Artificial Intelligence
How AI Could Help Breast Cancer Research
Scientists test GPT-4 in finding novel potential breast cancer treatments.
Posted June 17, 2025 Reviewed by Kaja Perina
A new study led by the University of Cambridge harnesses artificial intelligence (AI) large language model (LLM) GPT-4 to discover novel combinations of drugs not typically used to treat cancer to be repurposed as potential breast cancer treatment.
“We are already witnessing the emergence of AI scientists and AI-assisted researchers, signaling a shift in the way science is conducted,” wrote corresponding author University of Cambridge Professor Ross King in collaboration with co-authors Abbi Abdel-Rehim, Hector Zenil, Oghenejokpeme Orhobor, Marie Fisher, Ross Collins, Elizabeth Bourne, Gareth Fearnley, Emma Tate, Holly Smith, and Larisa Soldatova.
Chance discoveries versus the scientific method
Throughout history there are examples of scientific discoveries and inventions created by accident or by chance versus the scientific method. Yet serendipitous discoveries are the exception rather than the norm when it comes to how most major scientific discoveries have been conducted. In fact, the common scientific method accounts for 75% of all Nobel Prize and major science discoveries since the 1900s according to a 2024 report by Alexander Krauss published in PNAS Nexus.
The more productive approach is the scientific method, a technique for the formation and evaluation of a scientific hypothesis that is important in the formation of scientific theories. The steps of the common scientific method include making an observation, asking a question, forming hypothesis as a potential answer to the question, making a prediction based on the hypothesis, testing the prediction, and iterating as needed.
In this new study, the authors of the University of Cambridge-led study tasked the general purpose LLM GPT-4 to not only create scientific hypotheses but also evaluate its validity. Human researchers evaluated the drug combinations and individual efficacy in lab tests.
How LLMs could generate scientific hypotheses
The power of generative AI is widely known thanks to popular chatbots such as ChatGPT by OpenAI. What is not as commonly known is that LLMs may at times “hallucinate” and produce results that may seem plausible but are entirely made-up or are completely false. What seems like a bug may actually turn out to be more of a feature, at least when it comes to finding novel ways to perform drug discovery for cancer treatment.
“Despite the clear potential of LLMs for hypothesis generation, their utility for hypothesis generation has been little investigated,” wrote the study authors.
Applying AI for breast cancer research
Breast cancer in women is one of the most common cancers and leading causes of cancer mortality worldwide and is second only to lung cancer according to GLOBOCAN 2022. Nearly all breast cancer cases impact women, with only 0.5-1% of breast cancers occurring in men.
Breast cancer is a disease where abnormal cells of the breast grow out of control. The common forms of breast cancer are ductal carcinoma in situ (DCIS), invasive ductal carcinoma (IDC), lobular breast cancer, and the rarer forms include Paget's disease of the breast, inflammatory breast cancer (IBC), and triple-negative breast cancer (TNBC).
The subtypes of breast cancer are categorized by the state of the receptor cell. Examples of breast cancer subtypes include estrogen receptor-positive breast cancer (ER+) or ER-positive (ER+), hormone receptor-negative or HR-negative (HR-), hormone receptor-positive or HR-positive (HR+), = progesterone receptor-positive or PR-positive (PR+), and human epidermal growth factor-receptor 2 positive or HER2-positive (HER2+).
“We chose breast cancer as our test domain due to its critical importance in medical research, the vast body of existing literature, and our access to specialized equipment for studying tissue cultures as proxies for real patient tumors,” the scientists shared.
The scientists decided to focus on the MCF7 breast cancer cell line, one of the most commonly used and widely studied culture models for human breast cancer that was derived in 1973 and has since been used in roughly 25,000 scientific publications. The scientists also opted to use the MCF10A human mammary epithelial cell line, a commonly used in vitro model used for the study of normal breast cells, as the control.
The researchers used LLM GPT4 to analyze drug combinations by prompting ChatGPT4 to identify new affordable, synergistic drug combinations, preferably US Food and Drug Administration (FDA) approved, that would significantly impact the MCF7 breast cancer cell line while at the same time not harming the MCF10A normal breast cell line and at least one of the drugs in each pair should not be an antineoplastic drug. Antineoplastic drugs are anticancer medication that prevent the growth and spread of neoplasms (tumors) and may have harmful side effects to reproductive health and pregnant women. There are over 1,800 antineoplastic drugs and over 450 antineoplastic regimens in the National Cancer Institute (NCI) Seer*Rx Database.
“In the first round of laboratory experiments, GPT4 succeeded in discovering three drug combinations (out of 12 tested) with synergy scores above the positive controls,” the researchers reported.
As a next step, a second round was conducted to see if GPT-4 could improve its hypotheses if provided a summary from the first round and additional prompting that also factors in combination of drugs from the positive controls. The researchers reported that GPT-4 created three new combinations that had positive synergy scores out of the four tested. The study results suggest that LLMs are capable of creating novel scientific hypotheses.
“By leveraging the vast knowledge encoded in LLMs, scientists can explore regions of the hypothesis space that human researchers may miss or find more difficult to explore due to biases, exhaustion, or other factors,” conclude the scientists.
Copyright © 2025 Cami Rosso All rights reserved.
