AI Maps Whole Gene Networks That Cancers Cannot Live Without

Precision oncology has always promised more than it has delivered. The idea is elegant: sequence a patient’s tumor, find the mutated gene driving its growth, and prescribe a drug that shuts that gene down. In practice, only a small fraction of patients benefit, because most tumors do not carry a mutation that matches an available drug. A new study published in BMC Genomics argues that the problem lies not just in which genes we target, but in the very concept of targeting single genes at all. Instead of asking which individual genes a cancer cell needs to survive, researchers asked which entire biological processes the cell depends on — and then built a machine learning framework to find them at scale.

The work, led by Julia B. Curd and Gregory P. Way of the University of Colorado Anschutz, together with pediatric neuro-oncologist Adam L. Green and colleagues, builds on one of the largest datasets in cancer biology: the Cancer Dependency Map, or DepMap. DepMap systematically knocks out, one at a time, thousands of genes across hundreds of cancer cell lines and records which losses kill which cells. That effort dramatically expanded the catalog of potential drug targets beyond the handful of famous oncogenes. But Curd and her colleagues point to a fundamental weakness in the single-gene paradigm: drugs rarely act on only the molecule they are designed to hit. Off-target effects and polypharmacology — the tendency of one drug to interact with many proteins — can undermine therapies aimed at a lone gene, and some vulnerabilities are simply invisible when genes are considered in isolation.

The team’s solution is a framework called BioBombe, an artificial intelligence and machine learning approach that treats gene dependency data as a high-dimensional signal to be decomposed. Rather than fitting a single dimensionality reduction model, BioBombe fits many hundreds of them, spanning a large range of latent dimensionalities — essentially asking, at every level of resolution, what coherent patterns of gene dependency explain the survival of cancer cells. Low-dimensional models capture broad, cell-wide programs, while high-dimensional models resolve narrow, specific pathways. Running the full spectrum means that no signal is missed simply because a single model happened to be tuned to the wrong scale.

This multiple-model strategy proved decisively better than any individual model. By combining results across hundreds of decompositions, BioBombe uncovered far more gene process dependencies — groups of genes within a biological process that cells rely on for survival — than any single decomposition could reveal on its own. The researchers then annotated these discovered patterns using gene set enrichment analyses against two curated knowledge bases: Reactome, a comprehensive database of biological pathways, and CORUM, a repository of mammalian protein complexes. The result is what the authors call a landscape of gene process dependencies: a systematic map of which cellular processes different cancers cannot live without.

The landscape is striking in both its breadth and its specificity. Some dependencies are nearly universal, reflecting the basic machinery of cell division and metabolism. Mitotic regulation — the intricate process governing chromosome segregation during cell division — emerged as a broad vulnerability, as did the citric acid cycle, the mitochondrial engine that converts nutrients into cellular energy. These are processes that rapidly dividing tumor cells lean on heavily, and they have long been implicit targets of chemotherapy. What the new analysis adds is a principled, data-driven way to rank such processes and connect them to specific cancer types.

More clinically interesting are the cancer type-specific dependencies that surfaced. In gliomas, the aggressive brain tumors that include some of the deadliest forms of cancer in both adults and children, two themes stood out: TP53-related pathways and mitochondrial biology. TP53 is the most frequently mutated gene in human cancer, the guardian of the genome whose loss frees cells to accumulate damage unchecked. The finding that glioma cells with compromised TP53 function show process-level dependencies around that pathway — and around mitochondrial metabolism — suggests that the way in to killing these tumors may not be a single magic bullet but a coordinated attack on the networks the tumor has rewired to compensate for its broken defenses.

Turning dependencies into treatments requires connecting them to drugs, and the researchers did exactly that. By linking gene process dependency profiles with drug sensitivity scores measured on the same matched cell lines, they identified both established drug candidates — compounds already known to work through the relevant biology, which serves as an internal validation of the approach — and novel candidates that had not previously been associated with those vulnerabilities. Because the analysis operates at the level of processes rather than single protein targets, it naturally embraces drug polypharmacology rather than fighting it: a compound that hits multiple nodes in a dependent pathway may be an advantage, not a liability.

The team then pushed the framework one step further, toward the clinic. They developed a machine learning model that predicts gene process dependencies and their associated drug sensitivities directly from RNA-seq data — the routine gene expression profiles that can be generated from a patient’s tumor biopsy. In other words, a clinician could, in principle, feed a tumor’s transcriptome into the model and receive a ranked list of the biological processes that tumor depends on, along with drugs likely to exploit them. To test the concept, the researchers applied it to pediatric high-grade glioma, a devastating brain cancer with tragically limited treatment options. The model predicted that cladribine, an existing drug used for hairy cell leukemia and multiple sclerosis, might be effective against these tumors — and laboratory experiments confirmed that cladribine does indeed kill pediatric high-grade glioma cells.

That last result illustrates the most practical promise of the study: drug repurposing. Developing a new cancer drug from scratch takes a decade or more and costs well over a billion dollars, with most candidates failing in trials. Repurposing approved drugs, by contrast, offers compounds with known safety profiles that can move into new indications quickly. By predicting cladribine’s activity against a childhood brain cancer and validating it in the lab, the study demonstrated an end-to-end pipeline — from dependency discovery, through drug matching, to experimental confirmation — that could be rerun for many other tumor types.

Methodologically, the authors emphasize that BioBombe is both scalable and interpretable, two qualities that often pull in opposite directions in machine learning. Deep black-box models can be powerful but opaque; BioBombe’s dimensionality reduction approach, anchored to known pathways through enrichment analysis, yields patterns that biologists can read, question, and test. The work was supported by Alex’s Lemonade Stand Foundation, Tap Cancer Out, and the Morgan Adams Foundation, reflecting a research program with a strong pediatric focus. The authors report no conflicts of interest, and the paper is open access under a Creative Commons license.

The broader implication is a shift in how precision oncology might define its targets. For two decades, the field has been organized around the gene: find the driver mutation, drug the protein. This study suggests a complementary paradigm organized around the process: identify the biological program a tumor has become dependent on, and attack it with whatever agents — single drugs or combinations — disrupt that program most effectively. Such an approach could reach patients whose tumors lack any druggable single-gene lesion, which is to say, most patients. It also aligns with how drugs actually behave in cells, where promiscuity is the rule. The mapping of gene process dependencies across DepMap’s cell line panel, now published as a resource, gives the field a starting atlas — and the cladribine result gives it proof that the atlas can point to a bottle already on the pharmacy shelf.

Subject of Research: Machine learning analysis of gene process dependencies in cancer using DepMap data to identify pathway-level drug targets

Article Title: Characterizing the landscape of gene process dependencies in cancer

Article References: Characterizing the landscape of gene process dependencies in cancer. (n.d.). https://doi.org/10.1186/s12864-026-13457-5

Image Credits: AI Generated

DOI: 10.1186/s12864-026-13457-5

Keywords: DepMap, BioBombe, gene dependency, precision oncology, dimensionality reduction, glioma, pediatric high-grade glioma, drug repurposing, cladribine, TP53, mitochondrial metabolism, machine learning