USC scientists use AI to turn deadly brain cancer cells against themselves

USC researchers just trained AI to turn brain cancer cells against themselves.

In a new study published in Cancer Immunology Research, scientists at the Keck School of Medicine of USC used machine learning to reprogram glioblastoma cells into dendritic cells. Instead of trying to force immune cells across the blood-brain barrier, the therapy effectively recruits the tumor's own cells to sound the alarm.

In mouse models, this Trojan horse strategy boosted survival chances by up to 75% when combined with standard immunotherapy.

Turning tumors into immune factories

Glioblastoma is the deadliest and most common adult brain cancer, with a five-year survival rate below 10%. While immunotherapy has transformed treatments for skin and lung cancers, it rarely touches glioblastoma because the blood-brain barrier blocks most immune cells from ever reaching the tumor.

Lead author Dr. David Tran and his team realized they could bypass that barrier entirely by changing the tumor from the inside out.

Dendritic cells act as the immune system's intelligence officers. They sample cancer antigens and show them to T-cell armies to trigger an attack. The researchers wondered: what if the cancer cells inside the brain could be reprogrammed to act like dendritic cells instead?

Reprogramming a cell with 20,000 genes and millions of regulatory connections is nearly impossible using trial-and-error biology. USC's machine learning model sifted through those massive networks to identify the exact genetic switches that force glioblastoma cells to change identity without altering healthy brain cells.

When tested in mice, existing checkpoint inhibitors and classic dendritic cell vaccines did nothing on their own. But when paired with AI-driven reprogramming, the checkpoint combo improved survival rates by 75%, while pairing it with a dendritic cell vaccine doubled survival.

Why it matters

For patients and oncologists facing one of the most hopeless diagnoses in medicine, this represents a completely different strategic vector. It is not ready for the clinic tomorrow, as human trials are still several years away while researchers test viral delivery vectors in animals.

Across the biopharma landscape, this work shows AI graduating from molecular screening to full cell-fate engineering. Instead of just designing molecules that fit into receptor pockets, models are now mapping how to fundamentally rewire cellular identity.

We think the brilliance of this approach is that it stops fighting the blood-brain barrier. Rather than trying to ship massive immune armies into a hostile, inaccessible environment, it converts the enemy garrison into whistleblowers.

The obvious bottleneck remains delivery. Packaging these human gene combinations into viral vectors and ensuring they only rewrite tumor cells inside a human brain will be a massive regulatory and clinical hurdle. If it succeeds, the same concept could theoretically be adapted across other notoriously hard-to-treat solid tumors.