Search

Humanity’s Last Exam, a multi-modal benchmark at the frontier of human knowledge, is designed to be an expert-level closed-ended academic benchmark with broad subject coverage.

Suzgun et al. find that current large language models cannot reliably distinguish between belief, knowledge and fact, raising concerns for their use in healthcare, law and journalism, where such distinctions are critical.

A deep learning-based model, developed using the rich, multimodal data available from polysomnography-derived sleep recordings, performs well on common sleep analysis tasks and predicts future disease risk across a range of diseases.

Liang et al. estimate the prevalence of text modified by large language models in recent scientific papers and preprints, finding widespread use (up to 17.5% of papers in computer science).

Analyses of 2,658 whole genomes across 38 types of cancer identify the contribution of non-coding point mutations and structural variants to driving cancer.

Accurate estimations of the frequency distribution of rare variants are needed to quantify the discovery power and guide large-scale human sequencing projects. This study describes an algorithm called UnseenEst to estimate the distribution of genetic variations using tens of thousands of exomes.

The CMS Collaboration reports the measurement of the spin, parity, and charge conjugation properties of all-charm tetraquarks, exotic fleeting particles formed in proton–proton collisions at the Large Hadron Collider.

Understanding deregulation of biological pathways in cancer can provide insight into disease etiology and potential therapies. Here, as part of the PanCancer Analysis of Whole Genomes (PCAWG) consortium, the authors present pathway and network analysis of 2583 whole cancer genomes from 27 tumour types.

Analysis of cancer genome sequencing data has enabled the discovery of driver mutations. Here, as part of the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium the authors present DriverPower, a software package that identifies coding and non-coding driver mutations within cancer whole genomes via consideration of mutational burden and functional impact evidence.

Multi-omics datasets pose major challenges to data interpretation and hypothesis generation owing to their high-dimensional molecular profiles. Here, the authors develop ActivePathways method, which uses data fusion techniques for integrative pathway analysis of multi-omics data and candidate gene discovery.

InterPLM is a computational framework to extract and analyze interpretable features from protein language models using sparse autoencoders. By training sparse autoencoders on ESM-2 embeddings, this study identifies thousands of interpretable biological features learned by the different layers of the ESM-2 model.

Squidiff is a diffusion-based model to predict transcriptomic changes in response to perturbations.

Generative machine learning models are used in synthetic biology to find new structures such as DNA sequences, proteins and other macromolecules with applications in drug discovery, environmental treatment and manufacturing. Gupta and Zou propose and demonstrate in silico a feedback-loop architecture to optimize the output of a generative adversarial network that generates synthetic genes to produce ones specifically coding for antimicrobial peptides.

Artificial intelligence agents are autonomous systems that use large language models to reason and as such can perform complex, multistep tasks with minimal human oversight. This Review by Truhn et al. discusses how these agents — which have already been implemented in several industries — could transform cancer research and oncology, and looks at the challenges that need to be addressed before they can be efficiently and safely used.

In somatic cells the mechanisms maintaining the chromosome ends are normally inactivated; however, cancer cells can re-activate these pathways to support continuous growth. Here, the authors characterize the telomeric landscapes across tumour types and identify genomic alterations associated with different telomere maintenance mechanisms.

Whole-genome sequencing data from more than 2,500 cancers of 38 tumour types reveal 16 signatures that can be used to classify somatic structural variants, highlighting the diversity of genomic rearrangements in cancer.

A pan-cancer genomic analysis reports the effects of structural variations on chromatin domains (TADs). Most TAD disruptions do not result in appreciable changes in expression of nearby genes.

An analysis of 2,954 genomes from 38 cancer subtypes identified 19,166 retrotransposition events in 35% of samples. Aberrant LINE-1 retrotranspositions can lead to the deletion of tumor-suppressor genes as well as the amplification of oncogenes.

In this study the authors consider the structural variants (SVs) present within cancer cases of the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium. They report hundreds of genes, including known cancer-associated genes for which the nearby presence of a SV breakpoint is associated with altered expression.

Guan et al. report a simple yet multifunctional molecule which enables direct, photoresist-free patterning of quantum dots under ambient conditions. This advance leads to over 20% efficiency for patterned QLEDs and full-color active-matrix displays, offering a practical route to next-generation display manufacturing.

Fernando Pardo-Manuel de Villena and colleagues generate a 3 × 3 diallel cross of three inbred mouse lines and examine gene expression in multiple tissues. They identify allelic imbalance favoring the expression of the paternal allele across the genome.

There’s an emerging body of evidence to show how biological sex impacts cancer incidence, treatment and underlying biology. Here, using a large pan-cancer dataset, the authors further highlight how sex differences shape the cancer genome.

With the generation of large pan-cancer whole-exome and whole-genome sequencing projects, a question remains about how comparable these datasets are. Here, using The Cancer Genome Atlas samples analysed as part of the Pan-Cancer Analysis of Whole Genomes project, the authors explore the concordance of mutations called by whole exome sequencing and whole genome sequencing techniques.

The flagship paper of the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium describes the generation of the integrative analyses of 2,658 cancer whole genomes and their matching normal tissues across 38 tumour types, the structures for international data sharing and standardized analyses, and the main scientific findings from across the consortium studies.

Integrative analyses of transcriptome and whole-genome sequencing data for 1,188 tumours across 27 types of cancer are used to provide a comprehensive catalogue of RNA-level alterations in cancer.

Viral pathogen load in cancer genomes is estimated through analysis of sequencing data from 2,656 tumors across 35 cancer types using multiple pathogen-detection pipelines, identifying viruses in 382 genomic and 68 transcriptome datasets.

Whole-genome sequencing data for 2,778 cancer samples from 2,658 unique donors across 38 cancer types is used to reconstruct the evolutionary history of cancer, revealing that driver mutations can precede diagnosis by several years to decades.

Some cancer patients first present with metastases where the location of the primary is unidentified; these are difficult to treat. In this study, using machine learning, the authors develop a method to determine the tissue of origin of a cancer based on whole sequencing data.

The authors present SVclone, a computational method for inferring the cancer cell fraction of structural variants from whole-genome sequencing data.

Many tumours exhibit hypoxia (low oxygen) and hypoxic tumours often respond poorly to therapy. Here, the authors quantify hypoxia in 1188 tumours from 27 cancer types, showing elevated hypoxia links to increased mutational load, directing evolutionary trajectories.

The characterization of 4,645 whole-genome and 19,184 exome sequences, covering most types of cancer, identifies 81 single-base substitution, doublet-base substitution and small-insertion-and-deletion mutational signatures, providing a systematic overview of the mutational processes that contribute to cancer development.

Cancers evolve as they progress under differing selective pressures. Here, as part of the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium, the authors present the method TrackSig the estimates evolutionary trajectories of somatic mutational processes from single bulk tumour data.

Measuring sub-surface thermal conditions during 3D printing is crucial for microstructure evolution understanding and control. Authors use embedded fiber optic sensors to measure sub-surface temperatures and use machine learning to improve sensor resolution to 30 µm, providing detailed data for thermal modeling and prediction.

For a third year in a row, we followed up with authors of several recent Comments and Perspectives in Nature Machine Intelligence about what happened after their article was published: how did the topic they wrote about develop, did they gain new insights, and what are their hopes and expectations for AI in 2022?

Human collaboration with a team of artificial intelligence (AI) agents powered by large language models was used to efficiently design a complex interdisciplinary research project leading to the design of novel nanobodies against SARS-CoV-2 spike protein.

Here, the authors present an expanded version of the Cultivated Genome Reference (CGR), termed CGR2, a catalog that includes 3324 high-quality draft genomes based on gut bacterial isolates from Chinese individuals, and classifies 527 species from 8 phyla, including 179 previously unidentified species, and provides information of secondary metabolite biosynthetic gene clusters and gut phage-bacteria interactions.

Cell type labelling in single-cell datasets remains a major bottleneck. Here, the authors present AnnDictionary, an open-source toolkit that enables atlas-scale analysis and provides the first benchmark of LLMs for de novo cell type annotation from marker genes, showing high accuracy at low cost.

Generative artificial intelligence (AI) systems can be optimized using TextGrad, a framework that performs optimization by backpropagating large-language-model-generated feedback; TextGrad enables optimization across diverse tasks, including radiotherapy treatment plans and molecule generation.

Hematoxylin and eosin (H&E) staining is a widely used method in histopathology, but it cannot directly inform about specific molecular markers. Here, the authors present ROSIE, a deep-learning framework that computationally imputes the expression and localisation of dozens of proteins from H&E images.

Modulating mitochondrial NAD⁺ levels by changing the expression of the mitochondrial NAD⁺ transporter, SLC25A51, Mukherjee et al. demonstrate that mitochondrial, rather than cytosolic or nuclear, NAD⁺ levels are a key determinant of the rate of liver regeneration.

The antiviral dsRNA sensor PKR is regulated by PACT. This paper shows how PACT prevents aberrant PKR activation by endogenous dsRNAs like Alu. PACT disrupts PKR’s dsRNA scanning without blocking its binding, resetting its activation threshold to tolerate cellular dsRNA and preserve homeostasis.

The quark structure of the f₀(980) hadron is still unknown after 50 years of its discovery. Here, the CMS Collaboration reports a measurement of the elliptic flow of the f₀(980) state in proton-lead collisions at a nucleon-nucleon centre-of-mass energy of 8.16 TeV, providing strong evidence that the state is an ordinary meson.

A spatially resolved single-cell transcriptomics map of the mouse brain at different ages reveals signatures of ageing, rejuvenation and disease, including ageing effects associated with T cells and rejuvenation associated with neural stem cells.