A new study in Science Bulletin presents DVSTP, a deep learning system that integrates pathology images with spatial transcriptomics and proteomics to map intra-tumor heterogeneity. DVSTP predicts molecular profiles from routine pathology slides, making spatial multi-omics more accessible. Whole–tumor 3D reconstruction reveals that SRSF6 drives immune exclusion and is associated with poor clinical outcomes.

Large language models and autonomous agents have advanced rapidly, showing broad promise in medical imaging analysis, clinical diagnosis, and treatment planning. However, most existing medical AI systems still rely primarily on pre-trained knowledge and fixed workflows, making it difficult to learn continuously from long-term clinical feedback, patient outcomes, and prior treatment experience. This "static AI" architecture limits their value in complex real-world clinical settings.

To address this bottleneck, a team led by Dr. Lian Zhang from the First Hospital of Hebei Medical University, in collaboration with domestic and international research partners, has proposed VIBEMed, which is a self-evolving multi-agent framework for clinical decision support designed to enable dynamic learning and safe, traceable system evolution.