Tongwu Zhang

Pathology foundation models trained from whole-slide images alone treat tissue morphology as an autonomous visual phenotype, leaving learned representations only weakly anchored to the molecular processes that generate tissue architecture. Here we present Fuji, a multimodal pathology foundation model that grounds morphology in transcriptomic and mutational supervision through a two-stage pretraining strategy: multi-teacher contrastive distillation that harmonizes formalin-fixed paraffin-embedded and fresh-frozen specimens, followed by multimodal multi-task learning that jointly optimizes masked image reconstruction, transcriptomic embedding reconstruction, and attention-aligned mutational signature inference. Trained on 60,546 whole-slide images paired with 45,675 bulk RNA-seq profiles and 22,683 whole-genome sequencing samples drawn from TCGA, GTEx, and CPTAC, Fuji yields a protocol-invariant, genome-informed embedding that we treat as a quantitative coordinate system for probing how genomic states project into tissue architecture. Genome-scale instability phenotypes, microsatellite instability, chromosomal instability, homologous recombination deficiency, and whole-genome duplication, are recoverable from histology with AUCs up to 0.98, consistently exceeding existing pathology foundation models. Extrachromosomal DNA amplification leaves a distinct morphological footprint that stratifies survival independently of tissue lineage. Etiologic mutational processes such as tobacco exposure and mismatch repair deficiency localize to discrete architectural neighborhoods, validated in two independent population-based cohorts, whereas clock-like signatures remain diffuse, delineating the scope and limits of genomic visibility from routine histology. Fuji further enables direct inference of tumor mutational burden, neoantigen load, immunotherapy response, and FDA-actionable driver mutations from hematoxylin and eosin slides. By coupling pretraining to molecular and mutational supervision, Fuji establishes an engineering basis for genome-informed inference from the most widely performed assay in oncology.

The role of extrachromosomal DNA (ecDNA) in lung cancer, particularly in subjects who never smoked (LCINS), remains unclear. Examination of more than 1200 whole-genome-sequenced lung cancers identified ecDNA in 18.9% of patients. Recurrent amplification of MDM2 and other oncogenes via ecDNA possibly drives a LCINS subset. Tumors harboring ecDNA showed worse overall survival than tumors harboring other focal amplifications. A strong association with whole-genome doubling suggests most ecDNA reflects genomic instability in treatment-naïve lung cancer.

Air pollution is a leading global environmental health risk, responsible for millions of premature deaths worldwide each year. Although traditionally associated with respiratory and cardiovascular diseases, accumulating epidemiological and mechanistic evidence now demonstrates that air pollution acts as a systemic toxicant, contributing to a wide spectrum of pathologies across nearly all major organ systems. These include pulmonary and extrapulmonary malignancies, neurodegenerative disorders, cardiovascular and renal diseases, metabolic dysfunction-associated fatty liver disease, dermatological conditions, and sensory organ impairment. In this comprehensive review, we synthesize large-scale population studies, experimental models, and molecular evidence to define air pollution as a multi-organ pathogenic exposure driven by convergent biological mechanisms. Disease development across organs is driven by shared mechanisms such as oxidative stress, inflammation, epigenetic changes, and extracellular vesicle signaling, while organ-specific exposure pathways determine tissue vulnerability. Importantly, we emphasize that disease risk is not dictated by exposure alone. Genetic susceptibility plays a decisive role in shaping individual responses to air pollution, with polygenic risk scores and functional variants in detoxification, antioxidant, immune, and host-defense pathways modifying exposure–response relationships. Evidence from human cohorts and genetically defined animal models underscores gene–environment interactions as key drivers of interindividual heterogeneity in pollution-related disease. By integrating mechanistic convergence with organ-specific exposure biology and genetic susceptibility, this review reframes air pollution as a systemic driver of chronic disease. We further identify critical knowledge gaps related to real-world multipollutant exposure, biological resilience, and precision prevention, providing a conceptual foundation for targeted interventions and risk stratification in pollution-exposed populations.

Approximately 10% of cutaneous malignant melanoma cases are familial. Variants in CDKN2A account for up to 40% of melanoma-prone families, with an additional ~10% explained by other genes. Many CDKN2A mutation-negative families show linkage to chromosome-band 9p21, which harbors CDKN2A, suggesting non-coding variants may contribute to familial risk. Here, whole-genome sequencing revealed a novel 100 kb deletion mapping to 9p21 in a gene-desert region, 205 kb from CDKN2A, cosegregating in a four-case family from Genoa, Italy. The deletion overlaps melanocyte enhancers that interact with the promoters of CDKN2A p16 and p14 transcripts and is predicted to reduce p16 expression. Using a nearby rare exonic variant in MTAP (rs755147810) on the deletion haplotype, we searched for deletion carriers in WES data from high-risk melanoma patients and controls and identified 22 cases and a single control carrying rs755147810 and the deletion. The association with melanoma in case-control analysis was highly significant (3,319 cases and 5,680 controls; P=1.27x10-6; OR=27.50). We observe loss-of-heterozygosity of the wild-type allele in a carrier’s tumor sample. The founder haplotype with the most recent common ancestor dates approximately 26 generations, broadly overlapping the period when Italy was struck by devastating outbreaks of plague that decimated the population creating a genetic bottleneck. Our results provide evidence of a high-penetrance intergenic variant conferring melanoma susceptibility, with potential for genetic screening of high-risk individuals.

Background: Transcriptome-wide association studies (TWAS) integrate gene expression and genome-wide association studies (GWAS) to identify disease susceptibility genes. Because gene expression varies substantially across cell types within tissues, cell type–specific prediction models may enhance the power of TWAS. Methods: We conducted cell type–specific TWAS leveraging single-cell RNA sequencing data from the OneK1K cohort (14 immune cell types, 1.27 million cells) and GWAS summary statistics for 7 cancers (>290 000 cases in total). To improve prediction accuracy, we developed a modeling framework that incorporates shared gene expression effects across cell types. Results: At a false discovery rate of 5%, we identified 106 (Bonferroni 5%: 13) previously unreported loci for breast cancer, 51 (4) loci for prostate cancer, 11 (4) loci for lung cancer, 39 (5) loci for melanoma, 9 (1) loci for ovarian cancer, and 2 (1) loci for diffuse large B-cell lymphoma, with most genes exhibiting cell type specificity. Gene set analyses confirmed joint associations of unreported genes with breast and prostate cancer risk in UK Biobank data. Additional lung tissue single-cell RNA sequencing data with 113 individuals validated 18 of 32 (56.3%) statistically significant genes for lung cancer. Across cancers, 139 statistically significant genes were shared by at least 2 cancer types and were primarily enriched in specific immune cell types. Conclusion: Cell type–specific TWAS improve the identification of novel cancer susceptibility loci and provide insights into the immune landscape of cancer etiology.

Uterine corpus endometrial carcinoma (UCEC) is a common gynecologic malignancy with rising mortality, yet its genome-wide mutational architecture remains incompletely understood. Here we analyze deep whole-genome sequencing data from 440 UCEC tumors from The Cancer Genome Atlas, integrated with transcriptomic and clinical data, to define subtype-specific mutational processes and genomic architectures. We uncover pronounced molecular subtype-specific differences in endogenous mutational mechanisms, retrotransposition activity, structural variation, and chromosomal instability. LINE-1 retrotransposition emerges as a key contributor to genome instability in UCEC, acting as a prominent source of structural variation in copy-number stable tumors and associating with chromothripsis and ecDNA-mediated oncogene amplification in copy-number high tumors. Mutational signature SBS28 contributes significantly to mutations in POLE -deficient tumors and is strongly correlated with SBS10a and SBS10b. Mismatch repair-deficient tumors exhibit a previously unrecognized doublet base substitution signature and a defining imbalance between ID2 and ID1 indel processes, reflecting pervasive DNA template-strand replication slippage and associated with increased tumor immunogenicity. Copy-number low tumors follow a distinct low-mutagenesis evolutionary trajectory characterized by reduced replication stress, low proliferative activity, genomic stability and enrichment of SBS18 associated with oxidative damage. Notably, we identify a consistent inverse association between body mass index and tumor mutational burden in UCEC, suggesting that host metabolic state may influence fundamental cellular processes governing mutation accumulation, promoting tumor development through non-mutagenic mechanisms rather than elevated genomic instability. Together, these findings establish a mutagenesis-centric framework for UCEC that links endogenous mutational mechanisms to tumor architecture and host metabolic context, uncovering previously unrecognized subtype-specific genomic features with important implications for refined risk stratification and therapeutic strategies.

Supplementary Figure S23 shows heatmap columns that define enriched pathways and rows define signature genes. Enrichment significance is indicated by color intensity. The annotation bar on the top denotes the gene databases used for the enrichment test, and the annotation bar on the right indicates the subtypes of signature genes. Signature genes not associated with any significantly enriched pathway are listed at the bottom of the heatmap.

Supplementary Figure S21 shows (A) Estimation of the minimum gene set per NS-LUAD expression subtype by cross validation in the Sherlock cohort, evaluating 1 to 30 genes per subtype. Box plots show the number of genes per NS-LUAD subtype (x-axis) versus cross-validation error rates (y-axis). (B) Distribution of the error rates for predicting NS-LUAD expression subtypes using the 60-gene signature by cross validation in the Sherlock cohort. Training and testing set error rates are indicated by colors, and median values are marked by dashed lines.