Navigating cancer treatment with the help of AI

During autumn of 2021, Lisa Benetti was having trouble breathing. Her doctor diagnosed asthma. But by Christmas, she was still wheezing. An X-ray showed that her condition was much more serious. There were ominous looking spots on her lungs. A biopsy confirmed the shocking diagnosis: She had stage III non-small-cell lung cancer.

When Benetti was first diagnosed, as part of her routine work-up, doctors sequenced the DNA from her cancer tissue. They found that her tumors carried a mutation in the KRAS gene. But she wasn’t told – at least not immediately. The reason: there was no alternative to traditional chemotherapy and radiation available, or so her doctor thought. But after her treatment, she learned that wasn’t entirely true. A new type of immunotherapy – treatments that train the immune system to fight cancer – had become available just as she was undergoing treatment.

Had she known of her KRAS status, she might have requested immunotherapy instead, which has a higher success rate for her type of cancer. Today Benetti is cancer free, but she lives with the anxiety that it could return, and with the frustration at not having taken a more active role in her own care.

Benetti’s case is fictional. But the experience is unfortunately far too common, says Dr. Altuna Akalin, Group Leader of Bioinformatics and Omics Data Science at the Max Delbrück Center in Berlin. However, he is quick not blame doctors.

The number of diagnostic tests and available cancer therapies has skyrocketed over the past decade. On average, 46 new cancer therapies have been approved per year over the past ten years. While the new drugs have been a boon for cancer patients – they are living longer than ever before – the doctors who treat them face a sea of ever-changing care guidelines that have many oncologists scratching their heads when deciding on the best treatments for their patients.

In the growing complexity of diagnostic tests and myriad new cancer treatments, Akalin saw an opportunity. He has been working for years on mining genetic data collected from various types of cancers to find clues on how to better treat it. But he wanted to have a more immediate impact on patients. As machine learning tools have become more sophisticated, he came up with the idea of Onconaut – an online AI based tool for clinicians and patients that can help them better navigate personalized cancer therapies.

“Developing drugs and diagnostics are great scientific efforts. But to translate them into a useful product takes decades,” he says. “We wanted to develop a tool to help clinicians make the best informed decisions about their patients. We also wanted to help patients understand their options so they could better advocate for themselves.”

AI in medicine

By now, almost everyone has heard of the promise and the perils of Artificial Intelligence (AI) tools such as ChatGPT. Biomedical researchers, however, have been using machine learning tools – which are an application of AI – for several years already. As computers have become ever more powerful, data scientists have been working to develop mathematical models that can create knowledge from the petabytes of data generated from their research. Through targeted training, they “teach” the models to find patterns in the data to better understand complex relationships and make predictions. The more data is fed into the model, the better the tool becomes.

Much of this research is in early stages. But there are reports of how AI might soon benefit patients. The Human Radiome Project spearheaded by the Helmholtz Association in Germany, for example, aims to consolidate different types of 3D radiological images, such as MRIs and CT scans, into an AI model to deepen our understanding of human anatomy and to better distinguish diseases. Such AI tools can make the work of radiologists also more efficient.

Biomedical researchers are training AI models in myriad other ways. Some researchers are using it to design better drugs, others to find biological clues on how diseases progress. “AI can speed up basic science, our work with big data, all the way to research on therapeutics,” says Professor Uwe Ohler, Group Leader of the Computational and Regulatory Genomics Lab at Max Delbrück Center.

Biomarkers and targeted therapies

For Altuna, creating Onconaut was a way that that he could directly and more immediately improve the care of cancer patients.

The Onconaut interface presents users with a simple search function that belies the complexity behind it. On the homepage, one can type in specific information about a patient into the “Input query related to cancer biomarkers,” field. Entering “KRAS and lung cancer,” for example, produces the latest clinical guidelines, a list of available medications for cancers with KRAS mutations, risks of treatment, and statistics on outcomes. Most important for patients, it provides a list of clinical trials, all within seconds.

Biomarkers give information about health and disease. Some biomarkers are simple — blood pressure and heart rate, for example, serve as proxies for cardiovascular health.

In the cancer field, researchers have developed many therapies that are designed to treat cancers with specific mutations. These mutations are also called biomarkers and are usually identified by taking a biopsy of the tumor or a sample of cancer cells from biofluids and sequencing genetic material. Such biomarkers have become indispensable tools in personalizing cancer therapy.

How Onconaut was trained

Such tools can by no means replace doctors. But they can at least speed up the decision-making process. They can make experts more productive and give beginners more of an edge.

Akalin and his colleagues trained Onconaut on several types of information such as published medical studies and clinical guidelines issued by various official organizations such as the German Cancer Society or the American Society of Clinical Oncology (ASCO).

To further test and strengthen the tool, Akalin is currently working on a project with the Charité – Universitätsmedizin Berlin. He and his colleagues are training a model with data from cancer patients diagnosed at the hospital. They ask the model to recommend a course of treatment and compare the output to the treatment plan devised by the hospital’s tumor board – a team of experts who jointly provide a treatment plan for each patient. “On the limited number of cases we have looked at so far, there is a good match,” says Akalin.

“Such tools can by no means replace doctors,” Akalin says emphatically. “But they can at least speed up the decision-making process. They can make experts more productive and give beginners more of an edge.”

Second opinions, other diseases

He is also extending the tool to help diagnose other diseases. People can already type their queries into his AI powered website for a second opinion. They can upload the results of medical tests and ask: “I was diagnosed with X based on this information. What else could the diagnosis be?” Akalin explains.

“Some people with rare diseases can spend years in search of a diagnosis,” Akalin says. “We wanted to create a simple tool that could point patients to potential diagnoses based on their symptoms and tests.”

The model is still being refined. Akalin is currently training it with data from difficult medical cases published in the "New England Journal of Medicine” to see if it can identify them accurately. The project is close to his heart – his wife suffers from a rare disease that took far too long for doctors to correctly diagnose.

“Humans have a lot of blind spots. Our routines sometimes prevent us from thinking out of the box. Doctors suffer from this too,” Akalin says. “AI presents a way to get around this problem.”

Had such a tool been available to her, she might have been diagnosed more quickly and received appropriate treatment earlier, he says. He hopes it will help other people dealing with complex illnesses avoid the trouble he and his wife went through. In tests, he says, the tool has been doing fairly well: “So far, it is performing better than Dr. Google.”