MethylGPT unlocks DNA secrets for age and disease prediction

By harnessing advanced AI, MethylGPT decodes DNA methylation with unprecedented accuracy, offering new paths for age prediction, disease diagnosis, and personalized health interventions.

In a recent study posted to the bioRxiv preprint* server, researchers developed a transformer-based foundation model, MethylGPT, for the DNA methylome.

DNA methylation is a type of epigenetic modification that regulates gene expression via methyl-binding proteins and changes in chromatin accessibility. It also helps maintain genomic stability through transposable element repression. DNA methylation has features of an ideal biomarker, and studies have revealed distinct methylation signatures across pathological states, allowing for molecular diagnostics.

Nevertheless, several analytic challenges impede the implementation of diagnostics based on DNA methylation. Current approaches rely on simple statistical and linear models, which are limited in capturing complex, non-linear data. They also fail to account for context-specific effects such as higher-order interactions and regulatory networks. Therefore, a unified analytical framework that can model complex, non-linear patterns in various tissue and cell types is urgently needed.

Recent advances in foundation models and transformer architectures have revolutionized analyses of complex biological sequences. Foundation models have also been introduced for various omics layers, such as AlphaFold3 and ESM-3 for proteomics and Evo and Enformer for genomics. The achievements of the foundation models suggest that DNA methylation analyses could be transformed with a similar approach.

The study and findings

In the present study, researchers developed MethylGPT, a transformer-based foundation model for the DNA methylome. First, they acquired data on 226,555 human DNA methylation profiles spanning multiple tissue types from the EWAS Data Hub and Clockbase. Following deduplication and quality control, 154,063 samples were retained for pretraining. The model focused on 49,156 CpG sites, which were selected based on their known associations with various traits, as this would maximize their biological relevance.

The model was pre-trained using two complementary loss functions: masked language modeling (MLM) loss and profile reconstruction loss, enabling it to accurately predict methylation at masked CpG sites. The model achieved a mean squared error (MSE) of 0.014 and a Pearson correlation of 0.929 between predicted and actual methylation levels, indicating high predictive accuracy. Researchers also evaluated whether the model could capture biologically relevant features of DNA methylation. As such, they analyzed the learned representations of CpG sites in the embedding space.

They found that CpG sites clustered based on their genomic contexts, suggesting that the model learned the regulatory features of the methylome. In addition, there was a clear separation between autosomes and sex chromosomes, indicating that MethylGPT also captured higher-order chromosomal features. Next, the team analyzed zero-shot embedding spaces. This showed a clear biological organization, clustering by sex, tissue type, and genomic context.

Major tissue types formed well-defined clusters, indicating that the model learned methylation patterns specific to tissues without explicit supervision. Notably, MethylGPT also avoided batch effects, which often confound results in complex datasets. Besides, female and male samples demonstrated consistent separation, reflecting sex-specific differences. Next, the researchers assessed the ability of MethylGPT to predict chronological age from methylation patterns. To this end, they used a dataset of over 11,400 samples from diverse tissue types.

Fine-tuning for age prediction led to robust age-dependent clustering. Notably, intrinsic age-related organization was evident even before fine-tuning. Moreover, MethylGPT outperformed existing age prediction methods (e.g., Horvath’s clock and ElasticNet), achieving superior accuracy. Its median absolute error for age prediction was 4.45 years, further demonstrating its robustness. MethylGPT was also remarkably resilient to missing data. It exhibited stable performance with up to 70% missing data, outperforming multi-layer perceptron and ElasticNet approaches.

Analysis of methylation profiles during induced pluripotent stem cell (iPSC) reprogramming showed a clear rejuvenation trajectory; samples progressively transitioned to a younger methylation state over the course of reprogramming. The model was also able to identify the point during reprogramming (day 20) when cells began showing clear signs of epigenetic age reversal. Finally, the model’s ability to predict disease risk was assessed. The pre-trained model was fine-tuned to predict the risk of 60 diseases and mortality. The model achieved an area under the curve of 0.74 and 0.72 on validation and test sets, respectively.

In addition, they used this disease risk prediction framework to evaluate the impact of eight interventions on predicted disease incidence. Interventions included smoking cessation, high-intensity training, and the Mediterranean diet, among others, each of which showed varying degrees of effectiveness across disease categories. This showed distinct intervention-specific effects across disease categories, highlighting the potential of MethylGPT in predicting intervention-specific outcomes and optimizing tailored intervention strategies.

Conclusions

The findings illustrate that transformer architectures could effectively model DNA methylation patterns while preserving biological relevance. The organization of CpG sites based on regulatory features and genomic context suggests that the model captured fundamental aspects without explicit supervision. MethylGPT also demonstrated superior performance in age prediction across different tissues. Moreover, its robust performance in handling missing data (≤ 70%) underscores its potential utility in clinical and research applications.