多组学患者来源类器官嵌入可预测靶向治疗反应和KRAS抑制剂敏感性

Clinical trials remain a major bottleneck in contemporary drug development. Functional assays based on patient-derived organoids (PDOs) are a promising tool to derisk clinical trials, but their impact has been limited by small cohort sizes and a lack of systematic validation. Using the largest cohort of matched PDOs, longitudinal clinical data, transcriptomic data and WGS data, we generate multi-modal patient representations. We then evaluate how large collections of these representations can be used to characterize drug response phenotypes and to map how tumor-intrinsic features relate to sensitivity to targeted therapies, with a particular emphasis on KRAS inhibition. Methods: We assembled 135 PDOs from colorectal (CRC) and pancreatic (PDAC) cancers, derived from primary tumors and metastatic biopsies. Patients were 31-89 years old and had received a mean of 2.6 prior lines of treatment at the time of tissue sampling. Each PDO underwent WES and bulk RNA-seq and was profiled in functional drug screens with a panel of targeted agents, including KRAS inhibitors. We generated embeddings from omics features for each PDO. These embeddings were used as inputs to machine-learning models trained to predict drug AUC values, enabling in silico estimation of PDO drug sensitivity from baseline molecular profiles. Within the same framework, we analysed KRAS inhibitor response by identifying molecular neighborhoods in the embedding space enriched for sensitive or resistant PDOs and stratifying samples accordingly. Results: Embeddings captured major axes of biological variation across the PDO collection, including tumor type, key oncogenic alterations and clinically relevant subgroups, and enabled accurate prediction of AUC values for multiple targeted agents, including KRAS inhibitors. Predicted sensitivities were concordant with experimentally measured responses and reflected known dependencies associated with specific mutational backgrounds. We identified biomarkers and pathway-level signatures associated with response to different KRAS-targeted treatments. These signatures were reproducible across cross-validation procedures and aligned with reported mechanisms of KRAS inhibitor activity and escape. Conclusions: Large, clinically annotated PDO collections coupled with multi-omics profiling support the construction of embeddings that both predict ex vivo drug response and highlight molecular contexts associated with sensitivity to targeted agents. Integrating these predictive models with systematic pathway interpretation provides a rigorous framework to uncover transcriptomic and genomic markers of response and resistance. For drug developers, this approach shows how PDO-based functional genomics combined with AI driven analysis can be used to derisk development of targeted therapies and inform patient stratification and trial design.