Abstract
Single-cell sequencing is frequently marred by “interruptions” due to limitations in sequencing throughput, yet bulk RNA-seq may harbor these ostensibly “interrupted” cells. In response, we introduce the single cell trajectory blending from Bulk RNA-seq (BulkTrajBlend) algorithm, a component of the OmicVerse suite that leverages a Beta-Variational AutoEncoder for data deconvolution and graph neural networks for the discovery of overlapping community. This approach proficiently interpolates and restores the continuity of “interrupted” cells within single-cell RNA sequencing dataset. Furthermore, OmicVerse provides an extensive toolkit for bulk and single cell RNA-seq analysis, offering uniform access to diverse methodologies, streamlining computational processes, fostering exquisite data visualization, and facilitating the extraction of novel biological insights to advance scientific research.
Introduction
In transcriptomics, researchers typically navigate a trade-off between bulk RNA-seq, which offers a high-level average of gene expression across tissues, and single-cell RNA-seq (scRNA-seq), which provides high-resolution data on individual cells. However, scRNA-seq often suffers from "interruptions"—gaps in developmental trajectories caused by limited sequencing throughput that misses rare intermediate cell states. Interestingly, bulk RNA-seq data often contains these missing cells, albeit blended into a composite signal.
Furthermore, the bioinformatics landscape is currently fragmented across various languages (R, Python) and disconnected tools, complicating the creation of efficient, reproducible pipelines. OmicVerse addresses these challenges by providing a unified Python-based ecosystem.
The OmicVerse Framework and BulkTrajBlend
OmicVerse is designed as a comprehensive library with a standardized API for over a dozen transcriptomic methods. It integrates established tools like CellPhoneDB for cell-cell interaction and MOFA2 for multi-omics, optimized for computational efficiency and visualization within a Pythonic workflow.
A cornerstone of this suite is the BulkTrajBlend algorithm, which bridges the gap between bulk and single-cell data through a two-stage process:
- Deconvolution via $\beta$-VAE: A Beta-Variational AutoEncoder deconvolves bulk RNA-seq data into synthetic single-cell profiles.
- Community Discovery via GNN: Graph Neural Networks (GNNs) perform overlapping community discovery to identify "interrupted" cells and interpolate them back into the original scRNA-seq dataset, restoring trajectory continuity.
Datasets and Validation
The effectiveness of OmicVerse and BulkTrajBlend was validated across several biological contexts:
- Mouse Hippocampus (Dentate Gyrus): Testing trajectory restoration using paired scRNA-seq and bulk RNA-seq.
- Alzheimer’s Disease (AD): Verifying cell types identified in bulk analysis.
- Pancreatic Ductal Adenocarcinoma (PDAC): Validating deconvolution accuracy.
- Colorectal Cancer: Projecting patient data into a tumor microenvironment cell atlas.
Key Results: Healing the Transcriptomic Universe
The most significant achievement of BulkTrajBlend is its ability to "heal" broken trajectories. In the mouse dentate gyrus, standard scRNA-seq data often shows disconnected clusters. BulkTrajBlend successfully identified missing links, correctly positioning Oligodendrocyte Progenitor Cells (OPCs) on the differentiated branch of Neuronal Intermediate Progenitor Cells (nIPC).
Beyond trajectory restoration, OmicVerse demonstrates superior usability by supporting 14 different data formats (e.g., h5ad, csv) and providing a high-quality visualization suite. The pipeline also excels in multi-omic integration, successfully identifying variant factors across scRNA-seq, scATAC-seq, and spatial transcriptomics layers using the MOFA2 framework.
Discussion and Future Directions
OmicVerse represents a shift toward a more streamlined, "one-language" workflow in bioinformatics, leveraging modern machine learning libraries like PyTorch. By treating bulk data as a reservoir of missing cellular states rather than just a low-resolution alternative, it unlocks new potential in existing datasets.
However, as the field evolves, OmicVerse will require continuous updates to incorporate emerging algorithms. While the use of GNNs and VAEs provides powerful interpolation capabilities, it is more computationally intensive than traditional statistical methods. Future work will focus on refining these models and applying them to "dark" biological processes where ground truth lineages are yet to be discovered.
Summary updated