Drug-drug interactions (DDIs) can significantly impact drug efficacy and safety, potentially leading to severe adverse effects. Existing works on DDI event prediction have typically relied on labels of specific events for supervision, neglecting the importance of mining textual descriptions. This limits their ability to address two challenges: (1) the lack of observable data for new drugs, hindering meaningful feature extraction; (2) the highly imbalanced event distribution, which causes models to overfit to common categories and struggle with rare interactions. To address these challenges, we propose RADDI, a retrieval-augmented DDI prediction method. This approach improves prediction accuracy and adapts to the dynamic nature of new drug discovery. Specifically, to solve the first challenge, RADDI introduces a collaborative prediction strategy that integrates general knowledge transfer with specialized knowledge retrieval. This approach uses pretrained language models to generate embeddings for drug descriptions at a coarse level, enabling broad interaction classification. At a finer level, RADDI incorporates retrieval augmentation, using drug pair descriptions as retrieval keys and interaction categories as retrieval targets, thereby enhancing semantic understanding. For the second challenge, we design a class-aware probability distribution strategy to mitigate class imbalance. By leveraging the prior distribution of event categories, RADDI adjusts the retrieval sample weights and normalizes category probabilities, thereby improving the prediction accuracy for rare-class interactions while reducing over-reliance on high-frequency categories. Experiments on benchmark datasets demonstrate that RADDI excels in zero-shot DDI prediction scenarios, effectively balancing generalization to new drugs and maintaining high accuracy across various interaction categories.

Radiology report generation aims to produce textual reports automatically based on input images, a critical process that aids in accurate diagnoses and lightens the workload of radiologists. Following recent advances in Large Language Models (LLMs), several Retrieval-Augmented Generation (RAG) based report generation models have been proposed. Despite the continuously improved performance, these report generation models often suffer from two main limitations, i.e., interference of irrelevant information, and lack of alignment between the input image and the resulting generated report. In this study, we propose the Semantic feedback based RAG Radiology report generation model, namely RAGSemRad. RAGSemRad comprises two key components: the fine-grained semantic retrieval module and the semantic assessment module. The fine-grained semantic retrieval module is designed to retrieve adequate and relevant prompt information, while ignoring irrelevant interference. This is achieved by clustering the data at the semantic level and leveraging the domain knowledge within a large pre-trained visual-language model, thus alleviating the issues of hallucination and databias. Further, the semantic assessment module enhances the performance of the upper bound by enhancing the alignment between the input image and the resulting generated report, utilizing supervision signals derived from paired image-label data. Experimental evaluations are conducted on two benchmarks, IU X-Ray and MIMIC-CXR, to assess the performance of RAGSemRad. The results demonstrate RAGSemRad exhibits competitive performance compared to the state-of-the-art methods, showcasing its potential to advance automatic radiology report generation.