Messenger ribonucleic acid (mRNA) vaccines that overcame the novel coronavirus infection (COVID-19), CRISPR gene editing that can freely manipulate genetic information, and the discoveries of the Higgs particle and gravitational waves that unveiled the secrets of the universe are considered major achievements in 21st-century science. However, over the past 20 years, the most widely read and utilized papers had a different protagonist. They served as tools that aided other research rather than presenting new theories.
The international journal Nature announced on the 15th (local time) that it analyzed tens of millions of papers published since the 21st century based on five representative academic citation databases and selected the top 25 papers according to citation frequency.
The most cited paper over the past 20 years was research related to 'Deep Residual Network (ResNet)' published by a Microsoft (MS) research team in 2016. It dealt with a technology that allows artificial intelligence (AI) to deepen its neural network structure for learning. According to Google Scholar, it has been cited more than 250,000 times, and in the academic database 'Web of Science,' it has been cited over 100,000 times in other papers. This paper became a key foundation for AI systems such as AlphaGo for Go, AlphaFold for protein structure prediction, and ChatGPT for conversational AI.
Other AI-related papers also ranked highly. There is the AI model 'AlexNet' for image recognition published in 2012 and 'Transformer,' the core structure for language models published in 2017. They are studies that became the starting points for image recognition and large language models (LLM), respectively. The Transformer technology is a key technology that enables conversational AI like ChatGPT to understand and generate sentences.
Jeffrey Hinton, a professor at the University of Toronto in Canada who received the Nobel Prize in Physics last year, noted, "The reason that AI-related papers have a high citation count is that the AI field produces and utilizes papers much more rapidly than traditional sciences. Because it is widely applied in various sectors such as healthcare, finance, robotics, and translation, there is a natural tendency for citation counts to increase." Some researchers have also stated that most early AI machine learning papers are open source, making them accessible to anyone, which has led to higher citation counts.
Analytical tools used in life science and medical research also made the rankings. Papers outlining methods for determining gene quantities, known as polymerase chain reaction (PCR), a collection of programs analyzing X-ray scattering patterns to reveal molecular atomic structures, the World Health Organization (WHO) report summarizing global cancer incidence and mortality rates, and the DSM-5, which outlines diagnostic criteria for mental disorders, also ranked highly.
Statistical programs or software utilized by researchers worldwide were also included. These include the machine learning tool 'Scikit-learn' created with the Python programming language, the lme4 package for analyzing biological experimental data, and the G*Power tool, which calculates the minimum sample size required for experiments.
In this regard, Misha Teplitsky, a professor at the University of Michigan, analyzed, "Scientists say they prioritize new theories or discoveries, but in reality, they tend to cite useful tools and methods more frequently."
For reference, according to Nature, the paper with the highest citation count recorded was a study measuring the amount of protein in solution published in 1951 in the international journal 'Journal of Biological Chemistry.' That paper has recorded over 350,000 citations in Web of Science.
References
Nature (2025), DOI: https://doi.org/10.1038/d41586-025-01125-9