Researchers have developed DeepMet, a chemical language model designed to identify the "dark matter" of the metabolome by predicting the structures of previously unrecognized metabolites. By training on textual representations of known chemical structures, the model learns metabolic logic to generate plausible new molecules that existing databases often overlook. The study demonstrates that DeepMet can prioritize these hypothetical structures based on their generation frequency and successfully match them to mass spectrometry data from biological samples. This computational approach was validated by discovering dozens of new metabolites in mouse tissues and human biofluids through comparison with synthetic standards. Ultimately, the tool provides a systematic way to fill gaps in metabolic maps and improve the annotation of complex chemical datasets.

References:

• Qiang H, Wang F, Lu W, et al. Language model-guided anticipation and discovery of mammalian metabolites[J]. Nature, 2026: 1-10.

Scientists utilized deep learning and magnetic resonance imaging to analyze aortic valve function in nearly 60,000 participants from the UK Biobank. By measuring specific physical traits like peak velocity and aortic valve area, researchers identified 166 unique genetic loci linked to valve performance and the risk of aortic stenosis. The study discovered that many of the same genes influencing healthy valve function also drive the development of disease. Mendelian randomization further suggested that high levels of lipoproteins and blood pressure play a causal role in valve degradation. These findings highlight lipid-lowering strategies and hypertension management as critical pathways for preventing the narrowing of the aortic valve. Additionally, the development of a polygenic score allowed for the successful prediction of future heart complications in diverse external patient groups.

References:

• Kany S, Rämö J T, Hou C, et al. Multitrait analyses identify genetic variants associated with aortic valve function and aortic stenosis risk[J]. Nature Genetics, 2025: 1-10.

The paper describes the development and validation of popEVE, a deep generative model designed to interpret missense genetic variants on a proteome-wide scale. By integrating deep evolutionary data with human population variation, the model creates a calibrated scoring system that allows researchers to compare the severity of mutations across different proteins. This approach addresses a major limitation in clinical genomics where previous tools often overpredicted pathogenicity or failed to distinguish between mild and life-threatening conditions. In practical applications, popEVE identified 123 novel candidate genes for severe developmental disorders and successfully prioritized causal mutations using only a patient's exome. The researchers demonstrate that high-scoring variants frequently cluster at critical 3D interaction sites, such as ligand-binding pockets and protein interfaces. Ultimately, the framework improves diagnostic yields for rare diseases, particularly in cases where parental sequencing data is unavailable.

References:

• Orenbuch R, Shearer C A, Kollasch A W, et al. Proteome-wide model for human disease genetics[J]. Nature Genetics, 2025: 1-10.

Researchers utilized advanced single-cell and spatial multi-omics to map the cellular landscape of metabolic dysfunction-associated steatotic liver disease (MASLD) as it progresses toward more severe stages. A primary finding is the significant expansion of lipid-associated macrophages (LAMs), which are characterized by a specific phenotype driven by the MITF transcription factor. This regulatory pathway enhances fatty acid oxidation and mitochondrial function through the PGC1α–PPARγ axis, helping the liver manage lipid overload. Furthermore, the study suggests that these specialized macrophages exert a hepatoprotective effect by secreting growth factors that promote cell survival and limit tissue damage. By integrating transcriptomic and metabolic data, the authors identified potential therapeutic targets, such as the RSPO3–LGR6 signaling pair, to address inflammation and fibrosis. Overall, these findings provide a comprehensive spatiotemporal reference for understanding the immune and metabolic shifts that occur during chronic liver injury.

References:

• Li Z, Luo G, Gan C, et al. Spatially resolved multi-omics of human metabolic dysfunction-associated steatotic liver disease[J]. Nature Genetics, 2025: 1-14.

This research article examines how primary mismatch-repair-deficient high-grade gliomas are categorized into three distinct molecular subgroups based on specific secondary mutations. The study identifies that genomic instability mechanisms dictate whether a tumor develops through point mutations or copy number alterations, creating an inverse relationship between these two genetic events. Notably, the authors found that mutational signatures act as a semi-deterministic force, favoring the emergence of specific driver mutations in genes like TP53 and IDH1 while excluding others. These biological pathways significantly influence the tumor immune microenvironment, leading to varied patient outcomes. For instance, tumors with polymerase proofreading deficiency often show ultrahypermutation and respond well to immunotherapy, whereas IDH1-mutant varieties tend to be resistant. Ultimately, this framework provides a new model for risk stratification and the development of targeted therapeutic strategies for these aggressive brain cancers.

References:

• Fernandez N R, Chang Y, Nunes N M, et al. Patterns of hypermutation shape tumorigenesis and immunotherapy response in mismatch-repair-deficient glioma[J]. Nature genetics, 2025: 1-11.

The paper details the development of DNA-Diffusion, a generative artificial intelligence framework designed to engineer synthetic DNA regulatory elements. By utilizing a diffusion model trained on large-scale genomic data, researchers can now create compact sequences that dictate how genes are turned on or off in specific cell types. These lab-designed elements were shown to be functionally superior to natural sequences, offering higher activity and better precision across different tissues. The study validated these findings using large-scale screening and successfully edited a leukemia-protective gene within its natural environment. Ultimately, this technology provides a powerful new toolkit for precision medicine and the advancement of targeted gene therapies.

References:

• DaSilva L F, Senan S, Kribelbauer-Swietek J F, et al. Designing synthetic regulatory elements using the generative AI framework DNA-Diffusion[J]. Nature Genetics, 2025: 1-15.

This research explores how the inactivation of the p53 tumor-suppressor gene significantly increases the likelihood of breast cancer spreading to the brain. While mutations and chromosomal losses in p53 are common in many cancers, this study identifies a specific link to brain metastasis through the reprogramming of fatty acid metabolism. Scientists discovered that p53-deficient cells thrive in the brain by upregulating the SCD1 enzyme via the DEPDC1/SREBP1 pathway, allowing them to exploit lipids provided by astrocytes. Because these cancer cells become highly dependent on this metabolic shift, they are particularly vulnerable to FAS inhibitors, which successfully slowed tumor growth in experimental models. These findings suggest that screening for p53 alterations in primary tumors could better predict metastatic risk and help identify patients who might benefit from lipid-targeting therapies.

References:

• Laue K, Pozzi S, Zerbib J, et al. p53 inactivation drives breast cancer metastasis to the brain through SCD1 upregulation and increased fatty acid metabolism[J]. Nature Genetics, 2025: 1-16.

Spatial transcriptomics relies on sequencing and imaging to map cellular activities, but both methods face significant technical hurdles in accurately assigning RNA molecules to individual cells. The provided text highlights how cell segmentation errors and the limitations of two-dimensional mapping lead to "admixture," where transcripts from neighboring cells are incorrectly attributed to a specific profile. These errors create artifactual signals that can distort biological results, such as differential expression and ligand-receptor analyses. To address this, the researchers developed cellAdmix, a tool that utilizes computational models like non-negative matrix factorization and conditional random fields to identify and remove foreign transcripts. By filtering out these contaminating molecules, the method improves the clarity of gene set enrichment and more accurately reflects true biological roles, such as those of cancer-associated fibroblasts. The study demonstrates that such corrections are essential across various tissue types and technologies to ensure the integrity of spatial genomic

References:

• Mitchel, J., Gao, T., Petukhov, V. et al. Impact and correction of segmentation errors in spatial transcriptomics. Nat Genet (2026). doi.org

This research highlights how broad ethnic labels like "Hispanic" or "Latino" fail to account for the deep genetic diversity found across Mexico. By analyzing over 6,000 individuals in the Mexican Biobank, scientists discovered that the frequency of medically important gene variants—such as those affecting fentanyl and statin metabolism—varies significantly by region. These differences are driven not just by general European or African admixture, but specifically by distinct Indigenous ancestral groups like the Maya or Zapotec. The study introduces MexVar, an interactive platform that provides ancestry-specific data to help clinicians move toward more accurate precision medicine. Ultimately, the authors argue that localized genetic insights are essential for developing effective clinical guidelines across the Americas.

References:

• Barberena-Jonas C, Medina-Muñoz S G, Cedillo-Castelán V, et al. Clinical genetic variation across Hispanic populations in the Mexican Biobank[J]. Nature Medicine, 2026: 1-11.

The paper introduces MedHELM, a comprehensive framework designed to evaluate the performance of large language models across a broad spectrum of medical and operational tasks. Developed through collaboration with clinicians, this suite utilizes 37 diverse benchmarks and a hierarchical taxonomy to assess functions ranging from clinical decision support to administrative workflows. The research highlights that general-purpose AI rankings often fail to predict medical competence, as several top-tier models showed significant performance drops when faced with real-world healthcare data. To ensure scalable and nuanced assessment, the authors implemented an LLM-jury system that demonstrates strong agreement with human expert ratings while managing the high costs of professional clinical review. Ultimately, the study identifies DeepSeek R1 and o3-mini as current performance leaders, though models like Claude 3.5 Sonnet are noted for providing a superior balance between clinical accuracy and computational cost.

References:

• Bedi S, Cui H, Fuentes M, et al. Holistic evaluation of large language models for medical tasks with MedHELM[J]. Nature Medicine, 2026: 1-9.