Prioritizing missense mutations via a deep learning phosphorylation prediction model Authors: Yue Xu, Yuan Zhou, Kun Li, Shiao Zhou, Hai Yang, Dan Zhou Published: 2025 ABSTRACT AbstractBackgroundPhosphorylation is a crucial post-translational modification mechanism that enhances proteomic diversity, and its malfunction has been confirmed to be associated with complex traits, especially brain disorders. One of the factors contributing to this malfunction is the missense mutations given that they may alter the peptides flanking the phosphorylated residues. However, the specific effects of these missense mutations on phosphorylation remain unclear.MethodsTo ascertain these, a deep learning phosphorylation prediction model (DeepMEP), which is the first to be developed on a Chinese-brain-specific phosphorylation dataset (CBMAP), was established to bridge the phosphorylation and peptides. The impact of each missense mutation on phosphorylation was subsequently quantified based on the differences between the outputs of reference and mutant protein sequences. A permutation test adjusting for the confounding factors was finally employed to estimate the enrichment for high-impact mutations in disease-associated genomic loci.ResultsDeepMEP achieved superior predictive performance compared with other existing tools on both CBMAP and publicly available datasets. Enrichment analysis revealed the high-impact mutations were significantly enriched in GWAS signals for Alzheimer’s disease (AD) and Parkinson’s disease (PD). The corresponding genes of those missense mutations overlapping with GWAS included ABCA7, APOE, and PLCG2 for AD, and MMRN1 and TMEM175 for PD, which are disease-associated genes confirmed by other studies beyond GWAS.ConclusionOur study demonstrated that DeepMEP effectively captured the impact of missense mutations on phosphorylation and highlighted an enrichment of high-impact mutations in AD- and PD-associated genomic loci.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40246-025-00898-4.Keywords: Deep learning, Phosphorylation, Missense mutation, AD, PD, GWAS Introduction Phosphorylation is the most widely observed and well-studied post-translational modification (PTM). Several studies have confirmed the malfunction of phosphorylation is associated with Alzheimer’s disease (AD), Parkinson’s disease (PD) and other brain disorders [1, 2]. For instance, pathological tau, characterized by its abnormal hyperphosphorylation, is a hallmark in the progression of neurodegenerative diseases [3]. The dysfunction of G protein-coupled receptors (GPCR), one key mechanism of whose signaling is the GRK phosphorylation [4], is observed in nearly all neurodegenerative diseases [5]. Moreover, extensive phosphorylated residues are situated in the intrinsically disordered regions [6], which can alter the conformation and activity of protein. Therefore, understanding phosphorylation may provide more insights into the study of disease pathology and drug discovery. Phosphorylation typically occurs on serine [S], threonine [T], and tyrosine [Y] and is susceptible to the alteration to their flanking sequences [7, 8]. Given that such alterations in proteins usually lead to issues in protein folding, activity, or stability [9], we assumed that they could facilitate the malfunction of phosphorylation. And one source of that alteration is the missense mutations on the human genome that may alter the flanking sequences around the phosphorylated residues through protein translation. Because phosphorylation regulates a wide range of signaling pathways and is frequently perturbed in neurodegenerative diseases, prioritizing missense variants by their effect on phosphorylation may provide a biologically meaningful way to capture disease-relevant regulatory disruptions. But, the specific effect of missense mutation on phosphorylation is unclear. Deep learning phosphorylation prediction methods provide insights for this problem. The deep sequence model can infer the phosphorylation effect from protein sequences, thus the specific effect of missense mutation on phosphorylation can be quantified as the difference between the outputs of inputting reference and mutant sequences. Currently, many studies have been conducted to predict phosphorylation using deep learning and have obtained notable performance. These tools took the flanking sequences around the potential phosphorylated sites (serine, threonine, and tyrosine) as input and whether phosphorylation as label. These existing deep learning phosphorylation prediction tools include MusiteDeep [10], DeepPhos [11], DeepPPSite [12], EMBER [13], Phosformer [14], etc. All these tools have achieved good predictive performances, but they cannot be immediately applied to this work due to limitations of their training sets and applicable scenarios. Their training sets were primarily sourced from comprehensive public databases and fail to reflect the characteristics of specific diseases and tissues. Consequently, to reveal the association between brain disorders and missense mutations via phosphorylation, we had to develop a novel phosphorylation prediction model on a brain-specific dataset. Thus, we accessed the Chinese-brain-specific phosphorylation annotations from CBMAP [15], which is a database specializing in collecting multi-omics data related to Chinese brain. The phosphorylation annotations provided by it are phosphorylation abundance measured by Liquid Chromatography-Mass Spectrometry (LC–MS). Based on it, we developed a novel Deep learning model to quantify the Missense mutation Effect on Phosphorylation (DeepMEP). This model is the first to combine a pretrained protein language model with brain-specific phosphorylation patterns to predict the effect of missense mutation on phosphorylation. By prioritizing the missense mutations according to their phosphorylation effects, we estimated the enrichment of those missense mutations that highly affect phosphorylation in GWAS via a permutation test adjusting for the confounding factors. Results Development of DeepMEP The quantitative effect of missense mutation on phosphorylation is unclear. Thus, we developed a deep learning phosphorylation prediction model, DeepMEP, on existing phosphorylation annotations and utilized this model to measure the Missense mutation Effect on Phosphorylation (MEoP) based on reference and mutant sequences. The enrichment for those missense mutations with a high MEoP was finally estimated in disease-associated genomic loci via a permutation test adjusting for the confounding factors, to detect the association between phosphorylation and disease. Figure 1 provides an overview of the DeepMEP framework and illustrates how phosphorylation context is encoded and used for model training. As shown in Fig. 1A, the S, T, Y residues were viewed as potential phosphorylated sites (red points). For all potential phosphorylated sites, their flanking sequences were sliced, aligned and then fed into a pretrained protein language model (PLM) to get numeric representations. DeepMEP was trained based on these numeric representations, and the phosphorylation annotations used for model development were sourced from Uniprot [16], EPSD [17] and CBMAP [15] respectively. Schematic overview of the DeepMEP pipeline. A Training a deep learning phosphorylation prediction model. Red points are the potential phosphorylated residues (S, T, Y sites) and grey points are the residues. The input instances were first generated by slicing the flanking sequences around the potential phosphorylated residues, followed by inputting into a pretrained protein language model to get numeric representations. Based on these numeric representations, a deep learning phosphorylation prediction model was developed to predict the phosphorylation effect. The phosphorylation annotations were from CBMAP [15], Uniprot [16] and EPSD [17]. B Measuring the phosphorylation difference of each potential phosphorylated residue caused by missense mutation. Blue point on the DNA sequence is the missense mutation, blue point on the protein sequence is the mutant residue, and the red points on the protein sequence are the potential phosphorylated residues adjacent to the mutant residue. The phosphorylation difference of each site was measured as the absolute difference between outputs of reference and mutant sequences. C Extracting MEoP DeepMEP. (left) As each missense mutation may affect multiple residues, the MEoP was summarized across all affected residues by maximum or average (default maximum). (right) Using different thresholds to identify the significant missense mutations and estimating their enrichment in GWAS After developing the model, we quantified the MEoP for each missense mutation (as illustrated in Fig. 1B and C). Each mutant residue (blue point on the protein sequence) caused by missense mutation (blue point on the DNA sequence) was first located. Those adjacent potential phosphorylated sites (red points on the protein sequence) around mutant residue were then identified. MEoP was calculated as follows: Firstly, the reference and mutant sequences of each adjacent potential phosphorylated site were generated; Then, the phosphorylation difference at each site caused by missense mutation was calculated as the absolute difference between the outputs of the model on input reference and mutant sequence; Finally, the maximum or average phosphorylation difference (default maximum) across all these adjacent potential phosphorylated sites was identified as MEoP. By prioritizing missense mutations via MEoPs, we selected the significant missense mutations according to different thresholds, and estimated their enrichments in GWAS disease-associated variants via a permutation test that adjusts for the confounding factors, to detect the association between phosphorylation and disease. Benchmarking and phosphorylation prediction performance of DeepMEP The main purpose of this study was to unravel the association between phosphorylation and brain diseases. Therefore, DeepMEP was developed on a brain-specific dataset from CBMAP. The discrepancy of the phosphorylation annotations between Uniprot and CBMAP is shown in Supplementary Fig. 1. To ensure the comparison with existing deep learning phosphorylation prediction tools was fair, we also realized DeepMEP on public datasets and other tools on CBMAP. The prediction tasks on CBMAP, Uniprot and EPSD were corresponded to general phosphorylation prediction, kinase-family-specific phosphorylation prediction and phosphorylation intensity prediction, respectively. The first one was regression task and the last two were classification tasks. Phosphorylation effects were measured as the probability of class in classification and intensity values in regression. Model training consisted of 3 phrases: Firstly, training several basic predictors under different lengths of flanking sequence varying 5 to 16. Figure 2 presents the dataflow and architecture in each basic predictor. For each length, 6 basic predictors were trained using output of 6 different layers in ESM2, given that the embedding of different layers in ESM2 had different expressive spaces (Supplementary Fig. 2). Secondly, evaluating basic predictors on the validation set. The validation performances of basic predictors were shown in Supplementary Fig. 3. The optimal lengths of the flanking sequence were identified as 15, 7 and 12 for general prediction, kinase-family-specific prediction and intensity prediction, respectively. Finally, building an ensemble of 6 basic predictors under the optimal length according to validation performance to enhance the robustness. The ensemble took the average of the output of all predictors as the final output. Dataflow of DeepMEP. Each input instance was encoded into the numeric representation via ESM2, then flattened and transferred to multiple linear transformation modules. Each module was comprised of a dropout layer, a linear transformation layer, a batch normalization layer and a LeakyReLU activation function [18]. The output layer mapped the output value of the former layer to the target space according to the type of prediction task (general phosphorylation prediction, kinase-family-specific phosphorylation prediction or phosphorylation intensity prediction) using different activation functions After building the ensemble, we evaluated it on the testing set. To compare fairly, all tools were trained and tested on the same dataset. On general phosphorylation prediction, DeepMEP outperformed MusiteDeep, DeepPhos and DeepPPSite with ROC AUC of 0.884, PR AUC of 0.496 and sensitivity of 0.843 (Supplementary Fig. 4A). On kinase-family-specific phosphorylation prediction, DeepMEP has a comparable performance to other tools (Supplementary Table 2). Specifically, DeepMEP outperforms other tools on PKA and PIKK family in terms of ROC AUC and on SRC family in terms of sensitivity. On phosphorylation intensity prediction, DeepMEP outperformed other tools with R2 of 0.121 (Supplementary Fig. 4B). To further evaluate the predictive performance of DeepMEP, we conducted fivefold cross-validation, with detailed performance metrics summarized in Supplementary Table 3. Specifically, DeepMEP achieved an average ROC AUC of 0.8864 for general phosphorylation prediction and an average R2 of 0.1607 for phosphorylation intensity prediction across all folds, outperforming all comparative tools. Enrichment of high-MEoP missense mutations in AD and PD GWAS Following the development of DeepMEP, the MEoP of each missense mutation was calculated. The high-impact missense mutations were then selected according to a threshold and subsequently estimated the strength of enrichment in GWAS via a permutation test adjusting for the confounding factors (see Methods). Figure 3 depicts the statistical description of MEoP and the enrichment for selected mutations in disease-associated loci. High-MEoP missense mutations were enriched in GWAS AD- and PD-associated variants. A The distributions of MEoPs computed by the model trained on CBMAP. The red line is the threshold for identifying significant missense mutations (median value default). B The violin plot of MEoPs computed by the model trained on CBMAP at each distance. C The number of missense mutations in GWAS significant variants. D, E and F The enrichment for the missense mutation selected by baselines, Uniprot and CBMAP model. Each row represents a set of disease-associated variants and the abbreviations of diseases are shown in Supplementary Table 1. The text annotations and the intensity of the red color represent the fold change the P value of the permutation test respectively. The redder the color, the more significant the selected missense mutations in GWAS. D The enrichment for the missense mutations selected by three different baselines: (left) the missense mutations that directly alter the focused sites. (middle) the missense mutations whose SIFT score was more than the median. (right) the missense mutations whose pathogenicity score predicted by AlphaMissense was more than the median. (E) The enrichment for the missense mutations selected by DeepMEP (left: Uniprot model; right: CBMAP model). F The enrichment for the missense mutations that were not selected by DeepMEP (below the default threshold). G, H, I and J The distributions of permutations for the enrichment results of AD and PD that are shown in (D) Since this work focused on general phosphorylation effect caused by missense mutations rather than kinase-specific differences, the MEoPs were computed using models trained on Uniprot and CBMAP datasets for all single-substitution missense mutations, excluding the insertion and deletion. The distribution of the MEoPs is shown in Fig. 3A, with details provided in Supplementary Table 4 and 5. The MEoP decreases with increasing distance (Fig. 3B). Taking the median value as the threshold by default (see Methods), the high-MEoP missense mutations were identified by those whose MEoP was above the threshold. Their enrichments in GWAS were estimated with a permutation test adjusted for the linkage disequilibrium (LD) score and minor allele frequency (MAF; see Methods). The permutation test took all missense mutations as the sampling background, with the aim to estimate the enrichment for these significant missenses against the entire pool of missense mutations. Conventionally, when investigating the association between missense mutations highly affecting phosphorylation and disease without specific phosphorylation effects, we might regard the missense mutations that alter the S, T, or Y residues as high-impact [19]. To form a contrast, we therefore used the enrichment of missense mutations that directly alter the S/T/Y residues as the first baseline (Fig. 3D left), where there was no enrichment in any diseases. The evolutionary conservation-based predictors have always been the standard in interpreting missense mutations. Thus, the SIFT [20], which predicts the impact of missense mutation on protein function, was taken as the second baseline. However, it didn’t show a strong association with diseases (Fig. 3D middle). Moreover, we noticed the AlphaMissense, which is also a deep learning model that can predict the pathogenicity score of each amino acid subsititution, and took it as the third baseline. The enrichment analysis found that the mutations with high pathogenicity score strongly enriched in BMI (Fig. 3D right). The enrichment results for high-MEoP missense mutations identified by DeepMEP are shown in Fig. 3E, where we can clearly see that high-MEoP mutations were enriched in AD- and PD-GWAS. This result was consistent with the fact that the malfunction of phosphorylation is associated with AD and PD [1]. The high-MEoP missense mutations identified by CBMAP model got a P-value of 0.007 for AD (FDR-adjusted P = 0.042) with a fold change of 1.4 and a P-value of 0 for PD (FDR-adjusted P = 0) with a fold change of 1.84. To ensure the robustness of the results, we independently repeated the permutation tests for AD and PD 30 times, and calculated the 95% confidence intervals (CIs) of the P-values using a one-sample t-test. The mean P-value across the 30 replicates was 0.005 for AD (95% CI: 0.004–0.006) and 0.0002 for PD (95% CI: 0.0000–0.0003), respectively. The results of 30 repeated independent permutation tests are provided in Supplementary Table 6. Furthermore, we threw sights on those not high-impact missense mutations and explored their enrichment (Fig. 3F). Interestingly, the mutations that not highly affect phosphorylation enriched in some phenotypes, suggesting that the association between these brain phenotypes and mutations may be poorly explained by phosphorylation. Additionally, we tried the permutation test under different configurations, including different thresholds, adjusting for different confounding factors and using different methods for calculating MEoP. To compare logically, we took the results in Fig. 3E as the reference. The enrichment results using different quantiles as threshold are shown in Supplementary Fig. 5 and Supplementary Table 7. The enrichment results using a permutation test without adjusting for any confounding factors are shown in Supplementary Fig. 6, where there is hardly signaling, suggesting the necessity of adjusting for the confounding factors. We also tried to use the average to calculate the MEoPs, and the enrichment results using the MEoPs calculated by average are shown in Supplementary Fig. 7. Besides adjusting for the MAF and LD, we also tried to adjust for the gene length when performing the enrichment analysis and the results are presented in Supplementary Fig. 8. The enrichments for missense mutations, which was selected by CBMAP model, in AD were all observed in Supplementary Fig. 5, 7 and 8 under different thresholds (from 0.45 to 0.55). Notably, a low threshold led the enrichment to approximate the whole missense mutation set while a high threshold limited the statistical power as a result of the reduced number, thus the median was selected as the threshold by default. These supplementary analyses demonstrated the robustness of our results. We also assessed the enrichment for these significant missense mutations in rare trait-associated variants according to rare variant annotations from RAVAR [21] (Supplementary Table 9), after filtering out traits associated with fewer than 5 rare variants. Significant missense mutations identified by the CBMAP model—using the median MEoP as the threshold—enriched in thyroid disease, which were not significant in Uniprot model. However, some alterations in brain structures of cerebellar regions are induced by hyperthyroidism [22]. The high-MEoP missense mutations overlapping with GWAS The enrichment analysis shown in Fig. 3 illustrates the association between the high-MEoP missense mutations and brain disorders. Here, we investigate the details of the missense mutations overlapping with GWAS variants (Supplementary Table 8) and explore the potential roles of their corresponding genes on trait. Figure 4 shows the MEoP of each missense mutation calculated by Uniprot and CBMAP model, respectively, and marks the overlapped mutations. Manhattan plots for the MEoPs of the missense mutations. Each point is a missense mutation. The y-axis represents the MEoP and the red line is the median value of MEoP. The annotated points are the overlapped missense mutations with GWAS. The bordered points are the missense mutations uniquely identified by the corresponding database. A and B are the MEoPs computed by Uniprot model, while (C and D) are those computed by CBMAP model. A and C are for AD, and (B and D) are for PD Although genome-wide association studies identify extensive trait-associated loci, the underlying mechanisms of these associations are still elusive [23]. Thus, the overlapped missense mutations here suggested that they may be associated with traits through highly affecting phosphorylation. In total, CBMAP identified 34 overlapped missense mutations (or corresponding genes) in AD and 17 in PD while Uniprot identified 30 in AD and 16 in PD (overlapped missense mutations are points with text annotations in Fig. 4). In AD, 27 overlapped missense mutations were shared between CBMAP and Uniprot, and 10 of them have explicit evidence beyond GWAS for their association with AD, including ABCA7 [24], ABI3 [25], APH1B [26], ATP8B4 [27], MCM7 [28], MS4A6A [26], PILRA [29], SHARPIN [30, 31], SPPL2C [32], TREML2 [33, 34]. Notably, ABCA7 are AD-associated genes confirmed by many studies, and it is found that its deficiency causes neuronal dysregulation by altering mitochondrial lipid metabolism via a genome editing experiment [24]. The remaining 17 without other evidence were APOC4, BCL3, CBLC, CR1, GRIN3B, NKPD1, USP6. In PD, there were 14 overlapped missense mutations were shared between CBMAP and Uniprot, where MMRN1 [35], TMEM175 [36] having evidence beyond the GWAS. Then, we evaluated the difference in the application of brain diseases between Uniprot and CBMAP. The unique overlapped missenses that were not identified by another model are highlighted with borders in Fig. 4. For example, 7 CBMAP-identified overlapped missense mutations (bordered annotations in Fig. 4C) were missed by Uniprot model. They were rs117289933, rs12185233, rs189063316, rs5167, rs549046469, rs769452 and rs72824905 corresponding to genes EXOC3L2, PPP1R37, APOE, APOC4, POLR1G and PLCG2. Importantly, the APOE and PLCG2 are AD-associated genes. Specifically, APOE is a gene influencing both the average age of onset and the lifetime risk of developing AD [37], and the APLCG2 modulates TREM2 expression and signaling in response to AD pathology [38]. Similarly, the Uniprot identified rs16978738, rs143332484 and rs12373139 corresponding to gene ZNF225, TREM2, and SPPL2C uniquely, and SPPL2C is a gene associated with AD confirmed by an exome-wide age-of-onset analysis [32]. In terms of PD, CBMAP identified rs2270968 and rs35675346 corresponding to MCCC1 and FBXL19 uniquely while Uniprot identified the rs2230288 and rs35833914 corresponding to gene GBA and KANSL1. Sequence basis of phosphorylation learned by DeepMEP In this section, we show the sequence patterns of phosphorylation and evaluate that whether the model learns these patterns. By aligning the 10 residues around the focused site, Fig. 5A and B present the comparison of sequence patterns between phosphorylation and non-phosphorylation, as well as between high- and low-phosphorylation. We can clearly see that the flanking sequences around the phosphorylated or high-phosphorylation sites are significantly distinct and the same amino acid at the same position exhibits different frequencies between phosphorylation and non-phosphorylation, or between high- and low-phosphorylation, such as ‘P’, ‘L’ and ‘E’. To observe the specific frequency difference between phosphorylation and non-phosphorylation, or high- and low-phosphorylation for each residue excluding ‘S’, ‘T’ and ‘Y’, we further counted the sum of the absolute frequency difference at all positions (Fig. 5C). It’s notable that the frequency of ‘P’ has the maximum difference, thus we show the frequency difference of ‘P’ at each position in Fig. 5D to further investigate this difference, from which we can speculate that the main difference of ‘P’ is caused by the ‘P’ located on the right side of the focused site. Missense mutations altering ‘P’, ‘A’, ‘G’ and ‘N’ get high MEoP. A and B are the sequence basis of aligned instances in Uniprot (A) and CBMAP (B). The X-axis represents the flanking around the focused site. The height of residue is the frequency at this position. In (A), the upper panel are phosphorylated instances in Uniprot and the bottom panel are those non-phosphorylated. In (B), the upper panel shows the instances above the median intensity in CBMAP while the bottom panel shows those below. The Logo diagram was realized based on Logomaker [39]. C The comparison of residue frequency difference between phosphorylation and non-phosphorylation, or between high- and low-phosphorylation. For each residue, the frequency difference was calculated as the sum of absolute difference of frequency between the upper panel and the bottom panel in (A) and (B). D The absolute difference of residue ‘P’ at each position. E The proportion of amino acids altered by missenses whose MEoP ranks in top 10%, excluding the ‘S’, ‘T’ and ‘Y’ Next, we investigate whether our model learns these patterns. Figure 5E displays the proportion of amino acids altered by missenses whose MEoP ranks in top 10% excluding the ‘S’, ‘T’ and ‘Y’. In other words, it shows the missenses altering which residue obtain the high MEoP. Consistent with the sequence pattern, the missenses altering the ‘P’ definitely obtain the high MEoP in both Uniprot and CBMAP, and those altering ‘R’, ‘G’, ‘E’ and ‘L’ also take a large proportion, suggesting the existence of complex sequence pattern. Discussion In this work, we proposed DeepMEP, a deep learning phosphorylation prediction tool that can predict the phosphorylation effect on the focused site based on its flanking sequences. Architecturally, existing deep learning phosphorylation prediction tools predominantly employ convolutional neural network (CNN), attention mechanism and multi-layer perceptron (MLP) to construct the model. For instance, MusiteDeep [10], the first deep learning phosphorylation prediction tool, consists of a multi-layer CNN, dual attention modules and a MLP. DeepPhos [11] uses three densely-connected CNN blocks followed by a MLP. EMBER [13] is comprised of a self-trained embedding model and a prediction model, and both 2 models use the CNN and MLP. Phosformer [14] stands apart by adopting a transformer-based architecture. Compared with these tools, DeepMEP did not include computation modules like CNN and attention mechanism, and just consisted of a frozen pretrained protein language model (PLM) and an MLP. Methodologically, the first step in model training involves encoding the sequences into numeric representations. Previous studies used one-coding or self-trained embedding. Given that PLMs exhibit the superior abilities in extracting the protein comprehensive features [40], DeepMEP is the first to incorporate a PLM to predict brain-specific phosphorylation patterns. Considering that training large-scale PLMs is time- and computation-power-consuming, we frozen the pretrained PLM and used it directly rather than training de novo. Our results suggested that the phosphorylation could be modelled well from the embedding immediately. Moreover, we found that the representations from different layers in ESM2 had different expressive space. Thus, we integrated multiple models that used embeddings from different layers under the same optimal flanking sequence length into an ensemble to boost the accuracy. Existing studies summarized the phosphorylation prediction into general prediction and kinase-specific prediction. General prediction refers to judging whether the residue can be phosphorylated and kinase-specific prediction refers to answering which kinase can the site be phosphorylated by. Except Phosformer, either general prediction or kinase-specific prediction is abstracted into a classification task in these methods, where general prediction is a binary classification and kinase-specific is a multi-label classification. Phosformer innovatively reframed this task into a context-based question-answering task, where the flanking sequence serves as context, the kinase sequence as the question, and the probability of phosphorylation as the answer. Different with these tools, DeepMEP used phosphorylation intensity— the average abundance across all samples—as the label, rather than discrete classes. We assumed that the continuous intensity value has different patterns with class probabilities. Previous phosphorylation prediction tools did not extend their model to further application, restricting the practical value of model. In this work, we employed DeepMEP to quantify the missense mutation effect on phosphorylation (MEoP) and identified the significant missense mutations via it. The enrichment for these significant missenses in AD- and PD-GWAS suggested that they may be associated with AD and PD through highly affecting phosphorylation. In terms of AD and PD, the malfunction of phosphorylation has been confirmed to be associated with AD and PD by many studies, but no enrichment in AD or PD was observed in baselines. This is expected, as SIFT and AlphaMissense primarily capture evolutionary constraint or general deleteriousness rather than phosphorylation-specific. In addition, the mutation corresponding to APOE, which a single greatest genetic modulator of sporadic Alzheimer disease (AD) risk [37], got a high MEoP by CBMAP model. DeepMEP and the resulting MEoP scores may offer potential clinical utility by helping prioritize variants of unknown significance (VUS) in neurogenetic diagnostic pipelines, particularly when phosphorylation dysregulation is implicated. MEoP could complement existing pathogenicity tools by highlighting VUS that are more likely to perturb phosphorylation. However, we acknowledge that these associations detected by DeepMEP remain statistical in nature and do not yet establish causal relationships to disease mechanisms. Future work will be essential to provide functional validation. One direction is to perform targeted mutagenesis followed by phosphoproteomic profiling—particularly for top-ranked variants in genes such as APOE or PLCG2—to directly test whether DeepMEP-predicted phosphorylation changes occur in cellular or neuronal models. Additionally, validation using independent phosphoproteomics datasets, especially from human post-mortem brain tissues, would further strengthen the robustness of our findings. Incorporating case studies of variants with known functional evidence may also help contextualize DeepMEP predictions in well-characterized biological scenarios. Together, these follow-up experiments will be critical for bridging the gap between statistical association and biological causality in neurodegenerative disease pathology. As our training data are largely derived from brain phosphoproteomic profiles and phosphosite annotations remain incomplete and tissue-specific, the generalizability of DeepMEP to other tissues or unrelated disease contexts is limited. Consequently, MEoP scores may underperform for traits or disorders in which phosphorylation is not a major regulatory mechanism, and their interpretation should be restricted to biological contexts where phosphorylation regulation plays a central role. On the other hand, even if DeepMEP was just applied to phosphorylation here, it can be extended to other types of post-translational modifications, such as acetylation, ubiquitination and lactylation, and other diseases, provided that new datasets become available in the future. Furthermore, we also noticed that more and more novel machine learning algorithms were applied into bioinformatics research, such as matrix factorization [41], DeepWalk [42], neural collaborative filtering [43, 44], transfer learning with BioBERT [45], stacked variational autoencoder [46] and graph neural network [43, 47]. Based on these advances, we will explore more computational methods to enhance DeepMEP in the future. Methods Dataset description Phosphorylation datasets Public phosphorylation annotations were downloaded from Uniprot [48] (https://www.uniprot.org/uniprotkb) and EPSD [17] (https://epsd.biocuckoo.cn), and the Chinese-brain-specific phosphorylation annotations were accessed from CBMAP [15]. The annotations about which family the kinases belong to were loaded from KinBase [49] (http://kinase.com/web/current/kinbase/). Variant annotations The GWAS summary statistics of common neural diseases were loaded online according to the URLs provided in the references, including AD [50], BIP [51], BMI [52], EA [53], IQ [54], MDD [55], PD [56], PTSD [57], SCZ [58], TAAU-SmkInit and TAAU-CigDay [59] (see disease abbreviations in Supplementary Table 1). The disease-associated variants were identified by those whose P-value were less than 5e-8. The rare trait-associated variants were obtained from RAVAR [21] (http://www.ravar.bio/#/). The missense mutation annotations were loaded from dbSNP [60] (https://www.ncbi.nlm.nih.gov/snp/). The mutant protein sequences for each missense mutation required for the utility of DeepMEP were generated by translating from transcript sequences according to the missense mutation annotations. Data preprocessing We manually curated the dataset for model training and evaluation. Firstly, the input instance for each focused site (S, T, Y) was yielded by slicing its flanking sequences. The size of flanking sequence is a hyperparameter and was identified during model development. If the focused site was proximal to the N or C-terminus of the protein, the flanking sequences would be padded by ‘_’ to ensure that all instances have equal length. The label of each input instance was then generated according to the phosphorylation status of the central site. For phosphorylation intensity prediction, label was the phosphorylation intensity of central site according to CBMAP. Since the raw phosphorylation intensity annotations in CBMAP were population-level, the median intensity across all samples was calculated for each focused site. These median intensity values across all focused sites were then log₂-transformed and subsequently normalized to derive the final phosphorylation intensity. The focused sites without phosphorylation annotations were excluded. For general phosphorylation prediction, all focused sites were initially annotated as non-phosphorylation until phosphorylation evidences were found according to Uniprot. The phosphorylated sites were viewed as positive while non-phosphorylated sites were negative. For kinase-specific phosphorylation prediction, label was the family of the kinase that phosphorylated the central site. Each kinase family was deemed as a class and the sites that could not be phosphorylated by any kinases were categorized into a non-phosphorylation class. In addition, we removed the redundant and contradicted instances. Training, validation and testing sets The whole dataset was split into the training, validation and testing sets with a ratio of 70:15:15. Notably, the instances were split at protein level to ensure that the instances from the same protein were assigned into the same set to decrease the similarity between 3 subsets. The protein sets used for dataset splitting were consistent in Uniprot, EPSD and CBMAP. Table 1 summarizes the data size of each subset. To prevent the data leakage, the similar instances were excluded from the testing set based on Hamming distance scores. Specifically, on the one hand, the pairwise Hamming distance scores of instances in testing set were calculated to ensure all retained testing instances were sufficiently distinct from one another. If the Hamming similarity of a given sequence pair surpassed 0.6, one instance from the pair was eliminated from the test set. On the other hand, any testing set instance that exhibited a similarity above 0.6 to any instance in the training set was excluded to avoid cross-set data overlap. Through this scheme, we removed 5,880, 2,086, and 240 instances in Uniprot, EPSD and CBMAP testing set respectively. Table 1The size of the curated datasetGeneral phosphorylation prediction (Uniprot)ClassificationTraining instancesValidation instancesTesting instancesPositive27,62258605857Negative625,734127,837122,704Kinase-specific phosphorylation prediction (EPSD)Kinase FamilyTraining instancesValidation instancesTesting instancesPKA3887560AKT1774624CDK609142125MAPK58696110SRC474126127CK24008385PKC5619195PIKK3124751Non-phosphorylation191,59840,77236,034Phosphorylation intensity prediction (CBMAP)Training instancesValidation instancesTesting instances73,67114,93515,345The dataset was split at protein level, resulting in approximate (though not exact) allocation ratios of 70% training, 15% validation, and 15% testing data The size of the curated dataset The dataset was split at protein level, resulting in approximate (though not exact) allocation ratios of 70% training, 15% validation, and 15% testing data DeepMEP method Sequence representation Before being input into model, each instance would be transformed into a high-dimensional numeric representation employing a pretrained protein language model (PLM). We selected a pretrained PLM—the version of ‘ESM2_t6_8M_UR58D’ in ESM[40]—to represent the sequences directly using its frozen parameters rather than training de novo. The selected pretrained PLM is a stack of 6 transformer layers and can convert each residue into a 320-dimensional tensor. Model design and training Generally, we trained the model in three phases. Firstly, we trained multiple basic predictors under different flanking sequence lengths. We assumed that the embedding of different layers ESM2 had different expressive space. Thus, under each length, different basic predictors were trained respectively using embedding of different layers in ESM2. Then, we identified the optimal flanking sequence length according to validation set performance, specifically ROC AUC for classification and R square for regression. Finally, we built an ensemble by taking the average output of all basic predictors under the optimal length as the final output. The basic predictor of DeepMEP was a multi-layer perceptron (MLP), where each layer except the output layer was a combination of a dropout sublayer, a linear transformation sublayer, a batch normalization layer and an activation function. Due to the high dimensionality of the output of the pretrained PLM, the dropout layer was introduced before the linear sublayer to avoid overfitting. Batch normalization was added after the linear layer to stabilize the training, accelerate the convergence and prevent outlier from emerging. The LeakyReLU function[18] was used to activate the hidden value of each module to avoid the dying ReLU problem with a negative slope of 0.01. A schematic of the layers in the basic predictor for different tasks are shown in Fig. 2. DeepMEP was comprised of four layers: input layer, hidden layer1, hidden layer 2 and output layer respectively. Let denote the length of input instance, thus the flattened embeddings of the pretrained PLM were -dimensional. The dimensions of the embeddings in DeepMEP were condensed step by step. The input layer, hidden layer1 and hidden layer 2 condensed the dimension from to 2048, from 2048 to 512 and from 512 to 12 respectively. This configuration was applied for all prediction tasks. Subsequently, the output layer mapped the outputs of hidden layer2 to the target space according to different prediction tasks. Specifically, for classification, the dimension of embeddings was condensed to be equal to the number of labels (2 for general phosphorylation prediction, 9 for kinase-family-specific prediction), and the softmax function was used as activation function. For regression, the dimension was condensed to 1 and the linear output of the output layer was used directly. The weights of the basic predictor were optimized using Adam [61] algorithm with a learning rate of 0.0001, 0.001, 0.0006 and a weight decay of 0.0001, 0.001, 0.0008 for general prediction, kinase-family-specific prediction and intensity prediction respectively. The dropout rates in the input layer were 0.5 for all prediction tasks, in the hidden layer1 were 0.15, 0.5 and 0.5 for general, kinase-specific-family and intensity respectively, and in the hidden layer2 were 0.1, 0.15 and 0.15 for general, kinase-specific-family and intensity respectively. The epochs were 10, 30 and 20 for them respectively. Notably, we saved a checkpoint at each epoch. The model showing the best validation performance during training was selected as the final model. A grid search strategy was used to optimize model hyperparameters. We first fixed the sequence length at 16 and the number of embedding layers at 6, and performed grid search within a predefined range to identify the optimal hyperparameter settings. After obtaining the best-performing configuration, we extended this hyperparameter combination to all other sequence length and embedding layer settings. Learning of phosphorylation intensity For phosphorylation intensity prediction, we used Huber loss, which can be formalized as: where denotes the true intensity, denotes the predicted intensity, and is used for adjusting the extent between L1 and L2. This function can reduce the penalty level for outliers and improve the robustness. The was set to 1.0 by default. Learning of imbalanced dataset Paramount imbalance between the number of phosphorylation and non-phosphorylation instances exerted the prodigious difficulty on learning phosphorylation instances in classification. Here, we introduced the focal loss function [62] to prevent the model from solely focusing on non-phosphorylation, the function can be formalized as: where denotes the predicted probability of target class, is used for adjusting the weight of the target class, and γ determines the weight of complex samples. On general phosphorylation prediction, the of the non-phosphorylation was set to 0.03. On kinase-family-specific prediction, the of the non-phosphorylation was set to 0 to exclude the loss of non-phosphorylation instances during training. This approach was inspired by Phosformer [14], the author of which divided the non-phosphorylation instances into the hard and the easy. The hard refers to the instances that can be phosphorylated by other kinases while the easy refers to the instances that cannot be phosphorylated by any kinases. The exclusion of easy negative instances from training set can make model focus on hard samples. To compare fairly, the easy negative instances in testing set were not excluded. Furthermore, the γ was set to 2.0 by default. Performance evaluation metrics Common evaluation metrics [63] were employed in our study, including ROC AUC (the area under receiver operating characteristic curve), PR AUC (the area under precision-recall curve) and Recall for classification, as well as R2 for regression. For classification, although ROC AUC is a common classification metric for evaluating overall performance, it is not sensitive to imbalanced datasets. In this study, phosphorylated instances were extremely less, thus we couldn’t merely focus on the ROC AUC. The precision and recall are metrics about the positive samples, specifically, precision reflects the accuracy of positive samples that predicted by model and recall reflects the ability of model to hit the positive samples in real-scene. Moreover, PR AUC, a metric integrating the precision and recall, reflects model’s comprehensive performance on positive samples. Thus, all of these metrics were used for model evaluation. For regression, we used R2. Quantifying missense mutation effect on phosphorylation We firstly used annotations from dbSNP to locate the mutant residue caused by missense mutation on the protein sequence. Only single-substitution mutations were kept, excluding the insertion and deletion. As one missense mutation could affect multiple adjacent S/T/Y sites, we firstly calculated the phosphorylation difference for each affected site, and then selected the maximum difference to represent the missense mutation effect on phosphorylation. Given that the key purpose of this work was to reveal the association between phosphorylation malfunction and brain diseases, this work used the maximum phosphorylation difference as MEoP to highlight the changes in phosphorylation signals caused by missense mutations. The average was additionally used to form a contrast. The phosphorylation difference for each affected site was calculated as illustrated in Eq. (3): where is the output of DeepMEP inputting the reference flanking sequence of site and is the output of the mutant sequence. The Missense mutation Effect On Phosphorylation (MEoP) was finally defined as the maximum phosphorylation difference across all affected sites , as formulated in Eq. (4): Notably, if the missense mutation directly alters the S, T, Y sites into other residues, the will be set to minimum effect. Similarly, if the mutation creates a new S, T, Y site, will also be set to minimum effect. The missense mutations whose MEoP are above the threshold are identified as the missense mutations highly affecting phosphorylation. The threshold is set to the median of all MEoPs by default. A lower threshold includes a larger set of missense mutations, leading the enrichment results to approximate those obtained from the full mutation set. In contrast, a higher threshold restricts the analysis to mutations with strong predicted effects on phosphorylation, but the reduced number of variants limits statistical power. Therefore, we selected the median value as the threshold to balance effect size and sample size. Given that the conceptual distinction between phosphorylation probability (Uniprot) and phosphorylation intensity (CBMAP), we ranked mutations within each model separately and defined significant mutations according to threshold for each output type. This rank-based normalization avoids mixing scales and ensures that MEoP reflects relative phosphorylation impact rather than absolute model values. Estimating the strength of enrichment for the significant missenses in GWAS After obtaining the significant missense mutations, we explored whether they were enriched in GWAS. The GWAS disease-associated variants were identified as those whose P-value was less than 5e-8 in the summary statistics. Then, we used a permutation test adjusted for the confounding factors, such as linkage disequilibrium (LD) score and minor allele frequency (MAF), to estimate the strength of enrichment for these missense mutations in disease-associated variants. The null hypothesis of the permutation test is that these significant missense mutations are not more likely to be disease-associated than expected by chance given the background of all missense mutations. The P value and fold change of the permutation test, and , were calculated as follows: where is the set of missenses highly affecting phosphorylation, is the set of disease-associated variants, is the permutated set generated by sampling from all missense mutations, is the permutation times and is the indicator function. The value is calculated as the proportion of . The is calculated as the ratio of to the average of . The was set to 1000 by default. The P-values for disease enrichment analyses derived from UniProt and CBMAP datasets were adjusted independently using the false discovery rate (FDR) method. The main purpose of the permutation test was to explore the enrichment for the significant missense mutations against all missense mutations, but the inconsistency of MAF and LD score would disrupt the enrichment signal, so we considered them as confounding factors here. Some tricks were used to adjust for confounding factors to improve the result robustness. Firstly, we calculated the MAF and LD score for variants whose MAF was above 0.005 using GCTA [64]. Then, both MAF and LD score were each partitioned into 5 quantile-based intervals with equal sample sizes. The joint distribution of all missenses was thus represented as a 25-cell grid, where each cell corresponded to a unique combination of an MAF interval and an LD score interval. Additionally, those variants whose MAF was below 0.005 were categorized into the 26th cell. When generating dataset , we ensured that the elements in exhibit the same distribution across these 26 cells as observed in . The significant missense mutations with high LD were represented by its tag when performing permutation test. Notably, these tricks were not used for performing the permutation tests of rare variants. We additionally considered the gene length as the third confounding factor. When considering the gene length, each cell was partitioned into 5 intervals according to the quantiles of gene length and ensured that the elements in exhibit the same distribution across these cells as observed in . Comparing with other predictors for functional effects of missense mutation The SIFT [20] and AlphaMissense [9] were taken as another two baselines to form a contrast to DeepMEP. SIFT is a predictor predicting the potential impact of amino acid substitution on protein function based on sequence homology, while AlphaMissense is an adaptation of AlphaFold [65] combining structural context and evolutionary conservation that can predict the pathogenicity score of missense mutation. The SIFT scores were sourced from Ensembl Biomart (https://www.ensembl.org/biomart/martview) and the predictions of AlphaMissense were loaded from Zenodo according to the link provided by developers (10.5281/zenodo.10813168). The missense mutations were then prioritized by these scores and the enrichment in GWAS for these significant missense mutations were subsequently estimated to compare with DeepMEP. Notably, the SIFT scores were processed by “1 – SIFT score” when performing enrichment analysis given that the lower raw values indicate high risk. Supplementary Information Supplementary Material 1. Supplementary Material 2. Author contributions D.Z. designed the study. Y.X. wrote the manuscript. Y.X., Y.Z., K.L., S.Z. and H.Y. collected and processed the data. Y.X. performed the data analyses. D.Z. supervised and acquired funding for the study. All authors have read and approved the final version of the manuscript. Funding This work was supported by the National Natural Sciences Foundation of China No. 82204118 (D.Z.), the Healthy Zhejiang One Million People Cohort K-20230085 (D.Z.), the Key Laboratory of Intelligent Preventive Medicine of Zhejiang Province 2020E10004 (D.Z.). Data and code availability All public datasets can be obtained by the access shown in the main text. The phosphorylation annotations from CBMAP can be obtained by reaching out to the CBMAP consortium. The source code of this work and the model weights for reproducibility are available at https://github.com/zdangm/DeepMEP. An archived snapshot corresponding to the version used in this manuscript is deposited at Zenodo: 10.5281/zenodo.17906379. Declarations Ethics approval and consent to participate CBMAP: All donors or their families provided informed consent through voluntary donation agreements, permitting the use of their information and biospecimens for future studies. The study protocol was approved by the Ethics Committee of Zhejiang University School of Medicine (2020–005 and 2024–007). Approval 2020–005 authorized the collection of human brain tissue samples, and approval 2024–007 permitted the subsequent multi-omics sequencing and analysis of these previously collected samples. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement All public datasets can be obtained by the access shown in the main text. The phosphorylation annotations from CBMAP can be obtained by reaching out to the CBMAP consortium. The source code of this work and the model weights for reproducibility are available at https://github.com/zdangm/DeepMEP. An archived snapshot corresponding to the version used in this manuscript is deposited at Zenodo: 10.5281/zenodo.17906379.