Machines and Patient Power Move Biomarker Discoveries Forward

Andy Futreal, PhD

The biggest unanswered question in biomarker development for cancer is how to determine which patients will respond to targeted therapy or immunotherapy. Investigators around the world are trying to answer this through clinical trials and data mining, in addition to finding new tools to add to the mix.

One of those tools is deep learning, a form of artificial intelligence (AI) that uses artificial neural networks modeled after how the human brain processes information to learn and make predictions from large amounts of data. Another tool is crowdsourcing, or collecting large amounts of patient data from multiple sources.¹

“The holy grail in the pretreatment space is what is determining who will respond and who will not, so then you don’t expose patients to drugs that you don’t have to,” said Andy Futreal, PhD, chair of the Department of Genomic Medicine, Division of Cancer Medicine, at The University of Texas MD Anderson Cancer Center in Houston, during an interview with Targeted Therapies in Oncology. “There is this yin and yang thing where everyone is not only trying to understand the responder but also [understand] who may be at higher risk for toxicities.”

However, the challenge is how to generate data on all the patients who receive treatment once they move into the standard-of-care space, Futreal explained.

“Once you start giving approved agents to thousands of patients as opposed to the tens of patients on phase 1/2 [clinical] trials, that’s where you begin to get power and the data begin to emerge, and that’s where you begin to really understand [that there are] germline factors that are driving toxicity, and [eliciting] other questions,” Futreal said.

Delve Deep to Find Answers

Although the study of AI dates to the 1940s and 1950s, as explained in a study published in JCO Clinical Cancer Informatics by Nagy et al, advancements in recent years have brought the use of deep learning to fruition.¹

Deep learning has had the greatest impact on digital image analysis, according to Ash A. Alizadeh, MD, PhD, an associate professor of medicine (oncology) at Stanford University in California.

“If you look at the last decade in oncology, most drug approvals have focused on biomarker selection strategies for targeting patients with a mutation of interest, a genomic subgroup of interest,” he said. “But despite those successes, the number of failures seems to be larger. That doesn’t mean we should give up. But a key concept is the early hope of finding a biomarker that can work in general, to avoid false discovery in the front line and [aid] the identification of biomarkers, and this cuts across all the technologies, whether it’s imaging or molecular diagnostics or data mining from electronic records. They all have promise, but the key [goal] is to find signals so we can move the needle for any given patient [with cancer].”

Investigators have begun to explore the application of deep learning in cancer types such as melanoma and breast, colorectal, and metastatic non–small cell lung (NSCLC). For instance, a study conducted by Skrede et al at the Institute for Cancer Genetics and Informatics in Oslo University Hospital in Norway used deep learning to develop a clinically useful prognostic marker for colorectal cancer.²The investigators obtained slides of hematoxylin and eosin (H&E)–stained histopathology sections from patients following tumor resection. By training 10 convolutional neural networks on more than 12 million image tiles derived from slides of patients with distinctly good or poor disease outcomes, they developed a biomarker and tested it on 920 patients’ slides prepared in the United Kingdom, followed by independent validation in 1122 slides prepared in Norway from patients treated with capecitabine. The biomarker reliably predicted cancer-specific survival, stratifying patients with stage 2 or 3 disease into distinct prognostic groups. By identifying whether patients have a sufficiently low risk of recurrence so that adjuvant chemotherapy can be avoided, the biomarker could help guide treatment decisions.²

Similarly, investigators in South Korea worked with Lunit SCOPE, a deep learning–based H&E image analytics tool, using 1824 H&E images from immune checkpoint inhibitor (ICI)–naive patients with metastatic NSCLC. Using samples from ICI-treated patients (n = 189), an AI score was calculated that predicted clinical response and progression-free survival after ICI treatment independently of PD-L1 expression.³

To derive maximal benefit from deep learning, clinicians must accomplish 3 overall objectives, explained Futreal. First, clinicians need to be able to work in real time. “When you do large-scale data generation, you have to be able to give something back immediately because it’s important to the clinical trial to correlate the most important data. Did the patient respond? Did the tumor get bigger, smaller, or stay the same?” he said.

Next, clinicians must examine how this information applies to translational research. For example, Futreal explained, BRAF is mutated in multiple tumor types, but what are the implications of a clinician deciding which MEK inhibitor to use in a patient with lung or thyroid cancer?

“If you do both of these things right, and spend time on...data quality, flows, transformations, then you have high-quality data [that can be used in deep learning] because now...the machine can look for those correlations,” he said.

Finally, after the framework is built, the longterm repercussions for clinical biology need to be explored, Futreal said. “The more data we have, the more we can measure, and the more likely we are to hit upon the nuggets that can be pulled out and refined,” he added.

At MD Anderson, platforms have been built out over the last 9 years that were solely concerned with longitudinal sample collection from patients. Now they are focused on scaling that information, Futreal explained, and starting in the rare cancer space, where much of the infor-mation already exists.

In addition, a large-scale project called Patient Mosaic, which is being led by Futreal’s laboratory, will focus solely on patient populations receiving standard-of-care therapy to develop a precision-driven, personalized approach. Germline and tumor sequence data will be compared to determine how treatment leads to genetic changes that drive resistance over time, and patients’ microbiomes will be examined. Investigators will begin by evaluating how these factors affect immunotherapy and cell-based therapies and then move on to targeted therapies.

“We know with certain cancer types, the microbiome seems to play a role in terms of response, as well as patients’ body mass....It is a complicated space, and it may usher in a new era of more patient-centric complex biomarkers than the tumor-driven biomarkers we tend to think of,” Futreal said.

Turning to the Crowds

One key to the success of establishing new cancer biomarkers may lie in the hands of the public. Data sharing and crowdsourcing have been studied to speed up the process of piecing together information from images and annotating data.⁴

In a recent JCO Precision Oncology commentary, Benjamin G. Vincent, MD, and colleagues at the University of North Carolina School of Medicine, Chapel Hill, explained how crowd-sourcing can drive precision medicine with immunotherapy. Dialogue on Reverse Engineering and Assessment Methods (DREAM) Challenges are enabling collaboration among experts around the world.

Three key DREAM Challenges in oncology, according to Vincent et al, have demonstrated the value of crowdsourcing: the Prostate Cancer DREAM Challenge, the Multiple Myeloma DREAM Challenge, and the Digital Mammography DREAM Challenge.⁵

Novel, previously underreported prognostic biomarkers were identified through the Prostate Cancer DREAM Challenge. In the Multiple Myeloma DREAM Challenge, 171 models were assessed, which led to discovery of the epigenetic regulator PHF19 as a novel marker of aggressive disease and high risk of early progression. And the Digital Mammography DREAM Challenge showed how integrating AI and radiologist assessment can improve the interpretation of breast cancer screening mammograms.⁵

Investigators are now beginning the Anti–PD-1 Response Prediction Challenge to explore alternative biomarkers to PD-L1 expression and tumor mutational burden. Their focus is on developing models that predict which patients will benefit from nivolumab (Opdivo) monotherapy in NSCLC. Patient samples from a phase 3 clinical trial, CheckMate 026 (NCT02041533), will be used.⁵ “Successful completion of this DREAM Challenge will provide a model for future collaborations among industry, academia, and citizen scientists to discover determinants of response or resistance to immunotherapy, facilitating rapid development of clinically actionable biomarkers,” the authors wrote.

Challenges and Limitations

However, deep learning and crowdsourcing come with some hurdles. For instance, images may require extensive preprocessing because of the data size and amount of nontumor tissue in a typical histology slide, which Echle et al explained can dilute the information content.⁶

Patient privacy and confidentiality are also a concern with crowdsourcing. “We have to look at what role does empowering patients have,” Futreal said. “Patients have the capability to electronically say, ‘Boom. I want to give you my data, and these are the things I would like you to do with it.’ That changes the nature of the government and compliance piece, as well as the patient involvement, which is critical.”

Futreal also described the frustration and obstacles involved in capturing electronic health records (EHRs) where patient data are kept.

“The best time to ensure that the data are complete is at the initial collection, and that involves the EHR,” he said. “We need to make it not 18 clicks to get to high-quality data collection, but [down] to 3 clicks that doesn’t slow the team down.”

Filling the Glass

nce these areas are fine-tuned and the analytics are built out, the application of deep learning can extend beyond research at academic institutions and into community oncology groups. But that can only work, Futreal explained, if there is interinstitutional harmonization.

“If we make all of our apples Granny Smith, McIntosh, and Fujis, but another group is collecting apples in the community and calling them red or green apples, then we have to understand when they say red apples, they are including the 3 red apple types that we’re talking about so that you can extrapolate,” he said. “If you can’t extrapolate data, then that’s a missed opportunity.”

Leveraging this technological progress to advance clinical applications will make it useful for the everyday patient, Alizadeh explained.

“The glass is half full, and there are lots of opportunities to keep filling it, but how do we improve on bang for the buck in terms of dollars spent, time, and learning from individual patients?” Alizadeh said. “I don’t think anyone has a solution, but there are a number of creative strategies, such as finding exceptional responders and studying them or going back to clinical trials and studying those patients post hoc to identify biomarkers. There isn’t a 1-size-fits-all solution across the board.”

_References

_{1. Nagy M, Radakovich N, Nazha A. Machine learning in oncology: what should clinicians know? JCO Clin Cancer Inform. 2020;4:799-810. doi:10.1200/CCI.20.00049}

_{2. Skrede OJ, Raedt SD, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. Lancet. 2020;395(10221):350-360. doi:10.1016/S0140-6736(19)32998-8}

_{3. Park S, Ahn CH, Jung G, et al. Deep learning-based predictive biomarker for immune checkpoint inhibitor response in metastatic non-small cell lung cancer. J Clin Oncol. 2019;37(15):9094. doi:10.1200/JCO.2019.37.15_suppl.9094}

_{4. Amgad M, Elfandy H, Hussein H, et al. Structured crowdsourcing enables convolutional segmentation of histology images. Bioinformatics. 2019;35(18):3461–3467. doi:10.1093/bioinformatics/btz083}

_{5. Vincent BG, Szustakowski J, Doshi P, et el. Pursuing better biomarkers for immunotherapy response in cancer through a crowdsourced data challenge. JCO Precis Oncol. 2021;5:51-54. doi:10.1200/PO.20.00371}

_{6. Echle A, Rindtorff NT, Brinker TJ, et al. Deep learning in cancer pathology: a new generation of clinical biomarkers. Br J Cancer. 2021;124(4):686-696. doi:10.1038/s41416-020-01122-x}