Editors’ Selections from Relevant Scientific Publications

Plasma protein signature predicting lung cancer risk (by Pandya et al via Cell)A 14-protein plasma signature may predict lung cancer risk more than 5 years before diagnosis and select individuals for pharmacological cancer interception. Using machine learning to analyze data from the UK Biobank Pharma Proteomics Project, Pandya and colleagues identified 14 proteins associated with future lung cancer incidence and validated the signature across eight independent cohorts. The signature was elevated in individuals who smoke or those exposed to particulate matter (PM), and was linked to lung myeloid and alveolar cells. Mechanistic studies showed that the signature proteins were primarily produced by lung epithelial cells (AT2, club, and basal cells) in response to PM exposure, oncogenic EGFR signaling, and IL-1β activation, converging on a keratin8+/claudin4+ alveolar transitional cell (KAC) state required for malignant progression. Importantly, IL-1β inhibition suppressed PM-induced malignant potential of KACs in vitro and in vivo, and individuals with high levels of the signature benefited most from anti–IL-1β therapy in the CANTOS trial. By integrating machine learning-based analyses of population-scale proteomic data with evidence from preclinical and clinical studies, this work advances the development of precision cancer prevention through early risk stratification and targeted preventive interventions.Pandya T, … Swanton C. Cell. 2026 Jun 25;189(13):3903-3921.e26.Mutation frequencies in normal tissue and cancer genomes (by Cheek et al. via Nat Genet.)Patient age may help distinguish mutations that drive carcinogenesis from those that are merely positively selected during normal tissue aging. Cheek and colleagues developed a statistical framework to compare cancer and normal tissue genomes and estimate the effects of mutations on cancer risks in blood, esophageal, and colorectal tissues. They found that mutations with stronger carcinogenic effects tended to occur in cancers diagnosed at younger ages, whereas positively selected but non-carcinogenic mutations were enriched in older patients. Patient age distributions could be used to identify carcinogenic mutations even in the absence of normal tissue sequencing data. In acute myeloid leukemia, age-dependent patterns of purported driver mutations were largely attributable to normal blood evolution, challenging the long-standing assumption that childhood cancers require distinct mutations. These findings provide a framework for distinguishing cancer drivers from positively selected bystander mutations, improving cancer risk assessment and informing precision cancer prevention.Cheek D, … Naxerova K. Nat Genet. 2026 Jun;58(6):1320–1330.Distribution of clinical symptoms among pediatric patients referred for genetic testing (by Wang et al. via Nat Med.)Pathogenic germline variants contribute substantially to pediatric cancer risk and may inform cancer surveillance and early detection. Wang and colleagues analyzed exome sequencing data from 75,602 children younger than 3 years who were referred for genetic testing, primarily because of neurologic symptoms, metabolic disorders, or immune dysfunction. Focusing on 139 pediatric cancer-predisposition genes, the researchers identified 501 pathogenic or likely pathogenic (P/LP) variants, including 456 single-nucleotide variants and 45 copy number variants. Among 411 patients with tumors (203 diagnosed before genetic testing and 208 diagnosed during follow-up), 134 (32.6%) carried causative germline P/LP variants. The most frequently affected genes were NF1 (13.1%), TSC2 (5.8%), RB1 (4.6%) and WT1 (3.2%). Prospective follow-up for 64,187 patients without tumors at the time of genetic testing showed that carriers of P/LP variants had a significantly higher incidence of malignant tumors than noncarriers (3.23 vs. 0.272 per 1,000 person-years). These results underscore the clinical value of comprehensive genetic testing and support proactive genetic counseling and risk-adapted surveillance for children with inherited cancer predisposition.Wang H, … International Consortium in Digital Twins in Healthcare and Medicine. Nat Med. 2026 Jun;32(6):2238–2244.Image-derived AI-based model for breast cancer risk prediction.An image-derived artificial intelligence (AI) model may improve long-term breast cancer risk prediction. Using digital mammograms and an ensemble machine learning approach, Eriksson and colleagues developed a model to predict 10-year breast cancer risk. The model was trained in a case cohort sampled from the Swedish KARMA screening study and independently validated in two population-based case cohorts (KARMA and the US Olmsted cohort) comprising 8,696 individuals and 1,633 incident breast cancer cases, with additional validation in the US hospital-based EMBED nested case-control study. The model achieved a 10-year time-dependent area under the curve [AUC(t)] of 0.72 for invasive breast cancer with an expected-to-observed event ratio of 0.99 in both the KARMA and Olmsted cohorts, demonstrating excellent discriminatory performance and calibration. Similar discriminatory results were observed in EMBED. This risk model significantly outperformed other established ones, predicting 33% of future breast cancers among the top 10% of high-risk individuals in the KARMA cohort, compared with 23%, 20%, and 24% predicted by the Tyrer-Cuzick v8, BCSC v3, and Mirai models, respectively. The model could identify up to 40% of future breast cancers using current US clinical guidelines for high risk, supporting the use of image-derived AI tools to improve risk stratification.Eriksson M, … Vachon C. Sci Transl Med. 2026 May 20;18(850):eady7414.