Explainability has increasingly become a core requirement for intelligent medical devices. Current medical artificial intelligence (AI) technologies suffer from the 'interpretability gap' despite tremendous efforts for enhancing explainability. Here we propose class-association manifold learning, a generative approach that enhances explainability of medical AI models. Our method efficiently decouples common decision-related patterns from individual backgrounds, enabling us to represent global class-associated knowledge in a low-dimensional mapping while preserving near-perfect diagnostic accuracy. The extracted knowledge is further used to enable AI-generated modifications on arbitrary samples and visualize differential diagnosis rules. Moreover, we develop a topology map to model the entire decision rule set, so that the logic underlying black-box models can be intuitively explicated by traversing the map and generating virtual contrastive examples. Extensive experiments show that our method not only achieves higher accuracy in explaining the behaviour of medical AI models but also helps with extracting medical-compliant knowledge that are unknown during model training, thus providing a potential means of assisting clinical rule and medical knowledge discovery with AI techniques.

Dysfunctional autophagy, a key cellular cleaning process, is a key driver of brain ageing and neurodegenerative diseases such as Alzheimer's disease (AD). However, developing effective treatments by enhancing autophagy has been challenging, as most known compounds act through the broad mTOR pathway, risking side effects, and few can effectively penetrate the brain. To address this, we developed DeepDrugDiscovery-a mechanism-aware, AI-powered screening platform incorporating ADMET and blood-brain barrier penetrability predictions. Here we show that this platform successfully identified novel, mTOR-independent autophagy enhancers, with two lead compounds demonstrating an ability to cross the blood-brain barrier, clear AD-related protein aggregates and restore memory function in worm and mouse AD models. By releasing DeepDrugDiscovery as an open-source, modular tool, we offer a user-friendly AI platform that enables customized therapeutic screening. Our work establishes a scalable, AI-driven pipeline that integrates cross-species validation to rapidly discover mechanism-based therapeutics for diseases with high unmet medical need.

Despite their notable abundance in circulation and biological significance in the tumour microenvironment, direct genetic programming of neutrophils has been challenging. Here we develop a neutrophil-specific modified RNA translation platform, termed NeuSMRT, that enables the expression of chimeric antigen receptors (CARs) in primary neutrophils. NeuSMRT combines engineered extracellular vesicles or lipid nanoparticles for modified RNA delivery with a microRNA-responsive L7Ae:k-turn switch to restrict protein translation to neutrophils. In a syngeneic glioma model, CAR-neutrophils produced in vivo significantly inhibit tumour growth and prolong survival, accompanied by enhanced T cell recruitment and activation, and reduced immunosuppression of myeloid cells in the tumour microenvironment. CAR-neutrophils further enhance the efficacy of chemotherapy and CAR-T therapy. Furthermore, CAR-neutrophils display antitumour activities in a humanized glioblastoma mouse model. The feasibility and safety of NeuSMRT are also demonstrated in experimental dogs. These findings establish a programmable neutrophil platform for cancer immunotherapy.

Frameshift mutations, responsible for >20% of Mendelian inherited diseases, pose substantial therapeutic challenges. Here we developed Template-Independent Genome Editing for Restoration (TIGER), a platform for the efficient and precise correction of frameshift mutations across various models. By identifying reproducible nucleotide-level factors that influence therapeutic efficacy across cells and tissues, we developed a scoring system for guide RNA (gRNA)-Cas9 outcomes. Approximately 75% of deletion and 50% of insertion mutations produced ≥30% in-frame products, sufficient for phenotypic restoration, with 38% and 65% achieving wild-type correction, respectively. To expand the applicability of TIGER across species and genome wide, we retrained the inDelphi algorithm to predict therapeutic gRNAs for single-nucleotide frameshifts. In a mouse model of deafness, delivery of SpCas9 and optimal gRNA via dual adeno-associated virus restored hearing thresholds to wild-type levels, with ~90% of in-frame edits being wild type. TIGER provides a robust and broadly applicable strategy for in vivo correction of inherited frameshift diseases.

Current SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) vaccines have shown robust induction of neutralizing antibodies and CD4+ T cell activation; however, CD8+ responses are variable, and the duration of immunity and protection against variants are limited. Here we repurpose our DNA origami vaccine nanotechnology DoriVac to target infectious viruses, namely, SARS-CoV-2, HIV and Ebola. The DNA origami nanoparticle, conjugated with infectious-disease-specific heptad repeat 2 peptides, which act as highly conserved antigens, and CpG adjuvant at precise nanoscale spacing, induces neutralizing antibodies, Th1 CD4+ T cells and CD8+ T cells in naive mice, with significant improvement over a bolus control. Pre-clinical studies using lymph-node-on-a-chip systems validate that DoriVac, when conjugated with antigenic peptides or proteins, induces promising cellular and humoral immune responses in human cells. Moreover, DoriVac bearing full-length SARS-CoV-2 spike protein achieves immune responses comparable to current mRNA vaccine platforms while potentially reducing storage constraints. These results suggest that DoriVac holds potential as a versatile, modular vaccine platform, capable of inducing both humoral and cellular immunities, underscoring its potential future use.

Detecting antigen-specific lymphocytes is crucial for immune monitoring in vaccination, infection, cancer and autoimmunity. However, their low frequency and dispersed distribution make reliable detection challenging. We developed a strategy exploiting the functions of tissue-resident memory T cells (TRMs) to locally concentrate target circulating immune cells in the skin, enabling their non-invasive sampling using a microneedle (MN) skin patch. TRMs were first induced by antigen sensitization and subsequently restimulated by intradermal inoculation of the same antigen to trigger the 'alarm' and immune recruitment functions of these cells, leading to accumulation of antigen-specific T cells from the circulation over several days. In mouse vaccination models, MN patches with optimized hydrogel coatings effectively isolated thousands of live antigen-specific lymphocytes and innate immune cells. In a human participant with allergic contact dermatitis, allergen-induced TRM stimulation followed by MN application recovered diverse lymphocyte populations absent from untreated skin sites. These results establish TRM restimulation coupled with MN sampling as a non-invasive platform to monitor both local and systemic antigen-specific immune responses across disease or vaccination settings.

Large language models (LLMs) can generate impressive data visualizations from simple requests, yet their accuracy remains underexplored. Here we present a benchmark of 293 coding tasks derived from 39 studies across 7 biomedical research areas, including biomarkers, integrative analysis, genomic profiling, molecular characterization, therapeutic response, translational research and pan-cancer analysis. Benchmarking eight proprietary and eight open-source LLMs under various prompting strategies reveals an overall accuracy below 40%. This low accuracy raises serious concerns about the risk of propagating incorrect scientific findings when blindly relying on AI-generated analyses. Therefore, we develop an AI agent that begins with and iteratively refines an analysis plan before generating code, achieving 74% accuracy. We embody this insight in a platform that enables users to codevelop analysis plans with LLMs and execute them within an integrated environment. In a user study with five medical researchers, the platform enabled users to complete over 80% of the analysis code for three studies.

Understanding the dynamics of human gait, including both motions and forces, is vital to promote human mobility. While deep learning models may have advantages over costly laboratory-based experiments and physics-based simulations, existing models have been trained on small datasets with homogeneous demographics and focus on predicting a single output. We developed GaitDynamics, a generative foundation model trained on a large dataset of diverse gait patterns, which allows for flexible inputs, outputs and clinical applications. We illustrate the use of GaitDynamics for: (1) estimating ground reaction forces from kinematics with high accuracy even with missing kinematic data, (2) predicting the effects of gait modifications on knee loading without resource-intensive experiments and (3) predicting kinematic and force changes that occur with increasing running speeds. Our results demonstrate the accuracy and efficiency of GaitDynamics, showing its potential to assess and optimize gait for injury prevention, disease treatment and performance coaching. All data, code and trained models are publicly shared.

Adoptive T cell therapy using T cells engineered with novel T cell receptors (TCRs) targeting tumour-specific peptides is a promising immunotherapy. However, these TCR-T cells can cross-react with off-target peptides, leading to severe autoimmune toxicities. Current efforts focus on identifying TCRs with reduced cross-reactivity. Here we show that T cell cross-reactivity can be controlled by the co-signalling molecules CD5, CD8 and CD4, without modifying the TCR. We find the largest reduction in cytotoxic T cell cross-reactivity by knocking out CD8 and expressing CD4. Cytotoxic T cells engineered with a CD8→CD4 co-receptor switch show reduced cross-reactivity to random and positional scanning peptide libraries, as well as to self-peptides, while maintaining their on-target potency. Therefore, co-receptor switching generates super selective T cells that reduce the risk of lethal off-target cross-reactivity and offers a universal method to enhance the safety of T cell immunotherapies for potentially any TCR.