Passenger mutations and survival: AlphaGenome applied to 8,800 tumors
A UT Austin lab applied Google DeepMind's AlphaGenome and AlphaMissense to 8,800 tumors across 33 cancer types: calculated passenger-mutation damage associates with survival independently of total mutation burden, with the effect direction depending on the regulatory channel measured; findings are preliminary.
A September 16 preprint from the lab of Ilias Georgakopoulos-Soares (University of Texas at Austin) analyzed 8,800 tumors across 33 cancer types: calculated gene damage from "passenger" mutations was associated with patient survival. After statistically adjusting for total tumor mutation burden, the signal held.
Google DeepMind built the two models used in this analysis (Google's page on Eternal Search is linked at the end). The university lab ran the analysis on TCGA data (The Cancer Genome Atlas, a catalog of tumor genomes). Two instruments: AlphaMissense scores how much a specific mutation disrupts protein function; AlphaGenome takes a DNA sequence and predicts how changes in it affect gene activity, including how accessible the DNA is to regulatory proteins and how the RNA copy of the gene is assembled during splicing.
"Passenger" mutations are random DNA substitutions in a tumor genome, conventionally assumed to have no functional significance. "Driver" mutations recur at the same positions in known cancer genes and produce 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. This study tests whether the collective effect of passenger mutations can damage a gene, given that many of these mutations individually look harmless.
The central finding involves patients who carry no known hotspot mutation in a given gene (those who look clean by standard analysis). In these patients, higher calculated gene damage, measured through chromatin accessibility (the density of DNA packing, which determines how readily regulatory proteins can reach the genetic material), was associated with worse survival in 24 of 25 tested genes. For the regulatory protein binding score, the association reversed: in 22 of 25 genes, lower calculated damage corresponded to worse prognosis. The authors interpret this as the tumor losing programs of aggressive growth.
AlphaGenome also passed a separate test: it predicted microsatellite instability status (a condition where tumors accumulate mutations because of defective DNA repair) more accurately than raw mutation count in two of three cancer types.
In a separate set of 570 tumors with treatment data, the same damage measure was associated with therapy outcomes. In breast cancer patients, damage to the RB1 gene was accompanied by worse survival on DNA-damaging chemotherapy. In colorectal cancer patients on the same drug class, the gene TCF7L2 showed the same association with worse survival. The subgroups are small, and the authors call the findings preliminary. The next step is prospective testing and gene knockouts (deliberate gene inactivation) with activity measurement, to confirm whether these statistical associations reflect real biological effects.
The preprint uses Google DeepMind's AlphaGenome and AlphaMissense as its core analytical instruments, making the Google organization page in Eternal Search a concrete anchor: the tools are Google's, the analysis is a third-party university lab's, and the page lets readers examine the company through that specific research application rather than its consumer products.