Powerful AI Algorithm Provides New Insights into Deadly Glioblastoma and Other Cancers

Antonio Iavarone, MD

An international research group, led by scientists at Sylvester Comprehensive Cancer Center at the University of Miami Miller School of Medicine, has developed a new machine learning algorithm that can combine large datasets from multiple sources to pinpoint cancer-driving enzymes called kinases and potentially target them for treatment.

In research published on February 2 in Nature Cancer, the team developed and tested the Substrate PHosphosite based Inference for Network of KinaseS (SPHINKS) algorithm’s AI in glioblastoma, one of the deadliest cancers.

In the study, SPHINKS identified kinases (PKCδ and DNAPKcs) associated with progression in two glioblastoma subtypes. In addition, by using tumor organoids grown in the laboratory from patient samples, the team showed targeted drugs that interfere with the two kinases can block tumor growth. These aberrant kinases were also activated in tumor subtypes in breast, lung, and other cancers.

“This is translational science that offers immediate opportunities to change the way glioblastoma patients are routinely managed in the clinic,” said Antonio Iavarone, MD, deputy director of Sylvester Comprehensive Cancer Center and senior author on the study. “SPHINKS gives oncologists a new tool to apply the correct treatments in homogenous cancer subtypes.”

Despite breakthroughs in other cancers, glioblastoma patients face dismal prognoses — the 5-year survival rate is below 10%. While numerous drugs are being developed to help improve these numbers, clinicians have needed a way to identify the molecular mechanisms that drive each patient’s disease and personalize care.

In previous work, the Iavarone team had reported a new glioblastoma classification, capturing key tumor cell traits and grouping patients based on their chances of survival and their tumor’s vulnerability to drugs. In this Nature Cancer paper, these classifications were independently confirmed through several omics platforms: genomics (genes), proteomics (proteins) lipidomics (fat molecules), acetylomics (epigenetics), metabolomics (metabolites), and others.

SPHINKS leverages machine learning to refine these omics datasets and create an interactome (a complete set of biological interactions) to zero in on the kinases that generate aberrant growth and treatment resistance in each glioblastoma subtype. These findings show that multi-omics data can generate new algorithms that predict which specific targeted therapies can provide the best therapeutic options based on each patient’s glioblastoma subtype.

“We can now stratify glioblastoma patients based on biological features that are concurrent between different omics,” said Dr Iavarone. “Reading the genome alone has not been enough; we have needed more comprehensive data to identify tumor vulnerabilities.”

The SPHINKS algorithm, and associated methods, can be readily incorporated into molecular pathology labs. The paper includes a clinical classifier that can assign the appropriate glioblastoma subtype to each patient. The team has also established an online portal to access the algorithm. The authors believe this approach can produce actionable information that could benefit as many as 75% of glioblastoma patients.

“This classifier can be used in basically any lab,” said Anna Lasorella, MD, professor of biochemistry and molecular biology and co-senior author on the study. “By importing the omics information into the web portal, pathologists receive classification information for one tumor, ten tumors, whatever they import. These classifications can immediately be applied to patient care.”

While SPHINKS was first tested on glioblastoma, the algorithm is equally applicable in many cancers. The team found the same cancer-driving kinases in breast, lung, and other tumors. Drs Iavarone, Lasorella, and colleagues believe this information could even help create a new type of clinical trial.

“We are exploring the concept of basket trials,” said Dr Iavarone, “which would include patients with the same biological subtype but not necessarily the same cancer types. If patients with glioblastoma or breast or lung cancer have similar molecular features, they could be included in the same trial. Rather than doing multiple trials for a single agent, we could do one and potentially bring more effective drugs to patients faster.”