Artificial intelligence for comprehensive DNA methylation analysis: overview, challenges, and future directions

Abstract

This paper offers a comprehensive review of the synergy between artificial intelligence and DNA methylation analysis, encompassing machine learning, deep learning, natural language processing, and explainable artificial intelligence. In this study, we also highlighted the underexplored potential of signal processing and large language models-based models in DNA methylation research. Additionally, we discussed the challenges and limitations faced when managing and analyzing large and complex DNA methylation datasets. Furthermore, this article tries to shed light on the continuing evolution of this field and on the possible directions for future research.

Introduction

DNA methylation, a key epigenetic modification, plays a pivotal role in regulating gene expression and maintaining genomic integrity [1]. Abnormalities in DNA methylation patterns have been linked to various diseases, including cancer, neurodegenerative disorders abnormalities [2, 3]. As a result, accurate and comprehensive analysis of DNA methylation is crucial for understanding the molecular basis of these diseases, advancing early detection and developing targeted therapies [4]. Traditional methods for DNA methylation analysis, such as bisulfite sequencing, have provided significant insights but are often time-consuming, costly, and limited in terms of scale and accuracy [5–8]. Thus, with the rapid advancement of high-throughput sequencing technologies, the volume of epigenomic data has grown exponentially, creating an urgent need for novel computational approaches to analyze and interpret these datasets efficiently and effectively [3]. Artificial intelligence (AI) has shown great promise in addressing complex data analysis challenges across various fields [9]. The application of AI encompassing the use of deep learning (DL), machine learning (ML), explainable AI (XAI) and integrating signal processing, large language models (LLMs), and natural language processing (NLP) in detecting DNA methylation offers an unprecedented opportunity to enhance the precision, scalability, and depth of epigenomic studies. In fact, AI models have demonstrated remarkable success in capturing intricate patterns in large and heterogeneous datasets. These capabilities position AI as a transformative tool for comprehensive DNA methylation analysis, with the potential to uncover new biological insights, improve disease diagnostics, and facilitate personalized medicine. Yet, the field of AI and DNA methylation is ever growing and changing. Thus, if a comprehensive review is to be done properly, reviewing the previous works is still relevant to see what has been done, but also be updated with the current developments is more challenging. Unlike the few previous reviews [9–11] that are more generic and have explored several DL models for various epigenetic modifications understanding [10, 11], including DNA methylation, chromatin interactions, and histone modifications, or have been focused on cancer epigenetics understanding via DNA methylation and multi-omics analysis [9]. In this review paper, through more than 30 generic studies reviewed about this field, beyond overviewing existing DL- and ML- based methods for DNA methylation analysis.

• We will shed light on the less explored paths including signal processing and image processing.

• We highlight the transformative potential of integrating the LLMs into DNA methylation analysis.

• We evaluate and benchmark the performance and computational efficiency of the existing AI-based models.

• We explore the impact of the feature selection and encoding schemes on performances.

• We highlight the implication of XAI in DNA methylation detection.

• We reveal the importance of multi-omic data integration to provide a more holistic understanding of the role of DNA methylation in gene regulation and diseases.

General overview of the discussed technologies

Deep learning and machine learning based approaches

Up to now many researches have explored various machine and DL models to extract meaningful information from DNA methylation data. Introduced by Angermueller et al. [12], DeepCpG employs a convolutional neural network (CNN) architecture to discern DNA methylation patterns and elucidate epigenetic regulatory mechanisms. DeepCpG’s strength lies in its capacity to handle missing data. Through sophisticated imputation techniques, the model achieves precise predictions even in regions with incomplete methylation data, surpassing traditional linear models and other ML approaches. MethylNet [13] is another DL framework designed for DNA methylation analysis that integrates multiple tasks, including age prediction, identifying factors associated with smoking, and pan-cancer classification. The framework uses variational auto-encoders to extract biologically meaningful features and utilizes these encodings to make accurate predictions across common DNA methylation analysis tasks. Evaluated on 34 datasets from 9500 samples for various prediction tasks, MethylNet confirmed its superiority over other methods and demonstrated its ability to accurately predict age, estimate cellular proportions, and classify cancer subtypes. Paper [14] introduces Deep6mA, a DL framework designed to predict DNA 6mA methylation sites by integrating CNN and bidirectional long short-term memory networks (LSTM). The research reveals that 6mA methylation modifications are concentrated within TATA boxes of promoters and exhibit conservation across various species. Trained initially on rice data, the model achieves high prediction accuracy exceeding 90% for Arabidopsis thaliana, Fragaria vesca, and Rosa chinensis. Notably, the occurrence of 6mA is particularly associated with GAGG motifs, suggesting conservation of sequences adjacent to 6mA sites. The BiLSTM-5mC model [15] represents a significant advancement in the field of computational biology, specifically designed to accurately identify 5mC sites within genome-wide DNA promoters of small cell lung cancer (SCLC) cell lines. Methodologically, the BiLSTM-5mC model integrates one-hot and nucleotide property and frequency (NPF) encoding strategies to capture sequence-order and position-specific information from DNA sequences. Trained on a benchmark dataset associated with SCLC, results validate the BiLSTM-5mC model’s efficacy, consistently demonstrating competitive performance in identifying 5mC sites within genome-wide DNA promoters. The study [16] presents two computational models, LA6mA and AL6mA, designed to identify DNA N6-methyladenine (6mA) sites. These models utilize LSTM networks and attention mechanisms to capture both short- and long-range information, offering biologically meaningful insights. The models demonstrate outstanding performance, with LA6mA and AL6mA achieving area under the receiver operating characteristic curve (AUROC) values of 0.962 and 0.966, and 0.945 and 0.941, respectively, on two benchmark datasets. A key feature of these models is the attention layer, which enhances prediction accuracy by focusing on crucial nucleotide positions that contribute to the identification of 6mA sites. The study [17] introduces DeepTorrent, an approach for predicting 4mC (N4-methylcytosine) sites in genomic sequences. DeepTorrent employs a combination of four different feature encoding schemes to represent raw DNA sequences, which are then processed through a DL framework consisting of multi-layer CNNs with an inception module, BLSTMs, and an attention layer. This architecture enables the model to capture higher-order feature representations and optimally encode inputs while mitigating overfitting. The model’s hyperparameters are fine-tuned using Bayesian optimization, ensuring optimal performance. The results of the study demonstrate that DeepTorrent significantly excels in both cross-validation and species-specific models, showing improved performance across several key metrics. However, the study acknowledges some potential limitations, such as the complexity of the framework, and the challenges of preventing overfitting when using small datasets. In [18], another computational model, namely i4mC-w2vec, has been developed to identify N4-methylcytosine (4mC) sites in DNA sequences using advanced word embedding and CNN. The i4mC-w2vec model has shown its effectiveness across balanced and imbalanced class datasets, which is a notable improvement over other state of arts methods. A significant aspect of this model is the use of the word2vec technique, which has proven to be more effective than the one-hot encoding method in feature representation of 4mC sites. Despite its successes, the i4mC-w2vec model does have limitations. Its reliance on high-quality datasets for training and testing means that its performance may vary with different datasets or species. Furthermore, the study focused only on two species, F. vesca and R. chinensis, suggesting that the model’s generalizability to other species remains to be fully explored. Additionally, the use of word embeddings and feature representations requires careful tuning of hyperparameters to achieve optimal performance. The study [19] proposes an omics-level attention-based framework, namely moSCminer, for cell subtype prediction using single-cell multi-omics datasets. Three publicly available single-cell multi-omics datasets were used to evaluate the performance of moSCminer, accordingly the results achieved underscore the model’s ability to provide meaningful biological insights. In fact, the attention module and the integration of omics data from different biological perspectives help in improving prediction performance. Besides, the authors stated the computational cost and the lack of single-cell multi-omics datasets as a significant challenges. The study [20] explores the use of semi-supervised learning (SSL) techniques combined with multinomial logistic regression to improve the classification of DNA methylation data, particularly in the context of central nervous system tumors prediction. SSL models effectively leverage large amounts of publicly available, unlabeled methylation data to label or relabel samples, providing additional training examples for supervised models, especially for rare tumor types. Among the 11 developed models, the self-training with editing using support vector machine (SETRED-SVM) model, outperformed other SSL models in labeling methylation subclasses. While the SSL models improved classification accuracy and provided ground truth labels for validation, the study acknowledged limitations related to class imbalance and incomplete reference sets, particularly for rare tumor types. Within the same context, the study [21] proposes another computational strategy, known as the mixture regression model (MRM), to impute missing or unmeasured CpG methylation values. This approach integrates three kinds of information: neighboring CpGs, local methylation patterns across subjects, and local methylation patterns across multiple genomic regions within a subject. The model was tested on simulated whole genome bisulfite sequencing (WGBS) data and real data from the MuTHER cohort, showing better performance in identifying novel differentially methylated CpGs (DMCs), particularly in an epigenomic association study related to bone mineral density (BMD), where it identified 2459 significant DMCs, with 2452 being new discoveries. Moreover, MRM proved robust to noise and missing rate. However, the MRM method has some limitations, including the requirement for a large number of samples and the need for careful tuning of model parameters. Additionally, its performance may be influenced by the quality of the input data, and the computational complexity can be high for large datasets. In the same line, authors in [22] proposed StableDNAm, a DNA methylation prediction model that incorporates a feature fusion approach, an adaptive feature correction technology, and contrastive learning, using a transformer encoder to improve prediction accuracy and robustness. The model’s stability and robustness are attributed to its ability to learn robust feature representations from diverse class samples, effective fine-tuning based on pre-training, and the fusion of multiple features. The results demonstrated the performances of StableDNAm model, which excels in 12 out of 17 datasets. Besides, while the study does not discuss the limitations of the StableDNAm model explicitly, it implies that the model may still be affected by sparse datasets, although the contrastive learning strategy helps to mitigate this impact. Within the same setting, the paper [23] proposed another strategy to accurately predict DNA N4-methylcytosine (4mC) sites. The proposed method utilizes five types of sequence features: one-hot binary (OHB), sequential nucleotide frequency, k-nucleotide frequency, k-spectrum nucleotide pair frequency, and PseDNC. The analysis of feature importance score distributions across datasets from six species reveals that OHB and PseDNC features consistently have high average scores, indicating their significance in the prediction process. An XGBoost machine is trained to rank all the features based on their importance scores. This ranking is then used in a feature selection scheme to reduce the feature space dimension and modeling complexities. The selected features are subsequently used to train a support vector machine (SVM) prediction model. Conclusions drawn from the study highlight that the proposed method is more accurate in predicting 4mC sites in DNA sequences. However, the study acknowledges some limitations, such as the high cost. Additionally, the limited size of the training data and the focus on sequence content features rather than biological characteristics are noted as constraints. By the same token, the study [24] aimed to compare the performance of various ML workflows for analyzing high-throughput DNA methylation data. The ML algorithms evaluated include Random Forests (RF), Elastic Net (ELNET), SVM, and boosted ensemble trees. The study also assessed different calibration methods like Platt scaling, logistic regression, and Firth’s penalized logistic regression. The best overall two-stage workflow was identified as the multinomial ridge (MR)-calibrated SVM with linear kernels (LK), closely followed by the ridge-calibrated tuned RF. These findings offer crucial guidance on selecting and tuning ML workflows to generate well-calibrated class probability estimates for precision diagnostics. However, the study had several limitations, including evaluating only a limited range of ML algorithms and calibration methods, using a single dataset that may not generalize to all DNA methylation datasets, and not addressing the impact of batch effects. Additionally, extensive parameter tuning was required, and there was a potential bias towards the 450k methylation array. Similarly, the study [20] examines the effectiveness of ML techniques in predicting 5-methylcytosine (5mC) sites in genome-wide DNA promoters, employing a feature extraction method based on k-mers embeddings. The primary goal was to compare the performance of various ML algorithms, including XGBoost, RF, Deep Forest, and Deep Feedforward Neural Network, in accurately predicting these sites. The results indicated that the XGBoost model achieved the highest performance, and trinucleotide embeddings were found to be slightly better than di-nucleotide and nucleotide embeddings. The imbalance between negative and positive data is a common issue in modification site prediction, and this study is no exception. Additionally, the evaluation was conducted using a single dataset, and the study did not provide a detailed analysis of the biological significance of the predicted 5mC sites. It is worth to mention that 5-methylcytosine (5mC) plays a significant role in various biological processes and is associated with numerous human diseases. Therefore, accurately identifying the locations of 5mC in the genome is essential for understanding its biological functions and implications for health. Based on this, authors in [25] developed a computational approach called BERT-5mC, which leverages the power of ML to improve the efficiency and accuracy of 5mC detection. Another interesting approach called DeepMethylation, which leverages DL techniques to predict methylation sites is found in [26], DeepMethylation operates by first encoding DNA sequences using word embedding techniques and GloVe (global vectors for word representation). Next, the model employs dilated convolution and a transformer encoder to extract relevant features from the encoded sequences. These advanced techniques allow the model to capture complex patterns and dependencies within the DNA sequences. Finally, the model applies full connection and softmax operators to predict the methylation sites. DeepMethylation demonstrates impressive performance, achieving an accuracy of 97.8% on the 5mC dataset, and 95.8% on the m1A dataset.

L‌LMs and NLP based approaches

NLP can be a powerful tool for studying DNA methylation and can help researchers to identify patterns and relationships in large amounts of data that might not be apparent using traditional analytical methods [27]. Inspired by the inherent characteristics shared between natural languages and biological sequences [8, 28, 29] as well as the recent advances in AI [30], LLMs have shown great promise in deciphering the biological mechanisms underlying the epigenetics [31, 32], enabling impressive performances, better understanding of complex structures, and deeper insights into DNA methylation [33–37]. Owing to NLP and LLMs several approaches have been proposed in the literature. Based on BottleneckMLP and transformer model, paper [38] introduced an interesting strategy, to deal with the common issue encountered by most of models, when failing to generalize to new samples and unseen sites. Within the same context, of addressing the significant challenge of ensuring high generalizability across various tasks, paper [32] introduced CpGPT framework specifically dedicated to age and mortality prediction tasks. To this end, CpGPT is leveraged with the Transformer++ architecture based on the self-attention mechanism to effectively extract discriminative biological variation for an accurate classification. Another strength of CpGPT is its ability to address missing data. The study [39] introduces iDNA-ABF, a multi-scale biological language learning model that considers genomic sequences as ”biological texts.” The model is pre-trained with large-scale genomic sequences to learn biological contextual semantics and employs a multi-scale processing strategy to capture discriminative methylation information from different scales. Adversarial training and transfer learning are incorporated to enhance the predictive performance and robustness of the model. Moreover, by exploring local sequential characteristics through attention mechanisms, the iDNA-ABF model can learn functional semantics from millions of background genome sequences, thereby discovering biologically meaningful sequential regions. Another study for identifying RNA methylation sites using NLP and DL approaches is found in [26]. This approach, namely EMDLP, uses three encoding methods (RNA word embedding, one-hot encoding, and RGloVe) to represent RNA sequences, followed by a dilated convolutional bidirectional LSTM network (DCB) model to extract features for methylation site prediction. The three encoding methods are integrated by a soft vote to obtain better predictive performance. The EMDLP predictor achieves high performance in identifying RNA methylation sites, with AUROC values of 95.56% and 85.24% for m1A and m6A, respectively. The iDNA-OpenPrompt model [40] is another cutting-edge approach designed to identify DNA methylation sites. The iDNA-OpenPrompt model incorporates a DNA vocabulary library, a BERT tokenizer, and specific label words to enhance the precision of its predictions. The iDNA-OpenPrompt model has demonstrated its performances across 17 benchmark datasets that include various species and three types of DNA methylation modifications (4mC, 5hmC, 6mA). However, a key limitation of the iDNA-OpenPrompt model is the need for careful tuning of the DNA vocabulary and verbalizer label words to achieve optimal results. Moreover, the process of manually generating these vocabularies remains a challenge in bioinformatics. MuLan-Methyl [41] is a DL framework that leverages five popular transformer-based language models to predict DNA methylation sites. Hence, by collectively utilizing the predictions from all five models, MuLan-Methyl achieves a comprehensive and robust methylation status prediction. MuLan-Methyl works by identifying methylation sites for three different types of DNA methylation: N6-adenine, N4-cytosine, and 5-hydroxymethylcytosine. The performance and evaluation of MuLan-Methyl demonstrate its effectiveness, as it reports excellent results on a benchmark dataset. The model successfully captures characteristic differences between different species that are pertinent to methylation, showcasing its ability to generalize across diverse biological contexts. 4mCNLP-Deep [27], is another DL-based approach designed for predicting 4mC and non-4mC sites in the genome by performing the DNA sequence as NLP and using a CNN model. The model’s performance underscores the model’s superior capability in predicting 4mC sites with high accuracy. However, despite its promising results, the model’s performance may be limited by the quality and complexity of the genomic data used for training and testing. Additionally, the study did not explore the model’s performance on other datasets or organisms, indicating a need for further research to expand its applicability. Inspired by the image processing field, the iDNA-ITLM model proposed a novel strategy for identifying DNA methylation sites, in which the DNA sequence is represented as a 2D matrix, a form like image [42]. The main contribution of this paper is to identify the DNA methylation for three distinct modification types including 4mC, 5hmC, and 6mA, while most of approaches are restricted to a singular type of DNA methylation. For this purpose, this model integrates the masked language model (MML), the attention mechanism along with a BERT model for an accurate prediction of DNA and RNA methylation. This method was evaluated on 17 benchmark datasets encompassing different species showcasing its high performances and generalizability to other form of methylations

Signal processing based approaches

DNA methylation data can be also represented as signals where the signal values correspond to the degree of DNA methylation at different genomic locations. Therefore, signal processing techniques can be used to extract relevant features, which can be used to identify genomic regions that are associated with specific biological processes or clinical outcomes [43]. There are several benefits for using signal processing techniques for DNA methylation detection. It is important to note that DNA methylation data may contain noise. Consequently using signal processing techniques can help in denoising the data, enabling us to focus on the true underlying signals related to DNA methylation patterns [44]. Moreover, signal processing techniques can be used to visualize DNA methylation data in meaningful ways, such as heatmaps or frequency-domain plots, which can help researchers to identify patterns or relationships that might not be apparent using other analytical methods [45]. Many researches have explored the use of signal processing methods to extract meaningful information from DNA methylation data. The study [45] introduced a short-time Fourier transform (STFT) based approach for detecting CpG islands. The proposed algorithm employs STFT to extract periodicity features associated with CpG islands, converting DNA sequences into numerical representations to compute the power spectrum of windowed sequences. The algorithm was tested on various datasets, including a Homo sapiens chromosome X region (accession number L44140) consisting of 219 447 base pairs, as well as 100 DNA sequences from human, mouse, and fish species. The results were benchmarked against existing methods. In the initial test on the chromosome X sequence, the algorithm successfully identified all 17 CpG islands present, albeit with some false positives. On larger datasets, the overall findings indicate that the STFT-based algorithm provides robust and reliable CpG island detection across diverse species. DeepMod2 [46] is a cutting-edge methylation detection tool that leverages BiLSTM and Transformer-based deep neural networks to analyze 5mC methylation in Nanopore sequencing data from three genomes: HG002, HG003, and HG004. Using DeepMod2, researchers identified 19 out of 21 known imprinting control regions (ICRs) across these genomes, demonstrating substantial overlap with previously characterized ICRs. One of DeepMod2’s unique capabilities is its ability to analyze methylation from unmapped reads or unaligned read segments. It also demonstrates robustness in detecting methylation in targeted regions, particularly in imprinting control regions and differentially methylated sites. Despite its strengths, the study identified limitations. The DeepMod2 model may overestimate methylation at certain CpG sites, potentially due to biases in signal summary statistics. The analysis was restricted to specific genomes and datasets, and the tool’s performance on other species was not evaluated. The study [47] developed another method for DNA methylation site prediction by integrating four feature extraction techniques: k-gram, multivariate mutual information (MMI), discrete wavelet transform (DWT), and pseudo amino acid composition (PseAAC). A sparse Bayesian learning model was then trained to classify methylation states based on the extracted features. To address data imbalance, the SMOTE algorithm was used, and the method was evaluated through Jackknife cross-validation and holdout validation. Three datasets were employed in this study: a benchmark dataset from MethDB, containing 20 236 records of 5-methyl-cytosine data and 6312 individual methylation patterns, and two scBS-seq datasets profiling mouse embryonic stem cells (mESCs) under 2i and serum culture conditions. Feature selection experiments revealed that discrete wavelet transform (DWT) and pseudo amino acid composition (PseAAC) contributed most significantly to the classifier’s performance, while k-gram and MMI added complementary insights. This finding underscores the importance of combining diverse feature extraction methods to capture the complex patterns associated with DNA methylation. DeepSignal-plant [48] is another advanced pipeline based signal processing designed to detect and quantify DNA methylation. The primary objective of the study was to develop a pipeline capable of accurately calling methylation and calculating methylation frequencies across CpG, CHG, and CHH contexts. DeepSignal-plant was evaluated using data from two model plants, Arabidopsis thaliana and Oryza sativa, and validated with additional datasets. Features were extracted for training using bidirectional recurrent neural network (BRNN) with LSTM models. DeepSignal-plant represents a significant advancement in plant epigenetic research. Its ability to profile more 5mC sites than bisulfite sequencing and detect methylation in complex genomic regions makes it a valuable tool for understanding plant epigenetics. DeepSignal [49], a DL signal processing-based method, has been developed to detect DNA methylation states from Nanopore sequencing reads. DeepSignal excels in detecting 6mA and 5mC methylation states at both the read level and genome level. DeepSignal employs a CNN and a BRNN to construct features. These features are concatenated and fed into a fully connected neural network to predict methylation states. DeepSignal achieves higher performance at both read and genome levels in detecting 6mA and 5mC methylation states compared to previous methods, achieving 90% above accuracy for detecting 5mC and 6mA using only 2 coverage of reads. The study [43] introduces wavelet screening (WS), a wavelet-based method designed for large-scale epigenome-wide association studies (EWAS), which have become essential in understanding the role of DNA methylation (DNAm) in complex diseases. WS delivers remarkable computational efficiency, detecting differentially methylated regions (DMRs) with high accuracy while using minimal computational resources. The study utilized data from the Norway Facial Cleft Study, a case-control study on orofacial clefts (OFCs) that included 750 cases and 1100 controls. WS demonstrated exceptional coverage, analyzing 98% of CpGs, a significant improvement over existing methods. The method identified 95 DMRs at a false discovery rate (FDR) of 0.05, including 89 true-positive regions and six false positives. Among these, 82 regions were linked to genes and loci associated with OFCs, while others represented novel findings that merit further validation. The study acknowledged some limitations, including the relatively small sample size of the Brazilian discovery cohort, which may explain discrepancies with previous findings.

XAI-based approaches

In DNA methylation analysis, AI models may identify important methylation sites or predict methylation patterns related to diseases, but it is often difficult to understand why the model made such predictions. In fields like genomics and medicine, where human expertise is essential for decision-making, XAI can bridge the gap between AI models and biological understanding. Hence, by making AI models more transparent and interpretable, XAI enhances trust, facilitates collaboration between AI systems and human experts, and leads to more actionable insights in both research and clinical settings. Within this context, the objective of the study [50] was to evaluate the performance of different ML algorithms in predicting methylation sites in proteins and to provide insights into the decision-making process of these models using XAI. The proposed method uses a combination of ML algorithms, including RF, SVMs, naive Bayes classifier, K-nearest neighbors, and fuzzy SVMs, to predict arginine methylation sites. The RF model achieved the highest accuracy of 88.4% and an AUC score of 0.94 in predicting methylation sites. XAI techniques are employed to interpret the predictions. However, the proposed method relies on a single dataset of experimentally validated methylated and unmethylated protein sequences, which may not be comprehensive or representative of all possible protein sequences. With the main aim of developing an accurate and explainable neural network architecture for estimating biological age from DNA methylation data, and to identify biological pathways associated with aging, another interesting work, namely XAI-AGE model is found in [51]. Based on the obtained results, the XAI-AGE model not only accurately estimated biological age but also provided interpretable results, highlighting key biological pathways. It identified significant pathways associated with aging, such as DNA repair, chromatin organization, and reproduction. Additionally, the model predicted biological age and identified critical pathways in human cell aging and rejuvenation, including the circadian clock pathway. However, despite its promising results, the XAI-AGE model has some limitations. It performed poorly for specific tissue types like blood cord, bone marrow, and esophagus, and showed limitations in datasets such as umbilical cord plasma transfusion data. Further limitations include class imbalances of tissues and age groups, batch effects, and curation bias in the Reactome Pathway Database. Focusing on handling missing values and improving model interpretability, a comprehensive approach for classifying healthy individuals and patients using Parkinson’s disease and schizophrenia as examples based on whole-blood DNA methylation data has been proposed in [52]. This method encompasses data harmonization, ML classification models, dimensionality reduction, imputation of missing values, and XAI. The best accuracy results, exceeding 95% for Parkinson’s disease, were achieved using tree ensembles. Additionally, XAI approaches have been useful in explaining model predictions from both population-level and individual perspectives. However, the number of samples in the training set and the complexity of diseases like schizophrenia can affect classification accuracy. Additionally, missing values in DNA methylation data is still a challenge, and imputation methods may not always be effective. While harmonization improves accuracy, its benefits are limited when preprocessing of training and test data is consistent. Moreover, the constructed classification models may not be globally optimal in terms of quality metrics. Arginine methylation, a critical post-translational modification, plays a significant role in various cellular processes, including gene regulation and signal transduction. Predicting arginine methylation sites in proteins is crucial for understanding these biological processes and developing therapeutic strategies. Existing methods often rely on extensive evolutionary, disorder, and structural information, which can be challenging to obtain. To address these limitations, the authors in [53] proposed PRMxAI, which leverages sequence-based features and a random forest classifier to accurately predict arginine methylation sites. The PRMxAI model extracts various sequence-based features from primary protein sequences, including amino acid composition (AAC), dipeptide composition (DPC), physicochemical properties (PP), and information theory-based features (ITB). The PRMxAI model demonstrates remarkable performance, achieving 87.17% accuracy on mono-methylarginine datasets and 90.40% accuracy on di-methylarginine datasets. Despite its success, the PRMxAI model presents some limitations, including the reliance on sequence-based features alone, without considering evolutionary, disorder, or structural information. Additionally, the study did not explore the potential benefits of incorporating fractal dimension analysis, which could further enhance the detection of self-similarities within amino acid sequences.

Benchmarking analysis

To compare AI-based tools for DNA methylation analysis in a fair and consistent way, we selected 30 methods published between 2018 and 2025. This allowed us to show how research in this area has evolved in recent years (see Table 1 and Fig. 1). We focused on recent, data-driven studies with potential for real-world application, and only included those that clearly reported performance measures such as accuracy, sensitivity, and specificity. We identified relevant studies through structured searches using keywords like ”DNA methylation,” ”ML,” ”DL,” ”epigenomics,” ”CpG prediction,” ”XAI,” ”LLMs,” ”NLP,” ”signal processing,” and ”AI.” The final selection offers a broad view of current methods, making it easier for readers to compare their strengths and weaknesses. Table 1 lists each study along with its reference, tool name (if available), publication year, targeted methylation sites (CpG, 5mC, 6mA, 4mC), technologies used (ML, DL, LLM, XAI, or signal processing), and the type of data integration (single-cell omics, genomic data, or multi-omics). Figure 1 shows their reported performances. Together, these resources highlight the diversity of methods, their capabilities, and their limitations, helping readers quickly assess the scope and relevance of each approach.

Table 1.

A summarized overview of the approaches involved in this review.

Study	Name	Year	DNA methylation sites				Technology	Data Integration			Generalizability
			CpG	6mA	5mC	4mC		Multi-omics	Single-cell omics	Whole-genome
[7]	DeepCpG	2017					DL
[54]		2018					Signal processing
[35]		2020					Signal processing
[55]		2020					ML
[56]	MethylNet	2020					DL
[57]	Mixture regression	2020					ML
[58]	4mCNLP-Deep	2021					DL
[59]	Deep6mA	2021					DL
[20]	BiLSTM-5mC	2021					DL
[60]		2021					DL
[61]	DeepTorrent	2021					DL/attention/transfer learning
[62]		2021					ML
[48]		2021					DL
[46]		2022					XAI
[49]	PRMxAI	2023					XAI
[63]	MuLan-methyl	2023					Transformers/LLMs
[64]	DeepMethylation	2023					Transformers/LLMs
[21]	moSCminer	2024					DL/ ATTENTION
[28]	RMSxAI	2024					XAI
[5]	DeepMod2	2024					Signal processing/transformer
[65]		2024					XAI
[66]	MethylGPT	2024					Transformers/LLMs
[24]	CpGPT	2024					Transformers/LLMs
[67]	iDNA-ITLM	2024					Transformers/LLMs
[40]	MethylProphet	2024					Transformers/LLMs
[68]	DeepSF-4mC	2024					DL/Transfer Learning
[69]	scWGBS-GPT	2025					Transformers/LLMs
[70]	Methyl-GP	2025					Transformers/LLM
[71]	MethPriorGCN	2025					Knowledge graph
[72]	MORE	2025					Knowledge graph

Figure 1.

Performance analysis and comparison of the approaches involved in this review.

Discussion, challenges, and vision

The landscape of DNA methylation detection technologies is rich with a variety of methods, each offering distinct advantages and facing specific limitations. As the field progresses, continued research and development in AI-driven methodologies promise to revolutionize our understanding of DNA methylation. Ultimately, through this review, we concluded that the integration of AI in detecting DNA methylation represents a significant advancement in the field of epigenetics. Our exploration into more than 30 AI-driven methods for DNA methylation highlights the transformative potential of AI in analyzing and interpreting complex genomic data, leading to a deeper understanding of gene regulation, disease mechanisms, and personalized medicine. However, despite the growing potential of AI in biomedical research, its application to DNA methylation analysis remains underdeveloped and still faces several challenges.

Robust performance metrics for imbalanced and noisy datasets

Evaluating the performance of AI models in DNA methylation analysis relies heavily on metrics such as accuracy, specificity, sensitivity and precision. Sensitivity, referred to as recall, calculates the proportion of true positive predictions among all positive predictions [73]. Precision is another performance metric that accounts for the accuracy of correct classifications made by a model [73]. Specificity, quantifies the proportion of true negative predictions among all negative predictions [73]. So far, while these metrics offer quantitative benchmarks, their robustness and interpretability are highly context-dependent, particularly when faced with varying dataset sizes, class imbalances, and experimental noise. A common issue arises when dealing with class imbalance, characterized by an underrepresented class referred to as the minority class compared to the other, referred to as the majority class [74]. The main concern with the class imbalance is finding the best compromise between sensitivity and specificity [73, 74]. In such cases, accuracy may not be chosen to assess the model’s performance [56, 73], as it ignores the imbalanced nature of methylation, while it is sensitive to noise, which is inherent in real-world contexts [59, 74–76]. Thus, relying on accuracy solely as an evaluation metric will result in a misleading assessment [73, 75], reflecting the model’s bias toward prioritizing the majority while ignoring the minority class [69, 74]. This issue reflects a broader movement toward more transparent and reproducible evaluation practices. Salmi et al. underscore the importance of evaluating AI on realistic, clinically sourced datasets, rather than clean and balanced academic benchmarks, to ensure applicability in real-world contexts [74]. Therefore, ensuring a balanced performance across all classes is critical [69]. Under such conditions, researchers have been increasingly turned to looking for the reliable evaluation metrics, such as area under the precision–recall curve (AUPRC) [69, 74, 75] and AUROC also known as the AUC-ROC or AUC [73, 75–77]. The AUC allow a comprehensive assessment of the model’s performance on the imbalanced dataset [75]. AUC focuses on the minority class and better reflects a model’s ability to identify rare but biologically meaningful methylation patterns [74]. The main reason that makes metrics such as AUC more suitable to evaluate the model’s performance in case of unbalanced data, is that it is measured based on the trade-off between sensitivity and specificity through different thresholds [56, 73]. The Pearson Matthews correlation coefficient (MCC) is another well-known metric for evaluating the model’s performance, especially for imbalanced datasets [76, 78]. Besides, the F-value and balanced accuracy (BA) are also renowned to be suitable for classification tasks in the class imbalance scenarios [61, 69, 73, 74, 77]. The BA metric, measured by averaging the sensitivity and specificity [73], has underscored its ability to discriminate both majority and minority classes [69]. F-value, namely the F-measure, defined as a combination of precision and sensitivity, is another commonly used metric to evaluate the model’s performance for the imbalanced tasks, but is underexplored in the context of DNA methylation [61, 69, 73, 74]. Another set of quietly less-used metrics that are not affected by the imbalance between the classes contains G-mean [73, 74, 79]. G-mean is calculated as the geometric mean of sensitivity and specificity [73]. Complicating matters further, the problem of data imbalance often intersects with noise [80]. Considering the critical impact of noise on overall performance [81, 82]. In such contexts, traditional performance metrics may fail entirely [82]. Liu et al. argue that sample selection bias, exacerbated by noisy labels, can skew results and prevent fair assessment of a model’s true capability. Their proposed solution involves bias mitigation in sample selection, which helps stabilize performance across metrics under noisy conditions [80]. Within the same context, Zhang and his co-workers leveraged two metrics, namely, normalized reads concordance (NRC) and normalized CpGs concordance (NCC), exhibiting good performance in coping with noisy data of distinguishing different methylation patterns [82]. In the same line, to assess the effect of noise, a metric, namely noise parameter, has been defined in [83]. Besides, another challenge lies in dealing with both imbalanced and multi-class learning. Because methods are often evaluated on distinct tasks, each with different data distributions and performance expectations. Yet, although they have been proven useful in mitigating imbalance in binary class contexts, several metrics fail at assessing multi-class imbalanced datasets, including the ROC curve, the accuracy, and the F-score [79]. Therefore, a robust performance evaluation metric should move beyond a narrow focus on a binary dataset. Reporting practices should provide a global multi-class performance metric, which allows a more comprehensive assessment of the model’s performance independently of the dataset’s class distribution [56], so that overrepresented classes have no further effect on the overall performance results [79]. Recently, the imbalanced multi-class classification performance (IMCP) curve and also the area under the IMCP curve (AU(IMCP)) have solved this problem by reporting the model’s performance in case of multiclass imbalance independently of the class distribution [79].

Interpretability, reproducibility, and explainability concerns

Interpretability, the ability to explain the inner workings of AI-driven technologies, is of ultimate importance, not only for uncovering the biological mechanisms underlying traits and diseases [84], but also contributes to the model’s trustworthiness [85]. Hence, while achieving better predictions is crucial, interpretability remains a main concern in DNA methylation detection [32]. In response to this, MethylGPT demonstrated that transformer architectures, along with their integrated attention mechanism, could help in effective modeling of DNA methylation patterns while maintaining biological interpretability [30]. Within the same settings, the CpG Pretrained Transformer (CpGPT) addresses this by enabling the attention mechanism leveraged by the transformer-based deep neural network to highlight the methylation sites, offering insights into which methylated sites are most influential in a given sample’s methylation profile [32]. So far, the emergence of LLMs and transformers equipped with their built-in attention mechanisms has become a hotspot in the realm of AI [86]. Beyond their performances, their built-in attention mechanisms offer a more intuitive and accessible window into inner-working processes, potentially making LLMs more trustworthy [86, 87]. Basically, attention-based interpretations are inherently leveraged in transformer architectures, but can also be incorporated in graph neural network (GNN), through the use of a graph transformer architecture, which has been introduced as a promising new avenue for designing interpretable DL applications [85, 88]. However, beyond the disconnect between theoretical promise and practical usability, an issue that may arise is the ”reproducibility” of the interpretations [86]. Because without reproducibility mechanisms, such models risk being sidelined or misused in practice. In fact, interpretability is tightly linked to reproducibility: if researchers or practitioners cannot understand the inner workings of the model, it becomes nearly impossible to replicate its behavior across datasets, institutions, or populations. Hence, if the aim of a study is to understand the underlying biological mechanism, then it is crucial to evaluate the reproducibility of interpretations over multiple runs [85]. Therefore, failure to ensure reproducibility can lead to inconsistent or even dangerous results. Based on this, studies will need to assess the consistency of the results over multiple tests. However, most of studies lack this feature [85]. Consequently, AI developers must consider interpretability and reproducibility as primary design goals [87, 89]. Within this line, as transformer architectures and LLMs continue to revolutionize critical fields like bioinformatics, healthcare, and computational epigenetics [89, 90]. Yet, alongside these powerful advances comes another growing concern: the models often act as ”black boxes,” delivering impressive predictions with minimal insight into why those predictions are made. Yet, in critical settings, such as epigenomics, where interpretability is vital, transparency, also referred to as explainability, is also paramount [86]. In fact, understanding the inner workings of the model (the how) and the rationale behind a model’s prediction (the why) a model predicts a particular methylation pattern, for instance, is vital for reproducibility and validation [85, 91]. However, LLMs still struggle with transparency concerns that arise in critical applications, due to their intricate complexity and opacity in AI model decision making [87, 89, 90]. In fact, a lack of visibility on their training data and decision-making process makes their explainability challenging [87, 89]. Hence, there is a pressing concern towards elucidating their inherently opaque nature. This involves an ongoing effort to uncover and understand the underlying logic behind their decision-making process [89, 90]. In this context, XAI can help bridge this gap by incorporating explainability in the design of LLM-based frameworks [87, 89]. Considering this fact, developing XAI frameworks for LLMs should become a forthcoming need for building more efficient, trustworthy, and transparent AI systems [89–92]. XAI stands for a range of techniques that hold great promise in bridging the gap between complex LLMs models and human understanding [87, 88, 90, 92, 93]. Specifically, why a model reaches a particular decision [90, 91]. So far, while the transparency issues have not yet been fully fixed for classical AI models, the increased model complexity of LLMs has introduced novel challenges for XAI techniques [92]. In light of the above considerations, in the context of LLMs, two challenges are raised: (i) Leveraging XAI for enhancing interpretable LLMs, and (ii) Boosting XAI through interpretable LLMs [92]. Yet, while attention-based AI models hold great promise in enhancing interpretability, they specifically rely on identifying which patterns contribute most to the outcomes, but not necessarily how this information has been interpreted [86]. In such a way, attention mechanisms fail in establishing a correlation between the inner workings of the model and its outcomes [86]. In fact, this shortcoming hinders the practicability and usability of attention-based models [86]. To address this issue, causality also known as cause and effect, can be a promising path towards enhancing AI models interpretability and explainability, through understanding the correlation between cause and effect [71, 77, 86, 94]. Hence, understanding the causal relationships between the input data (the why) and the model’s outcome (the what) can lead to designing better AI models [71, 77, 86, 94]. However, incorporating causal mechanisms inside LLMs is challenging due to their complexity and the lack of visibility of training data [86]. Moreover, the adoption of XAI in LLMs and transformers is still challenging given the opacity of training data, the lack of robust XAI metrics, ground truths and benchmark datasets and studies, which necessitate more thorough efforts [86]. In summary, Our review analysis highlight that explainable and interpretable AI (XIAI) adoption in LLMs is not possible without a domain-AI experts collaboration [71, 86, 94]. The need for a more integrated collaboration and involvement that better aligns the expertise of domain experts (biologists, end-users, clinicians) along with the assistance of AI is a critical issue [95]. The emerging Interpretable AI techniques with attention mechanisms has become a trend as the most dominant approach for interpretation alongside XAI [86]. Another promising venue lies in combining LLMs and transformers with causal mechanisms, to better understand the inner-working of these models [86]. Within the same context, integrating XIAI into the design of LLMs is paramount [86].

Variability across different studies

An in-depth examination of the existing literature highlights substantial methodological variability among studies utilizing AI models for DNA methylation analysis. These differences significantly influence the performance, reproducibility, and generalizability of proposed frameworks. Factors like noisy data, missing data, batch effects, or technical artifacts can significantly impact data quality [52, 59, 96]. Pre-processing steps are required to alleviate these issues [52, 59, 96]. However, different pre-processing approaches will lead to varying transformations impacting the performances and final outcomes [59]. Hence, generating high-quality data still remains a challenging task [54, 97]. Yet, it is worth noting that methylation data are also prone to the issue of missing values [65]. Over the last few years, several tools have been explored for the purpose of DNA methylation missing values imputation [32, 52, 54, 65, 97–99]. However, imputing regions with high variability is still a critical issue [97]. The running time for imputing large datasets is also a key issue [98]. Another issue concerns cross-species and tissue types DNA methylation imputation [32, 99]. One notable source of heterogeneity lies in technical artifacts arising from variations in instrumentation, Laboratory settings, or any other experimental conditions [100–103]. These effects present a major challenge, which might otherwise interfere with the reliable analysis of DNA methylation data, or even misleading biological interpretations and downstream clinical applications if not corrected [100, 101, 103]. Hence, the correction of batch effects is still a hotspot area of research [100–103]. However, despite the wide availability of various approaches that have been devoted to mitigating batch effects, existing batch correction approaches often struggle in recovering the biologically meaningful patterns of DNA methylation data [102]. Sample size diversity is another significant contributing factor to variation [59]. Ranging from limits in acquiring a sufficient number of samples [52, 104] to the estimation of the optimal sample size [105]. Many studies have considered that variability in findings across studies is most likely linked to sample size [104–111]. Highlighting the need for further investigations and targeted approaches in limited sample size settings [104, 107, 108] Overall, while these different studies reported in this review reflect a rapidly evolving field characterized by increasing methodological and broader biological applicability, the lack of standardized datasets and benchmarking protocols introduces substantial variability that complicates the interpretation of the outcomes and comparative assessments [104]. Hence, future works should prioritize the development of benchmarking frameworks, support for multi-modal inputs, and clinically validated evaluation pipelines to ensure fair and reproducible progress in methylation-based AI systems.

Practical challenges

In addition to methodological innovations, practical considerations are essential for the real-world deployment of AI-powered DNA methylation tools. Large-scale DNA methylation datasets, including the integration of multi-omics data into DNA methylation research and the coverage of data across the genome, have emerged as a vital strategy to unravel more comprehensive insights into the biological mechanisms underlying disease development [3, 112, 113]. However, handling such large-scale datasets demands, raises many challenges, increasing the need for efficient computational platforms to store and analyze this data [3, 112–114]. In fact, scalability remains a pressing concern, particularly as datasets continue to grow in size and complexity [115]. Cloud-based platforms such as Google Cloud may bridge this gap by offering scalable and effective solutions for managing large-scale data [3, 112, 113, 116]. However, despite the great promise that cloud computing holds in providing scalable and large-scale data analysis capacity, these platforms pose accessibility challenges, especially in under-resourced laboratories [112, 114]. To fully realize the potential of cloud computing, new approaches are needed to help provide equitable access to these resources [112, 114]. Besides, federated learning holds great potential by enabling scalable and effective large-scale methylation data analysis from distributed sources [117]. However, privacy issues are a main concern [115]. Big data analytics involving high-performance distributed computing tools’ capabilities, such as Apache Spark, are among the best practices that have significantly contributed to offering scalability and leveraging efficient pipelines for handling large-scale genomic datasets [115, 116, 118]. However, in the realm of epigenetics, their integration remains limited [62, 118]. Therefore, Spark could be a prime choice to be used to develop scalable DL-based frameworks [118]. Moreover, successful solutions should be deployed on the Cloud-based platforms, with the ultimate aim of realizing accessibility to everyone [118].

Generalizability in DNA methylation

Generalizability, the capacity of a model to perform well across varying datasets, species, tissues, and experimental condition, is a crucial requirement in DNA methylation analysis, particularly for clinical translation and cross-species comparative studies. The field is moving beyond single-task models toward more versatile and scalable frameworks that incorporate multiple learning paradigms, including transfer learning, generative AI, foundation models, federated learning, multi-feature fusion, and knowledge graph integration. Transfer learning has emerged as a powerful strategy to boost model robustness and adaptability, especially in data-scarce settings. In DeepSF-4mC, transfer learning is leveraged to enhance the prediction of 4mC methylation sites by pretraining on large-scale sequence features and fine-tuning on more specific tasks [58]. Similarly, Luo and his co-workers demonstrated that transfer learning substantially improves the cross-cohort performance of epigenetic clocks by mitigating cohort-specific biases and learning shared epigenetic representations [119]. Federated settings further extend this capability: Wen and Li [120] demonstrate how federated transfer learning with differential privacy enables secure and scalable methylation modeling across distributed multi-omics datasets. Inspired by large-scale language models, foundation models such as CpGPT [32], MethylGPT [30], and scWGBS-GPT [121] are designed to encode vast methylome data into general-purpose embeddings that capture both short- and long-range CpG interactions. These models use pretraining on large genomic corpora and fine-tuning for task-specific methylation predictions. For example, CpGPT adapts GPT-style transformer architectures to DNA sequences, allowing zero-shot and few-shot learning on unseen datasets [32]. MethylProphet further extends this idea by incorporating gene context as a guiding factor for methylation prediction across different genomic regions [38]. In parallel, generative AI has been harnessed to simulate methylation landscapes and infer probabilistic methylation profiles. Yan et al. [100] presents a generative model that accounts for both local and distal CpG interactions, which enhances interpretability and provides data augmentation capabilities for underrepresented genomic regions. Combining heterogeneous data features can lead to more robust methylation predictions. Methyl-GP [36], e.g. integrates multiple sequence-derived and structural representations using transformer-based encoders. Likewise, MethPriorGCN [81] adopts a multi-feature fusion strategy by integrating sequence, epigenomic, and graph-derived features, improving performance on cell-specific and disease-relevant tasks. These models demonstrate that leveraging complementary feature spaces enables richer and more generalizable methylation modeling. Incorporating knowledge graphs into methylation prediction adds another dimension of interpretability and prior-informed learning. MethPriorGCN [81] utilizes graph convolutional networks built on curated biological knowledge, guiding the learning process with known gene interactions, pathways, and disease ontologies. This hybrid approach enhances generalization across biomedical domains by embedding structured domain knowledge directly into model architectures.

Genomic data integration in DNA methylation

Profiling the whole genome and integrating sequence-based features with additional genomic data has proven essential for enhancing the predictive performance of DNA methylation models as well [36, 72]. Xie and his collaborators have demonstrated that combining DNA sequence information with additional biological features, including contextual genomic features, significantly improves the generalizability and accuracy of methylation site prediction across datasets. In fact, this multi-feature fusion strategy allows models to capture complex patterns that are not evident from sequence data alone, leading to more accurate predictions [36]. In the field of neuro-oncology, authors in [64] have recognized the utility of genome-wide DNA methylation profiling, emphasizing its ability to reliably classify tumor subtypes [64]. These patterns serve as methylation signatures that guide clinical decision-making and contribute to more precise predictions [64]. Beyond classification, in broader clinical contexts, genome-wide DNAm profiling methylation have been used to build robust survival prediction models, underscoring the potential of genome-wide analysis of epigenomic modification in informing patient outcomes, offering valuable insight into disease prognosis and showcasing the translational potential of methylome analysis [122]. Besides, in [123], a comprehensive genome-wide methylation profiling has allowed researchers to discern novel hypermethylated loci across the whole genome, highlighting the importance of genome-wide DNA methylation analysis in biomarker discovery [123]. These findings highlight the translational potential of large-scale methylation analyses for biomarker discovery and personalized medicine. However, while these efforts have illuminated numerous disease-relevant methylation signatures, some key limitations still persist. Challenges remain in detecting methylation across all genomic regions with equal fidelity. Certain genomic regions remain difficult to interrogate due to challenges inherent in sequencing technologies and the intrinsic biases of methylation detection algorithms—particularly when long-read sequencing technologies are used [124, 125]. As has been noted by Sigurpalsdottir and his co-authors, a comparative study of long-read sequencing platforms reveals that variability in detection sensitivity and mapping efficiency can impact the reliability of methylation analysis, particularly in structurally complex regions of the genome [124]. Within the same context, Liu et al. [125] evaluated multiple sequencing platforms and highlighted inconsistencies in methylation detection due to some significant inherent variability in sequencing performances, reinforcing the need for standardization of sequencing platforms and data integration pipelines of cross-platform technologies in methylome studies [80]. Collectively, these insights advocate for hybrid approaches that combine advanced modeling techniques with cross-platform harmonization to ensure comprehensive, accurate, and clinically meaningful methylation analysis.

Multi-omics data integration in DNA methylation

Multi-omics integration holds transformative potential for improving the robustness, generalizability, and interpretability of predictive models [70]. Hence, by leveraging diverse biological modalities—such as genomics, epigenomics, transcriptomics, and proteomics—multi-omics approaches allow for a more comprehensive characterization of complex traits and disease mechanisms, while studying disease from various perspectives [70, 126]. One of the key challenges in DNA methylation prediction is the limited generalizability. To address this issue, as highlighted by Ni et al. [81], combining heterogeneous biological features, particularly DNA methylation and prior knowledge graphs, can guide personalized medicine and improve model reliability across different contexts. In fact, this multi-omics integration enables the capture of biological distinctive features that are often missed by single-omics analyses [126, 127]. This boosts new discoveries by unlocking hidden knowledge in multi-omics data [127]. In the same line, Yan et al. [70] and Wang [128] demonstrated that incorporating omics-specific knowledge by identifying biologically meaningful interactions across multi-omics data further strengthens biological interpretability. In the context of disease biomarker discovery, Yang et al. [129] show that integrating multiple omics modalities is expected to expand our knowledge of the etiology of diseases. Similarly, Zillich et al. [68] highlight the fact that integrating multi-omics into DNA methylation analysis would be of great interest for an in-depth understanding of underlying alterations associated with various complex traits, thereby offering new insights that may be remain obscured in single modalities. Another interesting work is found in [130], where Miao and his co-workers showcased that integrating multi-omics data can reveal more insights about disease subtypes. The advent of foundation models like CpGPT [32] and MethylGPT [30] also illustrates how multi-omic integration can support prediction in low-coverage methylation scenarios. Together, these examples substantiate the value of combining DNA methylation with other omics layers. Such integration not only mitigates the sparsity and noise associated with single-omics data but also enriches our capacity to model biological processes in a context-specific manner. Moving forward, models that can seamlessly incorporate these heterogeneous data types, while maintaining interpretability and scalability, are likely to drive the next wave of breakthroughs in systems biology and clinical translation. Moreover, MethylProphet [38] introduces a gene-contextual model that captures the relationship between DNA methylation and gene expression. This integrative approach provides not only higher predictive accuracy but also greater biological interpretability, particularly across diverse cancer types. Similarly, MOSCMiner [19] exemplifies the power of integrating three distinct omics datasets. Utilizing a cloud-based attention neural network, the model excels in classifying cell subtypes with higher precision showcasing the efficacy of multi-omics integration compared to single-omic approaches. Ruprecht and his collaborators further demonstrate that the multi-omics integration has proven to be extremely essential, especially in cancer research and precision medicine [113]. In summary, multi-omics integration, when supported by domain knowledge and powerful architectures such as graph-based architectures, substantially advances our ability to generate biologically relevant insights and build robust, generalizable predictive models across heterogeneous populations and disease contexts. However, despite its promise, multi-omics data integration faces several unresolved challenges that must be addressed to fully realize its potential [131]. Developing approaches to deal with the ”Big p, small n” problem raised when dealing with data with large dimensionality and small size has been a major concern in many fields in recent years [66, 120]. To deal with this problem, federated learning frameworks hold great promise for handling distributed and sensitive data sources, while maintaining performance to ensure the scalability and privacy-preserving potential of multi-omics modeling [120].

Single-cell methylation integration

Deciphering intercellular epigenetic modifications is critical for understanding cellular heterogeneity [57]. Recent advancements in single-cell DNA methylation technologies have revolutionized our understanding of cellular heterogeneity, epigenomics, and disease progression [72]. However, challenges such as data sparsity are of main concern. In fact, while technologies such as single-cell genome-wide bisulfite sequencing protocols offer powerful avenues for analyzing cellular heterogeneity and identifying functionally relevant methylation patterns, these technologies are limited by the data sparsity [57, 67, 132, 133]. Specifically, a large number of methylation patterns are missed () [132, 133]. In response to this challenge, recent innovations on the widely used sequencing tool ”Infinium BeadChip” now allow the application to ultra-low input samples, including single cells. Yet, this improved workflow has yielded reliable data even from as few as five cells, giving rise to new opportunities for applications with limited DNA [67]. Besides, single-cell whole genome bisulfite sequencing (scWGBS), as a powerful platform to uncover the methylome heterogeneity at the cellular level and genome-wide scale, has offered the opportunity to process a large number of single cells [121, 134]. However, handling the ultra-long genomic sequences with sparse single cells, still remains a significant challenge [121, 134]. This consolidates the increasing need of developments in the field, focusing on innovations in computational pipelines to address this issue. In response to this, based on a generative foundational language model, scWGBS-GPT has emerged as a powerful tool in single-cell epigenomics for high-resolution data analysis [121]. Hence, by leveraging a transformer-based generative language model with cross-attention mechanisms, scWGBS-GPT could automatically identify critical and imperviously renowned cellular epigenetic markers [121]. Pursuing these achievements, collectively, many advancements have deepened our understanding of DNA single-cell epigenomic variability [72]. Very recently, a focuses on single-cell epigenomic technologies has shed light on the epigenomic heterogeneity of tumors and has unlocked our ability to decipher some of their mechanisms and understand their development at each stage [135]. Specifically, powerful single-cell methylome analyses have revealed that cell-type-specific patterns of 5mC are particularly crucial in contexts like identifying type and stage of tumor [63]. Hence, measuring two modifications encompassing 5hmC and 5mC from the same molecules jointly, is expected to provide us with more insights about different tumors [63]. However, decoding the cell-type-specific 5hmC is a challenging task due to the lack of high-throughput single-cell 5hmC sequencing methods [63]. Another challenging task was to measure two modalities simultaneously from thousands of single cells [63]. It is within this context that recent focus has been devoted for joint analysis of 5mC and 5hmC from the same genome at the same single cell, highlighting their potential in exposing distinct regulatory signatures across tumor cell types and developmental states, offering a more nuanced understanding of DNA methylation dynamics in complex tissues [63]. Assessing epigenetic heterogeneity and detection of regions with discriminative methylation levels across cells, known as variably methylated regions, is a critical issue in the analysis of single-cell genome-wide bisulfite sequencing data [132]. In fact, these distinctive methylated regions may serve as epigenetic signatures for identifying the cell types and facilitate the integration of multi-omics single-cell analysis [132]. Efforts have been advocated for an accurate variable methylated regions selection through various approaches [132]. However, selecting the features from a pre-defined genomic region is a common shortcoming of these methods [132]. Identifying these regions is challenging as they may occur everywhere on the genome with different sizes and in various contexts [132]. While it is worth mentioning that only 5% of the genome is covered by single-cell genome-wide DNA methylation, which makes the identification of these regions an extremely challenging task, with up to 95% of missing values [136]. Moreover, beyond sparsity, this task is more challenging with the occurrence of noise and biases inherent in single-cell technologies [132, 136]. To address these shortcomings, an interesting approach to efficiently detect variably methylated regions with no prior knowledge of their sizes or location has been introduced in [132]. Within the same context, another interesting approach is found in [136], in which instead of the whole-genome DNA sequencing, which costs much, with MethylTree, a low genomic coverage of just 2% has been sufficient for an accurate lineage tracing, showcasing the potential of epimutation-based methods in tracing cellular history and function [136]. Being inspired by this high-throughput method, predicting the most informative distinctive methylation sets across the genome would be of great interest as a future direction [136]. Yet, detecting inter-cellular heterogeneity and methylation quantification have become a trend for studying tumor progression [137]. Within this context, an improved approach that allows better discriminative methylated regions for more accurate inter-cellular heterogeneity detection is developed in [60]. Shen and his collaborators proposed another approach that allows an accurate identification of inter-cellular heterogeneity from single-cell DNA methylation data [132]. However, a more challenging and underrepresented task that has caught particular attention very recently is the intra-cellular heterogeneity detection and methylation quantification within a given cluster [137]. Hence, with the advance of single-cell sequencing technologies, it has become possible to study the intra-cellular heterogeneity variation at an unprecedented scale, even at the multi-omics levels [137].

Recommendations for advancing DNA methylation modeling: experimental design, data integration, and computational methodologies

Advances in sequencing and computational technologies are reshaping DNA methylation research, yet key challenges remain. Based on recent developments in the field, our review analysis reveals that several key recommendations can be drawn. Building up DL based frameworks to train the signals generated from the nanopore sequencing, as well as owing to signal processing techniques to capture meaningful features has become a hotspot in methylation research [6, 46, 48]. Besides, the lack of studies that are able to handle multiple types of DNA methylation modifications is still a crucial issue. Notably, our review analysis reveals that only limited studies have managed to handle three methylation sites (4mC, 5hmC, and 6mA) [39, 41, 100, 138]. Showcasing the superiority of LLMs for this task. The explainability of AI models in the context of biological research is still a significant challenge, as many existing models are ”black boxes,” making them difficult to extract meaningful biological insights from their predictions. Hence, developing methods to enhance the explainability of models in the context of DNA methylation is an ongoing area of research. To overcome this issue, the use of XAI could be greatly beneficial. XAI aims to make AI models more transparent and understandable to humans, addressing the ”black-box” nature of these complex models. In the context of DNA methylation detection, XAI can be employed to make models more explainable, helping researchers understand how the model arrives at its predictions and ensuring that the insights derived are trustworthy and scientifically valid. Therefore, as the field progresses, integrating XAI could lead to more refined, scalable, and accessible approaches, deepening our understanding of epigenetic mechanisms and their implications for health and disease. Furthermore, there is a lack of standardized frameworks for integrating multi-omic data, which is essential for fully understanding the biological implications of methylation modifications. It is worth noting that the integration of multi-omics data, including genomics, transcriptomics, proteomics, and metabolomics, has the potential to significantly enhance DNA methylation analysis by providing a more holistic view of gene regulation and cellular function. However, effectively combining these diverse datasets requires advanced techniques to extract meaningful patterns from complex and high-dimensional data. The recent introduction of LLMs built on transformer and attention mechanism or Pretrained LLMs like GPT into the analysis of DNA methylation data is particularly promising [6, 32]. These advanced models excel at capturing complex patterns in large datasets, offering new possibilities for detecting methylation patterns with remarkable precision. Although still in the early stages of application, LLMs models have the potential to address some of the limitations of earlier methods, particularly in managing vast amounts of sequencing data [28, 55] and uncovering subtle epigenetic changes that might otherwise go undetected [6, 32]. Their ability to process and learn from large sequences of data makes them particularly well-suited for the complexities of genomic analysis. The quality and diversity of datasets play a foundational role in developing generalizable and biologically meaningful methylation models. In fact, current models could benefit from high-quality data but would be significantly enhanced by the integration of other single-cell omics technologies, which in fact would allow for a more holistic understanding of epigenetic regulation [57, 72, 121]. Future efforts should also focus on developing frameworks for analyzing cellular heterogeneity, methylation signals data, and benefiting from more genomic insights [72]. Additionally, integrating clinical metadata—such as individual patient demographics, imaging, and physiological parameters—could enhance the model’s ability to predict disease risk or treatment response at the individual level [121]. Future works should also consider the increasing number and varying modalities encompassing Multi-omics data integration [36, 66, 131, 139]. This suggest, integrating sequence-based features with additional features, including coverage of data across the genome, as well as expanding the use of diverse sequencing platforms by incorporating electrical signal data from third-generation technologies like nanopore sequencing [36]. Yet, it is worth mentioning that exploring input data from different perspectives will lead to a more reliable learning process and assist methylation detection [36]. Therefore, the availability of data is critical to developing more efficient methods [80]. However, privacy is of great concern because the availability of data might not always be released due to ethical constraints [100, 120, 140]. Thus, achieving the trade-off between data availability and privacy preserving is still a critical issue [120, 139, 140]. To address these issues, federated learning frameworks [120, 140] are adept at offering scalable, privacy-preserving avenues for integrating data from disparate sources. Hence, it enables better performance without exposing sensitive patient-level data. Therefore, considering these techniques for DNA methylation analysis significantly results in the availability of training data while ensuring ethical and legal constraints [120]. Besides, scalability and missing data are also key issues in DNA methylation identification that may affect the findings [100]. Thus, mitigating these problems require appropriate techniques. To this end, Generative AI has recently emerged as a powerful paradigm for methylation imputation by modeling DNA methylomes [100]. In fact, these models have excelled in silico generation of synthetic methylomes, leading to data augmentation which may in fact improve the prediction performances and generalizability of the model. Moreover, integrating a generative AI with LLMs suggests several benefits over other DL models, encompassing learning complex dependencies in the methylation landscape but also enabling improved generalization across datasets, and superior interpretation of multi-modal data distribution [86, 100, 141]. In complement to this, knowledge graph, the process of incorporating DNA methylation data with prior biological knowledge to guide predictions offers a promising avenue for enhancing model interpretability and prediction reliability [81, 128]. Therefore, these knowledge-guided architectures enable capturing associated methylation signatures that reflect both biological priors and data-driven patterns, showcasing how topological representations of biological relationships can improve both predictive power and mechanistic understanding [142].

Key Points

• We will shed light on the less explored paths including signal processing and image processing.

• We highlight the transformative potential of integrating the LLMs into DNA methylation analysis.

• We evaluate and benchmark the performance and computational efficiency of the existing AI-based models.

• We explore the impact of the feature selection and encoding schemes on performances.

• We highlight the implication of XAI in DNA methylation detection.

• We reveal the importance of multi-omic data integration to provide a more holistic understanding of the role of DNA methylation in gene regulation and diseases.

Conflict of interest: None declared.