Six proteomic aging clocks, applied to longitudinal serum proteomics from 42 participants with idiopathic pulmonary fibrosis in a phase 2a trial of rentosertib, an AI-designed TNIK inhibitor, all showed a statistically significant decrease in the pace of aging. Chronological-age-trained clocks tracked aging rate modulation more accurately than mortality-trained clocks, supporting multi-clock proteomic assessment as a trial endpoint for interventions targeting aging biology.
An LLM-based clinical assistant that orchestrates multiple tools, applied to sequential electronic health record data, can forecast maternal and infant conditions. The work suggests a window of opportunity to enhance risk-stratified care for mothers and infants, published in Nature Medicine.
A model-guided engineering strategy for sparse data, EvoMax, produced a compact eukaryotic Fanzor2 nuclease variant, FanzMAX v3-hLa, that reaches up to 97% editing efficiency at the best-performing endogenous locus and a mean of ~33% across 19 endogenous loci. It outperformed enNlovFz2 and enCnCas12f1 by more than 2.6-fold and edited hPCSK9 in humanized mice. EvoMax combines Gaussian process regression, protein language models and inverse folding with iterative experimental profiling.
The open-source platform AdaptiveFlow performs ultralarge virtual screening using Adaptive Target-Guided Virtual Screens, an 18-dimensional grid of molecular properties that focuses docking on promising regions of chemical space, plus an optional active learning layer. It supports over 1,500 docking protocols and runs on CPU and GPU architectures. Applied to FSP1 and PARP1, it identified nanomolar inhibitors confirmed by high-resolution co-crystal structures.
Gasdermin-D was engineered to carry viral protease-specific cleavage motifs, producing VID activators that trigger lytic cell death only in infected cells. Delivered as LNP-encapsulated mRNA in a hepatitis A virus model, they abolished viral replication and shedding in vivo and eased liver injury via a kill-and-alert mechanism, with activity also shown against Zika virus and SARS-CoV-2. Generative AI-designed cleavage motifs for the SARS-CoV-2 main protease improved antiviral potency.
A blinded, prospective benchmark (AIntibody) tested 511 AI-designed or predicted antibodies from 29 organizations across three tasks: affinity maturation, affinity ranking within HCDR3 clusters, and CDR design outside selection outputs. Several groups produced developable antibodies with affinities <100 pM, but successes did not transfer across tasks; affinity ranking was worse than random clone picking except for one model, and out-of-library design often failed to beat standard selections. The results show AI can optimize antibodies in defined, biologically grounded regimes while exposing gaps in affinity prediction, library-inspired design, and cross-task generalization.
A paper on the UniPert-G2CP framework, developed jointly by Tencent's AI for Life Sciences lab and Min Li's team at Central South University, has been accepted by Cell. It is the first AI virtual cell study from China published in the flagship journal Cell.
Why it matters: Its two-stage framework unifies genetic and chemical perturbations in a shared semantic space, offering a transferable path from molecular prediction to cell-level validation in AI drug discovery.
OptiPrime, a machine learning model grounded in prime editing mechanism, predicts PE efficiency and extends to PE3 and twinPE outcomes, outperforming prior models. It captures determinants of mammalian mismatch repair and nominates MMR-evasive silent edits that raise editing efficiency. In primary human and mouse cells and in a mouse model of KIF1A-associated neurological disorder, it guided efficient in vivo correction of a pathogenic brain mutation.
Deep learning tools achieved accurate de novo design of RNA pseudoknot secondary structures, with generative AI matching experienced human designers on most blind challenges in an Eterna competition involving 57 pseudoknots. Designs were evaluated by single nucleotide-resolution chemical mapping, compensatory mutagenesis, and cryo-electron microscopy, and AI-generated molecules with accurate secondary structures formed well-ordered 3D folds stabilized by noncanonical tertiary interactions not modeled during design. Success was guided by an RNet foundation model trained on prior chemical mapping data, indicating that difficult RNA design tasks may be tractable without first solving RNA 3D structure prediction.
A model-guided engineering pipeline (EvoMax) combining transfer-learning Gaussian process regression, ESM-2 and ESM-IF optimized Fanzor2 nucleases, compact eukaryotic genome editors, using sparse experimental data. Screening over 1,600 Fz2-like sequences yielded active orthologs including NaloFz2, and the resulting FanzMAX v3-hLa system reached up to 97% editing efficiency (mean ~33% across 19 loci) in bulk mammalian cells, roughly tenfold above baseline, with expanded TAM compatibility. In vivo editing of hPCSK9 in mice was achieved with single-AAV delivery.
Teams from Duke University and UC San Diego published the Raygun protein design framework in Nature, which can simultaneously perform amino acid substitutions, insertions and deletions on natural proteins, enabling miniaturization and sequence remodeling. In validation experiments, 6 of 8 miniaturized fluorescent proteins retained fluorescence, miniaturized TurboID retained ligase activity, and extended EGF showed higher EGFR binding affinity than the wild type.
Why it matters: The paper gives concrete data on 6 of 8 miniaturized fluorescent proteins retaining fluorescence and extended EGF exceeding wild-type affinity, letting readers judge the method's applicability to other proteins.
Stanford University published a study in Science using two genome language models, Evo 1 and Evo 2, to design viral genomes. Sixteen phages were synthesized in the lab that can infect, replicate in, and kill bacteria. The paper describes the two models simply as DNA versions of large language models.
PathPrism, an interpretable AI framework, encodes tissue architecture into pathologically informed spatial features to model prognosis, molecular alterations, and therapy response. Applied to 7,000 patients with colorectal cancer across 11 cohorts, it uncovered hundreds of biomarkers predictive of survival, MSI, BRAF, and TP53 mutations, and stratified chemotherapy benefit in stage II/III disease. It also uses large language models to generate hypotheses and includes VirtualWSI for semantic perturbation.
DeepSCan, a suite of deep learning models trained on ~45 datasets and translocation strength labels for ~310 cell surface display (CSD) elements, maps CSD sequence-function relationships and designs potent de novo modules for mRNA antigen display. Of 3,700 computationally designed CSDs, about 120 were experimentally validated, and 7 matched or exceeded the strongest natural CSDs; enhanced display worked across cell types and in a CAR-T cytotoxicity assay. This replaces lack of systematic CSD design rules with a model-guided pipeline.
A blinded, prospective competition benchmarked artificial-intelligence-generated and optimized antibody discovery against experimental measurements of affinity and developability. Participants submitted antibodies without knowledge of the eventual experimental readouts, and the benchmark anchored algorithm performance to those wet-lab results. The competition provides a template for evaluating in silico antibody discovery pipelines on experimentally grounded terms.
On August 13, 2026, a team led by Nicolas Thomä and Bruno Correia at EPFL (Swiss Federal Institute of Technology Lausanne) published a study titled "Proteome-wide identification of the druggable CRBN interactome" in Nature Biotechnology.
Why it matters: Reports AI mapping of the proteome-wide CRBN molecular glue binding landscape, shifting from degradation to a binding view for target discovery.
UniPert-G2CP is a two-stage deep learning framework that unifies multimodal molecular perturbagen representations to enable genetic-to-chemical perturbation phenotype transfer learning. Validated on large-scale genetic and chemical screening datasets, it supports simulation of multicellular and multidomain perturbation cause-effect spaces, and joint analysis of these spaces reveals cellular heterogeneity underlying drug responses. It offers a route toward unified modeling of genetic and chemical screens within the artificial intelligence virtual cell vision.
Science reports the first AI-designed bacteriophage virus, capable of surviving, infecting host cells, and combating antibiotic resistance. The original article notes that previous AI designs were mostly limited to single genes or proteins, while this work achieves genome-scale design, providing a blueprint for building diverse synthetic phages and practical biological systems.
Why it matters: Notes the key advance from single-gene design to whole-genome-scale phage design, useful for assessing the boundaries of AI synthetic biology.
Alphamab Oncology (Kangningjieri) has licensed ex-Greater China global rights to JSKN016, a first-in-class TROP2/HER3 bispecific antibody-drug conjugate, to AI drug discovery firm Pathos AI. Pathos AI will pay a $125 million upfront fee plus milestone payments, bringing the total deal value to $2.093 billion. JSKN016 has dosed its first patient in a Phase III trial in triple-negative breast cancer.
Yi Chengqi's group at Peking University and collaborators published a study in Nature describing ContactSeek, an AI framework based on AlphaFold3 contact probabilities that systematically identifies amino acid residues determining editing specificity in base editors. The work combined AlphaFold3-predicted contact probabilities with off-target editing data from high-throughput sequencing, then engineered Cas proteins and deaminases accordingly, markedly improving editing specificity and establishing a new paradigm for precision design of gene-editing tools.
Why it matters: The coupling of AlphaFold3 contact probabilities with high-throughput off-target data can be transferred to specificity engineering of other base editors or Cas proteins.
