Rare diseases affect more than 300 million people worldwide^1,2,3, yet timely and accurate diagnosis remains an urgent challenge^1,3,4,5. Patients often endure a prolonged ‘diagnostic odyssey’ exceeding 5 years, marked by repeated referrals, misdiagnoses and unnecessary interventions, leading to delayed treatment and substantial emotional and economic burden^4,5. Here we present DeepRare—a multi-agent system for rare disease differential diagnosis decision support^6,7,8 powered by large language models, integrating more than 40 specialized tools and up-to-date knowledge sources. DeepRare processes heterogeneous clinical inputs, including free-text descriptions, structured human phenotype ontology terms and genetic testing results to generate ranked diagnostic hypotheses with transparent reasoning linked to verifiable medical evidence. Evaluated across nine datasets from literature, case reports and clinical centres across Asia, North America and Europe spanning 14 medical specialties, DeepRare demonstrates exceptional performance on 2,919 diseases. In human-phenotype-ontology-based tasks, it achieves an average Recall@1 of 57.18%, outperforming the next best method by 23.79%; in multi-modal tests, it reaches 69.1% compared with Exomiser’s 55.9% on 168 cases. Expert review achieved 95.4% agreement on its reasoning chains, confirming their validity and traceability. Our work not only advances rare disease diagnosis but also demonstrates how the latest powerful large-language-model-driven agentic systems can reshape current clinical workflows.

Tumor-derived circulating cell-free DNA (cfDNA) provides critical clues for cancer early diagnosis, yet it often suffers from low sensitivity. Here, we present a cancer early diagnosis approach using tumor fractions deciphered from circulating cfDNA methylation signatures. We show that the estimated fractions of tumor-derived cfDNA from cancer patients increase significantly as cancer progresses in two independent datasets. Employing the predicted tumor fractions, we establish a Bayesian diagnostic model in which training samples are only derived from late-stage patients and healthy individuals. When validated on early-stage patients and healthy individuals, this model exhibits a sensitivity of 86.1% for cancer early detection and an average accuracy of 76.9% for tumor localization at a specificity of 94.7%. By highlighting the potential of tumor fractions on cancer early diagnosis, our approach can be further applied to cancer screening and tumor progression monitoring.

Rare cancers comprise 20-25% of malignancies (over 70% in pediatric oncology) but face major diagnostic challenges due to limited expert availability. While pathology vision-language models show promising zero-shot capabilities for common cancers, their performance on rare cancers remains limited. Existing multi-instance learning (MIL) methods rely solely on visual features, overlooking cross-modal knowledge and compromising interpretability critical for rare cancer diagnosis. To address this, we propose PathPT, a framework that exploits vision-language foundation models through spatially-aware visual aggregation and task-specific prompt tuning. PathPT converts WSI-level supervision into fine-grained tile-level guidance, preserving tumor localization and enabling cross-modal reasoning. Across eight rare and three common cancer datasets-spanning 56 subtypes and 3958 WSIs, PathPT consistently outperforms state-of-the-art methods under data-scarce settings. It achieves substantial gains in both subtyping accuracy and cancerous region grounding ability, providing a scalable, interpretable AI solution to improve rare cancer subtyping with limited access to specialized expertise.

Nuclei detection is a fundamental task in the field of histopathology image analysis and remains challenging due to cellular heterogeneity. Recent studies explore convolutional neural networks to either isolate them with sophisticated boundaries (segmentation-based methods) or locate the centroids of the nuclei (counting-based approaches). Although these two methods have demonstrated superior success, their fully supervised training demands considerable and laborious pixel-wise annotations manually labeled by pathology experts. To alleviate such tedious effort and reduce the annotation cost, we propose a novel local integral regression network (LIRNet) that allows both fully and weakly supervised learning (FSL/WSL) frameworks for nuclei detection. Furthermore, the LIRNet can output an exquisite density map of nuclei, in which the localization of each nucleus is barely affected by the post-processing algorithms. The quantitative experimental results demonstrate that the FSL version of the LIRNet achieves a state-of-the-art performance compared to other counterparts. In addition, the WSL version has exhibited a competitive detection performance and an effortless data annotation that requires only 17.5% of the annotation effort.

Object re-identification (ReID) aims to find instances with the same identity as the given probe from a large gallery. Pairwise losses play an important role in training a strong ReID network. Existing pairwise losses densely exploit each instance as an anchor and sample its triplets in a mini-batch. This dense sampling mechanism inevitably introduces positive pairs that share few visual similarities, which can be harmful to the training. To address this problem, we propose a novel loss paradigm termed Sparse Pairwise (SP) loss that only leverages few appropriate pairs for each class in a mini-batch, and empirically demonstrate that it is sufficient for the ReID tasks. Based on the proposed loss framework, we propose an adaptive positive mining strategy that can dynamically adapt to diverse intra-class variations. Extensive experiments show that SP loss and its adaptive variant AdaSP loss outperform other pairwise losses, and achieve state-of-the-art performance across several ReID benchmarks. Code is available at https://github.com/Astaxanthin/AdaSP.