A self-adaptive and versatile tool for eliminating multiple undesirable variations from large-scale transcriptomes

Accurate identification of true biological signals from diverse undesirable variations in large-scale transcriptomes is essential for downstream discoveries. Here we develop a universal deep neural network, called DeepAdapter, to eliminate various undesirable variations including batch, platform, purity and other unknown sources from transcriptomic data. Our approach automatically learns the corresponding denoising strategies to adapt to different situations. The data-driven strategies are flexible and highly attuned to the transcriptomic data that requires denoising, yielding reduced undesirable variation originating from batches, sequencing platforms and biosamples with varied purity beyond manually designed schemes. Comprehensive evaluations across multiple batches, different RNA measurement technologies and heterogeneous biosamples demonstrate that DeepAdapter can robustly correct diverse undesirable variations and accurately preserve biological signals, the faithful gene expression patterns that facilitate reliable biomarker discovery, transcriptomic network analysis and comprehensive biological characterization. Our findings indicate that DeepAdapter can act as a versatile tool for the comprehensive denoising of the large and heterogeneous transcriptome across a wide variety of application scenarios.