DESP: Demixing Cell State Profiles from Dynamic Bulk Molecular Measurements

Ahmed Youssef, Indranil Paul, Mark Crovella, Andrew EmilibioRxiv 2023 · DOI

TL;DR

This paper introduces DESP, a novel computational tool that can "demix" bulk molecular measurements (like proteomics) to reveal the distinct molecular profiles of individual cell states. This is crucial because traditional bulk measurements often obscure cell-specific changes, and DESP enables a deeper understanding of dynamic biological processes at the cellular level.

What Problem Does This Paper Address?

Understanding how cells change and adapt over time (phenotypic remodeling) is key to fields like development and disease. However, current technologies often measure molecules in bulk samples, which are mixtures of many different cell types and states. This "averaging" effect masks the unique molecular signatures of individual cell states, making it hard to study dynamic processes like cell differentiation or disease progression at the cellular resolution. While single-cell technologies can provide this resolution, they face technical challenges, especially for protein measurements. DESP aims to bridge this gap by leveraging existing bulk data and single-cell data to infer cell-state-specific molecular profiles.

Methods

DESP is a computational algorithm that uses a mathematical framework to "demix" bulk molecular data. It takes two main inputs:

  • Bulk 'omics data: This is the standard measurement from mixed cell populations (e.g., proteomics data).
  • Single-cell data: This data is used to estimate the proportions of different cell states within each bulk sample. DESP doesn't directly use the single-cell molecular profiles for prediction but rather their inferred cell state compositions.

DESP models the bulk measurements as a linear combination of the unknown cell state profiles, weighted by their proportions. To solve this underdetermined problem, it employs:

  • Nonnegativity constraints: Ensures that molecular abundances are not negative.
  • Tikhonov Regularization (Ridge Regression): Favors solutions with smaller norms, promoting more stable predictions.
  • Similarity Regularization: Incorporates information about the similarity between cell states (derived from single-cell data) to further constrain the solution and maintain biological relevance.

The core of DESP is solving a quadratic programming problem for each gene or protein independently.

Datasets

The authors validated DESP using several datasets:

  • In-vitro EMT model: This dataset, from a previous study (Paul et al., 2023), involved human mammary epithelial cells undergoing TGF-Beta-induced epithelial-to-mesenchymal transition (EMT). It included both bulk proteomics and scRNA-Seq data across eight timepoints. This served as the primary case study for DESP's application.
  • Human Cell Atlas (HCA) Benchmark: Publicly available scRNA-Seq data from the HCA was used to test DESP's performance on a large-scale, diverse dataset across 24 human tissues.
  • CITE-Seq data: This dataset (Hao et al., 2021) provided matched single-cell RNA and protein measurements from peripheral blood mononuclear cells (PBMCs) after an HIV vaccine. This was used to compare DESP's performance on transcriptomic versus proteomic data.

Key Insights and Ideas

  • Bridging the bulk-single-cell divide: DESP offers a powerful way to extract high-resolution, cell-state-specific information from readily available bulk data, overcoming the limitations of direct single-cell protein measurements.
  • Circumventing mRNA-protein correlation issues: By directly using bulk protein measurements and single-cell proportions, DESP bypasses the often poor correlation between mRNA and protein levels, which is a major hurdle in multi-omics integration.
  • Uncovering transient states: DESP can identify molecular signatures of transient cell states that might be diluted or missed in bulk measurements, providing insights into dynamic biological processes.
  • Generalizability: The framework is designed to be applicable to various 'omics data types (proteomics, metabolomics, etc.) and doesn't require pre-defined cell state structures, making it broadly useful.

Key Results

  • Accurate reconstruction: DESP accurately reconstructed cell state profiles from bulk measurements, showing strong overlap with ground truth data in validation studies. For the EMT case study, it recovered expected patterns of canonical EMT markers.
  • Performance on large datasets: DESP demonstrated impressive speed, demixing over 48,000 genes in under 6 seconds on a standard laptop, highlighting its scalability.
  • Robustness: DESP showed robustness to variations in its input parameters (λ and β), with prediction errors remaining stable across a wide range of values, reducing the need for extensive parameter tuning.
  • EMT insights: In the EMT study, DESP identified hundreds of differential markers and pathways not detected by other methods, including novel biological processes and key proteins involved in transient intermediate cell states. For example, it highlighted H2AFX as a marker of intermediate proteomics profiles with low abundance in scRNA, suggesting post-transcriptional regulation.
  • Proteomics vs. RNA: DESP predicted single-cell protein profiles with comparable or even superior accuracy to its predictions for RNA profiles when using matched single-cell RNA and protein data.

Limitations and Future Work

  • Pre-defined cell states: DESP requires pre-defined cell states (e.g., from clustering single-cell data) and does not identify these states itself. The quality of the output depends on the accuracy of the input cell state definitions.
  • Confounding factors: The authors note that cluster size and heterogeneity can be confounded and might influence prediction errors.
  • Future work: The authors envision future applications of DESP for demixing protein interaction networks and spatial transcriptomics data, further expanding its utility. They also provide DESP as an R package for community adoption.

Summary updated

Abstract

Abstract There is wide interest to determine the dynamic expression of proteins and other molecules that drive phenotypic remodeling in development and pathobiology, but due to technical limitations these systems remain largely unexplored at the foundational resolution of the underlying cell states. Here, we present DESP , a novel algorithm that leverages independent estimates of cell state proportions, such as from single-cell RNA-sequencing or cell sorting, to resolve the relative contributions of cell states to bulk molecular measurements, most notably quantitative proteomics, recorded in parallel. We applied DESP to an in-vitro model of the epithelial-to-mesenchymal transition and demonstrated its ability to accurately reconstruct cell state signatures from bulk-level measurements of both the proteome and transcriptome providing insights into transient regulatory mechanisms. DESP provides a generalizable computational framework for modeling the relationship between bulk and single-cell molecular measurements, enabling the study of proteomes and other molecular profiles at the cell state-level using established bulk-level workflows.
BibTeX
@misc{youssef2023desp,
  author = {Ahmed Youssef and Indranil Paul and Mark Crovella and Andrew Emili},
  title = {{DESP: Demixing Cell State Profiles from Dynamic Bulk Molecular Measurements}},
  year = {2023},
  doi = {10.1101/2023.01.19.524460},
  howpublished = {bioRxiv},
}
Figure