Synthetic lethality (SL) is a type of genetic interaction where the simultaneous mutation of two genes leads to cell death, while the mutation in either gene individually does not. It is promising to broaden the range of anti-cancer drug targets. Current machine learning methods primarily use either knowledge graphs or handcrafted features to represent genes for SL prediction, overlooking the effective integration of both data types. Furthermore, most of these methods were designed for pan-cancer SL prediction and only a few methods have been proposed for cancer-specific SL prediction. For these cancer-specific methods, a notable decrease in performance is observed when the training data is reduced. Cancer-specific SL prediction under low-data scenarios remains a challenging problem. To address these issues, we propose a novel model named Meta-CapSL to predict potential cancer-specific SL with few labeled data. First, we leverage a graph neural network with self-supervised learning to obtain the cancer-free gene embeddings from a biomedical knowledge graph. We design a strategy to get cancer-specific gene expressions by capturing the context of genes. Then, we use a self-attention-based fusion strategy to integrate the representations of KG embedding and gene expression. Moreover, a meta-learning framework is used to transfer meta-knowledge of SL interactions across diverse cancer types and obtain a well-initialized model which is capable of fast adaptation to a new cancer type with few samples. Extensive experiments demonstrate that Meta-CapSL significantly outperforms the best baseline and generalizes well under three low-data scenarios. Our code is available at https://github.com/JieZheng-ShanghaiTech/Meta-CapSL.