CMV-GLA: contrastive multi-view graph layer attention for predicting phosphorylation site-disease associations

Abstract

Protein phosphorylation is a critical posttranslational modification involved in cell signaling and metabolic regulation, and phosphorylation dysregulation is closely associated with various diseases. However, studies on phosphorylation site-disease associations remain scarce, and existing methods often fail to fully leverage complementary information across heterogeneous data sources, resulting in suboptimal predictive performance. We present CMV-GLA (Contrastive Multi-View Graph Layer Attention Network), a novel framework integrating protein sequences, disease semantics, and known associations into three complementary graph views (site similarity, disease similarity, association network). Built on a GAT backbone, CMV-GLA employs layer attention to mitigate over-smoothing and contrastive learning to maximize cross-view node embedding consistency while minimizing inter-node confusion, enhancing discriminative power. Specifically, the contrastive learning module encourages consistent representations of the same node across different graph views while explicitly separating embeddings of unrelated nodes, leading to more distinguishable and robust feature representations. Evaluated on benchmark data, CMV-GLA significantly outperforms state-of-the-art methods in AUC and AUPRC. Ablation studies confirm the critical roles of multi-view fusion and the contrastive module. Case studies demonstrate high-confidence, literature-supported predictions, highlighting CMV-GLA’s utility for elucidating phosphorylation mediated mechanisms and guiding therapeutic discovery. Code and dataset available at https://github.com/ljr078/CMV-GLA.

Introduction

Post-translational modifications (PTMs) refer to chemical alterations-including addition, removal, or modification of functional groups-ccurring on amino acid residues after protein translation [1, 2]. As pivotal regulators of physiological processes, PTMs govern critical biological functions such as protein synthesis, gene expression, and enzymatic activity modulation, profoundly influencing cellular behavior and disease pathogenesis [3, 4]. Among these, protein phosphorylation stands as a central PTM, orchestrating fundamental cellular activities including cell proliferation [5], signal transduction [6], tumorigenesis [7], and DNA repair [8]. Studies indicate that more than 25% of eukaryotic proteins undergo phosphorylation. This dynamic, reversible modification profoundly impacts biological processes by modulating protein conformation, activity, and interactions: It acts as an enzymatic "switch" where mutations at phosphorylation sites can disrupt kinase recognition or impede modification, causing protein dysfunction and pathway dysregulation [9, 10]. Phosphorylation homeostasis further regulates protein stability, enzyme activity, and subcellular localization; its disruption drives disease by amplifying pathogenic signals or suppressing protective pathways [11, 12]. Consequently, dysregulated phosphorylation constitutes a central mechanism underlying diverse pathological cascades and is mechanistically implicated in the pathogenesis of neurodegenerative disorders, cancers, and metabolic diseases [13, 14].

The centrality of phosphorylation dysregulation in disease is exemplified across multiple pathologies. In neurodegenerative disorders, Alzheimer’s disease (AD)-accounting for over 60% of global dementia cases and characterized by progressive cognitive decline. Aberrant phosphorylation of the tau protein represents a key pathological hallmark of this disease [15]. Hyperphosphorylation alters tau’s charge distribution, weakening microtubule binding and causing microtubule destabilization and impaired axonal transport [16]. It further promotes tau aggregation into neurofibrillary tangles (NFTs), pathological signatures directly linked to neuronal degeneration and cognitive loss [17, 18]. In oncology, dysregulated phosphorylation is equally critical [19]. For instance, in breast cancer, abnormal kinase activation induces site-specific hyperphosphorylation of key proteins [20, 21], disrupting signaling homeostasis to confer invasiveness, metastatic potential, and tumor progression [22, 23]. Thus, phosphorylation site specificity correlates with disease stage, molecular subtype, and treatment response [24, 25]. Beyond these examples, phosphorylation dysregulation has also been implicated in other major diseases. For example, abnormal phosphorylation of -synuclein at Ser129 is closely associated with Parkinson’s disease, contributing to protein aggregation and neuronal toxicity [26]. In metabolic disorders such as diabetes, dysregulated phosphorylation of insulin signaling components (e.g., IRS proteins) can impair glucose homeostasis [27]. These examples further highlight the widespread impact of phosphorylation abnormalities across diverse disease contexts. Additionally, mutations in residues adjacent to phosphorylation sites may disrupt kinase-recognized sequence motifs, reducing binding efficiency by altering local charge distribution or spatial conformation and thereby inhibiting phosphorylation reactions. The dynamic equilibrium of phosphorylation plays a crucial role in regulating protein stability, enzymatic activity, and subcellular localization [28, 29]. Its disruption may amplify pathogenic signals or suppress protective regulatory pathways, thereby promoting disease onset and progression. Therefore, systematically exploring potential associations between phosphorylation sites and diseases is crucial for elucidating the molecular mechanisms underlying disease onset and progression. It also provides a theoretical basis for screening early diagnostic markers and developing targeted therapeutic strategies. These examples underscore the imperative to elucidate phosphorylation site-disease relationships.

Nevertheless, direct research on phosphorylation site-disease associations remains scarce. Most efforts focus on phosphorylation site prediction-identifying modifiable residues within protein sequences. Approaches fall into traditional machine learning (ML) and deep learning (DL) categories. Early ML models enhanced predictions by integrating diverse features: Huang et al. employed physicochemical properties such as sequence conservation, hydrophobicity in an SVM framework for cross-species site prediction [30], while Fan et al. combined Gene Ontology (GO) annotations and protein-protein interaction networks within a random forest classifier to improve kinase-specific accuracy [31]. Recently, DL has superseded traditional methods by overcoming feature-engineering limitations. Wang et al. developed TransPhos, a Transformer-based model capturing long-range sequence dependencies via self-attention [32], and Luo et al. created DeepPhos using densely connected CNNs (DC-CNNs) for kinase-specific prediction [33]. Advances in site prediction, coupled with expanding experimental data, now provide the methodological and empirical foundation to explore site-disease linkages.

While experimentally identified phosphorylation sites-particularly those disease-associated-accumulate in databases, enabling computat-ional studies, their numbers remain limited relative to biological complexity. Thus, robust computational methods are urgently needed to prioritize candidate sites for experimental validation. Xu et al. pioneered disease-specific prediction using minimum redundancy-maximum relevance (mRMR) filtering with forward feature selection on 15-mer sequences, training SVMs for site classification [34]. Subsequently, they integrated disease-site-kinase relationships into heterogeneous networks for random walk-based inference [35]. However, their prediction results highly depend on known protein-disease interaction data, limiting their ability to discover novel associations. Moreover, the network construction is constrained by the completeness of existing databases, leading to relatively simple information capture. He et al. proposed a prediction model based on transfer learning and deep convolutional neural networks (CNN) to predict leukemia-related phosphorylation sites from protein primary sequences [36]. Using a simple CNN architecture, they automatically extracted features from raw sequences, avoiding the limitations of manual feature engineering in traditional machine learning. However, the model was only tested for leukemia, and its generalization ability was not validated. Later, Deng et al. proposed a computational method based on multi-similarity fusion and CNN to predict the association between protein phosphorylation sites and diseases [37]. This was the first method to integrate multiple similarity networks in phosphorylation site-disease association prediction. Nevertheless, the calculation of similarities between phosphorylation sites and diseases was overly simplistic, and the direct concatenation of different modalities of features from phosphorylation sites and diseases led to considerable information redundancy and insufficient use of each modality. Additionally, CNN’s ability to capture features is constrained by the size of convolutional kernels and network depth, requiring multiple layers to transmit information, which can lead to information loss. Graph neural networks have shown strong potential in related tasks such as lncRNA–disease and miRNA–disease association prediction. For example, Shi et al. [38] proposed VGAELDA, a variational graph autoencoder framework that integrates feature inference and label propagation for lncRNA–disease prediction. Similarly, Jin et al. [39] developed NIMGSA, which combines neural inductive matrix completion with graph autoencoders and self-attention mechanisms for miRNA–disease association prediction. These studies highlight the effectiveness of graph-based representation learning in capturing complex relationships in biomedical association tasks.

Whereas previous studies have explored graph neural networks for modeling phosphorylation site–disease associations, existing approaches often fall short in effectively integrating heterogeneous biological data, enhancing feature discriminability, and mitigating over-smoothing in deep architectures. To address these limitations, we propose CMV-GLA (Contrastive Multi-View Graph Layer Attention Network), which introduces several key innovations. First, we construct three biologically meaningful graph structures that capture complementary aspects of the data: phosphorylation site–disease associations, sequence-based site similarities, and ontology-based disease similarities. This multi-view graph design enables a more comprehensive and semantically enriched representation of the underlying biological landscape. Second, we enhance the graph attention network (GAT) architecture with a layer-wise attention mechanism, allowing the model to dynamically reweight information across different GNN layers. This helps preserve shallow-layer features while reducing the adverse effects of over-smoothing common in deep graph models. Most notably, we introduce a contrastive learning strategy tailored for multi-view representation learning, which encourages consistency in node embeddings across views while enforcing separation between different nodes. This innovation significantly improves the discriminative power of learned representations. Finally, a multilayer perceptron (MLP) scores associations using the fused embeddings. Tenfold cross-validation demonstrates CMV-GLA’s superiority over state-of-the-art methods in AUC and AUPRC. Ablation studies confirm the critical roles of multiview fusion and contrastive modules. Notably, case studies on Alzheimer’s disease, breast cancer, and neuroblastoma reveal high-confidence predictions strongly corroborated by literature evidence, attesting to the model’s biological validity and utility.

CMV-GLA framework

Model overview

Wang et al. [40] proposed GEHGAN, a heterogeneous graph attention network for predicting circular RNA–disease associations. This model leverages graph embedding to initialize node representations and employs heterogeneous attention mechanisms to integrate information from different types of neighboring nodes, enabling effective modeling of complex biological relationships. This work further demonstrates the capability of attention-based graph models in handling heterogeneous biomedical data.

Therefore, our proposed CMV-GLA (Contrastive Multi-View Graph Layer Attention Network) for phosphorylation site-disease association prediction. As illustrated in Fig. 1, CMV-GLA comprises four core modules:

• (A)Similarity matrix constructor: Builds the phosphorylation site-disease association network, phosphorylation site sequence similarity network, phosphorylation site GIP similarity network, disease semantic similarity network, and disease GIP similarity network.
• (B)View-specific similarity graph encoder: Aggregates phos-phorylation site and disease similarity networks using Graph Layer Attention (GLA) to generate view-specific embeddings.
• (C)Cross-entity heterogeneous graph encoder: Constructs site-disease heterogeneous graphs from association networks and applies GLA to derive embeddings under this view.
• (D)Contrastive feature fusion and association predictor: Fuses embeddings from (B) and (C), integrates adaptive and contrastive learning, and predicts association probabilities via MLP.

Fig. 1.

CMV-GLA model framework

Similarity matrix constructor

As outlined in Fig. 1A, we construct complementary similarity networks from multimodal biological data to comprehensively characterize phosphorylation sites and diseases, namely that Phosphorylation Site Sequence Similarity Network, Disease Semantic Similarity Network and Gaussian Interaction Profile (GIP) Kernel Similarity.

Phosphorylation site sequence similarity network

For each phosphorylation site, we extract a k-residue local amino acid sequence centered at the site. Features are encoded using ProtT5-a protein language model pre-trained on UniRef50 via masked language modeling and denoising autoencoding tasks [41]. In this study, ProtT5 is adopted due to its strong generalization ability and computational efficiency for large-scale embedding extraction, making it suitable for generating stable sequence representations as initial node features.

Let denote the ProtT5 embedding for site . The sequence similarity between sites and is computed as:

1

where and are the -th elements of embeddings and , and is the embedding dimensionality.

Disease semantic similarity network

We compute disease semantic similarity using directed acyclic graphs (DAGs) constructed from Medical Subject Headings (MeSH) descriptors, where nodes represent diseases and edges denote ontological relationships [42]. Adopting Fan et al.’s enhanced method [43] that considers both DAG hierarchy and disease specificity, we define the semantic similarity as follows.

For disease A, its DAG structure is represented as , where contains A and all its ancestor diseases, and comprises all connecting edges. The semantic contribution of any disease to A is calculated by:

2

where C(t) is the children set of t, is the semantic contribution factor of the edge connecting t and in , which is typically set to 0.5. is the contribution factor [43], defined by:

3

where D is the total disease count, counts DAGs containing t, and K is the complete MeSH disease set(note updates with MeSH releases). The total semantic value of A is then:

4

Finally, the semantic similarity between diseases A and B is given by Eq. (5):

5

Gaussian interaction profile (GIP) kernel similarity

The Gaussian Interaction Profile (GIP) kernel similarity effectively captures nonlinear relationships within association data by projecting interaction profiles into a high-dimensional feature space [44]. Demonstrated to outperform conventional similarity metrics in heterogeneous biological networks [45], this approach provides robust similarity characterization for both phosphorylation sites and diseases. For any pair of phosphorylation sites and , their GIP similarity is computed as:

6

where denotes the i-th row vector of the binary association matrix S(encoding known site-disease relationships), representing the complete interaction profile of site . The bandwidth parameter controls the kernel’s sensitivity, calculated as:

7

where is set to 1 following established practice [45], and is the total number of phosphorylation sites.

Similarly, disease GIP similarity between diseases and is defined as:

8

where is the i-th column vector of S, representing the complete interaction profile of disease . The bandwidth of the Gaussian kernel function is calculated by:

9

where is set to 1 and is the disease count, ensuring consistent scaling across entities.

During cross-validation, the GIP similarity matrices are constructed within each training fold using only the training association matrix.

View-specific similarity graph encoder

Similarity graph construction

As depicted in Fig. 1B, we construct two sparse graphs by integrating and filtering similarity matrices. First, disease semantic similarity () and phosphorylation site sequence similarity () are fused with their respective Gaussian Interaction Profile (GIP) similarities to mitigate sparsity:

10

11

where represents the integrated disease similarity between diseases and , and represents the integrated phosphorylation site similarity between and .

These integrated matrices are sparsified via k-nearest neighbor filtering: for each phosphorylation site , we retain edges only to its top-k most similar neighbors in PS, generating a binary adjacency matrix where:

12

Similarly, disease adjacency matrix is derived from DS. This yields sparse graphs and ready for feature encoding.

GLA-based feature encoding

The constructed graphs are processed by our Graph Layer Attention (GLA) block to generate view-specific embeddings:

13

14

where and are the initial node feature matrices directly derived from the integrated phosphorylation site similarity network (PS) and disease similarity network (DS), respectively. Specifically, each phosphorylation site (or disease) is represented by its similarity vector with all other nodes (i.e., the corresponding row in PS or DS). and denote the updated node embeddings after feature refinement.

The GLA mechanism operates in two stages:

Stage 1: Neighborhood aggregation

For a node i (Taking phosphorylation sites as an example) at layer , attention weights over neighbors are computed as follows:

15

is node i embeddings from the previous layer (). is a layer-specific weight matrix for transforming embeddings. The operator concatenates, and a is a learnable vector to weight the concatenated result. Then normalized as follows:

16

The updated embedding is then:

17

where is LeakyReLU activation ().

Stage 2: Cross-layer fusion

To counteract over-smoothing, features from all layers are adaptively fused:

18

where layer weights , u and V are learnable parameters. This yields the output embeddings (sites) and (diseases).

Cross-entity heterogeneous graph encoder

As shown in Fig. 1C, we apply the same Graph Layer Attention (GLA) mechanism to the phosphorylation site-disease association graph, constructing a heterogeneous network that integrates both entities. The adjacency matrix of this graph is defined as:

19

where PS and DS are the integrated matrices, and S is the known association matrix. Node features are initialized using the embeddings and obtained from the view-specific similarity encoder.

Processing this heterogeneous graph through the GLA block yields fused representations:

20

where , representing the concatenated feature vectors of phosphorylation site and disease nodes in the heterogeneous graph .

Contrastive feature fusion and association predictor

Shi et al. [44] demonstrate that contrastive learning can effectively address challenges such as high dimensionality, sample scarcity, and class imbalance, which are common in biomedical data analysis. Motivated by these advantages, our contrastive module is designed to align representations of the same entity across different graph views while separating those of distinct entities, thereby enhancing the discriminative power of learned embeddings in phosphorylation site–disease association prediction.

As depicted in Fig. 1D, our framework employs an adaptive fusion module to integrate embeddings from complementary modalities. The fused phosphorylation site and disease representations are computed through weighted combinations:

21

22

where and are learnable gating parameters that dynamically balance contributions between view-specific embeddings and heterogeneous graph embeddings . Specifically, and denote the embeddings derived from view-specific similarity graphs for phosphorylation sites and diseases, respectively, while and represent embeddings learned from the heterogeneous association graph. The final fused representations are denoted as P and D for phosphorylation sites and diseases, respectively.

These fused embeddings are then used to predict association scores via dot product:

23

The composite loss function comprises two key components:

1. Cross-entropy loss supervises the association prediction task:

   24

   where and denote positive and negative phosphorylation site-disease pairs respectively.
2. Contrastive learning objectives enforce representation con-sistency across modalities. The fundamental optimization goal is to satisfy:

   25

where represents positive pairs (same biological entity across different modalities) and denotes negative pairs (different entities).

• For phosphorylation sites, the contrastive loss is formulated as:

  26
• For diseases, the analogous loss is:

  27

The distance metric uses negative cosine similarity:

28

This penalizes dissimilar positive pairs while promoting discrimination between distinct entities.

The complete optimization objective combines these losses:

29

where hyperparameters and control the relative importance of contrastive regularization. This multi-task formulation simultaneously optimizes association prediction accuracy while enhancing representation discriminability through cross-modal consistency constraints.

In practice, we adopt a dynamic negative sampling strategy for the contrastive objective. For each anchor node (e.g., a view-specific embedding of a phosphorylation site or disease), its corresponding embedding from another view is treated as a positive pair, while embeddings of all other nodes are regarded as negative pairs. Importantly, negative samples are not fixed but dynamically resampled at each training epoch. Specifically, the training data are randomly shuffled to generate different negative pair combinations across epochs. This dynamic strategy prevents the model from overfitting to a fixed set of negatives and enables it to explore a more diverse set of contrasting instances during training, thereby improving the robustness and generalization ability of the learned representations.

Results and discussion

Dataset

We curated disease-associated phosphorylation sites from the PhosphoSitePlus [47] and PTMD [48] databases, specifically filtering for human diseases. Disease names were standardized using Medical Subject Headings (MeSH) terms [49], with unmapped entries excluded. The final dataset comprises 2441 phosphorylation site-disease associations spanning 1485 unique sites and 212 diseases. A binary association matrix was constructed where:

30

with sites and diseases explicitly defining the matrix dimensions.

Experimental setup

Experimental design

The dataset contains 1485 phosphorylation sites and 212 diseases, with a total of 2441 known site–disease associations. Due to the severe class imbalance-where unconfirmed associations greatly outnumber confirmed ones-all known associations are treated as positive samples. To construct a balanced dataset, an equal number of negative samples are randomly selected from the pool of unconfirmed associations. This results in 2441 positive and 2441 negative samples, totaling 4882 data points. To robustly evaluate model performance, we employ 10-fold cross-validation, where the dataset is partitioned into 10 subsets. In each round, nine subsets are used for training and one for testing, ensuring comprehensive assessment across different data splits.

Our experimental framework includes five key analyses:

1. Comparison with existing models: Benchmarking against the state‑of‑the‑art deep learning model MSF‑CNN [37] for phosphorylation site‑disease association prediction, as well as three advanced graph neural network models originally developed for miRNA‑disease association prediction: MGCNSS [50], DGAMDA [51], and AMHMDA [52]. All methods were evaluated under Evaluating random sampling and 10‑fold cross‑validation settings.

2. Negative sampling strategy analysis: Evaluating random sampling and Euclidean-based sampling.

3. Parameter optimization studies:

   • Phosphorylation site window size selection (residues from 15 to 35)
   • Sequence encoding method comparison (One-hot vs. Needleman-Wunsch vs. ProtT5)
   • Disease semantic similarity metric evaluation ( vs. )
   • Comparison with Layer Aggregation Methods (Residual connections vs. Simple layer concatenation vs. Ours )

4. Comprehensive ablation experiments: Validating contributions of the contrastive module, GLA block, and adaptive fusion.

5. Case studies: Biological validation of top predictions for Alzheimer’s disease, breast cancer, and neuroblastoma through literature evidence.

Implementation details

Experiments were conducted under the environment specified in Table 1. Key hyperparameters include: 450 training epochs, 0.2 dropout rate, 3-layer GAT architecture, and 128-dimensional embeddings. Neighborhood sizes were set to 100 for sites and 25 for diseases. Details of hyperparameter setting are listed in Table 2.

Table 1.

Experimental environment

Item	Configuration
Operating system	Ubuntu 20.04
CPU	48 vCPU, AMD TR 3960X, 24-Core
GPU	RTX 3090 (24GB)
Python	3.12
PyTorch	2.5.1
CUDA	12.4

Table 2.

Hyperparameter settings

Parameter	Value
Epochs	450
Dropout	0.2
GAT layers	3
Feature dimension	128
Learning rate	1e−3
	100
	25
	1
	1

We performed empirical tuning of key hyperparameters based on validation performance. The number of GAT layers was explored in the range of 1–5, where L = 3 provided a good balance between representation capacity and over-smoothing. The embedding dimension was evaluated from 32 to 512, and d = 128 achieved a favorable trade-off between predictive performance and computational efficiency.

The neighborhood sizes for phosphorylation sites () and diseases () were also examined, with = 100 and = 25 yielding the best results, which is consistent with the relative scale of the two node types. In addition, the contrastive loss weights (, ) were selected via grid search. We observed that setting = = 1.0 led to stable and competitive performance, while smaller values weakened the contrastive effect and larger values tended to degrade overall prediction accuracy.

Overall, the selected hyperparameters demonstrate stable performance within reasonable ranges.

Evaluation metrics

This study evaluates the model’s performance using the following seven metrics: Accuracy, Sensitivity, Specificity, Precision, MCC, AUC, and AUPRC. The first five metrics can be calculated using Equations (31–35). AUC refers to the area under the ROC curve, which is plotted using TPR and FPR at different classification thresholds. AUPRC refers to the area under the PR curve, which is plotted using Precision and Recall at different classification thresholds. The higher the AUC and AUPRC values, the better the model performance.

31

32

33

34

35

where TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives, respectively.

Performance benchmarking

Comparison with state-of-the-art models

To rigorously evaluate the performance of CMV-GLA, we compared it against four baseline methods: MSF-CNN [37], a state-of-the-art deep learning model specifically designed for phosphorylation site–disease association prediction, and three advanced graph neural network models originally developed for miRNA–disease association prediction, namely MGCNSS [50], DGAMDA [51], and AMHMDA [52]. All methods were evaluated under identical experimental settings, including the same 10-fold cross-validation protocol and random negative sampling strategy.

Table 3 reports the average performance of all methods across the ten folds. As shown, CMV-GLA consistently achieves the highest scores across all seven evaluation metrics. Specifically, CMV-GLA improves the MCC by approximately 2.4% over the best baseline (DGAMDA: 0.8352 vs. 0.8589) and increases the AUC by 1.5% (DGAMDA: 0.9516 vs. 0.9662). Notably, CMV-GLA also demonstrates a substantial advantage over MSF-CNN, with MCC increasing from 0.6935 to 0.8589 and AUC from 0.8952 to 0.9662. These results demonstrate the effectiveness of the proposed multi-view graph fusion and contrastive learning framework compared with both task-specific CNN-based methods and advanced graph neural network models.

Table 3.

Comparison of results from 10-fold cross-validation for advanced methods

Method	Acc	Sen	Spe	Pre	MCC	AUC	AUPRC
MGCNSS	0.9016	0.8916	0.8537	0.8413	0.8237	0.9273	0.8973
DGAMDA	0.8937	0.9053	0.8457	0.8657	0.8352	0.9516	0.9327
AMHMDA	0.8514	0.8562	0.8536	0.8642	0.8249	0.9218	0.9371
MSF-CNN	0.8276	0.8317	0.8341	0.8169	0.6935	0.8952	0.9015
CMV-GLA (Ours)	0.9506	0.9882	0.9132	0.9194	0.8589	0.9662	0.9431

Bold values indicate the best performance among all compared methods for each metric

To further validate whether the observed performance improvements are statistically significant, we conducted paired t-tests on the AUC scores obtained from the 10-fold cross-validation. For each fold, the AUC values of CMV-GLA and each baseline method were paired, and statistical significance was assessed accordingly.

The resulting p-values are presented in Table 4. As shown, all p-values are below 0.01 (and most are below 0.001), indicating that the improvements of CMV-GLA over all compared methods are statistically significant. These results confirm that the superior performance of CMV-GLA is not due to random variation across folds but reflects a consistent and reliable improvement.

Table 4.

Paired t-test p-values for AUC comparisons between CMV-GLA and baseline methods

Method	AUC p-value
MGCNSS
DGAMDA
AMHMDA
MSF-CNN

Impact of negative sampling strategies

In phosphorylation site–disease association prediction, known associations are typically treated as positive samples, while unlabeled pairs are used to construct negative samples. In this work, we adopt random negative sampling as the default setting in all main experiments, as it provides an unbiased evaluation protocol.

To further investigate the impact of different sampling strategies, we additionally compare random sampling with a Euclidean distance-based negative sampling method proposed in [53]. This analysis is intended as a supplementary study to better understand how sampling strategies influence model performance.

Formally, let U denote the set of unlabeled samples and represent the centroid of positive samples in the embedding space. Random negative sampling selects samples uniformly from U:

36

In contrast, Euclidean-based sampling selects samples based on their distance to the positive sample centroid:

37

A sample x is selected as a reliable negative if:

38

where denotes the average distance of unlabeled samples to the centroid. This strategy favors samples that are farther from the positive cluster and are thus less likely to be false negatives.

We evaluated both CMV-GLA and MSF-CNN under these two sampling strategies. The results are shown in Fig. 2. Three observations can be made.

Fig. 2.

Comparison of experimental results under different negative sample selection strategies

First, under random sampling, CMV-GLA significantly outperforms MSF-CNN across all evaluation metrics, demonstrating its robustness under a standard and unbiased evaluation protocol.

Second, both models benefit from Euclidean-based sampling, with noticeable improvements across multiple metrics. This suggests that filtering out potentially ambiguous samples can facilitate model training.

Third, CMV-GLA consistently achieves superior performance over MSF-CNN under both sampling strategies. Notably, the performance gap remains substantial, indicating that the advantages of CMV-GLA are not dependent on a specific sampling method.

Overall, these results show that while more selective sampling strategies can improve performance, CMV-GLA maintains strong and stable performance under both standard and distance-based sampling settings.

Evaluation on imbalanced data

To evaluate CMV-GLA under realistic biological conditions characterized by extreme sparsity, we conducted experiments in which the model was trained on a balanced dataset (1:1 positive-to-negative ratio) and evaluated on test sets with varying imbalance ratios. The test ratios range from 1:1 (balanced) to the natural distribution of the dataset (approximately 1:130), reflecting real-world scenarios where true associations are rare.

The results are summarized in Table 5. As shown, the AUC remains consistently high (above 0.96) across all imbalance levels, indicating that CMV-GLA preserves strong global discriminative ability even under highly skewed data distributions.

Table 5.

CMV-GLA performance under different test set imbalance ratios

Ratio (Pos:Neg)	AUC	AUPRC
1:1	0.9662	0.9431
1:5	0.9647	0.9287
1:10	0.9685	0.8975
1:50	0.9614	0.7056
Original (1:130)	0.9621	0.5275

As expected, the AUPRC decreases as the test set becomes more imbalanced due to the increasing rarity of positive samples. However, it is important to note that AUPRC should be interpreted relative to the baseline defined by the positive class proportion. In the most challenging setting (1:130), the random baseline AUPRC is approximately 0.0077, whereas CMV-GLA achieves an AUPRC of 0.5275, representing a substantial improvement of over 60-fold.

This result demonstrates that CMV-GLA maintains strong precision-recall performance even in extreme sparsity conditions, highlighting its practical applicability for identifying true phosphorylation site–disease associations in real-world high-false-positive-rate environments. This further suggests that CMV-GLA is well-suited for large-scale biological discovery tasks, where the vast majority of candidate associations are negative.

Ablation studies

Phosphorylation site window size

The sequence window size surrounding phosphorylation sites critically determines contextual information availability for the model. To identify the optimal size, we systematically evaluated window lengths from 13 to 35 residues using ProtT5-encoded features within the site similarity network. As shown in Fig. 3:

• Six metrics (Accuracy, Sensitivity, Specificity, Precision, AUC, AUPRC) exhibited minimal fluctuations across sizes.

• The Matthews Correlation Coefficient (MCC) peaked at 0.8569 with 21 residues.

• Concurrently, 21 residues achieved maximum AUC (0.9662) and AUPRC (0.9431)

Given the superior performance across all metrics-particularly the statistically significant MCC improvement-we selected 21 residues as the optimal window size.

Fig. 3.

Comparison of phosphorylation‑site window size

Optimization of sequence encoding methods

To identify the most effective method for encoding phosphorylation site sequences, we rigorously evaluated three distinct approaches: the One-hot encoding utilized in MSF-CNN [54], the widely adopted bioinformatics sequence alignment technique Needleman-Wunsch (NW) [55], and the ProtT5 protein language model employed in this study. The experimental results, summarized in Fig. 4, revealed clear performance differences. Overall, the One-hot encoding method demonstrated the weakest performance, yielding the lowest scores across all evaluation metrics. In contrast, the ProtT5 model consistently and significantly outperformed both the One-hot and NW encoding methods on every metric. Specifically, ProtT5 achieved the highest Matthews Correlation Coefficient (MCC) of 0.8569, compared to 0.8478 for One-hot encoding and 0.8506 for NW alignment. Similarly, for the Area Under the Precision-Recall Curve (AUPRC), ProtT5 reached 0.9431, surpassing One-hot (0.9356) and NW alignment (0.9403). Given this consistent and marked superiority, we selected the ProtT5 protein language model as the primary encoding framework for phosphorylation site sequences in our CMV-GLA approach.

Fig. 4.

Evaluation of phosphorylation‑site sequence encoding methods

Disease semantic similarity metric optimization

To evaluate the effectiveness of our proposed disease semantic similarity metric (), we performed a comparative analysis against the method () employed in MSF-CNN. The MSF-CNN metric is defined by Equation (36):

39

where and represent the MeSH terms for diseases A and B, respectively, and denotes the cardinality of the set.

Experimental results, as depicted in Fig. 5, demonstrate the superior performance of our metric (defined in Eq. 5) compared to . Specifically, achieved a significant increase of +1.43% in the Matthews Correlation Coefficient (MCC) and +1.26% in the Area Under the Precision-Recall Curve (AUC).

Fig. 5.

Assessment of disease‑semantic similarity measures

This consistent improvement validates the key design choices incorporated in over the simpler set overlap approach of :

1. Incorporation of DAG Topology: Unlike which relies solely on term set overlap, explicitly models the hierarchical relationships inherent in the MeSH Directed Acyclic Graph (DAG) structure, capturing richer semantic information.

2. Weighting by Term Specificity (Disease Rarity): Crucially, weights the contribution of shared MeSH terms based on their specificity. Terms associated with rarer diseases are assigned greater importance. This design acknowledges that sharing a specific term indicative of a rare disease phenotype often signifies a stronger and potentially more biologically relevant association than sharing a common term prevalent across many diseases.

The observed performance gains, particularly the improvement in MCC and AUPRC, underscore the importance of considering both the hierarchical structure of disease ontologies and the specificity (or rarity) of shared terms in computing disease semantic similarity. Our metric, by integrating these factors, provides a more nuanced and accurate measure of disease relatedness, ultimately leading to enhanced prediction of phosphorylation site-disease associations.

Comparison with layer aggregation methods

We compare our proposed Global Layer Attention (GLA) module with established layer aggregation techniques such as residual connections and simple layer concatenation. Specifically, we implement two widely used layer aggregation strategies:

• Residual connections (Res con): each GAT layer’s output is added to the output of the previous layer, forming skip connections.

• Simple layer concatenation (Sim con): the outputs of all GAT layers are directly concatenated along the feature dimension, which is analogous to the Jumping Knowledge (JK) network with concatenation mode.

The models were evaluated under the same 10 fold cross validation protocol. The results are summarized in Table 6. As shown, the proposed GLA module outperforms both residual connections (Res con) and simple layer concatenation (Sim con) across all evaluation metrics. Notably, while the performance of the two baseline methods (Res con and Sim con) is quite comparable, with only marginal differences, CMV-GLA (Ours) demonstrates substantial improvements, particularly in MCC (0.8569), where it outperforms Res con (0.7468) and Sim con (0.7083) by significant margins. These results highlight the superior performance of the GLA module in mitigating over-smoothing and preserving discriminative features. GLA’s attention-based aggregation approach proves to be more effective than standard residual or concatenation-based methods, offering a clear advantage in maintaining high-quality feature representation across layers.

Table 6.

Comparison of different layer aggregation strategies

Method	Acc	Sen	Spe	Pre	MCC	AUC	AUPRC
Res_con	0.9243	0.9814	0.8742	0.8794	0.7468	0.9356	0.9174
Sim_con	0.9135	0.9732	0.8679	0.8705	0.7083	0.9287	0.9048
CMV-GLA (Ours)	0.9506	0.9882	0.9132	0.9194	0.8569	0.9662	0.9431

Bold values indicate the best performance among all compared methods for each metric

Component contribution analysis

To quantitatively evaluate the necessity and contribution of each core component within the CMV-GLA model, we systematically designed five variant models for ablation studies:

1. No_ass: Removed the phosphorylation site-disease association network, relying solely on the phosphorylation site similarity network and disease similarity network for feature extraction.

2. No_sim: Removed the phosphorylation site similarity network and disease similarity network, using only the phosphorylation site-disease association network.

3. No_LA: Removed the inter-layer attention mechanism within the Graph Attention Network (GAT).

4. No_MAF: Removed the Multimodal Adaptive Fusion (MAF) module.

5. No_CL: Removed the multimodal contrastive learning (CL) module.

We conducted 10-fold cross-validation for all models, including the full CMV-GLA, under both the random negative sample selection strategy and the Euclidean distance-based negative sample selection strategy (detailed results presented in Table 7 and Table 8). The experimental results clearly demonstrate the critical role of each component in model performance:

Table 7.

Comparison of results from ablation experiments with random selection of negative samples across different models

Method	Acc	Sen	Spe	Pre	MCC	AUC	AUPRC
No_ass	0.8293	0.8872	0.8047	0.8107	0.6125	0.9032	0.8196
No_sim	0.9437	0.9764	0.9247	0.9184	0.8375	0.9582	0.9085
No_LA	0.8562	0.9414	0.7948	0.8076	0.6457	0.9076	0.8969
No_MAF	0.9462	0.9815	0.8986	0.9155	0.8427	0.9562	0.9309
No_CL	0.9459	0.9749	0.9087	0.9176	0.8379	0.9574	0.9378
CMV-GLA (Ours)	0.9506	0.9882	0.9132	0.9194	0.8569	0.9662	0.9431

Bold values indicate the best performance among all compared methods for each metric

Table 8.

Comparison of results from ablation experiments with different models under negative sample selection strategy

Method	Acc	Sen	Spe	Pre	MCC	AUC	AUPRC
No_ass	0.8397	0.9007	0.8094	0.8471	0.6724	0.9272	0.9257
No_sim	0.9546	0.9879	0.9461	0.9349	0.8573	0.9756	0.9679
No_LA	0.9465	0.9905	0.9472	0.9278	0.8356	0.9809	0.9768
No_MAF	0.9578	0.9864	0.9184	0.9372	0.8964	0.9824	0.9816
No_CL	0.9569	0.9846	0.9435	0.9437	0.8849	0.9847	0.9789
CMV-GLA (Ours)	0.9723	0.9929	0.9488	0.9511	0.9023	0.9889	0.9842

Bold values indicate the best performance among all compared methods for each metric

Association Information is Paramount: Regardless of the negative sampling strategy employed, the removal of either the association network (No_ass) or the similarity networks (No_sim) resulted in significant performance degradation. Notably, No_ass exhibited the most severe performance loss, under random negative sampling, its MCC value was only 0.6125, substantially lower than the full model’s 0.8569; under distance-based sampling, MCC was 0.6724 vs. 0.9023. This strongly supports that the association information between phosphorylation sites and diseases is the most critical factor for effective prediction.

Layer Attention Mitigates Over-Smoothing: Removal of the GAT inter-layer attention mechanism (No_LA) caused significant performance drops across all evaluation metrics, MCC decreased to 0.6457 under random negative sampling. This result validates the effectiveness of the layer attention mechanism in mitigating the node feature over-smoothing problem inherent in deep GAT architectures, thereby enabling the learning of more discriminative feature representations.

Contribution of Multimodal Fusion and Contrastive Learning: Removing either the Multimodal Adaptive Fusion module (No_MAF) or the multimodal contrastive learning module (No_CL) also resulted in measurable negative impacts on model performance. The most pronounced effect was observed on the Matthews Correlation Coefficient (MCC), which decreased by approximately 2%-3%, under random negative sampling, No_CL achieved an MCC of 0.8379 vs. 0.8569. Declines in other metrics were relatively smaller. This indicates that both the MAF and CL modules play important positive roles in optimizing the model, particularly for enhancing the MCC metric.

Collectively, across all tested scenarios under both negative sampling strategies, the full CMV-GLA model consistently demonstrated superior predictive performance. These ablation results robustly validate, from multiple perspectives, the indispensable contribution of each meticulously designed core component-leveraging association networks, the inter-layer attention mechanism, multimodal adaptive fusion, and multimodal contrastive learning-to the overall success of the CMV-GLA model.

Case study

To rigorously assess the practical predictive performance and novelty discovery potential of CMV-GLA, as well as its capability under inductive (cold-start) conditions, we conducted independent case studies on three distinct diseases: Alzheimer’s disease (AD), breast neoplasms (breast cancer), and neuroblastoma.

In this experimental setting, each target disease was completely excluded from the training graph by removing all its known phosphorylation site–disease associations prior to model training. Consequently, the model was required to predict associations for a previously unseen disease, thereby simulating a realistic inductive (cold-start) scenario.

We employed a Euclidean distance-based negative sample selection strategy, maintaining a 1:1 positive-to-negative ratio in the training data. For each disease, the top 30 predicted phosphorylation sites, ranked by descending prediction scores, were subjected to a systematic multi-level validation process:

1. Baseline database verification (PhosphoSitePlus and PTMD): Predictions were first cross-checked against the benchmark databases [47] and PTMD [48] used in model construction.

2. Updated database validation (PTMD 2.0): For candidates not validated in the baseline databases, further verification was performed using the updated PTMD 2.0 database (2025) [58] to assess their novelty and reliability.

3. Literature support for parent protein associations: For cases lacking validation in PTMD 2.0, literature mining was conducted to examine whether the parent proteins have documented associations with the target disease, thereby providing supporting biological evidence.

We further compared CMV-GLA with the baseline model MSF-CNN [37]. As reported for MSF-CNN, this model provides detailed predictions for the top 20 phosphorylation sites for Alzheimer’s disease and neuroblastoma. Accordingly, we conducted a comparative analysis between the two models on these diseases. Since results for breast cancer were not reported in MSF-CNN, no comparison was performed for this disease. Detailed results for each disease are presented in the following sections.

Alzheimer’s disease (AD)

AD is a progressive neurodegenerative disorder characterized by cognitive decline, with dysregulated protein phosphorylation-such as aberrant tau phosphorylation-being a central pathological mechanism [56, 57]. The top 30 AD-associated phosphosite predictions from CMV-GLA are summarized in Table 9. We further performed systematic analyses to validate the reliability of these predictions:

1. Baseline database coverage (PhosphoSitePlus & PTMD): Among CMV-GLA’s top 30 predicted phosphosites, 29 were validated within the PhosphoSitePlus [47] and PTMD [48] databases. Notably, the single unvalidated site was ranked 23rd, indicating full validation of the top 20 predictions. In comparison, MSF-CNN [37] reported 1 unvalidated phosphosite within its top 20 predictions.

2. Updated database validation (PTMD 2.0): The only CMV-GLA site unvalidated in the baseline databases (Q16555, T514) was subsequently confirmed in the newly released PTMD 2.0 (2025) [58] database. The site previously unvalidated in MSF-CNN also remained unverified in PTMD.

3. Literature support for parent protein associations: As all CMV-GLA predictions were validated in either the baseline or updated databases, no additional literature mining was necessary.

Collectively, as shown in Table 9, these results demonstrate CMV-GLA’s exceptional reliability in predicting AD-associated phosphosites. Its validation coverage (29/30) surpassed that of MSF-CNN’s top 20 predictions, and its sole initially unvalidated candidate was confirmed by PTMD 2.0.

Table 9.

Top 30 phosphorylation sites associated with alzheimer’s disease

Rank	Site	PSP & PTMD	PTMD 2.0	Literature
1	P10636–8,T217		–	–
2	P10636–8,T181		–	–
3	P49841,S9		–	–
4	P16949,S38		–	–
5	P16220,S133		–	–
6	P10636–8,S396		–	–
7	P12931,Y419		–	–
8	Q13315,S1981		–	–
9	P04049,S259		–	–
10	P10636–8,T231		–	–
11	P10636–8,S214		–	–
12	P10636–8,T205		–	–
13	O00429,S616		–	–
14	P10636–8,S404		–	–
15	P06241,Y420		–	–
16	P10636–8,T212		–	–
17	P10636–8,S202		–	–
18	P10636–8,T153		–	–
19	P10636–8,S199		–	–
20	P10636–8,T175		–	–
21	P04637,S15		–	–
22	P16949,S25		–	–
23	Q16555,T514			-
24	P10636–8,T149		–	–
25	P10636–8,S400		–	–
26	P10636–8,S262		–	–
27	P10636–8,S235		–	–
28	P10636,S713		–	–
29	Q16555,S522		–	–
30	P10636–8,S356		–	–

Bold values indicate the best performance among all compared methods for each metric

Breast neoplasms (breast cancer)

Breast cancer is one of the most prevalent malignancies among women, driven by complex signaling pathways in which protein phosphorylation serves a critical regulatory role [59]. The top 30 breast cancer-associated phosphosite predictions from CMV-GLA are summarized in Table 10. We performed systematic validation analyses to assess the reliability of these predictions:

1. Baseline database coverage (PhosphoSitePlus & PTMD): Of the top 30 predicted phosphosites, 29 were validated in the PhosphoSitePlus [47] and PTMD [48] databases. The single unvalidated site was ranked 25th, indicating complete validation of all top 24 predictions.

2. Updated database validation (PTMD 2.0): The unvalidated site (P27361, T202) was subsequently confirmed in the newly released PTMD 2.0 (2025) [58] database.

3. Literature support for parent protein associations: As all 30 predictions were validated in either the baseline or updated databases, no additional literature mining was necessary.

Collectively, as shown in Table 10, CMV-GLA successfully identified high-confidence breast cancer-associated phosphosites. The model’s high validation rate (96.7%) and the confirmation of its sole initially unvalidated site in PTMD 2.0 strongly highlight its predictive accuracy and its ability to uncover sites later recognized in updated knowledge bases.

Table 10.

Top 30 phosphorylation sites associated with breast cancer

Rank	Site	PSP & PTMD	PTMD 2.0	Literature
1	P01106,S77		–	–
2	P53985,S467		–	–
3	Q92934,S75		–	–
4	P04637,S15		–	–
5	P42345,S2448		–	–
6	Q7KZF4,S426		–	–
7	Q12778,S256		–	–
8	P21796,S104		–	–
9	P01106,T73		–	–
10	P40763,Y705		–	–
11	P31749,S473		–	–
12	Q92934,S99		–	–
13	P29317,S897		–	–
14	P98170,S87		–	–
15	Q96B36,T246		–	–
16	Q15118,T338		–	–
17	Q05397,Y397		–	–
18	P49841,S9		–	–
19	P01106,T58		–	–
20	P06400,S780		–	–
21	P01106,S62		–	–
22	P16949,S38		–	–
23	P16949,S25		–	–
24	P46527,T157		–	–
25	P27361,T202			–
26	P28482,Y187		–	–
27	P04792,S78		–	–
28	Q00613,S326		–	–
29	P60174–3,S58		–	–
30	Q00613,S303		–	–

Bold values indicate the best performance among all compared methods for each metric

Neuroblastoma

Neuroblastoma is a heterogeneous pediatric extracranial solid tumor, with aberrant activation of signaling pathways-often involving phosphorylation-being a key feature [60]. The top 30 neuroblastoma-associated phosphosite predictions from CMV-GLA are summarized in Table 11. We conducted systematic validation analyses as follows:

1. Baseline database coverage (PhosphoSitePlus & PTMD): Of CMV-GLA’s top 30 predictions, 21 phosphosites were validated in PhosphoSitePlus [47] and PTMD [48], leaving nine unvalidated sites (ranks 12, 13, 16, 17, 22, 24, 25, 26, 29). Within the top 20 predictions, 16 were validated and four remained unconfirmed-exactly mirroring the four unvalidated sites reported by MSF-CNN [37] among its top 20.

2. Updated database validation (PTMD 2.0): None of the nine CMV-GLA unvalidated sites nor the four MSF-CNN unvalidated sites were directly confirmed in PTMD 2.0 [58]. However, the parent protein of the site ranked 16 (P05198, S52) is present in PTMD 2.0, providing partial database support for that prediction.

3. Literature support for parent protein associations: For the nine CMV-GLA sites lacking database validation, targeted literature mining identified strong functional evidence for four phosphosites:

• Site P45983, T183 (Parent Protein: MAPK8/JNK1): Phosphorylation at T183 is a critical activation site of MAPK8 (JNK1). Activated JNK1 has been widely implicated in neuroblastoma-related processes, including neuronal differentiation and stress-response signaling. These pathways are closely associated with tumor progression and chemoresistance, suggesting that phosphorylation at T183 is functionally relevant in neuroblastoma [61, 62].

• Site Q07955, Y19 (Parent Protein: SRSF1): SRSF1 is frequently dysregulated in neuroblastoma and plays a key role in alternative splicing regulation. Aberrant splicing mediated by SRSF1 has been linked to tumor progression and angiogenesis. Therefore, phosphorylation at Y19 may influence SRSF1 activity and contribute to splicing-dependent oncogenic processes in neuroblastoma [63].

• Site O60885, Y97 (Parent Protein: BRD4): BRD4 is known to be overexpressed and hyperactivated in neuroblastoma. It is involved in critical oncogenic pathways, including MYCN transcriptional regulation and super-enhancer-driven gene expression. Phosphorylation at Y97 may modulate BRD4 activity, thereby affecting these transcriptional programs and contributing to neuroblastoma development [64, 65].

• Site P35637, S42 (Parent Protein: FUS): FUS is involved in RNA metabolism and has been shown to regulate snoRNA expression and rRNA modification in neuroblastoma cells. These processes are essential for neuronal differentiation and cellular homeostasis. Phosphorylation at S42 may affect FUS-mediated RNA regulatory functions, providing a plausible link to neuroblastoma pathogenesis [66].

Collectively, as shown in Table 11, CMV-GLA achieves an 80% baseline validation rate among its top 20 neuroblastoma predictions-on par with MSF-CNN–but offers clear advantages. First, it yields partial database support via parental protein presence for one additional site. Second, four of the remaining unvalidated sites enjoy direct literature backing, underlining CMV-GLA’s strong potential to uncover both known and novel phosphorylation site–disease associations.

Table 11.

Top 30 phosphorylation sites associated with neuroblastoma

Rank	Site	PSP & PTMD	PTMD 2.0	Literature
1	P49841,S9		–	–
2	P01106,T73		–	–
3	P04637,S15		–	–
4	P31749,S473		–	–
5	P31749,T308		–	–
6	P53778,Y185		–	–
7	Q9H6X2,Y425		–	–
8	P21802,Y656		–	–
9	P08670,Y61		–	–
10	P18077,Y34		–	–
11	Q08174,Y897		–	–
12	P45983,T183		–
13	Q9H9S0,T200		–	–
14	Q9UM73,Y1507		–	–
15	P04637,S20		–	–
16	P05198,S52			–
17	Q07955,Y19		–
18	P35580,Y1415		–	–
19	Q00535,Y15		–	–
20	Q3V6T2,Y1799		–	–
21	Q9BSJ8,Y822		–	–
22	Q16543,S13		–	–
23	Q13443,Y736		–	–
24	P12277,T133		–	–
25	O60885,Y97		–
26	P35637,S42		–
27	P15529–4,Y357		–	–
28	P06213,Y1185		–	–
29	P35611,S726		–	–
30	Q9UPS8,Y296		–	–

Bold values indicate the best performance among all compared methods for each metric

Conclusion

Elucidating the relationships between phosphorylation sites and human diseases remains an important challenge in precision medicine. The rapid growth of phosphoproteomic data highlights the need for effective computational methods to systematically identify phosphorylation site–disease associations. However, the integration of heterogeneous biological data remains a key challenge for improving predictive performance.

To address this issue, we propose CMV-GLA (Contrastive Multi-view Graph Layer Attention network), which integrates protein sequence features, disease semantic information, and known associations into a unified multi-view graph framework. By combining graph attention networks with layer-wise attention and contrastive learning, the model alleviates over-smoothing and enhances the discriminative ability of node representations.

Extensive experiments demonstrate that CMV-GLA consistently outperforms existing methods across multiple evaluation metrics. Ablation studies further verify the effectiveness of multi-view integration and contrastive learning. In addition, case studies on Alzheimer’s disease, breast cancer, and neuroblastoma show that the model can identify biologically meaningful and potentially novel associations, supported by database evidence and literature validation.

Despite these promising results, several limitations remain. First, the current framework does not explicitly incorporate kinase–substrate relationship information. Since phosphorylation events are mediated by specific kinases, incorporating kinase-related data (e.g., kinase recognition motifs or kinase–disease associations) could provide more mechanistic insights and further improve prediction accuracy.

Second, the model primarily relies on static sequence and semantic features, with limited exploration of more complex biological contexts. Future work may consider integrating additional data sources, such as multi-omics information, and developing more comprehensive representations that capture dynamic biological processes.

Overall, CMV-GLA provides an effective framework for phosphorylation site–disease association prediction and offers a useful foundation for future studies integrating richer biological knowledge.

Author contributions

L.J.R. was responsible for the design of algorithms and experiments, and drafted the main manuscript text; O.S.S. and Q.S. performed data analysis; M.Y.X. and G.S.Y. prepared figures and tables; S.X.H. supervised the research and finalized the manuscript. All authors reviewed and approved the final version.

Funding

This research was funded by NSFC (62272192).

Data availability

The secondary datasets used in this study were obtained from: PhosphoSitePlus (Hornbeck et al., 2019; available at: https://www.phosphosite.org/) PTMD – Post-Translational Modification Database (Xu et al., 2018; available at: http://ptmd.biocuckoo.org/) PTMD 2.0 (Ding et al., 2025; available at: https://academic.oup.com/nar/article/53/D1/D554/7779354) MeSH – Medical Subject Headings (U.S. National Library of Medicine; available at: https://www.nlm.nih.gov/mesh/) All code and processed datasets generated in this study are available at: https://github.com/ljr078/CMV-GLA

Declarations

Ethics approval and consent to participate

Not applicable.

Consent to publication

Not applicable.

Competing interests

The authors declare no competing interests.