DeepPTMPred: a multi-modal deep learning framework for accurate prediction of protein post-translational modification sites Authors: Chenkui Wang, Qianhui Jiang, Dan Yu, Jiahui Guan, Zimeng Chen, Xiaoling Lu, Bin Yan, Jing Qin, Yong Liu, Junwen Wang Published: 2026 ABSTRACT AbstractPost-translational modifications are important for regulating cellular functions. Although traditional experimental methods accurately identify PTM sites, they are time-consuming. In this study, we propose a novel model capable of predicting 17 types of PTMs through multi-modal integration and AlphaFold predictions. Our model employs an enhanced CNN-transformer architecture to capture local dependencies within the sequence, while incorporating structural features and evolutionary patterns to effectively capture complex spatial relationships and global contextual dependencies. Through rigorous cross-validation and testing, our model demonstrates exceptional performance, achieving area under the curve scores of 96.5%, 91.6%, 91.0%, and 89.5% for the prediction of hydroxylation, malonylation, O-linked glycosylation, and phosphorylation, respectively. Notably, our model accurately identified known phosphorylation sites on tau and two recently identified residues linked to pre-tangle stages and early Alzheimer’s disease pathology. This work not only deepens the understanding of PTMs but also holds promise for advancing future research in the prediction of PTM sites and functional annotation. Introduction Post-translational modifications (PTMs) are key regulatory mechanisms that dynamically modulate protein function by introducing chemical groups or inducing proteolytic cleavage at specific amino acid residues [1]. These modifications influence protein structure, stability, localization, and molecular interactions, thereby driving essential biological processes like signal transduction, cell cycle regulation, and transcriptional regulation [2–4]. Although hundreds of PTM types have been identified [5], many remain underexplored in terms of their functional roles and regulatory mechanisms. Dysregulation of PTMs has been linked to numerous human diseases. For example, aberrant acetylation and ubiquitination contribute to tumor progression in cancer [6], while defective hydroxylation impairs cardiovascular function [7, 8]. In neurodegenerative diseases like Alzheimer’s disease (AD), hyperphosphorylation of the tau protein is a hallmark of pathology [9]. This hyperphosphorylated tau aggregates into neurofibrillary tangles, disrupting microtubule stability and neuronal function, which are key drivers of cognitive decline [10]. Emerging evidence further pinpoints specific phosphorylation sites such as serine-262 and serine-356 as critical biomarkers of pre-tangle soluble tau assemblies, highlighting their potential in early AD pathology [11]. Identifying PTM sites is therefore essential for understanding disease mechanisms and developing targeted therapeutics. Traditional experimental methods for PTM site identification, such as chromatography and mass spectrometry, offer high specificity and resolution [12, 13]. However, they are often costly, time-consuming, and limited in throughput, particularly for low-abundance modifications [14]. Given the exponential growth of protein sequence data, there is a pressing need for computational methods that can complement experimental strategies by enabling efficient, high-throughput prediction of PTM sites. To meet this need, computational PTM site prediction has evolved significantly, driven by advancements in machine learning. Foundational approaches relied on classical machine learning algorithms, which typically utilize meticulously engineered sequence-derived features. Support vector machines (SVMs), for example, are effectively employed in tools like mLysPTMpred [15] for predicting multiple lysine PTM types. Similarly, GPS 6.0 [16] leverages a combination of sequence-derived features and functional annotations to predict protein kinase-specific phosphorylation sites, demonstrating the utility of feature-based methods. Building on this, ensemble methods, which aggregate predictions from multiple models, gained traction for their improved accuracy and robustness. This strategy is exemplified by tree-based models; iRice-MS [17], for instance, leverages an integrated XGBoost model to identify diverse PTM sites in rice, demonstrating the power of these methods in specialized research. More recently, deep learning has emerged as the dominant paradigm, leveraging complex neural networks to automatically learn hierarchical features from raw protein sequences. Recurrent neural networks (RNNs), well-suited for sequential data, are used by tools like DeepPPSite [18] to predict phosphorylation sites. Other novel architectures have also been explored, such as the capsule networks in CapsNet PTM [19] and the generative adversarial networks (GANs) applied to lysine modifications [20]. A key strategy enhancing deep learning models is the use of pretraining and transfer learning. By pretraining models on vast, unlabeled protein sequence databases, researchers can embed a fundamental understanding of protein language, which is then fine-tuned for specific PTM prediction tasks. This approach is central to the Pretrained ELECTRA model for kinase-specific phosphorylation site prediction [21]. Similarly, PhosTransfer employs deep transfer learning to hierarchically repurpose knowledge from well-annotated kinase groups to specific kinases, demonstrating how domain-specific representations can be effectively adapted for phosphorylation site prediction [22]. Recognizing the limitations of sequence-only models, a growing trend involves integrating diverse biological data sources. PhosIDN exemplifies this by combining protein sequences with protein–protein interaction (PPI) network data, allowing the model to leverage cellular context [23]. Despite these developments, several challenges remain. Many existing models rely on predefined features or single-scale sequence windows, which do not account for the long-distance dependencies and hierarchical contextual interactions inherent in protein sequences. Yet, studies have shown that phosphorylation events, for instance, can be influenced by residues located far from the modification site [24], highlighting the necessity of modeling both local and global sequence contexts. Furthermore, most models underutilize structural information, limiting their capacity to reflect the biochemical and geometric environments surrounding modification sites. To address these limitations, we present DeepPTMPred (Fig. 1), a novel multi-modal deep learning framework that integrates sequence context, evolutionary embeddings from Evolutionary Scale Modeling (ESM-2) [25], and predicted 3D structural features from AlphaFold [26] for PTM site prediction. Our model incorporates multi-scale sequence windows processed by a CNN–transformer hybrid module, enabling the capture of both local motifs and long-distance dependencies. To further enrich the representation, we embed global protein features derived from the ESM model and introduce a dynamic position-aware mechanism that adaptively adjusts attention weights. Beyond sequence information, DeepPTMPred integrates lightweight structural features—including backbone dihedral angles, solvent-accessible surface area (SASA), and predicted local distance difference test (pLDDT) scores—via a structure-aware module, thereby embedding geometric and physicochemical contexts. By fusing multi-scale sequence embeddings, pretrained language model features, and structural descriptors, DeepPTMPred achieves superior performance in PTM site prediction while maintaining biological interpretability. It is important to note that by incorporating structural features derived from AlphaFold, our model prioritizes regions with high structural confidence. Consequently, performance may be reduced for PTMs located in intrinsically disordered regions (IDRs) where accurate structural predictions are unavailable. Framework of DeepPTMPred for PTM site prediction. The model supports 17 types of PTMs predictions and integrates three types of features: (1) multi-scale sequence processing branch: three parallel channels process sequence windows of different sizes (21 amino acids, 33 amino acids, 51 amino acids) centered on the target residue. Each channel consists of (i) embedding layer converting amino acids to 64-dimensional vectors, (ii) learnable positional encoding, (iii) residual convolutional blocks for local pattern extraction, (iv) transformer encoder with MHSA for capturing long-range dependencies, and (v) attention pooling for generating fixed-size representations. (2) Evolutionary context branch: ESM-2 protein language model embeddings (1280-dimensional) provide evolutionary conservation information and global sequence context, processed through fully connected layers for dimensionality reduction. (3) Structural context branch: six structural descriptors (SASA, phi/psi dihedral angles, secondary structure, local and average pLDDT scores) capture the local physicochemical environment, encoded through dedicated neural layers. The processed features from all three modalities are concatenated and passed through an MLP with Swish activations, batch normalization, and dropout regularization, finally producing probability for a specific PTM type through a sigmoid output layer. Overall, DeepPTMPred establishes a robust and interpretable framework for predicting PTM sites. By integrating diverse feature modalities, our approach provides a comprehensive view of the sequence and structural determinants of PTMs, offering a valuable tool for advancing protein function annotation and disease-related biomarker discovery. Material and methods Dataset The primary dataset used to build the DeepPTMPred model was curated by the PTMPredictionMetric group and compiled mainly from the dbPTM database [27]. It contains various PTMs of human proteins, mainly focusing on phosphorylation (S/T), acetylation (K), and ubiquitination (K). The dataset was downloaded from the project’s official GitHub repository (https://github.com/meilerlab/PTMPrediction/tree/main/data) on January 15, 2025. Regarding data quality control, the source authors applied structural filtering (AlphaFold predicted pLDDT >50) and redundancy removal (CD-HIT at 90% sequence identity) to the majority of PTM types. However, as detailed in Supplementary Note S1, distinct curation protocols were applied to N-linked glycosylation and deamidation. Specifically, N-linked glycosylation sites were filtered based on glycan visibility in experimental structures rather than pLDDT scores, while deamidation data was included without CD-HIT or AlphaFold-based filtering due to data availability constraints. To further improve data quality, we carried out additional cleaning steps. We verified that the PTM-specific amino acid labels align with expected annotations. In addition, we used the ESM-2 model [28] to generate residue-level embedding features. Due to its limitation of processing protein sequences up to 1024 residues, we generated fixed-length embeddings of 1024 residues. Sequences were centered on the PTM-associated residue, with shorter sequences padded with ‘X’ characters on the left and/or right, and longer sequences truncated to 1024 residues. We assessed the coverage of valid ESM-2 features, filtering out invalid samples as needed. The final distribution of positive and negative sites remaining after these combined filtering steps is summarized in Supplementary Table S1 and Supplementary Fig. S1. Framework of the proposed model To predict protein modification sites, we developed a multi-modal deep learning model that integrated three distinct types of information: local sequence context, global evolutionary information, and predicted structural properties. The overall architecture was designed to process each feature type through specialized modules before fusing them for a final prediction. The complete workflow, from data preparation to feature fusion and classification, is outlined in Supplementary Algorithm S1. Sequence local feature To capture contextual information at different scales around the potential modification site, we employed a multi-scale approach using three parallel processing channels. Each channel analyzed a sliding window of a different size: 21, 33, and 51 amino acids. These windows corresponded to sequences encompassing ±10, ±16, and ± 25 amino acid residues, respectively, centered on the residue of interest. For a given window of length \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $L\in \left[\mathrm{21,33,51}\right]$\end{document}, the amino acid sequence was first converted into a numerical representation. The 20 standard amino acids were mapped to integers 0–19, and a special padding character (*), used for handling sequence boundaries, was encoded as 20 [29]. This numerical sequence was then transformed into one-hot vectors, resulting in an input matrix X ∈ RL × 21. This sparse matrix was projected into a continuous, low-dimensional space using an embedding layer: where E is the amino acid embedding matrix, and the embedding dimension dembed is set to 64. To equip the model with information regarding the relative positions of amino acids within the window, we incorporated a learnable position embedding [30]. This allows the model to place greater emphasis on the central modification site. The position embedding layer generated a unique vector for each position i in the sequence, which was adaptive to different window lengths: The final feature representation for the local sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${E}_p$\end{document} is obtained by summing the amino acid and position embeddings: This resulting matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${E}_p\in{{R}}^{{L}\times{{d}}_{{embed}}}\kern0em$\end{document} serves as the comprehensive local sequence representation. To effectively capture local contextual features from the input sequence, we designed a deep residual convolutional block. This architecture facilitates the training of deeper networks by improving gradient flow and mitigating the vanishing gradient problem [31, 32]. The block contains a main path with two 1D convolutional layers and a parallel shortcut connection. The input to the block, the positional embedding sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${E}_p\in{{R}}^{{L}\times{{d}}_{{embed}}}$\end{document} was processed through the main path as follows: First convolutional layer: Second convolutional layer: Here, padding = same was used in the Conv1D layers to ensure the output sequence length remains L. Batch normalization was applied post-activation to stabilize training dynamics. We used spatial dropout with a rate of P = .3, which regularized the model by dropping entire 1D feature maps, thus preventing co-adaptation between adjacent features. To preserve the original information, a parallel shortcut connection was used. The original input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${E}_p$\end{document} was passed through a linear projection, a 1D convolution layer with a kernel of size 1, which creates a learnable projection that allows for more flexible feature integration. Then the output of the main path \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${H}_2$\end{document} is added to the output of the shortcut connection to produce the block’s output, Z. Transformer-based sequence encoder. The local sequence embedding matrix Z was fed into a transformer-based encoder to model short and long-range dependencies between amino acids [33]. Our encoder consists of two stacked Transformer blocks. Each contains a multi-head self-attention (MHSA) block followed by a feed-forward network (FFN). Both sub-layers incorporate residual connections to facilitate stable training of the deep architecture. The operations within a Transformer block are defined as follows: The input Z was linearly projected into query (Q), key (K), and value (V) matrices using learnable linear transformations: The attention mechanism was computed as follows: We used a multi-head configuration with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${num}_{heads}$\end{document} = 4, where the dimension of each head’s key, dk​, was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${d}_{embed}/{num}_{heads}$\end{document} = 16. The MHSA output was passed through a dropout layer, and a residual connection was applied: The FFN applies a non-linear transformation using the Swish activation function to extract higher-level abstract features [34]: The output of the transformer block was the following: Finally, to convert variable-length sequence information into a fixed-size representation, an attention pooling layer was applied on the output of the Transformer block. Specifically, each residue representation was assigned an attention score, normalized by a softmax function [35], and the final local feature vector h was computed as a weighted sum of the position-wise outputs: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${W}_a\ and\ {W}_b$\end{document} were learnable weight matrices and α was the attention weight vector. Global evolutionary features. To complement the local sequence information, we incorporated global evolutionary features derived from the ESM-2 protein language model [25]. For each protein sequence, we used the pre-trained ESM-2 model to extract a 1280-dimensional feature vector for every residue. These embeddings, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${E}_{ESM}\in{\mathrm{R}}^{L\times 1280}$\end{document}, capture evolutionary conservation patterns and global semantic context learned from millions of diverse protein sequences. Structural features. The model’s input was further enhanced with structural information offered by AlphaFold database and processed by PyRosetta [26, 36]. For each central residue, a set of structural descriptors was extracted: SASA, dihedral angles (phi and psi), secondary structure that was represented as a 3-dimensional one-hot vector (helix, strand, coil), average pLDDT score, and local pLDDT score. By combining these descriptors, each central residue was converted into an 8-dimensional feature vector, denoted as F ∈ ℝ8. These features provide insights into the residue’s local chemical and physical environment. Feature fusion and prediction. The features from the three modalities were processed and fused before being passed to the final classifier. The 1280-dimensional ESM feature vector corresponding to the central residue was processed by a fully-connected layer, reducing its dimensionality to 256. The 8-dimensional structural feature vector was likewise transformed into a 128-dimensional representation. The fixed-size local context vector (64-dim) was then concatenated with the processed global and structural feature vectors. This concatenated multi-modal representation was passed to a prediction head, which was a multi-layer perceptron (MLP) with Swish activation functions and dropout layers for regularization. The output of the MLP was the probability of the site being modified. Class imbalance and loss function To address the imbalance of minority classes, we applied the Synthetic Minority Oversampling Technique (SMOTE) algorithm with adaptive sampling to the training set [37]. If the number of positive samples was below 2000, the target was set to 2000; otherwise, the original sample size was retained. To address the severe class imbalance problem and improve the model’s focus on hard-to-classify samples, we employed the focal loss function [38]. The focal loss is designed to address class imbalance by down-weighting easy examples and focusing training on hard-to-predict negative samples. The loss function is defined as follows: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${p}_t$\end{document} represents the model’s estimated probability for the true class, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${\alpha}_t$\end{document} = 0.25 is a balancing factor that adjusts the importance of positive/negative samples, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathrm{\gamma}$\end{document} = 2.0 is the focusing parameter that reduces the relative loss for well-classified samples. This formulation automatically handles class imbalance without requiring explicit class weight calculations and has demonstrated superior performance in highly imbalanced classification tasks. Model training and experimental setup The dataset was split into training and test set at a 4:1 ratio, maintaining the proportion of positive and negative modification site labels using stratified sampling. The training set was then subjected to stratified 10-fold cross-validation to guarantee robust performance evaluation while maintaining class distribution in each fold. Specifically, the training set was divided into 10 equal subsets, with one subset used as the validation set and the remaining nine combined as the training subset for each of the 10 rounds. We adopted a type-specific training approach, isolating samples for each PTM type based on dataset labels. For any given modification, the dataset was filtered to retain only the samples corresponding to that specific PTM type, which were then used to train a dedicated prediction model. This strategy ensures that each model is specialized for the unique characteristics of a single PTM type. Furthermore, to enhance the practical utility and robustness of our predictions, we designed a task-adaptive dynamic thresholding mechanism, moving beyond the limitations of a conventional fixed threshold. Recognizing that different PTMs may have different priorities in biological applications, this strategy allows for flexible adjustment of the decision boundary. The optimal threshold for each dedicated PTM model was dynamically determined through a data-driven process on the validation set. Specifically, we generated a series of candidate thresholds ranging from 0.1 to 0.9 and selected the one that best fulfilled a predefined performance objective tailored to the PTM’s biological priority. For PTM types requiring high specificity (e.g. phosphorylation), the threshold was chosen to deliver the highest possible true positive rate (TPR) while strictly constraining the false positive rate (FPR) to ≤0.18. For other PTMs where a balanced performance is desired, the threshold was selected to maximize the F1 score. This automated procedure ensures that the final decision boundary is precisely calibrated to each model’s output distribution and the specific practical requirements of the PTM task, thereby significantly improving predictive reliability (Supplementary Note S2). This sophisticated thresholding strategy, combined with our type-specific training, significantly improves the model’s performance and reliability. We trained our model for 300 epochs per fold in 10-fold cross-validation with an initial learning rate of 0.0005. A learning rate scheduler was applied, reducing the rate by 0.95 every epoch to promote optimal convergence and model fitting. The AdamW optimizer with a weight decay of 0.0001 for L2 regularization was employed to enhance model generalization, while an early stopping mechanism was implemented by halting the training process if no improvement in accuracy was observed after 40 epochs. Additionally, the dropout rate was set to 0.3 during training, minimizing the model’s reliance on specific features and further mitigating overfitting. A comprehensive list of all model parameters and their selection basis, including embedding dimensions, transformer layers, and optimization settings, is available in Supplementary Table S3. Furthermore, a parameter sensitivity analysis was conducted to validate the choice of key hyperparameters, with results detailed in Supplementary Table S4. Our experimental setup was developed using Python, with the PyTorch framework used to construct and train the model [39, 40]. The training process was carried out on RTX 4090D GPUs. A detailed summary of the runtime and computational resource utilization for data preprocessing and model training is provided in Supplementary Table S2. Evaluation metrics To evaluate the performance of the model, we selected area under the curve (AUC), Matthews correlation coefficient (MCC), F1 score, recall, FPR, and false negative rate (FNR) as the key evaluation metrics [41], they are defined as follows: AUC: MCC: F1 score: Recall (sensitivity/TPR): FPR: FNR: where TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively. These metrics evaluate the model’s performance from different angles, particularly when dealing with imbalanced datasets. Results Performance comparison For most PTM types, the proposed model demonstrates substantial improvements over the baseline model and other competing models (Table 1). To ensure a fair comparison, all competing methods were evaluated using their officially recommended default parameters and pre-trained weights. No fine-tuning was performed on competing models to simulate their standard usage scenario. Notably, for hydroxylation (P), the proposed model achieves a recall of 0.927, F1 of 0.876, MCC of 0.799, and AUC of 0.965, compared to the baseline’s MCC of 0.75 and AUC of 0.88, with reduced FP (0.116) and FN (0.072) rates. Similarly, for malonylation (K), the proposed model yields an MCC of 0.645 and AUC of 0.916, markedly outperforming the baseline’s MCC of 0.30 and AUC of 0.65. Furthermore, DeepPTMPred consistently outperformed other recent methods on PTM types where multiple predictors were available. For instance, it achieves superior performance in ubiquitination (K) and succinylation (K), with MCCs of 0.552 and 0.549 and AUCs of 0.864 and 0.859, respectively, surpassing MusiteDeep [42] and MTPrompt-PTM [43]. Comparable improvements are seen for acetylation (K), sumoylation (K), and O-linked glycosylation (S/T), where DeepPTMPred maintains top scores across MCC and AUC metrics, reinforcing its robustness and generalizability across diverse PTM categories. Table 1Performance comparison of PTMPrediction and ours across various PTM types.PTMAA typeMethodRecallF1MCCAUCFPRFNRHydroxylationPPTMPrediction——0.7500.8800.1200.120 DeepPTMPred 0.927 0.876 0.799 0.965 0.116 0.072 γ-carboxyglutamic-acidEPTMPrediction——0.6700.8300.170 0.170 DeepPTMPred 0.797 0.828 0.695 0.906 0.112 0.202Lys methylationKPTMPrediction——0.4100.7000.2200.370 DeepPTMPred 0.831 0.761 0.610 0.899 0.203 0.168 MalonylationKPTMPrediction——0.3000.6500.3400.360 DeepPTMPred 0.862 0.779 0.645 0.916 0.195 0.137 Arg methylationRPTMPrediction——0.4000.710 0.160 0.430 DeepPTMPred 0.840 0.660 0.562 0.909 0.198 0.159 CrotonylationKPTMPrediction—— 0.490 0.7400.440 0.080 DeepPTMPred 0.491 0.543 0.195 0.779 0.250 0.509UbiquitinationKPTMPrediction——0.2600.6300.3700.370MusiteDeep0.7010.3570.4180.6040.3210.299MTPrompt-PTM0.6820.7530.5420.7260.2840.318 DeepPTMPred 0.756 0.797 0.552 0.864 0.197 0.243 SuccinylationKPTMPrediction——0.2200.6100.3900.400MTPrompt-PTM0.7150.7520.4970.8100.2640.285 DeepPTMPred 0.748 0.767 0.549 0.859 0.201 0.251 GlutathionylationCPTMPrediction——0.2000.6000.4400.360 DeepPTMPred 0.795 0.752 0.578 0.874 0.208 0.204 SumoylationKPTMPrediction——0.3000.6600.2200.450MTPrompt-PTM0.7820.5810.4790.8480.2040.219 DeepPTMPred 0.811 0.604 0.517 0.887 0.191 0.189 S-nitrosylationCPTMPrediction——0.1700.5800.3400.500 DeepPTMPred 0.719 0.714 0.517 0.846 0.201 0.280 AcetylationKPTMPrediction——0.2200.6200.4900.280MTPrompt-PTM0.7610.8320.4760.8630.2280.239 DeepPTMPred 0.773 0.845 0.502 0.872 0.197 0.226 O-linked glycosylationS/TPTMPrediction——0.2700.7200.2000.350MTPrompt-PTM0.7950.3640.3710.8740.2790.205 DeepPTMPred 0.850 0.376 0.388 0.910 0.188 0.149 PhosphorylationS/TPTMPrediction——0.2400.720 0.160 0.390NetPhos3.10.7120.1550.1820.530.3250.288transphos0.6300.2150.2500.7800.2500.370Attenphos0.7700.2310.2600.8000.2200.230MusiteDeep0.7520.2520.2940.8210.2540.248MTPrompt-PTM0.7740.2370.2710.8190.2410.226 DeepPTMPred 0.827 0.263 0.305 0.895 0.196 0.172 GlutarylationKPTMPrediction——0.1800.6000.450 0.350 DeepPTMPred 0.570 0.589 0.379 0.769 0.199 0.429CitrullinationRPTMPrediction——0.1200.6100.2000.580 DeepPTMPred 0.540 0.287 0.253 0.778 0.153 0.460 N-linked glycosylationNPTMPrediction——0.2000.620 0.500 0.250 DeepPTMPred 0.592 0.725 0.251 0.616 0.227 0.409 Note: The best results are highlighted in bold. For evaluation metrics, higher values indicate better model performance in recall, F1 score, MCC, and AUC, while lower values indicate better performance in FPR and FNR. Performance comparison of PTMPrediction and ours across various PTM types. Note: The best results are highlighted in bold. For evaluation metrics, higher values indicate better model performance in recall, F1 score, MCC, and AUC, while lower values indicate better performance in FPR and FNR. The performance advantage was particularly evident for the widely studied phosphorylation (S/T) task, which included the largest number of competing methods. DeepPTMPred (recall 0.827, MCC 0.305, AUC 0.895) clearly outperformed all other predictors. It significantly surpassed the baseline PTMPrediction (MCC 0.240, AUC 0.720) and other established tools, including NetPhos3.1 [44] (MCC 0.182, AUC 0.530), transphos [45] (MCC 0.250, AUC 0.780), Attenphos [46] (MCC 0.260, AUC 0.800), MusiteDeep (MCC 0.294, AUC 0.821), and MTPrompt-PTM (MCC 0.271, AUC 0.819). However, performance varies for certain PTMs. For N-linked glycosylation, the proposed model shows an AUC of 0.616. Unlike other PTMs, N-linked glycosylation occurs exclusively within the conserved motif [47], and the curated dataset for this PTM type consists of positive and negative samples with highly similar local sequence motifs, focusing predictions on whether asparagine residues within these sequons are modified, rather than evaluating all asparagines. Thus, the limited diversity of negative samples may further prevent the model from learning discriminative features. Overall, the proposed model demonstrates superior predictive performance across most PTM types, with higher MCC and AUC values and lower error rates, highlighting its robustness and generalizability in PTM prediction tasks. Ablation experiments To validate the effectiveness of the proposed model and the contribution of each module to the overall performance, we conducted a systematic ablation experiment. The experiments covered three common PTM tasks: phosphorylation, ubiquitination, and acetylation, and we performed comparative analyses across multiple evaluation metrics, including MCC, AUC, recall, FPR, and FNR. The results are shown in Fig. 2. Ablation studies on the performance of three PTM types. Performance metrics (A: AUC, recall, B: MCC, FNR, and FPR) are shown for different model configurations tested on phosphorylation, ubiquitination, and acetylation prediction tasks. Each configuration reflects the removal (‘w/o’) of specific components: Window sizes (21 and 33), structural features, ESM embeddings, both ESM + structure, and the MHSA. The comparison demonstrates the contribution of individual components to overall model performance across PTM types. The multi-scale window approach proves crucial for model performance, as evidenced by the degradation observed when removing the 21 and 33 window sizes while retaining only the 51-window configuration. This ablation resulted in MCC drops of 1.0%, 2.2%, and 5.1% across the three tasks, with acetylation showing the most substantial decline. The corresponding FPRs increased from 0.173 to 0.182 for phosphorylation, 0.273 to 0.296 for ubiquitination, and 0.256 to 0.278 for acetylation, while FNRs rose from 0.173 to 0.203, 0.198 to 0.216, and 0.193 to 0.215, respectively, suggesting that the smaller window sizes capture important local sequence patterns that complement the broader contextual information from larger windows. The ESM features emerge as the most critical component across all tasks. Removing ESM features alone caused the most severe performance degradation, with MCC reductions of 2.9%, 8.6%, and 5.8% for phosphorylation, ubiquitination, and acetylation, respectively. The critical role of ESM features stems from their encapsulation of evolutionary constraints and long-range context, which are fundamental for defining functional sites. Notably, ubiquitination showed the largest sensitivity to ESM removal, with recall dropping from 0.802 to 0.600 and the FNR increasing from 0.198 to 0.400. This substantial impact underscores the importance of evolutionary information for PTM site prediction. Structural information contributes moderately but consistently across all tasks, with its removal leading to MCC decreases of 0.1%, 6.4%, and 7.2%. Interestingly, phosphorylation prediction showed minimal sensitivity to structural features, while acetylation and ubiquitination demonstrated greater dependence. The combined removal of both ESM and structural features produced compounding effects, particularly evident in phosphorylation where MCC dropped to 0.217, representing the largest degradation among the ablation settings for this task. Structural features provide a direct, biophysical basis for prediction by characterizing the local solvent accessibility and conformational flexibility of a residue. This suggests synergistic interactions between evolutionary and structural information in the model’s prediction process. We also evaluated the contribution of the MHSA mechanism. We found that ubiquitination and acetylation showed greater sensitivity to MHSA removal, with MCC drops of 8.3% and 8.2%, respectively, compared to phosphorylation, which showed only a 0.3% decrease. The attention mechanism’s variable importance across different PTM types likely reflects the distinct sequence patterns and long-range dependencies characteristic of each modification type. Overall, the ablation results validate our architecture and demonstrate that each component contributes essential information for accurate PTM prediction, with ESM features being the most critical and the remaining modules showing task-dependent importance. Model interpretation and feature importance analysis In this study, we employed SHAP (SHapley Additive exPlanations) to interpret the impact of individual features and better understand the key factors influencing the model’s predictions. Figure 3A illustrates the normalized contributions of each feature category, with ESM features exerting the most significant impact, accounting for 39.4% of the overall feature importance, followed by window 21 and structural features, contributing 21.7% and 21.3%, respectively. This result suggests that ESM features, which capture rich evolutionary information and sequence context, play a crucial role in PTM prediction. Feature importance and embedding visualization for PTM prediction. (A) Normalized contribution of feature categories. (B) Mean SHAP values of the top 20 features contributing most to the model predictions. (C and D) t-SNE visualization of protein embeddings before and after DeepPTMPred processing: (C) raw ESM embeddings of phosphorylation data and (D) learned embeddings after DeepPTMPred transformation. The visualization demonstrates how the model transforms the feature space to better separate PTM sites (positive samples) from non-PTM sites (negative samples). To further explore the specific features contributing to the model predictions, Fig. 3B highlights the top 20 features based on their mean SHAP values. Among these, win0_pos11, win0_pos7, and win0_pos8 exhibit the highest SHAP values, reflecting their significant impact on the model’s outputs. Notably, the prominence of these flanking positions over the central modification site (win0_pos10) aligns with the biological reality that PTM specificity is primarily determined by enzyme recognition motifs in the surrounding sequence context, with critical positions such as −3, −2, and +1 playing key roles in substrate-binding [48]. Additionally, pLDDT and esm_1160 provide complementary information, reinforcing the models’ predictive performance. To evaluate our model’s efficacy in feature representation, we used t-SNE (t-distributed stochastic neighbor embedding) method to visualize the learned representations of positive and negative samples [49]. As shown in Fig. 3C, the raw sequence embeddings of phosphorylation samples exhibit considerable overlap and clustering heterogeneity, with positive and negative instances appearing intermixed in the embedding space. In contrast, following processing through our proposed model, a marked separation emerges between positive and negative samples, with distinct clustering patterns becoming evident (Fig. 3D). This enhanced discriminative capacity in the learned feature space demonstrates that our model successfully captures salient biochemical patterns that differentiate phosphorylated from non-phosphorylated sites. The improved clustering coherence and inter-class separability directly correlate with the observed enhancement in classification performance, validating the model’s capacity to capture informative patterns for phosphorylation site prediction. Prediction of phosphorylation sites in soluble tau protein aggregates in Alzheimer’s disease In AD, neurofibrillary tangles are one of the most important pathological features, primarily formed by the abnormal aggregation of tau protein [10]. Under normal conditions, tau protein, a microtubule-associated protein, supports neuronal microtubule stability and facilitates material transport [50]. In AD, tau protein undergoes excessive and sustained phosphorylation, leading to its detachment from microtubules and forming toxic neurofibrillary tangles [51, 52]. These tangles disrupt the transport system and structural stability of neurons, representing one of the key molecular mechanisms driving disease progression [53]. To validate our model’s performance on a clinically significant protein not present in the training data, we predicted phosphorylation sites on the human tau protein. The results, presented in a multi-layered analysis in Fig. 4, confirm the model’s generalizability and provide significant biological insights. The model first demonstrated its baseline accuracy by successfully identifying numerous traditional phosphorylation biomarkers, such as p-tau181 and p-tau231 [54, 55], with high confidence scores (Fig. 4A). Multi-layered computational analysis and validation of predicted phosphorylation sites on the human tau protein. (A) Phosphorylation probabilities predicted by DeepPTMPred for each potential site across the 441-amino-acid tau sequence. Known AD biomarkers and novel high-confidence predictions, including S262 and S356, are indicated by distinct marker categories. The dashed line represents the confidence threshold. (B) Schematic representation of the functional domains of the human tau protein (2N4R isoform), aligning with the sequence positions in panels A and C. (C) Evolutionary conservation profile of the tau protein obtained from ConSurf [56]. The y-axis score is inverted for visualization, and higher peaks indicate more highly conserved residues. The locations of key biomarkers and the novel predicted sites are marked, highlighting that S262 and S356 lie within prominent, highly conserved regions. As a significant validation of the model’s predictive accuracy, it also assigned high phosphorylation probabilities to residues S262 and S356, a finding in strong concordance with recent experimental work by Islam et al. [11], establishing these sites as novel biomarkers for early-stage AD. The importance of this computational prediction is reinforced by our multi-layered analysis. First, both sites are situated directly within the microtubule-binding repeats (MTBRs), the core region for both microtubule stabilization and pathological tangle formation (Fig. 4B). Second, evolutionary analysis reveals that S262 and S356 are located within highly conserved regions, corresponding to prominent peaks in the conservation score [56] (Fig. 4C). This convergence of a high-confidence prediction, a critical functional location, and strong evolutionary conservation highlights our model’s ability to accurately identify key pathological PTMs, marking these sites as robust candidates for further investigation as early diagnostic biomarkers. To further validate the model’s performance, we cross-referenced our predictions with the comprehensive PhosphoSitePlus database. This comparison confirmed that all of our high-confidence predictions, including the canonical biomarkers (p-tau181, p-tau231) and the novel early-stage sites (S262, S356), are experimentally validated entries in the database. However, this analysis also revealed that the model did not assign high-confidence scores to several other known phosphorylation sites cataloged in PhosphoSitePlus (e.g. T175, S185, S208, S238, and S305). This finding is an expected and insightful consequence of our model’s specialized training. As detailed in the Methods, we trained a model highly specialized in identifying PTMs where local 3D structure is a key predictive feature. Since tau is a largely intrinsically disordered protein (IDP), many of its known phosphorylation sites, including those listed above, fall into this disordered category. Therefore, the model’s lower confidence in these sites is not a failure, but rather a validation of its specialization. It correctly identifies that these sites lack the specific structural-context features it was trained to weigh heavily. Conversely, the model’s high success on canonical and novel biomarker sites (T181, T231, S262, S356) suggests these sites do possess structural or evolutionary features that align with the model’s learned patterns. This analysis highlights a critical trade-off: our model is optimized for high precision in structure-dependent PTM prediction, while PTMs in highly disordered contexts (a distinct biological challenge) represent a different predictive domain. Prediction of phosphorylation sites in AKT1 RAC-alpha serine/threonine-protein kinase (AKT1), a central node in PI3K/AKT signaling pathway, is pivotal in regulating diverse cellular processes, including cell survival, metabolism, proliferation, growth, and angiogenesis [57, 58]. Overexpression or hyperactivation of AKT1 is frequently observed in various cancers, including breast, prostate, lung, pancreatic, liver, ovarian, and colorectal malignancies, making it an important therapeutic target [59]. To demonstrate the practical utility and generalizability of our model, we applied it to AKT1, a critical oncogene that was not included in our training set for phosphorylation data. The model’s performance was evaluated using a multi-layered validation approach (Fig. 5). The overall domain structure of AKT1, including its PH, protein kinase, and AGC-kinase C-terminal domains, is illustrated in Fig. 5C, providing a structural context for the predicted sites. Predictions were first benchmarked against experimentally validated phosphorylation sites from PhosphoSitePlus [60]. Multi-layered computational analysis and validation of predicted phosphorylation sites on the human protein kinase AKT1. (A) Close-up view of the allosteric inhibitor binding pocket showing key protein-ligand interactions. The AKT protein is shown as cartoon representation, with the allosteric inhibitor displayed as sticks. Critical binding residues are highlighted as sticks. Hydrogen bonds are depicted as dashed lines. Water molecules involved in binding are shown as spheres. Structure was visualized using PyMOL based on PDB entry 3O96. (B) Phosphorylation probabilities for each potential site on the AKT1 sequence as predicted by DeepPTMPred. Sites are categorized as true positives, putative novel sites, or low-probability predictions. The dashed line represents the high-confidence threshold of 0.6. (C) Schematic representation of the functional domains of the human AKT1 protein, including the PH domain, linker, protein kinase domain, and the AGC-kinase C-terminal domain. (D) Locations of pathogenic missense mutations from the ClinVar database are mapped onto the AKT1 sequence. Shaded windows indicate the biologically justified ±6 amino acid regions around putative novel sites that successfully capture a known pathogenic variant. The model demonstrated high accuracy by identifying the majority of true positive sites (Fig. 5B). Notably, these correctly identified sites include those with canonical roles in cancer: T308 and S473 are activation sites whose phosphorylation drives oncogenic processes [61]. S129 phosphorylation promotes AKT1 hyperactivation [62], and T450 phosphorylation is crucial for protein stability [63]. Beyond this validation, the model identified 13 high-probability sites not yet cataloged in the database, hereafter referred to as putative novel sites. To investigate the biological significance of these 13 putative novel sites, we performed a comprehensive validation analysis. First, an evolutionary conservation analysis revealed that sites T195 and T312 received the highest possible conservation grade (conservation scores of 7 and 9, respectively), suggesting a preserved functional role. Second, we found that S205 and K268 interact with an allosteric inhibitor of AKT1 (PDB: 3O96), while S205 and the proximal S266 are among our predicted putative sites, suggesting a potential regulatory function (Fig. 5A). Next, we cross-referenced the sites with the ClinVar [64] database of pathogenic mutations. This showed that a distinct set of three novel sites (S21, S431, and S433) are located within a ±6 amino acid window (which approximates the length of a typical kinase recognition motif) of known pathogenic variants, linking their positions to established clinical pathology (Fig. 5D). In total, these independent lines of evidence provide strong support for 7 of the 13 (53.8%) putative novel sites. The remaining putative sites are supported by our model’s high-confidence prediction score alone. While they lack additional validation from this analysis, they may represent functionally important sites in less-conserved regions or reflect novel disease mechanisms not yet captured in clinical databases. Overall, our model’s ability to identify a large cohort of novel PTM candidates—the majority of which are substantiated by evolutionary and/or clinical genetic evidence—underscores its utility as a powerful tool for identifying uncharacterized but functionally critical PTMs. Discussion In this study, we developed a highly accurate predictor for PTM sites, DeepPTMPred, by applying a multi-modal feature fusion approach. The results showed that DeepPTMPred achieves exceptional performance, with AUC scores surpassing 91% for malonylation and O-linked glycosylation and reaching 96.5% for Hydroxylation. These findings are consistent with our initial hypothesis that integrating sequence, structural, and evolutionary information provides a more comprehensive biological context for prediction than single-modality methods. Our model advances the field by demonstrating a sophisticated synergy between a CNN-transformer’s ability to capture local syntax and the global context provided by structural and evolutionary features. This integrated approach allows our model to outperform specialized predictors which, while effective, often neglect the interplay between different biological data layers. Our findings hold significant implications for both computational biology and translational medicine, as they offer a new, more holistic strategy for identifying functionally critical PTMs. This is particularly evident in our model’s ability to validate known phosphorylation sites on the tau protein and identify novel sites linked to early AD pathology, potentially accelerating research into neurodegenerative diseases. However, we acknowledge existing limitations. First, the model’s performance is inherently dependent on the accuracy of the input data. For example, if the predicted protein structures are inaccurate, these structural inaccuracies can propagate as prediction errors. Second, the predictive power for rare PTM types (e.g. citrullination, crotonylation) remains constrained by the scarcity of high-quality training data, which could hinder the broader applicability of our findings across the full range of PTM types. Furthermore, while our model integrates sequence, structure, and evolutionary information, it does not yet incorporate other critical contextual data, such as protein–protein interaction networks, which could provide crucial insights into the regulatory crosstalk between PTMs. Future research should focus on incorporating uncertainty estimation for structural inputs and exploring advanced computational techniques like few-shot or transfer learning to mitigate data scarcity issues. Additionally, future iterations could integrate other data modalities, such as protein–protein interaction networks and subcellular localization data, to better understand the complex regulatory crosstalk between different PTMs and further advance the development of predictive proteomics. In summary, this study has revealed the profound efficacy of multi-modal feature fusion for PTM site prediction and has emphasized that the synthesis of diverse data types is key to deciphering complex biological systems. These findings not only enrich the computational toolkit available to proteomic researchers but also provide clear guidance for the future development of next-generation predictive models. Key Points We propose DeepPTMPred, a novel multi-modal deep learning framework that employs an enhanced CNN-transformer architecture to predict protein post-translational modification sites by integrating multi-scale sequence features, evolutionary embeddings from ESM-2, and structural properties from AlphaFold, with feature fusion and adaptive thresholding to handle class imbalance and improve prediction accuracy. We propose DeepPTMPred, a novel multi-modal deep learning framework that employs an enhanced CNN-transformer architecture to predict protein post-translational modification sites by integrating multi-scale sequence features, evolutionary embeddings from ESM-2, and structural properties from AlphaFold, with feature fusion and adaptive thresholding to handle class imbalance and improve prediction accuracy. Through cross-validation and testing, DeepPTMPred achieves superior performance with AUC scores of 96.5% for hydroxylation, 91.6% for malonylation, 91.0% for O-linked glycosylation, and 89.5% for phosphorylation, outperforming baseline and competing models. Through cross-validation and testing, DeepPTMPred achieves superior performance with AUC scores of 96.5% for hydroxylation, 91.6% for malonylation, 91.0% for O-linked glycosylation, and 89.5% for phosphorylation, outperforming baseline and competing models. The model accurately identifies known phosphorylation sites on tau protein and novel residues linked to early Alzheimer’s disease, demonstrating strong generalizability and potential for biomarker discovery in disease research. The model accurately identifies known phosphorylation sites on tau protein and novel residues linked to early Alzheimer’s disease, demonstrating strong generalizability and potential for biomarker discovery in disease research. Supplementary Material