130] To further validate the potential palmityltransferase activity of the aforementioned proteins at the experimental level, the inventors overexpressed A0A0D9SEX5 and A0A1W2PRJ8 in normal human bladder epithelial cells (Urotsa) and performed palmitoylation analysis as described in the "Methodology" section. This method can systematically detect almost all palmitoylated peptides and their corresponding sites ( Figure 6 A). It is worth noting that overexpressing the two target genes mentioned above ( Figure 6 B and F can lead to a significant increase in the number of peptides with upregulated palmitoylation levels. Figure 6C, G). These changes mainly occur in the intrinsic membrane regions (C, G). Figure 6 (D, H), which is consistent with the known role of palmitoylation in maintaining membrane protein homeostasis.

[0131] Of particular note, in A0A0D9SEX5 overexpressing cells, the most significantly altered peptide, palmitoylation, was upregulated by 121-fold, and the second most significantly upregulated by 60-fold. Figure 6 E, Figure 7 This strongly supports its function of promoting palmitoylation.

[0132] Finally, the inventors summarized the top five peptides with the most significant palmitoylation changes after overexpression of A0A0D9SEX5 and A0A1W2PRJ8, and listed their corresponding proteins, modification sites, and related domains. Figure 6 E, I, Figure 7 and Figure 8 These results provide crucial clues to elucidating the downstream regulatory mechanisms of these two candidate palmitoyltransferases.

[0133] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the program to implement the above-described protein palmitoylation transferase prediction method based on a multi-branch deep convolutional neural network.

[0134] The electronic device can also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be achieved through input / output (I / O) interfaces. Furthermore, the electronic device can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0135] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0136] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting protein palmitoylated transferases based on a multi-branch deep convolutional neural network.

[0137] The readable storage medium may include a readable medium in the form of volatile memory, such as random access memory (RAM) and / or cache memory, and may further include read-only memory (ROM). Program code is stored thereon, which is executed by a processor to cause the processor to perform the steps of the embodiments described in this specification.

[0138] The memory may also include programs / utilities having a set (at least one) of program modules, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0139] A bus can represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus that uses any of the various bus structures.

[0140] The embodiments should not be regarded as limitations on the present invention, but any improvements made based on the spirit of the present invention should be within the protection scope of the present invention.

Claims

1. A method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks, characterized in that, It includes the following steps: S1. Obtain the protein sequence of the protein to be predicted; S2. Input the protein sequence into a pre-trained palmitoyltransferase prediction model; S3. Based on the output of the palmitoyltransferase prediction model, determine whether the protein is a palmitoyltransferase. The palmitoyltransferase prediction model includes: The encoding module is used to encode the protein sequence into a one-hot encoding matrix with a length of 1000 and a width of 21, where 20 dimensions correspond to standard amino acids and 1 dimension is used for filling gaps or unknown symbols; Two parallel convolutional branches are used to extract multi-scale local sequence features from the one-hot encoding matrix; The feature fusion module is used to concatenate the feature vectors output by the two convolutional branches; The classification module is used for palmitoyltransferase prediction based on the fused features. Each convolutional branch includes several two-dimensional convolutional layers of different sizes, and after the convolutional layers are stacked, an SE module is connected to perform adaptive recalibration of the feature channels of the convolutional output.

2. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 1, characterized in that: The feature fusion module performs the following operations in sequence: Global pooling and flattening are performed on the feature maps output by the two convolutional branches respectively to obtain two 64-dimensional feature vectors. The two feature vectors are concatenated into a 128-dimensional vector; The data is projected through a fully connected layer to 512-dimensional and 128-dimensional latent feature spaces for subsequent classification.

3. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 1, characterized in that: Each convolutional branch contains three two-dimensional convolutional layers. The kernel sizes of the first branch are (3,21), (3,1), and (3,1) respectively; the kernel sizes of the second branch are (5,21), (5,1), and (5,1) respectively. After each convolutional layer, a batch normalization layer and a ReLU activation layer are connected in sequence, and an SE module is set at the end of the branch to perform weighting and calibration of the feature channels.

4. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 1, characterized in that: The SE module performs the following operations in sequence: Global average pooling is performed on the feature map to extract overall sequence information; The weights for each channel are generated using two fully connected layers and a Sigmoid function. The generated weights are multiplied channel by channel with the original feature map to obtain a weighted output feature map.

5. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 1, characterized in that: The classification module includes at least two fully connected layers, and after the fully connected layers are sequentially connected a batch normalization layer, a ReLU activation function layer, and a random zeroing layer.

6. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 5, characterized in that: The fully connected layer structure includes residual connections to enable cross-layer transfer of features from previous layers, thereby improving the training stability of the network and enhancing the feature propagation effect.

7. The method for predicting protein palmitoylation transferases based on multi-branch deep convolutional neural networks according to claim 1, characterized in that: The palmitoyltransferase prediction model was obtained through the following method: 12,788 palmitoyltransferase sequences were obtained from the UniProt database as positive samples, and 38,364 sequences were proportionally extracted from 383,640 candidate non-palmitoyltransferase sequences to form a negative sample dataset. The positive and negative sample datasets were divided into training, validation and test sets in a ratio of 8:1:1 using stratified sampling. The palmitoyltransferase prediction model was trained and its hyperparameters optimized using the training and validation sets.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that, When the processor executes the program, it implements the protein palmitoylation transferase prediction method based on a multi-branch deep convolutional neural network as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the protein palmitoylation transferase prediction method based on a multi-branch deep convolutional neural network as described in any one of claims 1 to 7.

Images