Turning ‘Unknown’ Cancer Mutations into Personalized Treatment Strategies

While genetic testing has exploded over the past decade, the ability to fully translate that data to target diseases—such as cancer—has lagged behind. Oftentimes, genetic tests identify mutations whose functional impact and clinical relevance remain unclear, leaving clinicians without clear guidance on whether, or how, to act on them. 

New research led by Columbia’s Andrea Califano, Dr, introduces a computational framework designed to close this gap by systematically classifying so-called variants of unknown functional significance (VUFS), genetic alterations whose biological effects are poorly understood. By classifying these variants based on how they actually alter cellular behavior, the approach reveals new, clinically relevant insights that could help guide more precise cancer treatment. 

A flood of genetic data, but limited answers 

The challenge is vast. The Cancer Genome Atlas, a National Cancer Institute program that molecularly characterized over 20,000 cancer and normal samples, contains more than ten million VUFS, representing a largely untapped reservoir of potential therapeutic insight. 

While the prevalence of known actionable mutations varies widely, from roughly 10% in ovarian cancer to more than 70% in endometrial cancer, the majority of mutations identified through clinical sequencing remain difficult to interpret. Laboratory-based functional assays can characterize individual variants, but scaling those experiments to millions of mutations would be prohibitively expensive and impractical for real-time clinical decision-making. 

“Right now, we’re generating far more genomic data than we can realistically interpret,” says Califano. “As a result, many patients receive sequencing reports that identify mutations, but don’t provide clear guidance about whether those mutations actually matter for treatment.” 

A computational approach to a biological problem 

Rather than relying on the lab-based experiments, Califano’s team turned to computation. They hypothesized that mutations leave behind distinct “signatures” in the activity of transcription factors, key proteins that regulate gene expression. By learning these signatures from mutations whose effects had already been characterized, the researchers could infer the functional impact of unknown variants. 

Using the VIPER algorithm, previously developed in the Califano lab, the team identified transcription factor activity signatures associated with known oncogenic mutations in The Cancer Genome Atlas. Building on this work, the researchers developed a new framework called PHNToM (Protein-activity based identification of Hypermorphic, Hypomorphic, Neomorphic effecTors and therapeutically relevant Mutations).  

PHNToM works by learning the transcription factor activity patterns associated with well-characterized  mutations, then using those patterns to classify VUFS based on their functional effects. These include gain-of-function mutations that increase the activity of oncogenes, loss-of-function mutations that reduce the activity of tumor suppressors, neutral mutations with no measurable effect, and neomorphic mutations, which disrupt protein function in new or unexpected ways. 

Revealing the hidden function of cancer mutations 

As proof of concept, the team first focused on two of the most extensively studied cancer genes, PIK3CA and TP53, in breast cancer. Despite decades of research on these genes, only about 10% of the observed mutations in this cohort currently have known functional significance. PHNToM helped classify many previously ambiguous variants, substantially expanding the set of interpretable mutations. 

Extending the analysis across cancer types, the researchers applied PHNToM to the entire repertoire of oncogenes and tumor suppressors with a number of characterized mutations sufficient to establish reliable activity signatures. In total, they classified more than 500,000 genetic ‘events’—distinct DNA alterations across the tumor genome—dramatically increasing the number of cancer mutations with inferred functional and therapeutic relevance.  

Several findings stood out. The analysis revealed that neomorphic mutations—long thought to be rare—are far more prevalent than previously recognized, and in some cases exert stronger effects than well-established cancer driver mutations. The framework also uncovered instances of mutational mimicry, in which mutations in different genes produce similar cellular signaling effects, suggesting that tumors may share therapeutic vulnerabilities even when their genetic alterations differ. 

“What’s exciting is that this approach doesn’t just tell us whether a mutation is important,” Califano says. “It tells us how it’s important—whether it increases activity, shuts a pathway down, or rewires signaling in a completely new way. That level of resolution is critical for matching patients to the right therapies.” 

Translating genomic insight into personalized treatment 

Importantly, the study went beyond computational prediction. The researchers validated PHNToM’s classifications using independent experimental assays, confirming the functional effects of the vast majority of tested gain-, loss-, neomorphic, and neutral mutations. The framework also captured the context-dependent nature of mutation function, showing that the same mutation can behave differently in different tumor types.  

Equally important, PHNToM accurately identified neutral mutations that do not meaningfully affect protein activity, raising the possibility of sparing patients from targeted therapies unlikely to provide benefit. 

Designed with clinical application in mind, PHNToM leverages the same VIPER-based framework used by OncoTarget and OncoTreat—two VIPER based, CLIA-compliant laboratory tests approved by the CA and NY Dept. of Health and available via Columbia Pathology—thus supporting its clinical application. In principle, dozens of unknown mutations identified in a newly diagnosed patient’s tumor could be functionally classified at once, even a mutation that has never previously been observed, helping clinicians prioritize effective treatments and avoid unnecessary toxicity. 

“This is about making genomic data actionable at scale,” Califano adds. “Instead of testing one mutation at a time, we can computationally assess thousands, in the vast majority of diagnosed tumors, and do it fast enough to matter for patient care.”