A new preprint applies AlphaGenome and AlphaMissense to somatic mutations in 8800 patients across 33 cancer types: gene damage from passenger mutations is associated with survival independently of total tumor mutation burden
A new preprint applies AlphaGenome and AlphaMissense to somatic mutations in 8800 patients across 33 cancer types: gene damage from passenger mutations is associated with survival independently of total tumor mutation burden
On September 16, the laboratory of Ilias Georgakopoulos-Soares at the University of Texas at Austin posted a preprint on medRxiv. They applied AlphaGenome and AlphaMissense, both Google DeepMind models, to 8800 tumors spanning 33 cancer types from The Cancer Genome Atlas (TCGA) in order to measure gene damage from somatic mutations. In patients who carried no known hotspot mutations, this damage was associated with survival, and with treatment outcome in an independent cohort of 570 tumors.
Cancer genomics divides tumor mutations into "drivers," recurrent hotspots in known cancer genes, and "passengers," random substitutions presumed to have no functional significance. Between the two lies a blind spot: many rare mutations look harmless individually, yet collectively impair gene function, and assessing that kind of diffuse damage by hand is not feasible. The laboratory approached this blind spot in June 2026: in a smaller sample, mutation statistics showed that passengers can accumulate to the point of silencing a gene entirely. The new preprint, for the first time, measures this damage with a model that reads the DNA sequence itself.
The authors used AlphaMissense, a DeepMind model that scores how much a mutation harms the resulting protein, and AlphaGenome (Ukhvat previously covered its genome-wide prediction atlas), which takes a DNA sequence and predicts seven measures of gene activity, from how accessible the DNA is to reader proteins to how the RNA copy of the gene is assembled during splicing. Hotspot mutations typically hit the same position in the protein, producing a predictably large effect on protein function. Passenger substitutions are scattered across the gene and more often affect noncoding regulatory regions that classical hotspot detection does not capture. These are therefore two distinct, non-interchangeable channels of damage. In one glioblastoma patient, AlphaGenome even identified which mutation disrupts a splice site in the PTEN gene and in which stretch of DNA.
Before linking damage to survival, the authors checked whether the effect could be explained by mutation count alone, and regressed it out of each damage measure. The signal persisted: even after this adjustment, AlphaGenome predicted microsatellite instability status (a condition where the tumor accumulates mutations because of defective DNA repair) more accurately than raw mutation count in two of three cancer types, and without the adjustment it was nearly error-free in all three.
The central finding concerns patients who carry no known hotspot mutation in a given gene, those who appear "clean" of drivers by standard analysis. Among these patients, higher damage as measured by chromatin accessibility (the density of DNA packing that determines how readily reader proteins can reach it) was accompanied by worse survival in 24 of 25 tested cancer genes. The regulatory protein binding score showed the opposite pattern: in 22 of 25 genes, low damage was associated with worse outcomes. The authors interpret this as the tumor losing programs of aggressive growth rather than sustaining universal harm.
In an independent set of 570 tumors with treatment data, the same damage measure was associated with outcome within specific therapy classes: in breast cancer patients, damage to the RB1 gene was accompanied by worse survival on DNA-damaging chemotherapy, and in colorectal cancer on the same drug class, the gene TCF7L2 showed the association with worse survival.
The authors describe their findings as preliminary because of small subgroup sizes. The next step is to test the effect prospectively and functionally by knocking out the gene and measuring its activity.