Cancer researchers have spent decades cataloging mutations found in tumors, building databases that now hold millions of entries. The problem was never finding variants. It was figuring out which ones actually drive the disease and which are just along for the ride. In May 2025, a team at Google DeepMind, led by senior researcher Ziga Avsec and colleagues, offered a new way to attack that bottleneck: a deep-learning model called AlphaGenome, published in Nature, that reads long stretches of DNA and predicts how single-letter genetic changes ripple through the regulatory machinery that keeps cells healthy. Now, as of June 2026, the model’s predictions are being matched against some of the largest gene-editing experiments ever conducted, and the early results suggest pharmaceutical companies may finally have a reliable way to rank tens of thousands of cancer mutations by the ones worth targeting first.
What AlphaGenome actually does
Most AI tools that analyze DNA variants focus on one layer of biology at a time: Does this mutation change a protein’s shape? Does it sit near a known regulatory element? Earlier models such as Enformer, also developed at Google DeepMind, could predict gene expression from sequence but were limited to shorter input windows and a narrower set of regulatory readouts. Other tools like the Sei framework from the Troyanskaya lab at Princeton offered chromatin-level annotations but did not integrate as many output modalities simultaneously. AlphaGenome works differently. It processes sequences up to one million base pairs long and simultaneously predicts effects across multiple regulatory layers, including gene expression levels, chromatin accessibility, and transcription factor binding. That matters because cancer-driving mutations rarely do just one thing.
Consider a single nucleotide change in a stretch of non-coding DNA near a tumor suppressor gene. That change might weaken a transcription factor’s grip on a nearby binding site, which in turn closes off a patch of chromatin, which reduces expression of the gene that was keeping cell growth in check. AlphaGenome’s architecture is built to model these cascading effects and collapse them into a unified score reflecting how damaging a variant is likely to be. The result is a ranked list: variants predicted to be strong drivers at the top, likely benign passengers at the bottom.
The editing experiments that serve as ground truth
Predictions are only as useful as the data that tests them. Over the past two years, several research groups have run gene-editing screens at a scale that would have been unthinkable a decade ago, physically installing thousands of cancer-linked mutations into cells and measuring what happens.
One base-editing screen, published in Nature Genetics, tested 32,476 variants across 11 major cancer genes, mapping which mutations confer drug resistance and which leave cell behavior unchanged. A separate prime-editing effort, reported in Nature Biotechnology, generated 54,007 unique guide RNAs and compiled 3,191 unique EGFR-region variants drawn from the COSMIC somatic mutation database and the ClinVar clinical variant archive. Another campaign used prime editing sensor libraries to evaluate over 1,000 endogenous TP53 variants, linking each mutation’s identity to a measured functional outcome. And a saturation mutagenesis study, building on the approach described by Giacomelli et al., covered 94.5 percent of cancer-associated TP53 missense mutations, directly measuring their effects on tumor-cell fitness.
These datasets function as ground truth. They tell researchers what each mutation actually does in living cells, not what a model thinks it might do. When AlphaGenome’s predictions are benchmarked against such measurements, a feedback loop emerges: the model predicts, the editing screen confirms or refutes, and the model’s next iteration improves. For drug makers, that loop compresses the timeline between discovering a mutation in a patient’s tumor and knowing whether it is worth building a therapy around.
Clinical signals that go beyond the lab bench
Lab screens can show that a mutation disrupts a protein or confers resistance to a drug. But the question oncologists care about is whether those findings translate to real patients. Early validation work has started to address that gap. AI-predicted driver and pathogenic variant classifications were tested against outcomes in non-small-cell lung cancer survival cohorts and in a separate pan-cancer tumor profiling cohort. Variants that AlphaGenome labeled as likely drivers tended to appear in tumors with poorer prognosis, while variants predicted to be benign were more common in patients with comparatively better outcomes.
That correlation is encouraging, but it comes with an important caveat: retrospective survival analysis is not the same as a prospective trial. The data show that the AI’s classifications track with how patients fared, but they do not yet prove that using AlphaGenome to select drug targets leads to better therapies. That kind of evidence will require clinical studies designed specifically to test AI-guided target selection, and those studies are still in the planning stages.
The non-coding frontier
Most clinical oncology still focuses on mutations that change a protein’s amino-acid sequence. But a growing body of experimental work is revealing that variants in non-coding DNA, regions that do not encode proteins but regulate how much of a protein gets made, can be just as consequential for tumor growth.
Recent screens have experimentally evaluated tens of thousands of non-coding variants affecting mRNA abundance in cancer driver genes. In several cases, subtle changes in the 3-prime untranslated regions of genes altered how stable an mRNA molecule was, effectively changing how much protein the cell produced without touching the protein’s structure. Select variants were confirmed through targeted prime editing to have cancer-driving potential.
For drug developers, these findings open a new category of targets. But they also introduce uncertainty. The connection between mRNA-level changes and clinical outcomes is less well established than for protein-altering mutations. Companies considering non-coding targets will likely need to demonstrate that modulating the same regulatory elements with small molecules or antisense oligonucleotides yields meaningful tumor responses in preclinical models before committing to expensive clinical programs.
What has not been settled
The published studies establish two things independently: AlphaGenome can predict regulatory consequences of DNA variants, and large editing screens can measure variant effects at industrial scale. What has not yet appeared in a single, unified analysis is a direct head-to-head comparison of AlphaGenome’s predictions against the functional outcomes measured in the 32,476-variant base-editing screen. The two lines of evidence are complementary, but linking them explicitly would strengthen confidence that the AI can prioritize wet-lab testing rather than simply supplement it.
Without that comparison, practical questions linger. How often would AlphaGenome misclassify a variant that base-editing experiments show to be strongly resistance-conferring? How many variants would the AI flag as high-risk that turn out to be functionally neutral? For regulators evaluating companion diagnostics and for companies deciding where to invest, those error rates will determine how much additional experimental confirmation is needed before committing resources.
Exact performance metrics on held-out clinical cohorts, including false-positive rates and sensitivity thresholds, have not been fully characterized beyond the benchmarks reported in the original Nature paper. Independent validation linking the newly prioritized non-coding UTR variants to patient outcomes outside the original COSMIC and ClinVar training sets also remains an open question.
One hypothesis worth watching: integrating AlphaGenome scores with prime-editing sensor data could reveal a subset of rare regulatory variants that currently appear neutral in mutation databases but actually drive resistance when cells are exposed to standard kinase inhibitors. If confirmed, that would mean existing databases are systematically missing a class of resistance mechanisms that only emerge under drug pressure, creating a blind spot with direct implications for how oncologists choose second-line therapies. Testing the idea will require prospective experiments that expose edited cell libraries to targeted drugs and then compare survivors’ variants against AI predictions.
Why drug pipelines built on AI-plus-editing convergence will move fastest
The most defensible reading of the current evidence is that AI models like AlphaGenome are already strong enough to narrow the search space for cancer drivers and to rank which mutations deserve immediate experimental follow-up. When combined with large-scale editing data, the technology compresses years of variant-by-variant functional work into months. That acceleration matters for drug pipelines, where identifying the right target early can shave years and hundreds of millions of dollars off development timelines.
But the field has not reached the point where AI predictions alone can replace functional assays or clinical trials. AlphaGenome is best understood as a triage tool: it tells researchers where to look first, not what they will find. The next critical milestones will be prospective studies that test whether AI-guided target selection actually produces better drugs, and direct benchmarks that pit model predictions against the growing library of editing-screen results variant by variant.
For pharmaceutical companies watching this space, the practical takeaway is straightforward. The combination of genome-scale AI prediction and high-throughput gene editing is creating, for the first time, a systematic and evidence-backed method for sorting cancer mutations into those that matter and those that do not. The companies that build pipelines around that convergence will have a measurable head start over those still working through variants one at a time.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
