SCREP: Towards Single-Cell Drug Response Prediction by Pharmacogenomic Embedding Enhanced Meta-Pretraining and Few-Shot Transfer Learning

ObjectiveSingle-cell pharmacogenomic data is crucial for identifying biomarkers and understanding resistance mechanisms. However, the limited availability of those data poses significant challenges for efficient pre-training and thus hinders the generalization capability of the model. This paper aims to develop an efficient predictive model for single-cell drug response inference by leveraging knowledge from a bulk dataset. MethodsTo translate knowledge from bulk cell lines to single-cell analysis, this paper proposes a meta-pretraining and few-shot transfer learning frame-work based on pharmacogenomic embeddings. To enhance feature representation and alignment, genomic information is processed by a position-based feature extraction network to extract contextual features. Simultaneously, a graph-aware Transformer is developed to capture interatomic relations for drug information representation. Underlying this framework, key drug action pathways can be identified at the cellular level through the proposed gene gradient attribution algorithm. ResultsThis model integrates drug response data from 223 drugs across 14 tissues for meta pre-training, followed by transfer and testing on seven single-cell datasets. In comparison to other models, the proposed framework achieves an average accuracy increase of 4.58% for pre-trained drugs and demonstrates a 20% improvement in generalization performance for previously unseen drugs. Case studies further illustrate its capability to differentiate between resistance genes. ConclusionThe proposed framework exhibits enhanced generalization capabilities in predicting cellular responses across multiple drugs, including unseen drugs. This model serves as an effective tool for investigating drug action pathways and elucidating resistance mechanisms from single-cell insights, with significant implications for guiding clinical medication practices.