Artificial intelligence (AI) may be all over the news now, but it’s been in oncology clinics for more than three decades, helping oncologists analyze mammograms and other imaging scans to detect signs of cancer—albeit with some limitations. With recent technological advances, however, cancer researchers are now exploring new possibilities for AI, improving existing applications and developing additional ones to expand AI’s role in the clinic.
At the AACR Special Conference on Artificial Intelligence and Machine Learning and the AACR webinar on AI and Machine Learning in IO, presenters discussed how AI can be applied across the cancer research spectrum, highlighting its promise to improve our understanding of cancer biology, accelerate drug development, and improve patient care. In Part 1 of this blog post, we dove into some of the innovative applications for discovery science and translational cancer research discussed at these events.
Here, we will examine some of the promising clinical applications of AI, including its ability to enhance cancer diagnosis, guide precision medicine, and even facilitate the development of other clinical AI tools.
Improving Cancer Diagnoses
Accurate cancer diagnosis is key to ensuring patients receive the appropriate treatment for their cancer. This entails not only identifying cancer type (the organ where the cancer originated), but in many cases also determining the cancer subtype (which can be based on the cell or tissue it affects, the genetic alterations involved, and/or its morphologic features).
The complex process of diagnosing cancer subtypes typically relies on the expertise of pathologists to identify and interpret characteristics from hematoxylin and eosin (H&E)-stained tumor tissue, sometimes supplemented by data from sophisticated and expensive clinical assays like tumor genome sequencing.
AI, however, could help make this process more efficient and accessible, according to research presented by Kevin Boehm, MD, PhD, of Memorial Sloan Kettering Cancer Center, during the AACR Special Conference.
“An area where AI has shown superhuman performance is in pattern recognition and scalability,” said Boehm, noting that AI has the potential to quickly review thousands of H&E images, recognize established histologic patterns, and even identify novel, clinically relevant patterns.
Boehm presented an AI architecture designed to infer subtype and genomic information from digitized images of H&E-stained tumor tissue. The first step of the architecture was an AI-based model called AEON, which analyzed H&E images from roughly 80,000 samples to identify patterns that it then used, in combination with information from the open-source cancer classification system OncoTree, to classify the histologic subtype of each tumor.
AEON, combined with OncoTree, classified cancer subtypes with 78% accuracy, Boehm reported, and it was able to reclassify tumors into a more granular subtype than they had been assigned by a pathologist. This included reclassifying renal cell carcinomas (RCC) not otherwise specified as either clear cell RCC or papillary RCC subtypes, as well as assigning a cancer type to tumors previously diagnosed as cancers of unknown primary. In most cases, the overall survival of patients with reclassified cancers of unknown primary was consistent with what would be expected for the newly assigned cancer type.
In the second step of the AI architecture, a model called Paladin, also developed by Boehm and colleagues, integrated the granular subtype classifications from AEON with the digitized H&E images to infer genomic properties of each subtype based on patterns captured in the H&E images. Boehm reported that about 5% of the nearly 4,000 variants they examined could be reliably inferred through histologic patterns.
Because AEON identified granular subtypes that are often lumped together, the researchers used Paladin to uncover subtype-specific genotype-phenotype relationships that had been previously masked. MEN1 variants, for example, had been reported to drive pancreatic cancers, but Paladin found that these variants were not as relevant for the neuroendocrine tumor subtype of pancreatic cancer.
This has implications for precision medicine, as targeting certain variants may not be effective against all the subtypes of a particular cancer—underscoring the importance of accurate diagnoses.
“We can’t just lump all of these histologies together and infer genomic features,” Boehm said. “Each granular subtype must be considered separately.”
Because the AI architecture relies on H&E images instead of expensive sequencing methods, Boehm suggested that, with further research, it also “has the potential to help extend access to precision oncology to centers where it’s not logistically or financially feasible to run DNA sequencing on a large cohort of patients.”
(In another recent blog post, we discussed how AI-driven analysis of digital H&E images enhanced the diagnosis of pediatric sarcomas.)
Predicting Treatment Responses
Another emerging clinical application of AI is to predict how a patient’s cancer will respond to a particular treatment. This can help physicians decide which treatment to pursue so that patients can avoid undergoing therapies that are unlikely to be effective.
During the AACR webinar, Arnav Mehta, MD, PhD, of Stanford University, shared recent research that illustrated how AI can help researchers identify biomarkers or intercellular interactions associated with immunotherapy responses and integrate multiple types of data to improve response predictions.
“We’re in a really exciting era in how we use machine learning and artificial [intelligence] tools, not just in basic science or at the level of target identification, but really across the entire drug discovery cycle with … better clinical development and patient stratification,” Mehta said.
Using ‘Synthetic’ Patients to Guide Precision Medicine
A major obstacle to using AI to predict treatment outcomes, however, is that the development of AI models for this purpose requires large swaths of patient data that are not always available or representative of the patient population.
Research from Hanna Hieromnimon, of the University of Chicago, presented at the AACR Special Conference, suggests that AI-generated patient data—that is, artificial data from synthetic patients—could be one way to overcome this hurdle.
“Currently, researchers rely on limited public datasets that may not represent the full diversity of cancer patients,” said Hieromnimon. “Our goal was to create a method that can generate realistic ‘synthetic’ patients—complete with both digital pathology images and clinical data—that could help researchers build better multimodal AI models for cancer diagnosis and treatment [response] prediction.”
Hieromnimon and colleagues developed an AI-based tool that was trained on clinical information and digitized histology images from real patients. The model used these inputs to learn the connections between the provided clinical and histologic data and develop a reference map that plotted real patients based on their similarities to one another—the more similar that two patients are, the shorter the distance between them on the reference map. Finally, the model used the reference map as a guide for generating realistic synthetic patients.
When an AI model was trained on data from 1,000 synthetic lung cancer patients, it predicted immunotherapy responses with similar accuracy as when it was trained on data from 1,630 real patients (68.3% vs. 67.9% accuracy, respectively). And regardless of whether the model was trained on real or synthetic data, it correctly identified bone or liver metastases and lactate dehydrogenase (LDH) levels as prognostic factors.
“Synthetic patients were remarkably faithful to real patient data,” Hieromnimon said. She explained that the use of synthetic patients could accelerate the development of AI models by allowing investigators to augment their training data and impute missing data to improve the performance of their models. Synthetic patients could also enhance scientific collaboration, she added, explaining that since synthetic data does not have the same privacy concerns as real patient data, it can be shared more easily—which could help expand the reach of AI-based tools that advance precision medicine.
“Precision cancer treatment relies on expensive molecular tests that aren’t available globally,” Hieromnimon noted. “AI excels at finding subtle patterns across multiple data types simultaneously … If we can develop better AI models that work with standard tissue images, more patients could benefit from personalized treatment regardless of their geographic location or economic circumstances.”
With the growing influence of AI in cancer research and clinical care, it will be necessary to carefully consider the pros and cons of new AI tools, promote open science, and enhance collaboration, said panelists during the “AI Synergy Forum: Fostering Innovation Through Collaboration” session at the AACR Special Conference.
Panelist Elana J. Fertig, PhD, of the University of Mayland School of Medicine, noted that researchers should consider whether their AI model solves a new problem or uncovers a novel aspect of biology, whether it is easier to use than conventional methods, and how it performs relative to other methods. Additionally, understanding the conditions in which the model no longer works and scrutinizing performance data are key to ensuring that a model isn’t “too good to be true.”
Once effective models are developed, ensuring that others can access the data and code that led to their development is an important step towards open science, said panelist Shirin A. Enger, PhD, of McGill University in Montreal. This will require efforts to protect patient privacy when sharing data, as well as proactive code annotation so that other researchers can reuse a code.
Ultimately, the goal of using AI tools is to use existing data in a way that benefits patients. Panelist Skye Bork, from PACT AI, emphasized the mutually beneficial power of collaboration between startup companies and academic centers to drive faster science and better care. As an example, she highlighted how her company has collaborated with academic centers to develop and deploy AI models to prescreen patients for clinical trials using patient data from electronic health records and information from the clinicaltrials.gov website.
With these considerations, AI has the potential to transform cancer research and patient care. As Bork noted, “Our goal here is better and faster care for patients. I think we’re united in that.”
To learn about more AI applications across the cancer research spectrum, check out the special series on Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI in the AACR journal Cancer Research. The series includes additional examples of clinical applications of AI, including: