An estimated 610,000 people in the US died from cancer last year. That’s almost the same amount of people who perished in the country’s four-year civil war. At least two million more people were diagnosed with some form of cancer in 2024, a figure that’s climbed in recent years. Early detection remains one of the single biggest factors that determine whether or not someone ultimately survives cancer and, luckily, advances in medical treatment can help. Researchers and medical scientists believe artificial intelligence models could play a key role in that early detection process. Though AI still can’t substitute for a doctor’s real-world medical expertise or even produce a true medical diagnosis, it can serve as a critical tool to make them more effective.  AI’s ability to parse through dense troves of data and seek out patterns may make it well-suited to look for irregularities in images of organs and tissue and spot cancerous cells before they metastasize. A study published today in the journal Nature by researchers at Columbia University described a new medical AI model that they say can accurately predict the activity of genes at the cellular level. In theory, this level of granularity could open up new paths for researchers to understand the gene mutations that cause cancers to occur in the first place.  “Having the ability to accurately predict a cell’s activities would transform our understanding of fundamental biological processes,” paper senior author and Columbia professor Raul Rabadan said in a statement. “It would turn biology from a science that describes seemingly random processes into one that can predict the underlying systems that govern cell behavior.”  Today, doctors are already using AI to help spot tumors and expedite diagnoses. Scientists and pharmaceutics companies are similarly using the tech in varying degrees to assist with the creation of new cancer-fighting therapeutics. And while AI almost certainly won’t replace trained oncologists anytime soon, all signs are pointing toward a near future where these models play an increasingly present role in combating cancer, from the earliest moments to late-stage treatment.  AI gives researchers a glimpse at how cancer begins at the cellular level  The Columbia researchers developing the AI capable of predicting gene activity, referred to as GET (general expression transformer) say they trained their model on images of 1.3 million human cells. The researchers compared this process of injecting large training data of both diseased and healthy genes as similar to the way Open AI’s ChatGPT large language model ingests a vast corpus of the written internet. Once the medical AI model had learned the “grammar in many different cellular states” Rabadan notes, it could then be directed to predict patterns based on that information. When they tested the AI, researchers said it was able to predict certain gene expressions in cell types it had never seen before.  “These methods can effectively conduct large-scale computational experiments, boosting and guiding traditional experimental approaches,” Rabadan added.  The paper comes only a few months after scientists from Harvard Medical School described another cancer-related AI detection tool, also Nature. In that example, researchers trained their model to detect signs of 19 different types of tumors after observing medical patient images. The model was reportedly able to detect cancer and predict a tumor’s molecular profile all based on cellular features included in its training data. It could also forecast a patient’s survival potential across different cancer types. The model, called CHIEF (Clinical Histopathology Imaging Evaluation Foundation) was trained on 60,000 whole-slide images of tissues from lungs, prostates, colons, and other organs. Researchers said CHIEF went a step further than other cancer-detecting AI models due to its broad training data which lets it interpret a medical image more holistically than other more specialized models.  “If validated further and deployed widely, our approach, and approaches similar to ours, could identify early on cancer patients who may benefit from experimental treatments targeting certain molecular variations, a capability that is not uniformly available across the world,” Harvard Medical School professor and study senior author Kun-Hsing Yu said in a statement. AI is being used in every stage of cancer research  The promise of AI for cancer treatment broadly falls into five categories: prediction, detection, drug discovery, and treatment implementation. On the detection front, radiologists and other doctors are already using AI tools to help spot tumors. Just this week, a new study published in Nature Medicine involving nearly 500,000 patients in Germany found that doctors using an AI detection model confirmed more cases of breast cancer than doctors acting on their own. Specifically, doctors using the AI achieved a cancer detection rate 17.6% higher than those who didn’t. The FDA has also already approved marketing for an AI software design to help identify signs of prostate cancer.  A separate AI model created by researchers at the National Institutes of Health (NIH) called LORIS (logistic regression-based immunotherapy-response score) demonstrated the ability to predict which group of cancer patients might benefit best from certain immunotherapy treatments. That approach, which uses the body’s immune system to target cancer cells, is less invasive than more traditional cancer-fighting treatments like chemotherapy and radiotherapy but is only effective for a subset of people. Models like LORIS could help doctors better detect those therapies for patients who may benefit and simultaneously avoid exposing others to unnecessary treatments.  On the discovery front, researchers from the University of Chicago Medicine Comprehensive Cancer Center (UCCCC) recently received $16 million from the federal government as part of a project to use powerful machine learning models to comb through large medical datasets and look for patterns that could spark the development of new treatments for drug-resistant cancers. The hope, according to those involved with the efforts, is that advances in AI can fast-track the time it takes to find new drugs, hopefully in time for patients who may need them in the near future.  “Patients with cancer don’t have time to wait for new treatments, so there is a strong need to compress the drug discovery timeline and we aim to do that with novel synergistic approaches that take advantage of [The Department of Energy’s] supercomputing capabilities,” UCCCC Director Kunle Odunsi said in a statement.  AI tools aren’t a silver bullet  At the same time, there’s a risk of placing too much faith in AI screening and detection tools too quickly. Several of the models noted earlier are still in research phases and will require more testing before they are deployed in healthcare facilities at scale. There’s also the risk of an opportunist taking advantage of the overly broad umbrella term “AI” to pitch far less tested models as more effective than they actually are. There are already numerous cases of people receiving wrong and potentially dangerously incorrect diagnoses after interacting with popular large language models. One study published in JAMA Pediatrics last year found that OpenAI’s ChatGPT incorrectly diagnosed 83% of pediatric case studies it was presented with. Models like these are also prone to occasionally hallucinating false facts and doing so with a confident tone. That can lead to funny results when asking it to come up with a cake recipe, but those same inaccuracies can prove dangerous when someone uses them to self-medicate.  [Related: Will we ever be able to trust health advice from an AI?] And even as AI models (likely) improve their ability to detect different cancers in the years to come, they still fundamentally aren’t performing the same job as a trained physician. As New York University journalism professor Meredith Broussard notes in her 2023 book More Than a Glitch, even the most advanced AI models are essentially comparing a static image against a corpus of other images already labeled by humans and quickly seeing if there are mathematical similarities in the two. That can lead to impressive results, but that process is ultimately a prediction which isn’t the same as a diagnosis. A diagnosis still requires a human doctor who can look over evidence and draw their own expert conclusion based on years of real-world experience. We’re already living in a world where doctors can use these tools to bolster their own abilities. It’s less clear though whether or not AI will ever be reliable enough to remove doctors from that dynamic entirely. The post AI is already changing the ways we fight cancer appeared first on Popular Science.