Genetics

Evo 2 foundation DNA model boosts prediction of variant effects and genome design

Evo 2 is a large DNA foundation model trained on 9 trillion base pairs spanning all domains of life with single-nucleotide resolution. Without task-specific fine-tuning, it accurately predicts functional impacts of variants, including noncoding pathogenic mutations and clinically important BRCA1 variants. Model representations capture exon–intron boundaries, transcription factor binding sites, protein structural elements, and prophage regions, supporting mechanistic interpretability. Evo 2 can generate full mitochondrial, prokaryotic, and eukaryotic genomes and chromatin accessibility patterns with greater naturalness than prior methods. All model parameters, code, and the OpenGenome2 dataset are openly released, enabling translation into clinical variant interpretation workflows. 1

cf-EpiTracing profiles histone marks on plasma cfDNA to map disease tissues and responses

cf-EpiTracing measures histone modifications on cell-free DNA from as little as 50 μl plasma and integrates results with machine learning. In 2,417 profiles from 125 healthy individuals and 549 patients, it deconvolved cell type of origin across inflammatory bowel disease, colorectal cancer, coronary heart disease, and lymphoma. The platform identified primary diseased tissues, additional organ involvement, and stratified B-cell lymphoma subtypes with distinct genetic and epigenetic features. It detected early-stage diseases or lesions and traced transformation from follicular lymphoma to diffuse large B-cell lymphoma. Holistic epigenetic signatures predicted recurrence risk and therapy response, offering a noninvasive framework for diagnosis, subtyping, and prognostication. 2

gBRCA2 and RB1 hemizygosity channel breast cancer resistance toward RB1 loss under CDK4/6 inhibition

An integrated clinicogenomic study of more than 5,800 breast cancer patients showed germline pathogenic variants shape acquired resistance trajectories. gBRCA2-associated tumors were uniquely prone to acquired RB1 loss-of-function under standard frontline CDK4/6 inhibitor combinations, with poor clinical outcomes. This was driven by baseline RB1 hemizygosity lowering the barrier to biallelic inactivation and homologous recombination deficiency promoting RB1 loss under drug pressure. Preclinical gBRCA2 models showed near-uniform resistance to CDK4/6 inhibitors with consistent post-treatment Rb loss, while PARP inhibitors consistently outperformed CDK4/6 inhibitors. These data support prioritizing PARP inhibition in gBRCA2 carriers to intercept RB1-loss trajectories and illustrate a framework for forecasting resistance using germline and somatic features. 3

DNMT3A R882 is critical in preleukemia but largely dispensable in overt AML

Allele-specific CRISPR/Cas9 correction in human AML samples was used to dissect stage-specific functions of DNMT3A R882 mutations. DNMT3A R882 was required to maintain self-renewal and inflammatory programs in preleukemic cells but became largely dispensable once frank leukemia developed. Despite this, the mutation continued to influence leukemia stem cell frequency in established AML. These findings suggest the optimal therapeutic window for targeting DNMT3A R882–mutant clones is early in leukemogenesis or clonal hematopoiesis. 4