SHERLOCK is an interpretable deep generative framework that represents genetic, pharmacological, and experimental perturbations as structured interventions on a latent baseline cellular state within a structural causal model. It learns correlated and sparse perturbation representations, organizes perturbations by shared transcriptional responses, and supports counterfactual estimation of downstream transcriptional effects, condition-dependent response analysis, and compositional modeling of combinatorial perturbations. Across genome-scale CRISPR, chemical, and spatial perturbation datasets, it recovers perturbation relationships concordant with known pathways and pharmacological properties and predicts held-out perturbation combinations.
Two robustness measures, the classification stability score and prediction specificity score, quantify the reliability of cell-type labels from label-transfer tools in scRNA-seq data. Applied to human fetal kidney and four iPSC-derived kidney organoid protocols with a mouse fetal kidney reference, scores were high for most fetal kidney cell types except immature epithelial structures, high for organoid cap mesenchyme, and variable for epithelial lineages depending on protocol. A cell-removal simulation confirmed the measures flag unreliable classifications.
In-cell cryo-electron tomography combined with proteomics, integrative structural modeling, and bioinformatics uncovered an uncharacterized protein complex in Mycoplasma pneumoniae. The complex links Sec-based protein translocation with extracellular folding and supports both co-translational and post-translational pathways, functionally annotating a previously unknown structure in a minimal bacterium.
The material available contains no experimental results, methods description, or findings from the study itself, so the work cannot be summarized beyond its stated topic.
A software framework called the Model Hardware Standard (MHS), developed by Anthropic and the Howard Hughes Medical Institute's Janelia Research Campus, connects laboratory instruments from different vendors into a single system that AI agents can control to run experiments. A team testing it at Carnegie Mellon University set up a functioning experiment in hours, a process that can easily take months. Unlike the SiLA consortium's existing standard, MHS links instruments directly to AI agents.
unZipro is a geometric deep learning framework that adapts general protein design models to specific protein families using meta-learning to navigate family-specific fitness landscapes. It enabled rapid in silico protein evolution across gene editing, luciferase reporting, transcriptional regulation, and antiviral resistance with minimal experimental screening. The work simplifies protein evolution by cutting the screening burden required to tailor designs to a chosen family.
An expanded DNA methylation-based sarcoma classifier (v13.1) was trained on 4377 methylation profiles covering 116 tumor sub-classes, 4 control groups and 93 distinct methylation classes, and validated in 1547 mesenchymal tumors from four independent cohorts. Five-fold nested cross-validation yielded 98% class-level accuracy with a Brier score of 0.017. Predictions were obtained in 73% of validation cases, 91% matched the histopathology diagnosis, and 6% of inconclusive or ambiguous cases gained a definitive molecular diagnosis or reclassification.
The AlphaFold Protein Structure Database has added more than 8,000 virus protein dimers from 23 virus families that infect humans, including monkeypox, measles and hepatitis B viruses. Researchers analysed sequences of 41,774 proteins from around 2,800 viruses and used AlphaFold2 to predict roughly 40,746 homodimers and nearly 1.7 million heterodimers, of which only 2,749 homodimers and 5,279 heterodimers met accuracy standards for inclusion. The update, launched on 24 September with a pandemic preparedness portal, fills a viral blind spot in a resource with more than three million users.
Esophageal cancer detection through noncontrast computed tomography is enhanced by a new artificial intelligence tool called Esophageal AI-Guided malignant Lesion Evaluation. In a large-scale study, the tool achieved high sensitivity and specificity across diverse settings.
A Nature Biotechnology commentary argues that transcript annotations should not be split into 'known' versus 'novel' categories. It proposes instead a multidimensional framework for interpreting transcript diversity that weighs technical credibility, contextual prevalence, reproducibility, molecular consequence and phenotypic relevance. The piece was published online on 22 September 2026.
A commentary in Gut Microbes highlights the MIND framework (Microbial Interaction and Niche Determination), introduced by Moyne et al. in Cell. MIND uses translational efficiency to map resource competition and niche partitioning among microbes, moving beyond the correlative limits of metagenomics. It provides a mechanistic blueprint for causal, rational engineering of microbial communities.
CROWN is a universal visual foundation model for computational cytopathology, benchmarked on more than 200 real-world cytology tasks across cohorts, including lymph node metastasis and cervical screening datasets. It gives cytopathology a single pretrained model that can be evaluated across many clinical cytology settings rather than task-specific systems.
LLM judges match clinician ratings for safety screening and factual verification of patient communication about trial informed consent (Spearman ρ > 0.80), but agree less on communication quality (ρ < 0.60). Using a seven-criterion rubric validated on ICON-Bench, 19 LLM judges were benchmarked; safety-specialized guard models underperformed general-purpose models, and patient advocates rated communication quality lower than both clinicians and LLM judges. The results support LLM judges for scalable oversight, with recalibration needed to patient-centered standards.
TnpB-family proteins have been engineered into minimal de novo RNA-guided transposition systems. Nuclease-deficient TnpB-family proteins mediate three-component DNA integration with high efficiency, target-site immunity, orientation control, and programmable insertion into the bacterial chromosome. Mutations at the TniQ-TnsC interface in wild-type V-K CAST also attenuate off-site transposition.
A transplantation platform generates xenocortical mice by genetically depleting glutamatergic neurons from mouse neocortex and hippocampus, then neonatally engrafting the cortical cavity with human stem-cell-derived cortical organoids. The organoids occupy most of the cortical volume, yield diverse human cortical cell types including layer 5 extratelencephalic projection neurons, and integrate with the mouse nervous system, showing organized activity resembling developing circuits. Host mice retain broadly preserved locomotion with selective differences in limb coordination, and the platform yields behavioural readouts in an injury model of developing human cortical cells.
Why it matters: It provides circuit- and behaviour-level readouts from integrated human cortical neurons in living mice, enabling disease modelling and therapy testing.
The Pasca lab at Stanford reports in Nature that human brain organoids were implanted into mice that had been genetically engineered and had large brain regions removed, producing neural chimeras that are half-human by volume. In 3-month-old chimeric mice, the researchers detected human pyramidal projection neurons extending into the spinal cord, as well as specialized large neurons involved in social cognition that occur only in large mammals such as elephants, whales, and primates. The model offers a new route to studying hard-to-obtain human brain tissue, but it also raises increasingly complex ethical concerns.
Fluorescent protein ticker tapes (FPTT) record multiple transcriptional histories in single living cells by combining multispectral fluorescent proteins with self-assembling protein fibers. FPTT logged dose-dependent, reversible endogenous cFos transcription in hippocampal neurons at 3-hour resolution over 8 days, and variants were engineered for NF-κB, STAT3, mTOR, NFAT and cAMP signaling, enabling simultaneous analysis of four pathways during T cell activation. The tool revealed NFAT/STAT3 cross-talk and cell cycle-dependent mTOR oscillations.
A nanopore sensor built by modifying a bacterial porin identifies amino acids, nucleoside monophosphates, saccharides and peptides simultaneously, with 98.7% classification accuracy when coupled with machine learning. The platform also enables compositional analysis of native glycopeptides, extending nanopore detection beyond a single analyte class.
Gen-COMPAS, a generative committor-guided path-sampling framework, reconstructs biomolecular transition pathways without predefined collective variables and at acceptable computational cost. It combines a denoising diffusion probabilistic model producing structurally plausible intermediate targets with committor-based filtering, yielding transition-region ensembles at nanosecond-to-submicrosecond scales where conventional approaches need orders of magnitude more sampling. Applied to systems from a miniprotein to a pentameric ligand-gated ion channel, it recovers committors, transition states and free-energy landscapes from end-point structures alone.
Wutou decoction (WTD) treats rheumatoid arthritis (RA) through multi-component, multi-target action; its active ingredients benzoic acid, quercetin, apigenin and decanoic acid reduce macrophage inflammatory responses by inhibiting the PI3K/AKT and MAPK pathways. The study built a component-target-pathology network covering 440 active ingredients and 2638 targets, and identified 114 core functional component groups. In vivo experiments confirmed that WTD alleviated RA-induced arthritis and bone destruction.
