The AI boom in cancer: Game-changer or just hype?

The promise of AI in biotech and in oncology is vast. Lately, AI utilization in the health industry has seen signs of a market correction to the point where we wondered whether we were in an AI bubble about to burst. The experts we talked to tempered our fears leaning toward the explanation of a mismatch between particularly high expectations for the technology and its actual contributions to the industry. But what about AI in oncology specifically? 

“Artificial intelligence is not just a buzzword – it is a revolutionary element that changes our approach to cancer diagnosis, treatment personalization, and drug discovery. It’s an exciting time to be involved in this intersection between technology and healthcare,” said Ghazenfer Mansoor, founder and chief executive officer (CEO) of Technology Rivers.

The integration of AI into oncology is not just about speeding up processes, it’s about changing how we understand, detect, and treat one of the world’s most complex diseases.

We’ve previously explored AI’s potential in cancer treatment in 2019, hinting that it would play an important role in early diagnosis and drug discovery. Since then, the field has seen remarkable growth and new developments, as we discussed in early 2023. However, with recent breakthroughs and new applications of AI, now is the perfect time to revisit this topic with a more comprehensive look at how AI is transforming cancer care today.

Is AI making a difference in cancer detection?

According to Ryan Schoenfeld, CEO of the Mark Foundation for Cancer Research AI’s biggest leap in oncology so far has been in diagnostics, more specifically in radiology where advanced image analytics are transforming how we detect and diagnose cancer. “AI can now analyze scans faster and with greater accuracy, helping doctors catch cancer earlier,” he said.

Philip Lieberman, founder and president at Analog Informatics tempers Schoenfeld’s statement. In his opinion, AI is technically slightly less accurate than a typical trained technician in radiologic and dermatologic diagnosis. However, he does note that it is more consistent than humans and able to see artifacts that are not discernible to the human eye.

“For what it is worth, AI for image analysis is cheaper, faster, and immediately deployable. As an adjunct technology for technicians and for areas where there are no technicians – rural and 3rd world countries – this technology is wonderful. It can also dramatically reduce costs for insurance companies and governments while improving outcomes for those that cannot afford or wait for a technician to read their images,” explained Lieberman.

In addition to diagnosis, Jeffery Sorenson co-founder & CEO at Yunu, sees AI’s potential expanding into predictive medicine. “Most of today’s highest potential applications are those that use AI to help sift through healthcare’s messy unstructured data sources and create an efficient way to group patients and make sense of complex data sets. Interestingly, we seem to be leapfrogging over diagnosis as a goal and making incredible progress in predictive analytics, with the aim of determining which patients are most likely to respond to various treatment options.”

And AI is showing progress in this area every day. Recently, researchers at Mayo Clinic developed a new AI-based tool – hypothesis-driven AI – aimed at transforming cancer diagnostics and treatment personalization. The innovation revolves around a new class of AI that sifts through complex, unstructured healthcare data – such as clinical notes, radiology scans, and genomic data – to improve accuracy in diagnosing cancers and predicting patient outcomes. This tool can identify patterns in patient data that are often missed by traditional diagnostic methods, ultimately improving early cancer detection and the precision of treatment strategies.

One of the major advantages of Mayo’s AI system is its ability to process radiological images faster and more accurately than conventional methods. The AI can detect subtle abnormalities in cancerous tissues, which can lead to earlier cancer detection.

In addition to diagnostics, the Mayo Clinic’s AI system enhances predictive analytics, offering a better understanding of how patients will respond to treatments. This is particularly important in oncology, where outcomes can vary significantly based on tumor characteristics and patient genetics.

AI is also playing a crucial role in advancing liquid biopsy technology, which allows for the non-invasive detection of cancer biomarkers in blood samples. Recent developments have shown improvements in both the sensitivity and accuracy of these biopsies, particularly in detecting circulating tumor DNA (ctDNA).

One standout example is a study led by Weill Cornell Medicine, where AI was used to improve the detection of cancer recurrence in patients by identifying tumor DNA in the bloodstream. This new AI-powered liquid biopsy method, called MRD-EDGE, demonstrated the ability to detect cancer recurrence months or even years earlier than traditional methods. It showed success in cancers such as lung, colorectal, and breast cancer​.

Additionally, Johns Hopkins researchers have developed the DELFI-Pro liquid biopsy test, which combines AI with cell-free DNA analysis to screen for ovarian cancer. The test was able to detect early-stage ovarian cancer with higher accuracy than traditional methods, with almost no false positives​.

Diagnosing cancer as early as possible is crucial in the patient’s pathway but this is not the only area where AI shines – drug discovery is another significant contribution AI is making in oncology.

AI’s potential in cancer drug discovery and development

Drug discovery is probably where our imagination goes first when we think of AI in the biotech industry. Traditionally, drug discovery is a lengthy process, but AI has been able to streamline it by analyzing complex biological data, finding novel drug targets, and even repurposing existing drugs.

“AI and machine learning are supercharging drug discovery and development in cancer treatment. Tools like DeepMind’s AlphaFold have transformed how we study disease-related proteins, speeding up the identification of new therapeutic targets,” said Schoenfeld.

Indeed, AlphaFold has had a transformative impact on drug discovery, particularly in oncology. Its AI-based system can accurately predict the 3D structures of proteins to identify potential drug targets. Protein structures are often difficult to solve using traditional methods, which can take years. However, AlphaFold has dramatically reduced this timeline. For example, in 2022, it predicted the structures of over 200 million proteins, providing a massive dataset that researchers can use to find druggable targets in cancer and other diseases.

AlphaFold’s open-source database, made available to the research community, further accelerates this process by allowing researchers worldwide to incorporate this data into their drug discovery programs.

Several drugs discovered or optimized using AI are currently in the oncology pipeline. A prominent example is BBO-8520, a cancer drug targeting KRAS mutations, developed through a partnership between Lawrence Livermore National Laboratory (LLNL), the Frederick National Laboratory, and BridgeBio. Using AI-driven simulations and advanced computing platforms, the drug reached clinical trials in record time after 3 years of development and is now in phase 1.

Another example is Exscientia, a drug discovery company using AI to develop novel oncology treatments. The AI-driven platform can predict new small molecules and identify effective drug candidates at a pace previously unachievable with conventional methods​.

In 2021, Nature reported that Evotec announced a phase 1 clinical trial for an oncology candidate developed with the AI drug discovery company. While the traditional process of discovering a drug usually takes 4 to 5 years, Evotec and Exscienta accomplished it in only 8 months. The candidate EXS-21546 entered phase 2 in 2022.

Mansoor also noted that AI can allow for drug repurposing. Similar to how GLP-1 agonists – initially developed for diabetes – have attracted a second wave of investment and approvals for obesity, already approved drugs could find new indications in oncology thanks to AI. 

TxGNN – the AI model designed at Harvard Medical School – identifies connections between well-understood disease mechanisms and poorly understood conditions. In the context of cancer treatment, this means that AI can help discover how drugs initially developed for one type of cancer – or even another disease – may be effective against other cancers.

But patients are all different and cancer treatments often have to be tailored to individual profiles, and AI can help here too.

Moving toward AI precision oncology

AI’s capacity to analyze enormous datasets, including patient genetics, clinical history, and tumor profiles, could allow for highly personalized treatment plans tailored to individual patients. This holds the potential to move beyond traditional, one-size-fits-all treatment methods.

While Jason Williams, chief of interventional oncology & immunotherapy at Williams Cancer Institute, thinks we are not quite there yet, he admits it holds promise for the future.

“AI will be able to compile all of the patient’s data and from the cancer. This would include genetics, gut microbiome, tumor genetics, radiological imaging, and cytokine panels. Using this, it would predict the best treatment options, which could also go beyond standard approved agents, including supplements, off-label drugs, and experimental medications.”

As Sorenson puts it, cancer isn’t one single thing, it’s instead hundreds of different variations – variations in the cancers itself but also in patients’ genetic profiles. And what’s AI if not a tool to calculate multiple possibilities? Just as AI is now used to invent new protein structures and discover genetic signatures, we are now able to match patients’ genetics with new drug molecules and even make genetically modified therapies that fit each cancer like a custom-made puzzle piece.

“Huang et al., (2018), demonstrated the promising role of AI in treatment response prediction to chemotherapy in 175 cancer patients, based on genetic profiles used to train a machine learning model. A key benefit of this approach to patients is that targeted treatments seek out specific cancer cell types and therefore avoid some of the common problems with toxicity,” said Sorenson.

More recently, an Oxford University study published in June 2024, demonstrated how its AI-driven approach could tailor cancer therapies more effectively, particularly in preventing patient relapse.

The researchers applied deep reinforcement learning (DRL), a form of AI, to create adaptive therapy schedules for prostate cancer patients. The AI analyzed data from mathematical models, simulating how individual patients’ cancers would respond to various treatments. The results were promising, with the new AI-driven schedules more than doubling the time to relapse compared to conventional treatment methods. This is crucial for metastatic cancers where drug resistance often emerges. The AI models not only provided personalized treatment plans but also offered insights that were interpretable by clinicians, a key challenge in integrating AI into medical practice.

Maybe even more concretely, Evaxion Biotech recently presented one-year data from their phase 2 trial of EVX-01, an AI-designed personalized cancer vaccine, at the European Society for Medical Oncology (ESMO) Congress 2024. The trial combining EVX-01 with Keytruda (pembrolizumab) for patients with advanced melanoma, showed impressive results.

Key outcomes included a 69% overall response rate (ORR), with 15 out of 16 patients experiencing a reduction in tumor size, and an immunogenicity rate of 79%, demonstrating that Evaxion’s AI platform accurately predicted which neoantigens would trigger an immune response.

While it is true the expectation towards AI might have been slightly too high and could benefit from being tempered, oncology is starting to see some very concrete AI contributions. It will take time for AI to reach its full potential in cancer treatment and overcome the challenges that remain but we are on the right path.

What are the challenges?

While traditional medical devices and treatments are regulated with clear guidelines, AI-driven tools do not always fit neatly into these frameworks. The European Commission is working on regulations, such as the AI Liability Directive, which would apply to “high-risk AI systems” like those used in healthcare. However, there’s still uncertainty around how these regulations will be enforced, especially when AI models are continuously evolving after being deployed, raising questions of accountability and liability​.

In the U.S., the Food and Drug Administration’s (FDA) AI/ML Action Plan provides some guidance for AI-based medical devices but acknowledges the need for better international cooperation to ensure the safe deployment of AI systems in oncology. This regulatory landscape is still in flux, and developers of AI tools need to navigate evolving requirements, which could slow the pace of adoption​.

Although it is important to give a frame to AI to ensure its safety, Mansoor notes the risk of regulation not keeping up with innovation. “I believe it is essential that regulators create frameworks that promote innovation. We need regulations that keep pace with rapid technological advances, allowing us to harness the full potential of AI in cancer treatment.”

Beyond the framework itself, AI is a rather complex – and in a way – opaque tool. Indeed, this technology serves one purpose, optimization, and while humans created it, it’s often difficult even for AI experts to understand how AI reached its conclusion. Jason Alan Snyder co-founder of supertruth, noted there’s currently a lot of talk about AI as a “black box” to represent this idea.

“One of the biggest hurdles is ensuring that AI models are explainable and transparent. If we want AI to be fully trusted and integrated into cancer care, regulatory bodies must set more explicit standards around data transparency and algorithmic accountability,” said Snyder.

There are also ethical concerns around bias in AI models. If AI systems are trained on non-representative data, they could lead to biased treatment recommendations, potentially disadvantaging certain groups. For example, an AI model trained predominantly on data from Western populations may not perform as well for patients from other ethnic backgrounds.

These challenges will need to be addressed in the years to come, especially as AI is evolving at a fast pace and the experts we talked to certainly look forward to seeing what AI will bring to oncology.

The best is yet to come

While we’ve already seen remarkable advancements in diagnostics, drug discovery, and personalized treatment, the next decade will likely bring even more transformative changes.

According to Schoenfeld, we’re just beginning to tap into AI’s potential in oncology, and to him, three areas are particularly promising:

• Predicting immune responses: “AI will predict which unique T cell receptors will bind to specific antigens at a structural level. This could revolutionize cancer immunotherapies, broadening the reach of cell-based therapies and other treatments.
• Designer proteins: “AI is enabling us to create designer mutations in key proteins like cytokines, turning them into therapeutic agents. These engineered proteins can be combined with immunotherapies for highly personalized treatment plans.
• Large language models for predictive analytics: “Unlike earlier disappointments like Watson (IBM’s Watson is an AI system that was heavily promoted as a revolutionary tool in cancer diagnosis and treatment but ultimately failed to meet its high expectations in clinical practice), LLMs have the potential to transform clinical record analysis, finally delivering on AI’s promise to revolutionize how we understand and predict patient outcomes.”
  
  

To Snyder, the future in oncology lies where AI meets biology. “We’re on the cusp of using biocomputing to process biological signals directly, which could allow us to make real-time treatment decisions based on live data from the patient’s body. That’s a game-changer. The next wave of innovation will come from integrating AI with biological systems, pushing us into an era where medicine becomes even more personalized, precise, and predictive. The companies that can combine these technologies meaningfully are the ones to watch.”

More along the lines of tempering expectations, Sorenson thinks health systems cannot bear the cost of AI innovation as we have largely attempted to implement them so far.

“I believe the most successful oncology AI companies will be those that work closely with drug and device developers. There is a different risk profile with pharma-related investments because of the high failure rates of new molecules and the slow adoption of new techniques along the path to becoming a standard of care. However, the economic rewards of pharma-related technology investments can be expected to outpace those on the provider side due to low budgets and fragmented systems that are used there.”

Naturally, I had to ask the most famous AI – ChatGPT – if it thought AI would cure cancer. Its answer? “I’m flattered you think so highly of me, but curing cancer is a team effort. AI will play a big role in the diagnosis, treatment, and discovery of therapies, but we’ll need humans and biology to finish the job.” So, it seems AI is also going for tempered expectations.