A Comprehensive Comparative Analysis of Sequence-Based Deep Learning Models for Single-Cell Genomics

To streamline the application of sequence-based DL methods in single-cell genomics, we established a two-layer CNN as our baseline model. We focus our benchmark on how data characteristics, hyperparameter optimization, and advanced model architectures impact performance across sequence-to-expression and sequence-to-regulation tasks. A key contribution of our study is the exploration of multi-task learning (MTL) frameworks for mitigate technical sparsity. We demonstrated that MTL significantly enhances the modeling of cellular heterogeneity, evaluating the effectiveness of task grouping and balancing strategies, with particular focus on the prediction of rare cell types. Our comprehensive comparative analysis provided an actionable framework and valuable insights for guiding future research endeavors and facilitating the development of the sequence-based DL models capable of superior predictive performance in single-cell genomics.