An Automatic Recommendation Method for Single-Cell DNA Variant Callers Based on Meta-Learning Framework

The rapid expansion of single-cell sequencing-based research has motivated a proliferation of variant callers on the sequencing data. Due to the differences on calling strategies, these callers often exhibit varying performance when applied across heterogeneous sequencing samples. Selecting a suitable caller that fits for the data on-hand becomes an overwhelming task for researchers in this field. Thus, in this study, an automatic recommendation method for single-cell DNA (scDNA) variant callers is proposed. This recommender is designed on meta-learning framework. It explores the underlying associations between scDNA data features and the optimal variant caller on specific performance metric. The recommender is trained by benchmark sequencing datasets, and base on this, recommend appropriate caller for new sequencing data. A series of experiments on different datasets and various configurations have been conducted to validate the proposed method. The results demonstrate that the average performance of this recommendation method outperforms fixed and random strategies.