scDLC: a deep learning framework to classify large sample single-cell RNA-seq data.

Area of Science:

• Genomics
• Computational Biology
• Biomedical Data Science

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for disease diagnosis in medical research.
• Existing statistical methods like PLDA, NBLDA, and ZIPLDA struggle with large scRNA-seq datasets and violated distribution assumptions.

Purpose of the Study:

• To develop a novel deep learning classifier for accurate disease diagnosis using large-scale scRNA-seq data.
• To address the limitations of current methods in handling complex data distributions and large sample sizes.

Main Methods:

• A deep learning classifier, scDLC, was developed using long short-term memory (LSTM) recurrent neural networks.
• scDLC does not require prior data distribution knowledge and models dependencies among feature genes.
• LSTMs are employed for their ability to learn long-term dependencies within sequential data.

Main Results:

• Simulation studies demonstrated superior performance of scDLC compared to existing methods across various settings with large sample sizes.
• Analysis of four real-world scRNA-seq datasets confirmed scDLC's consistently best performance.
• The scDLC method shows robustness even when distribution assumptions of traditional methods are not met.

Conclusions:

• The proposed scDLC deep learning classifier offers a significant advancement for disease diagnosis using large scRNA-seq datasets.
• scDLC provides a more accurate and robust approach than traditional statistical methods, particularly for complex biological data.
• The scDLC code is publicly available, facilitating its adoption and further research in the field.