A large language model-based resource, MSK-Tox, captured the incidence, timing, and grade of anti-cancer therapy toxicity across more than 50,000 patients. Analysis of six representative adverse events showed toxicity risk varies by cancer type, regimen, and clinical context, is predictable from pretreatment clinical features, and has two germline modes: an organ-intrinsic mode, exemplified by a regulatory variant near FOXE1 linked to hypothyroidism, and an immune-mediated mode in which HLA-DRB1*15 drives adrenal insufficiency after immune checkpoint inhibition. The work supports pretreatment risk assessment for harm on the same principles used to predict benefit.
A fully on-premise clinical AI agent, combining local operational control with a multi-perspective reliability framework, achieved 90.04% accuracy on a seven-disease MIMIC-IV-derived benchmark and 83.8% on a four-disease task, approaching a cloud baseline. Behavioral consistency best discriminated correct diagnoses (AUC = 0.860; 0.875 under stress testing); at a 0.90 threshold, 49.4% of cases were retained at 98.9% accuracy. The framework lets institutions route a lower-risk subset to autonomy and defer the rest for review.
MoChiAgent, an LLM-based clinical assistant built on the MoChiFormer model, forecasts maternal and infant diseases from routine longitudinal EHR data, using 4,401,599 clinical visits for development and external validation in 263,452 maternal and 23,192 infant visits. It identified placental abruption and premature rupture of membranes at AUROCs of 0.89 and preterm labour at 0.91, and linked specific maternal clusters to neonatal jaundice (HR 2.81) and haematological diseases (HR 2.83) in infants. Pairing maternal gestational records with infant records improved prediction of infant conditions such as chromosomal abnormalities and respiratory disorders.
A study published in Nature Medicine built sex-specific biological aging clocks spanning multiple organs and omics layers, finding that female and male aging patterns can show disease-relevant differences at the organ level.
AdaptiveFlow, an open-source platform for ultralarge virtual screening, offers a ready-to-dock version of Enamine REAL Space comprising 69 billion compounds and integrates more than 1,500 docking protocols, scaling near-linearly on up to 5.6 million CPUs. Screening with an 18-dimensional property grid and optional active learning cut computational costs by orders of magnitude and yielded nanomolar inhibitors of FSP1 and PARP1, with co-crystal structures explaining FSP1 inhibition. It lowers the cost and scalability barriers that have limited billion-scale screening.
AI-designed protein assemblies combined with natural protein domains yielded more than 100 bottom-up RNA transfer vehicles with unique sizes and shapes, surpassing the RNA transfer efficiency of widely used delivery vehicles by several orders of magnitude. Tropism was programmed with computationally designed peptide binders, enabling delivery of therapeutically relevant cargo RNAs across cellular models. One vehicle's in vivo biodistribution was mapped in a mouse at near-single-cell resolution, confirmed as safe, and used for a gene-editing treatment strategy for Duchenne muscular dystrophy in patient-derived cells and a pig.
GPN-Star, a phylogeny-aware genomic language model that uses whole-genome alignments and species trees, outperforms previous methods across variant effect prediction tasks in coding and non-coding human genome regions. Trained on vertebrate, mammal and primate alignments, it improves prioritization of pathogenic and fine-mapped GWAS variants, yields strong complex trait heritability enrichments, and improves rare variant association testing power. Versions trained for five model organisms, from mouse to Arabidopsis, show the framework generalizes.
A multimodal explainable AI model predicted immunotherapy outcomes in non-small cell lung cancer better than established biomarkers, in a large international real-world study. The tool also improved physician decision-making. It offers an explainable alternative to current biomarkers for guiding immunotherapy choices.
DeepSomatic, a deep-learning method, detects somatic small nucleotide variations and insertions and deletions from both short-read and long-read data, with modes for whole-genome and whole-exome sequencing and support for tumor-normal, tumor-only and FFPE samples. To enable training and benchmarking, the authors generated the openly available CASTLE dataset of six matched tumor-normal cell line pairs sequenced with Illumina, PacBio HiFi and Oxford Nanopore Technologies. Across cell line and patient-derived samples and across sequencing technologies, DeepSomatic consistently outperforms existing callers.
GUTchetp predicts gut microbial biotransformation products using an ensemble of monolingual and multilingual neural machine translation models, reaching a Top-20 BLEU score of 66.59% and MaxFrag accuracy of 67.19% on a benchmark of six compound categories. A second component identifies associated microbial enzymes and species, tested on 95 metabolism events, and raises accuracy by 34.38 percentage points over existing tools via a re-ranking rule combining EC-classifier consistency with reaction similarity. Together the components capture 68.42% of known gut microbial biotransformation profiles, outperforming prior computational tools.
A machine-learning classifier detected early-stage ovarian cancer from untargeted metabolomic profiles with cross-validated sensitivity and specificity of 99.0% and 99.8% in serum (59 cancer, 32 controls) and 97.7% and 99.7% in plasma (63 cancer, 20 controls). Both cohorts shared 239 concordantly altered features across lipid, amino-acid, steroid, central-carbon, and redox metabolism, supporting a distributed metabolic 'Systemotype' detectable in blood.
In the I3LUNG study of 2,396 patients with non-small cell lung cancer, machine learning and deep learning models using clinical and blood data predicted immunotherapy outcomes with AUC up to 0.77 in the test set and surpassed PD-L1, ECOG performance status, NLR, LDH and LIPI scores. External validation gave AUC of 0.55 to 0.72. An explainable AI tool improved predictions by lung expert and nonexpert physicians, while multimodal fusion added no consistent benefit over clinical and blood data alone. Prospective validation in more than 2,000 patients is underway.
ELF integrates five pretrained pathology foundation models into unified slide-level representations, trained on 53,699 whole-slide images across 20 anatomical sites. It outperformed its constituent and slide-level foundation models on disease classification, biomarker detection, and anticancer and immunotherapy response prediction across multiple cancer types. Its architecture allows data-efficient downstream evaluation, including settings with limited data such as therapeutic response prediction.
CRISP, a vision-based pathology foundation model trained exclusively on frozen section slides, outperformed current foundation models and showed extensive validation, including a prospective cohort, for supporting treatment decisions across the surgical workflow. It brings intraoperative pathology decision support closer to clinical use.
A suite of large pre-trained language models extracts computable clinical data from unstructured electronic health records and integrates the findings into knowledge graphs, alongside structured EHR data, to build longitudinal patient journeys. The output is intended to help understand patient trajectories and treatment responses in the real world. The work was published in Nature Medicine on 10 September 2026.
A U.K. commission issued 44 recommendations for regulating artificial intelligence in medicine, aiming to balance AI's potential, such as reviewing millions of diabetes eye scans each year, with safety and equitable treatment. The proposals include staged authorization of new AI models and a system for providers to report how tools perform in clinics, including malfunctions, as models adapt over time.
ELF, a framework that ensembles five pathology foundation models into unified whole-slide image representations, outperforms any individual foundation model at disease classification, biomarker detection, and prediction of response to chemotherapy, targeted therapy, and immunotherapy. It was trained on nearly 54,000 slides spanning 20 anatomical sites.
An ensemble approach integrating five pretrained pathology foundation models into a single slide-level representation improves tumor subtyping, biomarker prediction, and therapy-response prediction. The gains come from exploiting complementary information across the five models, reported by Luo et al. in Cancer Cell. It points computational pathology toward ensemble and multimodal methods rather than single-model pipelines.
UniCure, a multi-modal foundation model, predicts drug-induced transcriptomic responses across diverse cellular contexts. It bridges cancer cell line data with patient-derived models to support personalized drug prioritization and patient stratification, and its predicted therapeutic rankings correlate strongly with clinical survival outcomes in real-world cohorts.
Rentosertib, an Insilico Medicine candidate whose target and molecular structure were derived with generative AI, has become the first AI-designed drug to complete clinical evaluation of anti-aging effects. In a 12-week trial, proteomics data from 42 patients with idiopathic pulmonary fibrosis were analyzed by six aging clocks developed independently by different teams, and all six indicated reduced biological age after rentosertib treatment.
