Machine Learning Maps the Exploding World of Epigenetic Epidemiology Research

Epigenetics has traveled a remarkable distance since the term first entered the scientific vocabulary more than 75 years ago. What began as an abstract question about how genes and their products interact to shape organisms has matured into one of the most dynamic arenas of biomedical research, with applications spanning cancer biology, the biology of aging, and the study of how environmental exposures alter gene expression. The arrival of high-throughput technologies capable of measuring epigenetic markers across the genome has been transformative, because it has allowed researchers to move beyond laboratory models and into population-based studies. That shift gave birth to a distinct discipline, epigenetic epidemiology, which examines epigenetic associations from a population perspective to generate insights into disease risk, prevention, and progression. Unlike fixed genetic variants, epigenetic markers are dynamic, and that plasticity offers epidemiologists an entirely new way to connect early-life circumstances and environmental exposures with the risk of disease decades later.

The scientific literature documenting this field has grown explosively. Over the past two decades, publications on epigenetic epidemiology have multiplied in both number and variety, and the scope of the research now extends well beyond the DNA methylation markers that once dominated the field. Histone modifications and non-coding RNAs have joined the toolbox, and the study designs on display range from epigenome-wide association studies, commonly known as EWAS, to candidate gene studies and clinical trials. Epigenetic markers appear in the literature in two distinct roles. Sometimes they are investigated as risk factors, as in studies of DNA methylation patterns associated with the incidence of type 2 diabetes. Sometimes they are investigated as outcomes, as in research documenting methylation changes in response to air pollution. Keeping track of this expanding and heterogeneous body of work has become a genuine challenge, and that challenge motivated a team of researchers affiliated with the Centers for Disease Control and Prevention and Emory University to build a purpose-made solution.

The result is the Epigenetic Epidemiology Publications Database, or EEPD, a publicly available, web-based application described in a recent correspondence published in Epigenetics Communications. The database is designed to offer a user-friendly way to explore the literature at the intersection of epigenetics, epidemiology, and public health. The scope of the collection is carefully defined. The developers aimed to include population-based studies of epigenetics in association with an exposure or outcome of interest, systematic reviews related to epigenetic associations, and published descriptions of epigenetic resources such as cohorts, protocols, and methods. Just as important is what the database excludes: articles focused on gene expression, including transcriptomics, non-population-based laboratory studies, non-human studies, and non-systematic reviews are all filtered out. That boundary keeps the collection focused on the population perspective that defines the field.

The technical machinery behind the database is where the project becomes genuinely interesting for anyone following the application of machine learning to scientific curation. Manually screening the biomedical literature is notoriously labor-intensive, and the volume of relevant publications makes hand-curation impractical at scale. The team therefore developed an automatic screening process that begins with a PubMed query and then applies a machine learning method, a support vector machine, combined with text mining techniques. Support vector machines are supervised learning algorithms that classify items into categories after being trained on labeled examples, and they have a long track record in text classification tasks. To evaluate how well their pipeline performed, the researchers assessed it against a random sample of 1,999 manually annotated abstracts. Of these, 285 positive records and 114 negative records were reserved for testing. The results were striking: the screening process achieved an estimated area under the curve of 96.4 percent, with 90 percent sensitivity and 90 percent specificity. In practical terms, the automated system caught the overwhelming majority of relevant articles while correctly discarding most irrelevant ones, a level of performance that rivals careful human screening at a fraction of the effort.

Once the screening pipeline is in place, the database essentially runs itself. Records are automatically selected and uploaded on a daily basis, which means the collection stays current without a team of curators reading every new abstract. Each record is indexed with genes, diseases, and other factors, creating a structured layer of metadata on top of the raw publication records. The system also automatically subsets the data according to 14 special topics, ranging from cancer and diabetes to environmental health, health equity, infectious diseases, neurological disorders, pharmacogenomics, rare diseases, and reproductive and child health. These subsets are generated by searching the database with topic-relevant terms, so a researcher interested specifically in, say, heart, lung, blood and sleep disorders can work within a curated slice of the literature rather than the full collection.

The payoff from all this automation is best illustrated by a direct comparison the authors conducted. On September 28, 2023, a search for cancer-related epigenetic epidemiology on EEPD returned 8,942 records. A search on PubMed using the terms cancer, epigenetics, and epidemiology returned only 1,198 records, a dramatic undercount that reflects how often relevant studies fail to use all three terms in their indexing. Removing the epidemiology term from the PubMed search pushed the result to 17,927 records, but at the cost of flooding the searcher with irrelevant material. The EEPD sits in the sweet spot between those two failures, capturing relevant work that conventional keyword searches miss while excluding the noise that broader searches admit. For researchers conducting literature reviews, that difference is not a convenience; it can determine whether a review is comprehensive or fundamentally incomplete.

Using the database is deliberately simple. Users type a query of interest into a search box at the top of the page, and results are returned as a list of related publications. From there, filtering functions allow progressive refinement using indexed factors including gene, disease, publication year, publication journal, and study design. Filters can be applied continuously until the desired result set is achieved, and the final results can be downloaded as an Excel spreadsheet with a single click on the download button. Alternatively, users can subset the entire database to one of the 14 special topics through a dropdown menu in the search bar, and all of the navigation functions remain available within that subset. The design philosophy is clear: the database should spare investigators from having to construct complicated PubMed queries while still offering the precision that expert searchers demand.

The scale of the collection as of August 18, 2023, underscores how much material the automated pipeline has already accumulated. The database contained 21,699 PubMed records, indexed against 9,378 genes and 1,019 disease terms. The annual number of publications has increased markedly since 2000, reflecting the field’s rapid maturation. Cancer dominates the collection as the most studied topic, followed by rare diseases. The five most frequently indexed disease terms are breast neoplasm, colorectal neoplasms, squamous cell carcinoma, lung neoplasm, and stomach neoplasms, a roster that mirrors the prominence of oncology in epigenetic research more broadly. On the gene side, the five most frequently indexed entries are CDKN2A, RASSF1, DNMT1, MGMT, and CDH1, names that will be immediately familiar to anyone who has followed methylation research in tumor biology, since several of these genes are classic targets of epigenetic silencing in cancer.

The authors are candid about the limitations of their tool. The sensitivity of EEPD is not perfect, and the search algorithms and indexing processes of both PubMed and the database have inherent constraints in a field that is rapidly evolving and not fully defined. Because EEPD relies exclusively on PubMed as its data source, potentially relevant articles not indexed there are simply not included, which means the database cannot claim exhaustive coverage of the global literature. These caveats matter, and the authors do not shy away from them. Even so, they argue that EEPD can serve as a valuable complement to traditional search methods for literature review, offering investigators an efficient way to gain an overview of available research on a particular topic as well as to identify articles relevant to their own projects without labor-intensive screening.

The roadmap for the future is equally pragmatic. The authors hope to expand the data sources feeding the database, drawing on resources such as Scopus and Embase to capture material that PubMed misses. Further tuning of the PubMed search query and of the database’s own algorithms could enhance both the sensitivity and specificity of data collection and retrieval, pushing the already strong performance metrics even higher. For a field that sits at the crossroads of epigenetics, epidemiology, and public health, and that produces new publications at an accelerating pace, tools like EEPD may prove essential infrastructure. The promise of epigenetic epidemiology, linking the environments and experiences of populations to the molecular marks that shape disease risk, depends on researchers being able to find, synthesize, and build upon the work of their predecessors. By automating the discovery of that work, the Epigenetic Epidemiology Publications Database aims to make sure that no important study is lost in the sheer volume of an exploding literature.

Subject of Research: A machine learning-curated public database for navigating epigenetic epidemiology literature

Article Title: Navigating epigenetic epidemiology publications

Article References: Yu, W., Drzymalla, E., Gyorfy, M. F., Khoury, M. J., Sun, Y. V., & Gwinn, M. (2023). Navigating epigenetic epidemiology publications. Epigenetics Communications, 3(1), Article 8. https://doi.org/10.1186/s43682-023-00023-3

Image Credits: AI Generated

DOI: 10.1186/s43682-023-00023-3

Keywords: epigenetics, epidemiology, DNA methylation, machine learning, support vector machine, PubMed, database, EWAS, cancer, public health, text mining, literature curation