Cancer starts as a writing problem. Somewhere in the genome, letters get swapped, deleted, duplicated, amplified, rearranged, or otherwise treated like a shared Google Doc with no version control. Some of those typos matter. Many do not. The hard part is knowing which ones are steering the tumor and which are just weird little passengers enjoying the ride.
That is the challenge behind a new Cancer Discovery study by Kong and colleagues: how do you turn a tumor’s messy mutation profile into a useful treatment prediction? Their answer is a model called MutationProjector, which learns from genomic alterations across more than 30,000 tumors and connects mutation patterns to altered molecular pathways and treatment response (Kong et al., 2026).
In normal precision oncology, a clinician might look for famous genetic culprits: EGFR, BRAF, ALK, BRCA, MSI-high, high tumor mutational burden, and the usual suspicious characters. TP53 is usually standing nearby wearing sunglasses and pretending it was not involved. But many tumors do not have one clean, obvious target. They have a whole committee of genomic nonsense.
MutationProjector tries to read the committee minutes.
The Model Is Basically a Genomic Translator
A foundation model is an AI model trained broadly first, then adapted to specific tasks. In language, that means learning enough grammar and context to answer different questions. In cancer genomics, the idea is similar, except the “sentences” are mutation patterns and the “grammar” is biology.
MutationProjector was pretrained on tumor genotypes, then layered with molecular knowledge so it could project a tumor’s mutations into pathway-level biology. That matters because genes rarely act alone. A mutation in one gene can tug on a pathway, irritate a protein network, change gene expression volume knobs, or make the tumor microenvironment act like a sketchy neighborhood where immune cells keep losing their keys.
The authors tested whether the model could reconstruct held-out mutations, which is a sanity check for whether it learned real structure rather than memorizing a genomic phone book. Then they asked the more clinically juicy question: can it predict response or resistance to immunotherapy and chemotherapy across cancer cohorts?
It performed as well as or better than state-of-the-art approaches in the tested settings. That is the scientific version of “okay, now you have our attention.”
Biomarkers Are Great Until They Get Weird
Cancer biomarkers are often treated like light switches: mutation present, drug works; mutation absent, drug does not. Real tumors, because they apparently enjoy paperwork, behave more like old mixing boards with 70 sliders and one knob labeled “surprise.”
Take immunotherapy. Tumor mutational burden, or TMB, counts how many mutations a tumor carries. The basic logic is sensible: more mutations can mean more abnormal proteins, which can give immune cells more targets to notice. But TMB is not magic. Reviews in Nature Cancer and Nature Reviews Clinical Oncology have emphasized that TMB can help, but its usefulness varies by cancer type, assay, cutoff, tumor context, and immune biology (Anagnostou et al., 2022; Budczies et al., 2024).
That is why models that combine genomic patterns with pathway context are interesting. A 2022 Nature Communications study showed that network-based machine learning could improve immunotherapy response prediction by looking beyond single markers and toward pathway neighborhoods (Kim et al., 2022). MutationProjector pushes that same spirit into a broader genotype foundation model.
The model also surfaced less obvious signals. The paper reports KMT2D mutation as associated with immunotherapy sensitivity and the joint alteration of SMARCA4 and STK11 as associated with resistance. Translation: sometimes the genomic subplot you almost skipped turns out to be the villain’s calendar invite.
Why This Could Matter Outside the Spreadsheet
Sequencing is already common in cancer care, but interpretation remains the bottleneck. UC San Diego’s coverage of the study notes that only about 8% of cases are currently matched to an FDA-approved therapy based on genetics. That is not because sequencing is useless. It is because a tumor genome can contain a blizzard of alterations, and only a limited number have clear, approved action attached.
If MutationProjector and similar systems keep validating in larger, prospective, diverse patient groups, they could help doctors do several useful things: identify patients more likely to benefit from immunotherapy, flag likely resistance earlier, generate biomarker hypotheses for clinical trials, and explain tumor behavior through pathways rather than just gene names.
That last part matters. Clinicians do not need a black box whispering “trust me, bro” in binary. They need models that show their work, because cancer treatment decisions involve real toxicity, cost, timing, and anxiety. AI in oncology needs interpretability the way PCR needs primers.
A 2026 review in Nature Reviews Cancer makes the same broader point: machine learning may help extract more value from next-generation sequencing, but clinical adoption requires serious evaluation, responsible implementation, and workflow fit (Reardon et al., 2026).
The Sensible Hype Level: Cautious but Awake
This is not a “your oncologist has been replaced by a laptop” moment. Please do not let anyone put that on a conference tote bag.
It is more like this: tumor DNA contains more treatment clues than we currently know how to read. MutationProjector suggests that if we train models on enough tumor genomes and anchor them in molecular biology, we may get better at translating mutation chaos into therapeutic strategy.
Cancer is still a genome editing accident with a survival instinct. But studies like this make the instruction manual a little less illegible.
References
-
Kong J, Lee I, Boecher D, et al. A foundation model of cancer genotype enables precise predictions of therapeutic response. Cancer Discovery. 2026. https://doi.org/10.1158/2159-8290.CD-25-1735
-
Anagnostou V, Bardelli A, Chan TA, Turajlic S. The status of tumor mutational burden and immunotherapy. Nature Cancer. 2022;3:652-656. https://doi.org/10.1038/s43018-022-00382-1
-
Budczies J, Kazdal D, Menzel M, et al. Tumour mutational burden: clinical utility, challenges and emerging improvements. Nature Reviews Clinical Oncology. 2024;21:725-742. https://doi.org/10.1038/s41571-024-00932-9
-
Kim E, Kim JY, Smith MA, Haura EB, Anderson ARA. Network-based machine learning approach to predict immunotherapy response in cancer patients. Nature Communications. 2022;13:3703. https://doi.org/10.1038/s41467-022-31535-6
-
Reardon B, Culhane AC, Van Allen EM. Convergence of machine learning and genomics for precision oncology. Nature Reviews Cancer. 2026;26:217-229. https://doi.org/10.1038/s41568-025-00897-6
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.