Abstract
Background
Protein post-translational modification (PTM) is a critical biological process that occurs after protein synthesis and has key roles in several biological processes. Among these, lysine modifications include multiple types and have received considerable attention. Most existing computational models predict whether a specific lysine site in a protein sequence corresponds to a lysine modification type by extracting features from a short peptide segment centered on that site. Therefore, information from the full protein sequence is not used.
Results
In this study, we gave a different direction for investigating lysine modifications. A computational model, PLysPTM-HGNN, was designed to identify lysine modification types at the protein level. Full protein sequence information was used to derive three feature types: gene ontology features, large language model features, and position-specific scoring matrix features. These features were refined separately through a linear transformation, a hybrid graph neural network, and a convolutional neural network combiner, after which they were concatenated and passed into a fully connected layer for prediction. Cross-validation results showed that the AUROC and AUPR were approximately 0.84 and 0.68, respectively, indicating strong predictive performance.
Conclusions
PLysPTM-HGNN outperformed several existing protein subcellular localization models and models based on traditional multi-label classification algorithms. This model provides a useful tool for studies of lysine modifications.
Keywords: Post-translation modification, Lysine modification, Hybrid graph neural network, Large language model, Gene ontology
Introduction
Protein post-translational modification (PTM) refers to the chemical modification of amino acid residues in a protein after its synthesis, typically through enzymatic processes [1]. Among the various PTM types, protein lysine modifications (PLMs) represent a critical category that has key roles in multiple biological processes, including the regulation of protein function, cellular signal transduction, and the onset and progression of diseases. Accurate identification of lysine modification sites is therefore essential for elucidating protein function and clarifying their mechanisms of action in pathological conditions.
With the increasing use of detection technologies in laboratories, a growing number of lysine modification types have been identified, including acetylation [2–6], biotinylation [7], butyrylation [8, 9], crotonylation [10–13], methylation [14–16], propionylation [17, 18], succinylation [19–21], and ubiquitination [22, 23]. At present, lysine modifications can be experimentally confirmed through mass spectrometry (MS) techniques [24]. However, these methods require specialized equipment and can only be performed in well-equipped laboratories, making the process time-consuming and resource-intensive. Some researchers also use modification-specific antibodies, such as those targeting phosphorylation, to detect modification types. These approaches, however, are associated with high financial costs, particularly when reliable and reproducible results are needed. Faster and more reliable detection methods for lysine modifications are still necessary.
As lysine modification data continue to accumulate and computational approaches advance, designing efficient computational models to identify lysine modifications in protein sequences has become an important supplement to experimental methods. Over the past two decades, numerous computational models have been developed to identify various lysine modification types. Some models detect only a single lysine modification type [2–8, 10–13, 15–23], whereas others handle multiple types and can identify them simultaneously [25–31]. In general, these models treat amino acid residues as samples and extract features from short peptide segments surrounding the residues. As these segments are often very short, the information they contain is limited. The exclusion of full protein sequence information further reduces predictive performance. For reliable identification of lysine modifications in a protein sequence, it is also necessary to determine which lysine modification types could be assigned to residues across the entire sequence. Using full-sequence information may therefore improve prediction accuracy.
In this study, we propose a computational model, PLysPTM-HGNN, to predict the lysine modification types that a protein may contain. A well-defined dataset was constructed using lysine modification data from the compendium of protein lysine modifications (CPLM) [32]. Three feature types were extracted from full protein sequences: gene ontology (GO) features, large language model (LLM) features, and position-specific scoring matrix (PSSM) features, each reflecting different protein properties. Inspired by the successful use of deep learning in developing various prediction models [33–36], these feature types were processed separately through a deep learning framework consisting of a linear transformation, a hybrid graph neural network (GNN), and a convolutional neural network (CNN) combiner to generate high-level protein features. All features were concatenated and passed through a fully connected layer for prediction. Cross-validation results showed strong performance, with AUROC and AUPR values of approximately 0.84 and 0.68, respectively. These results exceeded those of models using traditional multi-label classification algorithms and those developed for predicting protein subcellular localization. Additional analyses included ablation tests and performance evaluation across different protein groups.
Materials and methods
Benchmark dataset
The PLM data used in this study were obtained from CPLM (version 4.0, http://cplm.biocuckoo.cn/) [32], a well-annotated resource for PLMs. We extracted lysine modification data for Homo sapiens, yielding 24,327 human proteins covering 24 lysine modification types. Several modification types annotated only a small number of human proteins. As limited data do not support reliable pattern learning, the 14 lysine modification types represented by fewer than 1000 human proteins were removed. The remaining 10 lysine modification types included 2-hydroxyisobutyrylation, acetylation, crotonylation, glycation, malonylation, methylation, succinylation, sumoylation, ubiquitination, and β-hydroxybutyrylation (Table 1). These types covered 24,037 human proteins. To rigorously evaluate the model, the 24,037 proteins were filtered using the Cluster Database at High Identity with Tolerance (CD-HIT) [37] with a sequence identity threshold of 40%. In addition, proteins lacking GO annotations were removed as the proposed model relies on this information. A total of 10,272 human proteins remained and constituted dataset S. These proteins were annotated with one or more of the 10 lysine modification types. The number of proteins associated with each modification type is shown in Table 1. The total number of annotations (31,582) greatly exceeded the number of distinct proteins (10,272), indicating that many proteins carry multiple lysine modification types. We counted the number of proteins annotated with exact 1, 2, 3, …, 10 lysine modification types (Fig. 1). A total of 2758 proteins were annotated with exact one type, whereas the remaining proteins carried two or more types. Accordingly, predicting lysine modification types at the protein level represents a multi-label classification problem.
Table 1.
Number of proteins for each lysine modification type
| Lysine modification type | Number of human proteins |
|---|---|
| 2-Hydroxyisobutyrylation | 1867 |
| Acetylation | 6740 |
| Crotonylation | 2065 |
| Glycation | 2165 |
| Malonylation | 1427 |
| Methylation | 2114 |
| Succinylation | 1297 |
| Sumoylation | 4378 |
| Ubiquitination | 8557 |
| β-Hydroxybutyrylation | 972 |
| Total | 31,582 |
Fig. 1.
Bar chart showing the number of human proteins annotated with exact 1, 2, …, 10 lysine modification types. Only 2758 proteins carry exact one modification type, whereas the remainder have two or more modification types
Feature extraction
Effective prediction models require informative features that can represent the samples comprehensively. Three feature types were extracted to characterize each human protein.
GO features
GO is widely used in protein analysis as it represents functional annotations comprehensively [38]. Here, GO annotations were used as the first feature type. GO terms for the 10,272 human proteins were retrieved from https://geneontology.org. This yielded 5442 GO terms. Based on these terms, each protein p was encoded as a 5442-dimensional vector:
 |
1 |
where 
is set to 1 if p is annotated with the i-th GO term, and 0 otherwise. These features are referred to as GO features.
Language model features
In recent years, LLMs have achieved major progress in natural language understanding, demonstrating the ability to learn deep semantic representations from large volumes of unlabeled text [39]. Supported by deep learning architectures, these models capture complex dependencies and structural patterns within language. These developments have motivated researchers to apply similar methods to protein sequence analysis.
Protein language models (pLMs) treat amino acid sequences as linguistic data and use language modeling techniques to extract latent information. Trained on large-scale protein databases such as UniProt [40], pLMs identify hidden structural properties and generate informative embeddings. These embeddings have shown broad utility in tasks such as protein structure prediction, functional annotation, and interaction analysis.
Among the available models, ProtT5-XL-UniRef50 [41] (ProtT5) is widely used. In this study, we used the pretrained ProtT5 model to obtain biologically meaningful representations of protein sequences. Sequence embeddings were extracted from the final hidden layer of the ProtT5 encoder, resulting in an L × 1024 embedding matrix for a protein sequence of length L. To derive a unified representation, an average operation was applied to the embedding matrix, yielding a 1024-dimensional vector. Each protein p was represented as:
 |
2 |
where 
is the mean value of the i-th column of the ProtT5 embedding matrix. These features are referred to as LLM features.
PSSM features
Evolutionary information has a central role in biological sequence analysis and is widely used in protein classification tasks. The PSSM [42] is one of the most commonly used forms of evolutionary information. In this study, PSSMs were generated using PSI-BLAST [43] with three iterations against the Swiss-Prot database [44], employing an E-value threshold of 0.001. For a protein sequence of length L, the resulting PSSM is an L × 20 matrix in which each value reflects the frequency of observing a given amino acid at a specific position during sequence alignment. Because the size of a PSSM depends on sequence length, the matrices vary among proteins, making uniform processing difficult. To address this issue, the Bigram method [45] was used to convert each PSSM into a 20 × 20 matrix. This matrix was then transformed into a 400-dimensional vector containing all matrix elements. Thus, each protein p was encoded as:
 |
3 |
where 
represents an element of the Bigram-derived matrix. These features are referred to as PSSM features.
In this study, three feature types were constructed to represent each protein. Their characteristics are summarized in Table 2.
Table 2.
Information on three feature types
| Feature type | Number of features |
|---|---|
| GO feature | 5442 |
| LLM feature | 1024 |
| PSSM feature | 400 |
Protein network
In recent years, network-based approaches have been widely applied as they allow system-level organization of biological entities and enable the use of powerful network representation learning algorithms to generate high-level features. In this study, a protein network was constructed for the 10,272 human proteins, in which each protein served as a node. Edges in the network were defined using protein–protein interactions (PPIs) reported in STRING (version 11.0, https://cn.string-db.org/) [46], a public database containing approximately 24.6 million proteins from 5090 organisms and over two billion interactions. These interactions represent direct (physical) and indirect (functional) associations between proteins. PPIs corresponding to the 10,272 human proteins were extracted and used as edges to construct the protein network, denoted as G.
PLysPTM-HGNN
In this study, we proposed a computational model for predicting lysine modification types of proteins. The construction process consisted of several stages, illustrated in Fig. 2. The model was trained on dataset S described in Sect. “Benchmark dataset”. For each protein, three raw feature types (GO, LLM, and PSSM features) were extracted. Each feature type was then processed through a linear transformation, a hybrid GNN, and a CNN combiner to generate high-level protein features. The hybrid GNN used the protein network built from PPIs in STRING. The high-level features were concatenated to obtain the final protein representations, which were passed through a fully connected layer for prediction. This section provides a detailed description of the construction steps.
Fig. 2.
Construction process of PLysPTM-HGNN. A protein dataset is prepared using lysine modification data from the CPLM. Three feature types (GO, LLM, and PSSM) are extracted for each protein. Each feature type is separately processed by a linear transformation, a hybrid GNN, and a CNN combiner. This process uses the protein network constructed from STRING PPIs. The processed features are concatenated and passed into a fully connected layer for prediction
Linear transformation layer
Three feature types were employed to characterize the proteins. However, their dimensions were high (≥ 400). A linear transformation layer was applied to obtain a low-dimensional representation of each feature type. For GO, LLM, and PSSM features of a protein p, the low-dimensional representations were computed as:
 |
4 |
where 
, 
, and 
are trainable weight matrices, and 
, 
, and 
are bias terms. For each feature type, the resulting representations for all proteins were collected into initial representation matrices 
, 
, and 
, where N is the number of proteins and m is the feature dimension.
Hybrid GNN layer
The linear transformation layer generated low-dimensional representations that captured essential protein properties. As described in Sect. “Protein network”, a protein network was constructed to represent relationships between proteins. Combining essential features with network information can yield more informative protein representations. A hybrid GNN was applied to each feature type to generate high-level protein features. This framework consisted of two graph convolutional network (GCN) layers and one graph attention network (GAT) layer. The initial representation matrix was processed sequentially by the first GCN layer, the GAT layer, and the second GCN layer. For a network N with n nodes, GCN and GAT update node representations in the following manner.
Let 
be the initial node representation matrix. GCN updates node representations by:
 |
5 |
where 
(
is the adjacency matrix of network N and 
is the identity matrix),
is the degree matrix of 
, 
is the trainable weight matrix, and 
is the ReLU activation function.
In GCN, the importance of all neighbors of a given node is treated equally. In contrast, GAT uses an attention mechanism to assign different weights to neighboring nodes when aggregating their representations. Suppose there are k attention mechanisms. From the initial representation matrix H, the representation of the i-th node can be refined, denoted by 
, using the following manner:
 |
6 |
 |
7 |
where 
is the l-th normalized attention coefficient of the j-th node to the i-th node, 
stands for the trainable weight matrix for the l-th attention mechanism, 
is the shared attention vector, 
is the neighbor set of the i-th node. For later formulation, the procedure for generating new representation matrix 
by applying GAT to N and H is denoted by:
 |
8 |
For the initial representation matrix H and network N, the hybrid GNN updated the node representation matrix as follows:
 |
9 |
where 
.
There were three GCN or GAT layers in the hybrid GNN. Multiple graph layers may cause vanishing gradients. To address this, a residual module was added to each GCN or GAT layer, allowing deeper layers to access earlier outputs:
 |
10 |
where 
represented GCN or GAT procedure. Accordingly, Eq. 9 was updated as
 |
11 |
Three initial representation matrices 
were passed separately through the hybrid GNN using the protein network G:
 |
12 |
 |
13 |
 |
14 |
where 
denote the outputs of the three GCN or GAT layers. To reduce computational complexity, only the top K weighted edges were selected for each node instead of all neighbors. The parameter K was determined through cross-validation.
CNN combiner layer
Each feature type was processed through the hybrid GNN layer, yielding three new representations. The outputs of different GCN and GAT layers capture protein information at different levels, which may support the prediction of lysine modification types. Here, we selected the representations produced by the two GCN layers and fused them using a CNN combiner. Let 
and 
denote the outputs of the first and third GCN layers, respectively, where 
represents the feature type. The CNN combiner output, denoted as 
, can be obtained by:
 |
15 |
 |
16 |
where wk
is the convolution filter belonging to the set Wout = 
, denotes the convolution operator, 
represented the embeddings of the k-th output channel, 
denoted the stacking operator.
The CNN combiner was applied to each feature type after hybrid GNN processing, yielding three protein representation matrices: 
, 
, and 
. These matrices were concatenated to generate the final protein representation matrix:
 |
17 |
where 
denoted the matrix concatenation operation.
Prediction layer
Based on the final protein representation matrix 
, predictions were made using a fully connected layer consisting of an input layer, a hidden layer, and an output layer. The input layer contained 768 (256 
3) neurons. The hidden layer contained 256 neurons with a LeakyReLU activation function. The output layer contained ten neurons corresponding to the ten lysine modification types.
To measure prediction quality, we used the binary cross-entropy with logits loss, defined as:
 |
18 |
where N and C denote the numbers of proteins and lysine modification types, respectively, 
was the output of the model for the i-th sample on the j-th lysine modification type, 
was the observed label for the i-th sample on the j-th lysine modification type, and 
was the Sigmoid function. The trainable parameters were optimized through the Adam optimizer [47].
Evaluation metrics
Cross-validation is widely used to evaluate predictive model performance [48–53]. In this approach, samples are divided into several parts. Each part is used once as the test set, with the remaining parts forming the training set. The model trained on the training set is then applied to the test set to assess performance. In this study, we used tenfold cross-validation, dividing all proteins into ten parts and reporting the average performance across the ten folds. Standard tenfold cross-validation assigns samples randomly, which may lead to differences in the distributions of proteins in the training and test sets. These differences may bias model evaluation. To address this, we also used stratified tenfold cross-validation, which reduces such discrepancies. Specifically, proteins were first divided into nineteen groups. For the 2758 proteins annotated with exact one lysine modification type, each modification type formed one group, yielding ten groups. Proteins annotated with exact two, three, and so on, lysine modification types formed the remaining nine groups. Proteins in each group were then randomly divided into ten parts, and parts with the same index across all groups were combined to form one fold. This process reduced distributional differences among the ten folds.
Receiver operating characteristic (ROC) and Precision–Recall (PR) curves are commonly used to evaluate predicted results [54–62]. Originally designed for binary classification, these curves can be adapted for multi-label settings. For each label (i.e., lysine modification type), samples carrying that label are treated as positive and all others as negative. Under varying thresholds, the true positive rate (TPR) and false positive rate (FPR) are computed to generate the ROC curve, where TPR is plotted on the Y-axis and FPR on the X-axis. PR curves are generated analogously, using precision on the Y-axis and recall (equivalent to TPR) on the X-axis. The areas under these curves, AUROC and AUPR, provide quantitative measures of predictive performance. In this study, AUROC and AUPR were computed for each lysine modification type, and the average values were reported as the overall model performance.
Results and discussion
Parameter setting
In general, several parameters must be tuned to build an efficient computational model. The optimal settings were determined by evaluating model performance under standard tenfold cross-validation. For PLysPTM-HGNN, the main parameter settings were as follows.
In the linear transformation layer, the dimension parameter m controls the dimensionality of protein representations. We tested values of 256 and 384 and selected 256, as this setting achieved better performance. In the hybrid GNN layer, the feature dimension was fixed, and the trainable weight matrices were set to 
. In the prediction layer, the numbers of neurons in the hidden layer were set to 256, 384, and 512; cross-validation results indicated that 256 was the best choice.
During model training, the batch size was 128 and the number of training epochs was fixed at 200. The learning rate followed cosine annealing with an initial value of 0.001. Different learning rates were applied to the three feature types, ultimately set to 0.0001 (GO features), 0.0006 (LLM features), and 0.0025 (PSSM features). Gradient clipping was used to prevent gradient explosion during backpropagation.
Performance of PLysPTM-HGNN
The proposed PLysPTM-HGNN model, using the parameter settings described in Sect. “Parameter setting”, was first evaluated using standard tenfold cross-validation. The ROC and PR curves for each fold are shown in Figs. 3(A) and (B), and the corresponding AUROC and AUPR values are listed in Table 3. The AUROC values ranged from 0.83–0.86, and the AUPR values ranged from 0.65–0.72. The average AUROC and AUPR were 0.8456 and 0.6888, respectively, indicating strong model performance. The AUROC and AUPR values showed small variation across folds, with standard deviations of 0.0047 and 0.0133, suggesting good stability. The model predicts ten lysine modification types. We calculated the AUROC and AUPR values for each type, and the ROC and PR curves are displayed in Figs. 4(A) and (B). Most AUROC values were high, with nine exceeding 0.8. For AUPR, two values were above 0.9 and six exceeded 0.6. These results demonstrate that PLysPTM-HGNN performs well across many lysine modification types. Because PR curves are more sensitive than ROC curves to data imbalance, AUPR values were lower than AUROC values for eight lysine modification types. This is expected, as negative samples greatly outnumbered positive samples for these types. For example, β-hydroxybutyrylation included 972 positive proteins, whereas the remaining ~ 9300 proteins were negative, resulting in an AUPR of 0.5131, which was approximately 0.39 lower than its corresponding AUROC. For ubiquitination and acetylation, which had many positive samples, AUPR exceeded AUROC.
Fig. 3.
ROC and PR curves of PLysPTM-HGNN under two tenfold cross-validation strategies. A ROC curves for standard tenfold cross-validation; B PR curves for standard tenfold cross-validation; C ROC curves for stratified tenfold cross-validation; D PR curves for stratified tenfold cross-validation. AUROC values are approximately 0.84 and AUPR values are about 0.68
Table 3.
Performance of PLysPTM-HGNN on ten folds under two tenfold cross-validation strategies
| Fold index | Standard tenfold cross-validation | Stratified tenfold cross-validation | ||
|---|---|---|---|---|
| AUROC | AUPR | AUROC | AUPR | |
| 1 | 0.8501 | 0.7102 | 0.8350 | 0.6867 |
| 2 | 0.8450 | 0.6793 | 0.8461 | 0.6782 |
| 3 | 0.8480 | 0.6833 | 0.8430 | 0.6736 |
| 4 | 0.8492 | 0.6882 | 0.8382 | 0.6834 |
| 5 | 0.8399 | 0.6902 | 0.8450 | 0.6977 |
| 6 | 0.8484 | 0.6981 | 0.8441 | 0.6869 |
| 7 | 0.8499 | 0.6992 | 0.8542 | 0.6955 |
| 8 | 0.8435 | 0.6858 | 0.8491 | 0.6868 |
| 9 | 0.8347 | 0.6580 | 0.8498 | 0.6902 |
| 10 | 0.8469 | 0.6954 | 0.8458 | 0.6980 |
| Average | 0.8456 | 0.6888 | 0.8450 | 0.6877 |
| Standard deviation | 0.0047 | 0.0133 | 0.0053 | 0.0076 |
Fig. 4.
Performance of PLysPTM-HGNN on ten lysine modification types under two tenfold cross-validation strategies. A ROC curves under standard tenfold cross-validation; B PR curves under standard tenfold cross-validation; C ROC curves under stratified tenfold cross-validation; D PR curves under stratified tenfold cross-validation. Most AUROC values exceed 0.8 and several AUPR values exceed 0.6
We further evaluated PLysPTM-HGNN using stratified tenfold cross-validation. The AUROC and AUPR values for each fold are also listed in Table 3, and the corresponding ROC and PR curves are illustrated in Figs. 3(C) and (D). These values were similar to those obtained using standard tenfold cross-validation. The ROC and PR curves for the ten lysine modification types under this strategy are presented in Figs. 4(C) and (D), and showed similar trends. Taken together, the results from both cross-validation strategies indicate that PLysPTM-HGNN performs well in predicting PLM types.
Ablation tests
This study developed a computational model that incorporates three feature types and several modules (e.g., linear transformation, hybrid GNN). Several ablation tests were conducted to assess the importance of these feature types and modules.
First, feature-level ablation tests were performed. The original model used three feature types—GO, LLM, and PSSM features. In addition to the full model (PLysPTM-HGNN), we constructed models representing all possible combinations of feature types:
PLysPTM-HGNN-GO: This model used GO features only.
PLysPTM-HGNN-LLM: This model used LLM features only.
PLysPTM-HGNN-PSSM: This model used PSSM features only.
PLysPTM-HGNN-GO&LLM: This model used GO and LLM features.
PLysPTM-HGNN-GO&PSSM: This model used GO and PSSM features.
PLysPTM-HGNN-PSSM&LLM: This model used PSSM and LLM features.
These models were evaluated using standard tenfold cross-validation. The average AUROC and AUPR values are listed in Table 4. Compared with the performance of PLysPTM-HGNN (AUROC: 0.8456 ± 0.0047; AUPR: 0.6888 ± 0.0133), all models using only partial feature types showed reduced performance. This finding indicates that all three feature types contributed positively to the predictive ability of PLysPTM-HGNN. However, the contributions of the three feature types were not equivalent. Among the single-feature models, PLysPTM-HGNN-LLM achieved the highest AUROC and AUPR, followed by PLysPTM-HGNN-PSSM, and then PLysPTM-HGNN-GO. These results demonstrate that LLM features were most informative, which is reasonable because LLM features are derived from models trained on very large protein databases and therefore capture rich sequence information.
Table 4.
Performance of PLysPTM-HGNN and its variants (Bold numbers indicate the highest values in the corresponding columns)
| Model | AUROC | AUPR |
|---|---|---|
| PLysPTM-HGNN-GO | 0.7632 ± 0.0078 | 0.5711 ± 0.0126 |
| PLysPTM-HGNN-LLM | 0.8357 ± 0.0050 | 0.6695 ± 0.0074 |
| PLysPTM-HGNN-PSSM | 0.7896 ± 0.0090 | 0.6145 ± 0.0130 |
| PLysPTM-HGNN-GO&LLM | 0.8381 ± 0.0041 | 0.6715 ± 0.0107 |
| PLysPTM-HGNN-GO&PSSM | 0.8025 ± 0.0053 | 0.6145 ± 0.0130 |
| PLysPTM-HGNN-LLM&PSSM | 0.8417 ± 0.0036 | 0.6785 ± 0.0081 |
| PLysPTM-HGNN-GCN | 0.8437 ± 0.0044 | 0.6825 ± 0.0085 |
| PLysPTM-HGNN-GAT | 0.8408 ± 0.0055 | 0.6784 ± 0.0117 |
| PLysPTM-HGNN-CNN | 0.8445 ± 0.0045 | 0.6839 ± 0.0118 |
| PLysPTM-HGNN-Linear | 0.8174 ± 0.0041 | 0.6332 ± 0.0149 |
| PLysPTM-HGNN-Res | 0.8183 ± 0.0041 | 0.6348 ± 0.0132 |
| PLysPTM-HGNN | 0.8456 ± 0.0047 | 0.6888 ± 0.0133 |
Second, module-level ablation tests were conducted. We evaluated the importance of several modules by removing or replacing them. The following five models were constructed:
PLysPTM-HGNN-GCN: Hybrid GNN (GCN + GAT + GCN) replaced with GCN + GCN + GCN.
PLysPTM-HGNN-GAT: Hybrid GNN replaced with GAT + GAT + GAT.
PLysPTM-HGNN-CNN: CNN combiner replaced with simple concatenation.
PLysPTM-HGNN-Linear: Linear transformation layer removed.
PLysPTM-HGNN-Res: Residual module removed.
These models were evaluated under standard tenfold cross-validation, and the results are listed in Table 4. All showed lower performance than PLysPTM-HGNN, indicating that the hybrid GNN, CNN combiner, linear transformation, and residual modules all contributed positively to the model’s effectiveness. Notably, PLysPTM-HGNN-Linear and PLysPTM-HGNN-Res exhibited the largest performance decreases, highlighting the key roles of the linear transformation and residual modules in PLysPTM-HGNN.
Effectiveness of separately processing three feature types
Models that employ multiple feature types often combine them at the beginning of the pipeline. In PLysPTM-HGNN, GO, LLM, and PSSM features were instead processed separately by the linear transformation, hybrid GNN, and CNN combiner layers. To examine the rationale for this design, we constructed a model that first combined the three feature types and then applied the linear transformation, hybrid GNN, CNN combiner, and prediction layers. This model was evaluated using standard tenfold cross-validation and achieved an average AUROC of 0.8369 and an average AUPR of 0.6747. These values were lower than those of PLysPTM-HGNN. By processing the feature types separately, PLysPTM-HGNN improved AUROC and AUPR by approximately 0.01. Since the linear transformation, hybrid GNN, and CNN combiner contain trainable parameters, processing each feature type independently allows these parameters to be optimized separately. This yields more informative high-level representations and improves predictive performance.
Comparison with other models
As described above, PLysPTM-HGNN demonstrated strong performance in identifying lysine modification types of proteins. Because this is a pioneer model for prediction of lysine modification types at protein level, no direct comparison models exist. To partly evaluate its effectiveness, we constructed or employed several alternative models.
First, we constructed two models using traditional multi-label classification algorithms: Binary Relevance (BR) [63] and RAndom k-labELsets (RAkEL) [64]. To facilitate implementation, we employed the BR and RAkEL tools in MEKA (http://waikato.github.io/meka/) [65], a software suite containing multiple multi-label classification algorithms. The three feature types were directly input to these tools using their default parameters. For convenience, the resulting models are referred to as BR and RAkEL.
Second, we employed two multi-label protein subcellular localization models (ProtLoc [66] and DeepLoc 2.0 [67]):
ProtLoc [66]: Represents proteins as graphs and performs multi-label subcellular localization prediction using GraphSAGE and multi-head attention mechanisms.
DeepLoc 2.0 [67]: A multi-label subcellular localization model based on pLMs that uses only sequence information and has shown strong performance in previous studies.
These two models were directly employed to predict lysine modification types. For simplicity, they are referred to as ProtLoc and DeepLoc 2.0.
All four comparison models were evaluated using standard tenfold cross-validation. Their average AUROC and AUPR values are listed in Table 5. BR produced the best performance among them (AUROC: 0.7940; AUPR: 0.5920). ProtLoc and DeepLoc 2.0 yielded similar performance (AUROC ≈ 0.74; AUPR ≈ 0.54). RAkEL produced the lowest performance (AUROC < 0.6; AUPR < 0.4). For comparison, PLysPTM-HGNN is also included in Table 5. Clearly, both AUROC and AUPR of PLysPTM-HGNN exceeded those of all four models. Compared with the best baseline model (BR), PLysPTM-HGNN improved AUROC by 6.5% and AUPR by 16.4%, demonstrating its superiority.
Table 5.
Performance of different models (Bold numbers indicate the highest values in the corresponding columns)
| Model | AUROC | AUPR |
|---|---|---|
| BR | 0.7940 | 0.5920 |
| RAkEL | 0.5830 | 0.3870 |
| ProtLoc | 0.7447 | 0.5428 |
| DeepLoc 2.0 | 0.7434 | 0.5399 |
| PLysPTM-HGNN | 0.8456 | 0.6888 |
Analysis of model performance on different protein groups
As shown in Fig. 1, some proteins were annotated with many lysine modification types, whereas others had only one or two. It was necessary to investigate whether the number of annotated modification types affected the performance of PLysPTM-HGNN. For this purpose, all proteins were divided into three groups: group 1 contained 5 031 proteins annotated with 1–2 lysine modification types; group 2 consisted of 3025 proteins annotated with 3–4 types; and group 3 contained 2216 proteins annotated with five or more types. The performance of PLysPTM-HGNN on these groups was evaluated using standard tenfold cross-validation. The AUROC and AUPR values are shown in Table 6. Group 3 achieved the highest AUROC and AUPR, followed by groups 2 and 1. This pattern suggests that PLysPTM-HGNN produced more reliable predictions for proteins annotated with more lysine modification types. This result is reasonable because proteins with multiple annotated modification types have generally been studied more extensively, providing more accurate information, from which informative features can be extracted.
Table 6.
Performance of PLysPTM-HGNN on different protein groups
| Group index | Protein | AUROC | AUPR |
|---|---|---|---|
| 1 | Proteins with 1–2 Lysine modification types | 0.7024 ± 0.0286 | 0.2699 ± 0.0272 |
| 2 | Proteins with 3–4 Lysine modification types | 0.7511 ± 0.0203 | 0.5255 ± 0.0161 |
| 3 | Proteins with at least 5 Lysine modification types | 0.7837 ± 0.0246 | 0.8058 ± 0.0164 |
Analysis of the imbalanced problem on PLysPTM-HGNN
As shown in Table 1, dataset S was imbalanced. The largest class (ubiquitination) contained 8557 proteins, which was 8.8 times larger than the smallest class (β-hydroxybutyrylation). Models trained on imbalanced datasets often exhibit bias toward the majority classes, yielding higher performance on these classes and lower performance on minority classes. For PLysPTM-HGNN, we examined AUPR for each lysine modification type because AUPR is more sensitive than AUROC to class imbalance. The AUPR values for ubiquitination and acetylation, which were the two largest classes, were 0.9576 and 0.9032, respectively, and these were the highest values across all ten types. In contrast, the AUPR for β-hydroxybutyrylation, the smallest class, was 0.5131. These results are consistent with expected behavior on imbalanced datasets, indicating that class imbalance affected PLysPTM-HGNN. In future work, oversampling or undersampling strategies or modified loss functions (e.g., Focal Loss [68]) will be considered to reduce the influence of imbalance and further improve predictive performance.
Conclusions
This study presents a different approach for investigating lysine modifications. Unlike previous models that focused on predicting individual lysine modification sites, PLysPTM-HGNN predicts which lysine modification types can be annotated to lysine residues within an entire protein sequence. By integrating features derived from full protein sequences and applying deep learning algorithms, the model achieved strong predictive performance. As a pioneer model designed to identify lysine modification types at the protein level, this work may encourage further research in this direction and support the development of more efficient computational models. The data and code used in this study are available at https://github.com/barryjim-y/PLysPTM-HGNN.
Acknowledgements
Not applicable.
Abbreviations
- PTM
Post-translational modification
- PLM
Protein lysine modification
- MS
Mass spectrometry
- CPLM
Compendium of protein lysine modifications
- GO
Gene ontology
- LLM
Large language model
- PSSM
Position-specific scoring matrix
- CNN
Convolutional neural network
- CD-HIT
Cluster database at high identity with tolerance
- pLM
Protein language model
- PPI
Protein–protein interaction
- GCN
Graph convolutional network
- GAT
Graph attention network
- ROC
Receiver operating characteristic
- PR
Precision–recall
- TPR
True positive rate
- FPR
False positive rate
- BR
Binary relevance
- RAkEL
RAndom k-labELsets
Author contributions
L.C. and Y.D.C. designed the research; L.C., J.Y. and B.Z. conducted the experiments; J.Y., B.Z. and Y.D.C. analyzed the results; L.C. and J.Y. wrote the manuscript. All authors have read and approved the manuscript.
Funding
Not applicable.
Data availability
The data underlying this study are openly available in CPLM at http://cplm.biocuckoo.cn/. The codes and refined data are available at https://github.com/barryjim-y/PLysPTM-HGNN.
Declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Contributor Information
Lei Chen, Email: lchen@shmtu.edu.cn.
Yu-Dong Cai, Email: cai_yud@126.com.
References
- 1.Walsh CT, Garneau-Tsodikova S, Gatto GJ Jr. Protein posttranslational modifications: the chemistry of proteome diversifications. Angew Chem Int Ed Engl. 2005;44(45):7342–72. [DOI] [PubMed] [Google Scholar]
- 2.Liu Y, Zhou M, Bu Y, Qin L, Zhang Y, Shao S, et al. Lysine acetylation regulates the AT-rich DNA possession ability of H-NS. Nucleic Acids Res. 2024;52(4):1645–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Li S, Li H, Li M, Shyr Y, Xie L, Li Y. Improved prediction of lysine acetylation by support vector machines. Protein Pept Lett. 2009;16(8):977–83. [DOI] [PubMed] [Google Scholar]
- 4.Meng L, Chen X, Cheng K, Chen N, Zheng Z, Wang F, et al. Transptm: a transformer-based model for non-histone acetylation site prediction. Brief Bioinform. 2024;25(3):bbae219. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Wang H, Yan Z, Liu D, Zhao H, Zhao J. MDC-Kace: a model for predicting lysine acetylation sites based on modular densely connected convolutional networks. IEEE Access. 2020;8:214469–80. [Google Scholar]
- 6.Basith S, Chang HJ, Nithiyanandam S, Shin TH, Manavalan B, Lee G. Recent trends on the development of machine learning approaches for the prediction of lysine acetylation sites. Curr Med Chem. 2022;29(2):235–50. [DOI] [PubMed] [Google Scholar]
- 7.Kim DI, Cutler JA, Na CH, Reckel S, Renuse S, Madugundu AK, et al. BioSITe: a method for direct detection and quantitation of site-specific biotinylation. J Proteome Res. 2018;17(2):759–69. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Huang G, Zheng Y, Wu Y-Q, Han G-S, Yu Z-G. An information entropy-based approach for computationally identifying histone lysine butyrylation. Front Genet. 2020;10:1325. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Xu J-Y, Xu Z, Liu X, Tan M, Ye B-C. Protein acetylation and butyrylation regulate the phenotype and metabolic shifts of the endospore-forming Clostridium acetobutylicum. Mol Cell Proteomics. 2018;17(6):1156–69. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Chen Y-Z, Wang Z-Z, Wang Y, Ying G, Chen Z, Song J. NhKcr: a new bioinformatics tool for predicting crotonylation sites on human nonhistone proteins based on deep learning. Brief Bioinform. 2021;22(6):bbab146. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Li Z, Fang J, Wang S, Zhang L, Chen Y, Pian C. Adapt-Kcr: a novel deep learning framework for accurate prediction of lysine crotonylation sites based on learning embedding features and attention architecture. Brief Bioinform. 2022;23(2):bbac037. [DOI] [PubMed] [Google Scholar]
- 12.Lv H, Dao F-Y, Guan Z-X, Yang H, Li Y-W, Lin H. Deep-Kcr: accurate detection of lysine crotonylation sites using deep learning method. Brief Bioinform. 2021;22(4):bbaa255. [DOI] [PubMed] [Google Scholar]
- 13.Wei X, Sha Y, Zhao Y, He N, Li L. DeepKcrot: a deep-learning architecture for general and species-specific lysine crotonylation site prediction. IEEE Access. 2021;9:49504–13. [Google Scholar]
- 14.Shi S-P, Qiu J-D, Sun X-Y, Suo S-B, Huang S-Y, Liang R-P. PLMLA: prediction of lysine methylation and lysine acetylation by combining multiple features. Mol Biosyst. 2012;8(5):1520–7. [DOI] [PubMed] [Google Scholar]
- 15.Spadaro A, Sharma A, Dehzangi I. Predicting lysine methylation sites using a convolutional neural network. Methods. 2024;226:127–32. [DOI] [PubMed] [Google Scholar]
- 16.Ju Z, Cao J-Z, Gu H. ILM-2L: a two-level predictor for identifying protein lysine methylation sites and their methylation degrees by incorporating K-gap amino acid pairs into Chou’ s general PseAAC. J Theor Biol. 2015;385:50–7. [DOI] [PubMed] [Google Scholar]
- 17.Li A, Deng Y, Tan Y, Chen M. A transfer learning-based approach for lysine propionylation prediction. Front Physiol. 2021;12:658633. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Ju Z, He J-J. Prediction of lysine propionylation sites using biased SVM and incorporating four different sequence features into Chou’s PseAAC. J Mol Graph Model. 2017;76:356–63. [DOI] [PubMed] [Google Scholar]
- 19.Jia J, Liu Z, Xiao X, Liu B, Chou K-C. PSuc-Lys: predict lysine succinylation sites in proteins with PseAAC and ensemble random forest approach. J Theor Biol. 2016;394:223–30. [DOI] [PubMed] [Google Scholar]
- 20.Xu Y, Ding Y-X, Ding J, Lei Y-H, Wu L-Y, Deng N-Y. ISuc-PseAAC: predicting lysine succinylation in proteins by incorporating peptide position-specific propensity. Sci Rep. 2015;5(1):10184. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Zhu Y, Jia C, Li F, Song J. Inspector: a lysine succinylation predictor based on edited nearest-neighbor undersampling and adaptive synthetic oversampling. Anal Biochem. 2020;593:113592. [DOI] [PubMed] [Google Scholar]
- 22.Cai Y, Huang T, Hu L, Shi X, Xie L, Li Y. Prediction of lysine ubiquitination with mRMR feature selection and analysis. Amino Acids. 2012;42(4):1387–95. [DOI] [PubMed] [Google Scholar]
- 23.Qiu W-R, Xiao X, Lin W-Z, Chou K-C. IUbiq-Lys: prediction of lysine ubiquitination sites in proteins by extracting sequence evolution information via a gray system model. J Biomol Struct Dyn. 2015;33(8):1731–42. [DOI] [PubMed] [Google Scholar]
- 24.Wang H, Xiao D, Li W, Zhang X, Gao X, Yong J, et al. A simple and fast quantitative analysis of quinolizidine alkaloids and their biosynthetic precursor, lysine, in Sophora alopecuroides by hydrophilic interaction chromatography coupled with triple-quadrupole tandem mass spectroscopy. Phytochem Anal. 2018;29(5):500–6. [DOI] [PubMed] [Google Scholar]
- 25.Qiu W-R, Sun B-Q, Xiao X, Xu Z-C, Chou K-C. IPTM-mLys: identifying multiple lysine PTM sites and their different types. Bioinformatics. 2016;32(20):3116–23. [DOI] [PubMed] [Google Scholar]
- 26.Ahmed S, Rahman A, Hasan MAM, Rahman J, Islam MKB, Ahmad S. PredML-Site: predicting multiple lysine PTM sites with optimal feature representation and data imbalance minimization. IEEE ACM Trans Comput Biol Bioinform. 2022;19(6):3624–34. [DOI] [PubMed] [Google Scholar]
- 27.Hasan MAM, Ahmad S. mlysptmpred: multiple lysine ptm site prediction using combination of svm with resolving data imbalance issue. Nat Sci. 2018;10(9):370–84. [Google Scholar]
- 28.Chen L, Chen Y. Rmtlysptm: recognizing multiple types of lysine PTM sites by deep analysis on sequences. Brief Bioinform. 2024;25(1):bbad450. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Ahmed S, Rahman A, Hasan MAM, Ahmad S, Shovan SM. Computational identification of multiple lysine PTM sites by analyzing the instance hardness and feature importance. Sci Rep. 2021;11(1):18882. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Sua JN, Lim SY, Yulius MH, Su XT, Yapp EKY, Le NQK, et al. Incorporating convolutional neural networks and sequence graph transform for identifying multilabel protein Lysine PTM sites. Chemometr Intell Lab Syst. 2020. 10.1016/j.chemolab.2020.104171. [Google Scholar]
- 31.Zuo Y, Hong Y, Zeng X, Zhang Q, Liu X. Mlyspred: graph-based multi-view clustering and multi-dimensional normal distribution resampling techniques to predict multiple lysine sites. Brief Bioinform. 2022;23(5):bbac277. [DOI] [PubMed] [Google Scholar]
- 32.Zhang W, Tan X, Lin S, Gou Y, Han C, Zhang C, et al. CPLM 4.0: an updated database with rich annotations for protein lysine modifications. Nucleic Acids Res. 2022;50(D1):D451–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Yang B, Bao W, Wang J. Active disease-related compound identification based on capsule network. Brief Bioinform. 2022;23(1):bbab462. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Bao W, Liu Y, Chen B. Oral_voting_transfer: classification of oral microorganisms’ function proteins with voting transfer model. Front Microbiol. 2024;14:1277121. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Yuan L, Xu Z, Meng B, Ye L. ScAMZI: attention-based deep autoencoder with zero-inflated layer for clustering scRNA-seq data. BMC Genomics. 2025;26(1):350. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Chen B, Li N, Bao W. CLPr_in_ML: cleft lip and palate reconstructed features with machine learning. Curr Bioinform. 2025;20(2):179–93. [Google Scholar]
- 37.Fu L, Niu B, Zhu Z, Wu S, Li W. CD-HIT: accelerated for clustering the next-generation sequencing data. Bioinformatics. 2012;28(23):3150–2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, et al. Gene ontology: tool for the unification of biology. Nat Genet. 2000;25(1):25–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Mikolov T, Sutskever I, Chen K, Corrado GS, Dean J. Distributed representations of words and phrases and their compositionality. Adv neural inf process syst. 2013. 10.48550/arXiv.1310.4546. [Google Scholar]
- 40.Boutet E, Lieberherr D, Tognolli M, Schneider M, Bairoch A. UniProtKB/Swiss-Prot: the manually annotated section of the UniProt KnowledgeBase. In: Plant bioinformatics: methods and protocols. Totowa, NJ: Humana Press; 2007. p. 89–112. [Google Scholar]
- 41.Elnaggar A, Heinzinger M, Dallago C, Rehawi G, Wang Y, Jones L, et al. Prottrans: toward understanding the language of life through self-supervised learning. IEEE Trans Pattern Anal Mach Intell. 2022;44(10):7112–27. [DOI] [PubMed] [Google Scholar]
- 42.Guigo R. An introduction to position specific scoring matrices. Bioinformatica upf edu. 2016.
- 43.Altschul SF, Madden TL, Schäffer AA, Zhang J, Zhang Z, Miller W, et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res. 1997;25(17):3389–402. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Bairoch A, Apweiler R. The SWISS-PROT protein sequence data bank and its supplement TrEMBL in 1999. Nucleic Acids Res. 1999;27(1):49–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Chowdhury SY, Shatabda S, Dehzangi A. IDNAProt-ES: identification of DNA-binding proteins using evolutionary and structural features. Sci Rep. 2017;7(1):14938. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Szklarczyk D, Gable AL, Lyon D, Junge A, Wyder S, Huerta-Cepas J, et al. STRING v11: protein–protein association networks with increased coverage, supporting functional discovery in genome-wide experimental datasets. Nucleic Acids Res. 2019;47(D1):D607–13. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Kingma DP, Ba J. Adam: a method for stochastic optimization. In: 3rd international conference on learning representations: 7–9 May 2015; Louisiana, USA. 2019.
- 48.Kohavi R A study of cross-validation and bootstrap for accuracy estimation and model selection. In: International joint Conference on artificial intelligence: 1995. Lawrence Erlbaum Associates Ltd. pp 1137–1145.
- 49.Bao Y, Ma Q, Chen L, Feng K, Guo W, Huang T, et al. Recognizing SARS-CoV-2 infection of nasopharyngeal tissue at the single-cell level by machine learning method. Mol Immunol. 2025;177:44–61. [DOI] [PubMed] [Google Scholar]
- 50.Liao H, Ma Q, Chen L, Guo W, Feng K, Bao Y, et al. Machine learning analysis of CD4+ T cell gene expression in diverse diseases: insights from cancer, metabolic, respiratory, and digestive disorders. Cancer Genet. 2025;290–291:56–60. [DOI] [PubMed] [Google Scholar]
- 51.Ma Q, Zhang Y, Chen L, Bao Y, Guo W, Feng K, et al. Machine learning-driven discovery of essential binding preference in anti-CRISPR proteins. PROTEOMICS–Clin Appl. 2025;19(4):e70013. [DOI] [PubMed] [Google Scholar]
- 52.Ouyang D, Liang Y, Wang J, Li L, Ai N, Feng J, et al. HGCLAMIR: hypergraph contrastive learning with attention mechanism and integrated multi-view representation for predicting miRNA-disease associations. PLoS Comput Biol. 2024;20(4):e1011927. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 53.Yu Z, Huang F, Zhao X, Xiao W, Zhang W. Predicting drug-disease associations through layer attention graph convolutional network. Brief Bioinform. 2021;22(4):bbaa243. [DOI] [PubMed] [Google Scholar]
- 54.Bradley AP. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recogn. 1997;30(7):1145–59. [Google Scholar]
- 55.Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. In: The 23rd international conference on machine learning: 2006. pp 233–240.
- 56.Chen L, Li J. PDTDAHN: Predicting drug-target-disease associations using a heterogeneous network. Current Bioinformatics (in press). 2025.
- 57.Chen L, Zhang S, Zhou B. Herb-disease association prediction model based on network consistency projection. Sci Rep. 2025;15:3328. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Chen L, Lu Y, Xu J, Zhou B. Prediction of drug’s anatomical therapeutic chemical (ATC) code by constructing biological profiles of ATC codes. BMC Bioinformatics. 2025;26:86. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Chen L, Zhu W, Chen D. An End-to-End 3D graph neural network for predicting drug-target-disease associations. Current Bioinform. 2025.
- 60.Wang Y, Lu P. GEHGAN: circRNA–disease association prediction via graph embedding and heterogeneous graph attention network. Comput Biol Chem. 2024;110:108079. [DOI] [PubMed] [Google Scholar]
- 61.Lin S, Wang Y, Zhang L, Chu Y, Liu Y, Fang Y, et al. MDF-SA-DDI: predicting drug-drug interaction events based on multi-source drug fusion, multi-source feature fusion and transformer self-attention mechanism. Brief Bioinform. 2022;23(1):bbab421. [DOI] [PubMed] [Google Scholar]
- 62.Chen L, Chen Y, Zhou B. HCLAMCMI: prediction of circRNA-miRNA interactions based on hypergraph contrastive learning and an attention mechanism. J Chem Inf Model. 2025;65(21):12099–115. [DOI] [PubMed] [Google Scholar]
- 63.Zhang M-L, Li Y-K, Liu X-Y, Geng X. Binary relevance for multi-label learning: an overview. Front Comput Sci. 2018;12(2):191–202. [Google Scholar]
- 64.Tsoumakas G, Katakis I, Vlahavas I. Random k-labelsets for multilabel classification. IEEE Trans Knowl Data Eng. 2011;23(7):1079–89. [Google Scholar]
- 65.Read J, Reutemann P, Pfahringer B, Holmes G. MEKA: a multi-label/multi-target extension to WEKA. J Mach Learn Res. 2016;17(21):1–5. [Google Scholar]
- 66.Liang Q, Qiu W, Lin W. Predicting protein subcellular localization with multi-label using GraphSAGE and multi-head attention mechanism. In: proceedings of the 2024 4th international conference on bioinformatics and intelligent computing: 2024. pp 414–419.
- 67.Thumuluri V, Almagro Armenteros JJ, Johansen AR, Nielsen H, Winther O. Deeploc 2.0: multi-label subcellular localization prediction using protein language models. Nucleic Acids Res. 2022;50(W1):W228–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 68.Lin T-Y, Goyal P, Girshick R, He K, Dollár P. Focal loss for dense object detection. In: proceedings of the IEEE international conference on computer vision: 2017. pp 2980–2988.
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data underlying this study are openly available in CPLM at http://cplm.biocuckoo.cn/. The codes and refined data are available at https://github.com/barryjim-y/PLysPTM-HGNN.