[Reading Science] AI Advances in Cancer Conquest... Also Predicts the Possibility of Normal Cells Turning Cancerous

Columbia University Researchers Announce Cancer Research AI 'GET'
Predicts the Likelihood of Normal Cells Turning Cancerous
Blind Reliance on AI Should Be Avoided

According to the National Cancer Information Center, 85,271 people died from cancer in South Korea in 2024. Cancer deaths accounted for 24.2% of all deaths (352,511). Compared to the 35,605 deaths caused by COVID-19 as of August 30, 2023, the number of cancer deaths highlights how significant the impact of cancer is. In the United States, about 610,000 people died from cancer last year alone, a number comparable to the casualties of the Civil War. The number of people diagnosed with cancer also reached 2 million.

Although revolutionary breakthroughs in cancer treatment have been made through new drugs such as targeted anticancer agents and robotic surgery, many still lose loved ones to cancer.

Hope has emerged for those who wish to overcome cancer even by borrowing divine power: artificial intelligence (AI). Doctors are already using AI to detect tumors and speed up diagnosis. Scientists and pharmaceutical companies are also using AI to develop new anticancer drugs, but there have been cases where AI has provided hints for cancer treatment through new approaches.

Recently, researchers at Columbia University in the United States published a paper in the international scientific journal Nature titled "A foundation model of transcription across human cell types," which has attracted attention as a new medical AI model capable of accurately predicting gene activity at the cellular level. It has become possible to predict human cellular activity.

The General Expression Transformer (GET), which predicts protein structures, is also compared to AlphaFold by Google DeepMind, which won the Nobel Prize in Chemistry last year. The scientific community agrees that both systems provide revolutionary insights into understanding the fundamental mechanisms of life.

GET is an AI model trained in a manner similar to ChatGPT. The research team trained it using more than 1.3 million normal cell data obtained from 213 different human cell types. This approach differs from previous studies that mainly focused on abnormal cells such as cancer cells.

Raul Rabadan, a professor at Columbia University and the author of the paper, said, "The ability to accurately predict cellular activity will bring innovation to understanding basic biological processes. This will transform biology from a science that explains seemingly random processes to a science that predicts the fundamental systems governing cellular behavior."

Professor Rabadan also stated, "After GET learned situations in various cellular states and was instructed to predict cancer occurrence patterns based on this, it was able to predict specific gene expressions even in cell types never seen before."

Cancer typically involves more than 1,000 gene mutations, and the combinations of these mutations exceed the number of atoms in the universe. According to the research team, GET can identify the most important combinations among these, greatly improving research efficiency.

Professor Yang Li of the University of Washington School of Medicine added about GET's potential, "We want to learn the grammar of genes and identify elements that play key roles in various cell types because many human diseases arise from disruptions in this grammar."

Professor Rabadan also said, "It is possible to design gene therapies that deliver genes expressed only in specific cell types. If we can predict which genes are turned on, off, amplified, or reduced in various cells, it will help identify the origins of diseases."

AI is active in cancer treatment in areas such as prediction, detection, new drug development, and treatment execution. GET belongs to the prediction field. In diagnostics, radiologists already use AI tools to detect tumors.

According to a paper published earlier this month in Nature Medicine titled "Nationwide real-world implementation of AI for cancer detection in population-based mammography screening," a study involving about 500,000 patients in Germany found that doctors using AI diagnostic models had a 17.6% higher breast cancer diagnosis rate than those who did not. The U.S. Food and Drug Administration (FDA) has already approved the promotion of AI software that identifies signs of prostate cancer.

However, AI tools cannot be a panacea for cancer treatment. Overreliance on AI screening and diagnostic tools can be risky. The science media Popular Science pointed out that many of the AI models mentioned are still in the research stage and require additional testing before large-scale use in medical settings.

There is also the risk of overhyping insufficiently validated models under the name of AI. While AI may become increasingly proficient at diagnosing cancer, it cannot replace the role of trained physicians.

Meredith Broussard, a professor in the Journalism Department at New York University, stated in her book, "Even the most advanced AI essentially finds mathematical similarities by comparing fixed images labeled by humans (the process of adding identifiable criteria to data so AI can judge it) with other images." She argued that "this can produce impressive results but is only predictive, not diagnostic."