Recent advancements in single-cell RNA sequencing (scRNA-Seq) have transformed how scientists study individual cells, providing unprecedented insights into the molecular mechanisms underlying various diseases and conditions. By analyzing gene expression at the single-cell level, researchers can explore how different cell types function, how they respond to stimuli, and how diseases manifest. However, analyzing these vast amounts of data presents significant challenges that require specialized statistical methods and algorithms.

The Challenge with Standard Classification Methods

Traditionally, classification methods used in gene expression studies focus on grouping samples based on patterns in the gene activity. While this works well for many types of data, scRNA-Seq data is inherently more complex due to the high variability between individual cells. The unique nature of single-cell data requires new, more sophisticated approaches to accurately classify and interpret these patterns.

Researchers at the Centre for Genomics and Oncological Research (GENYO) have developed singleDeep, an end-to-end pipeline developed to analyze scRNA-Seq data by training deep neural networks. The pipeline is designed to make robust predictions about sample phenotypes—whether the data comes from healthy individuals or those with specific conditions. It offers a significant improvement over traditional machine learning methods, making it a powerful tool for researchers in the field of precision medicine.

Scheme of the singleDeep pipeline

Blue boxes represent input and generated data, while orange boxes depict the scripts constituting the software. Green boxes represent the outputs. The workflow starts from standard scRNA-Seq objects processed with PrepareData. R to extract and store the expression matrix from each annotated cell population, along with its associated metadata and sample information. These files serve as input for singleDeep.Py, which trains a dedicated neural network for each cell population and performs a nested CV for internal validation. This step provides training reports, gene contributions, and performance metrics. Lastly, singleDeep_pretrained.Py is available to predict sample phenotypes on new external data using the trained networks from the previous step. These external data consist of additional experiments using samples with the same potential phenotypes as those in the training dataset and the same annotated cell types.

Using singleDeep to Make Predictions

In real-world applications, singleDeep has been used to analyze scRNA-Seq data from various diseases, including systemic lupus erythematosus, Alzheimer’s disease, and COVID-19. Its performance has been validated both internally and externally, with strong diagnostic results that outperformed existing methods. But singleDeep doesn’t just predict outcomes—it also offers valuable insights into the biology of the conditions it analyzes.

For example, in the case of lupus, singleDeep helped identify specific interferon signature genes that were consistently linked to autoimmunity, regardless of the immune cell type. Similarly, in Alzheimer’s disease, it highlighted the role of genes like APOE in brain cell populations such as astrocytes, shedding light on the mechanisms behind dementia.

Why this is Important

The ability to classify and predict phenotypes at the single-cell level brings us closer to understanding the complexities of diseases at their most fundamental level. With singleDeep, scientists can identify biomarkers that may be crucial for diagnosing conditions like lupus, Alzheimer’s, or even COVID-19. Additionally, the insights provided by singleDeep into gene importance and cell type relevance can open doors to more personalized treatments, making this tool not only a technical achievement but also a potential game-changer in medicine.

In summary, singleDeep offers a cutting-edge solution for analyzing scRNA-Seq data, providing both accurate predictions and deeper biological insights. Its ability to tackle the unique challenges of single-cell data analysis is poised to enhance our understanding of disease and drive advances in precision medicine.

Availability – Software, detailed documentation, and help pages for user guidance are publicly available at https://github.com/GENyO-BioInformatics/singleDeep.