Abstract
Neuron-specific morphology and function are fundamentally tied to differences in gene expression across the nervous system. We previously generated a single cell RNA-seq (scRNA-Seq) dataset for every anatomical neuron class in the C. elegans hermaphrodite. Here we present a complementary set of bulk RNA-seq samples for 52 of the 118 canonical neuron classes in C. elegans. We show that the bulk RNA-seq dataset captures both lowly expressed and noncoding RNAs that are not detected in the scRNA-Seq profile, but also includes false positives due to contamination by other cell types. We present an analytical strategy that integrates the two datasets, preserving both the specificity of scRNA-Seq data and the sensitivity of bulk RNA-Seq. We show that this integrated dataset enhances the sensitivity and accuracy of transcript detection and differential gene analysis. In addition, we show that the bulk RNA-Seq data set detects differentially expressed non-coding RNAs across neuron types, including multiple families of non-polyadenylated transcripts. We propose that our approach provides a new strategy for interrogating gene expression by bridging the gap between bulk and single cell methodologies for transcriptomic studies. We suggest that these datasets advance the goal of delineating the mechanisms that define morphology and connectivity in the nervous system.
Introduction
In the quest to map every gene expressed in every neuron, scientists usually face a trade-off between sensitivity and specificity. Single-cell RNA-seq (scRNA-seq) is the gold standard for specificity, identifying exactly which cell a transcript originated from. However, it is often "shallow," missing lowly expressed genes (dropout) and, because most protocols target poly-adenylated (polyA) tails, it ignores most non-coding RNAs.
Conversely, bulk RNA-seq is highly sensitive and can capture non-polyA transcripts. However, isolating specific neuron types via Fluorescence-Activated Cell Sorting (FACS) inevitably introduces "contaminating" transcripts from neighboring cells, leading to false positives. This study bridges this gap, creating a refined transcriptomic atlas for the C. elegans nervous system that retains the sensitivity of bulk data without the noise of contamination.
Methods
The researchers isolated 41 of the 118 distinct neuron classes in the C. elegans hermaphrodite using FACS and performed bulk RNA-seq using a ribodepletion and random priming strategy. This method allows for the detection of non-coding RNAs and provides uniform coverage across gene bodies.
To integrate this with existing scRNA-seq data from the CeNGEN project, they employed a "pseudobulk" approach. They summed scRNA-seq reads from identical cell types to create a clean version of a bulk sample. They then integrated these datasets using a geometric mean calculation to filter out genes appearing in contaminated bulk data but absent in specific scRNA-seq clusters.
Additionally, they used Non-Negative Least Squares (NNLS) regression to estimate the proportion of non-neuronal contamination in each bulk sample, filtering out genes that correlated too strongly with contaminant tissues.
Datasets
The study utilized several key data sources:
- New Bulk RNA-seq: 160 samples covering 41 individual neuron types.
- CeNGEN scRNA-seq: An existing atlas covering all 118 neuron classes.
- Ground-Truth Sets: A positive set of 160 genes with high-confidence known expression patterns and a negative set of 445 genes known to be expressed exclusively outside the nervous system.
Key Results and Insights
Sensitivity and Accuracy
The researchers found that scRNA-seq sensitivity peaks at about 92%, missing nearly 10% of the transcriptome. The integrated dataset pushed this closer to the 100% sensitivity of bulk data while maintaining low false-positive rates. For rare neuron types (small clusters), the bulk data provided a vital safety net, detecting between 5 and 169 genes per cell type that scRNA-seq missed. These genes were enriched for critical neuronal functions like neuropeptide signaling and synaptic morphology.
Decontamination and Differential Expression
The integration strategy significantly reduced the False Positive Rate (FPR) for non-neuronal genes. In a comparison between I5 and BAG neurons, bulk data called 26.3% of non-neuronal genes as "enriched," while the integrated data reduced that to just 4.5%. Using the Matthew’s Correlation Coefficient (MCC), the integrated approach showed a significant mean improvement (+0.089) over bulk alone in pairwise differential expression tests.
The Non-Coding Frontier
Due to the random priming method, the bulk data revealed a massive world of non-coding RNAs (ncRNAs). The team identified 561 cell-type-specific ncRNAs, including snoRNAs and snRNAs, which likely play unappreciated roles in neuronal identity.
Discussion
While the integration currently covers 41 of the 118 neuron classes in C. elegans, expanding this bulk library to the remaining classes is a clear next step. This integrated atlas provides a high-resolution lens for studying mRNA splicing and the functional roles of non-coding RNAs in specific neurons—areas where scRNA-seq has traditionally been limited. This approach provides a new strategy for interrogating gene expression by bridging the gap between bulk and single-cell methodologies.
Summary updated