Predicting single-cell drug sensitivity utilizing adaptive weighted features for deep transfer learning

Abstract The advancement of single-cell sequencing technology had promoted the generation of a large amount of single-cell transcriptional profiles, providing valuable opportunities to identify drug-resistant cell populations in a tumor. However, the drug sensitivity data at single-cell level is still scarce to date, posing an urgent but tough challenge for the computational prediction of drug sensitivity in individual cells. This paper proposed scAdaDrug, a deep transfer learning model with adaptive weighted features to predict single-cell drug sensitivity. We used an autoencoder to extract domain-invariant features related to drug sensitivity from multiple source domains by exploiting adversarial domain adaptation. Especially, we introduced an adaptive weight generator to produce importance-aware and mutually independent weights, which could adaptively modulate the embedding of each sample in dimension-level for both source and target domains. Extensive experimental results showed that our model achieved state-of-the-art performance in predicting drug sensitivity on multiple independent single-cell datasets derived from cell lines, patient-derived xenografts (PDX) models, and patient tumor tissues. Moreover, the ablation experiments demonstrated our model effectively captured the underlying patterns determining drug sensitivity from multiple source domains.