Metric learning enables synthesis of heterogeneous single-cell modalities

Rohit Singh, Brian Hie, Ashwin Narayan, Bonnie BergerbioRxiv 2019 · DOI

TL;DR

This paper introduces "Schema," a novel metric learning framework that effectively synthesizes information from multiple, heterogeneous single-cell data modalities. Schema allows for more accurate biological insights by integrating diverse data types, overcoming limitations of single-modality analyses and enabling more robust and interpretable discoveries.

What Problem Does This Paper Address?

Understanding complex biological processes requires integrating information from various sources, such as gene expression, chromatin accessibility, cell type, and even spatial location. Historically, single-cell experiments could only capture one of these "modalities" at a time, forcing researchers to integrate data from separate experiments, which introduces noise and biases. The advent of multi-modal single-cell assays allows for simultaneous measurement of these diverse properties, but this new data paradigm demands new analytical tools. The challenge is to effectively synthesize these disparate data streams to reveal deeper biological insights, rather than just combining them in a superficial way.

Methods

The core of Schema is a principled metric learning strategy. Instead of simply combining data, Schema treats each modality as defining a "distance metric" between cells. The algorithm then learns a transformation of a "primary" modality (chosen by the researcher, often the most reliable) so that its learned distances align with the distances implied by the "secondary" modalities. This transformation is constrained to limit distortion of the primary modality, ensuring that high-confidence information is preserved. Schema uses a quadratic programming optimization framework, which is computationally efficient and allows for the incorporation of an arbitrary number of modalities. A key feature is its ability to return feature weights, indicating the importance of different features in the primary modality for aligning with secondary modalities, which aids in biological interpretation.

Datasets

The authors evaluated Schema on several publicly available datasets:

  • Mouse kidney cells: Integrating RNA-seq and ATAC-seq for cell type inference.
  • Mouse gastrulation epiblast cells: Analyzing differential gene expression while accounting for batch effects and developmental age.
  • Drosophila melanogaster neurons: Visualizing age-related structure in RNA-seq data.
  • Mouse dentate gyrus cells: Synthesizing spliced and unspliced mRNA counts for cell differentiation analysis.
  • Mouse cerebellum tissue (Slide-seq): Investigating spatial transcriptomics and cell density.
  • Human A549 cells: Linking chromatin accessibility to gene expression variability.
  • T-cell receptor (TCR) data: Integrating peptide sequence and binding specificity.

Key Insights and Ideas

  • Metric learning as a unifying framework: The core insight is that different biological modalities can be viewed as defining different ways to measure similarity (or distance) between cells. Schema leverages this by learning how to reconcile these different "views" of cellular similarity.
  • Constrained optimization for robustness: By limiting the distortion of the primary modality, Schema prevents overfitting to noisy or sparse secondary modalities, ensuring that the most reliable information is prioritized.
  • Feature weighting for interpretability: Schema's ability to output feature weights provides direct biological insights into which genes or genomic regions are most important for integrating different data types.
  • Scalability to massive datasets: The use of quadratic programming and a sampling approach allows Schema to handle the ever-increasing size of single-cell datasets efficiently.

Key Results

  • Improved cell type inference: Schema achieved an 11% increase in accuracy for inferring cell types from RNA-seq and ATAC-seq data compared to previous methods.
  • Robust differential expression analysis: Schema successfully identified genes with monotonically changing expression along developmental time while accounting for batch effects, outperforming MOFA+ in identifying stage-dependent genes.
  • Enhanced visualization: Schema enabled more informative visualizations of cellular trajectories, such as developmental age in neurons and cell differentiation from spliced/unspliced RNA counts, by incorporating metadata.
  • Identification of biologically relevant features: Schema identified sequence locations crucial for T-cell receptor binding specificity and genomic regions important for gene expression variability.
  • Computational efficiency: Schema processed a large Slide-seq dataset (3 modalities, ~20k cells, ~17k genes) in 34 minutes using less than 5GB of RAM, significantly faster than other methods like SpatialDE and Trendsceek.

Limitations and Future Work

  • Alternative formulations: The authors suggest exploring alternative formulations of the Schema objective, potentially including more complex nonlinearities than their current quadratic-program approach.
  • Experimental guidance: Schema can guide future experiments by identifying highly weighted features that can be profiled using targeted assays.
  • Hyperparameter tuning: While Schema offers more intuitive hyperparameters than some methods, principled selection of parameters like the correlation constraint (s) is still an area for user consideration.
  • Generalization of distance metrics: While Schema accepts arbitrary distance metrics for secondary modalities, the primary modality is limited to Euclidean distances. Future work could explore more general transformations.

Summary updated

Abstract

Abstract A complete understanding of biological processes requires synthesizing information across heterogeneous modalities, such as age, disease status, or gene/protein expression. Until recently, single-cell profiling experiments could measure only a single modality, leading to analysis focused on integrating information across separate experiments. However, researchers can now measure multiple modalities simultaneously in a single experiment, providing a new data paradigm that enables biological discovery but also requires new conceptual and analytic models. We therefore present Schema, an algorithm that leverages a principled metric learning strategy to synthesize multimodal information from the same experiment. To demonstrate the flexibility and power of our approach, we use Schema to infer cell types by integrating gene expression and chromatin accessibility data, perform differential gene expression analysis while accounting for batch effects and developmental age, estimate evolutionary pressure on peptide sequences, and synthesize spliced and unspliced mRNA data to infer cell differentiation. Schema can synthesize arbitrarily many modalities and capture sophisticated relationships between them, is computationally efficient, and provides a valuable conceptual model for exploring and understanding complex biology.
BibTeX
@misc{singh2019metric,
  author = {Rohit Singh and Brian Hie and Ashwin Narayan and Bonnie Berger},
  title = {{Metric learning enables synthesis of heterogeneous single-cell modalities}},
  year = {2019},
  doi = {10.1101/834549},
  howpublished = {bioRxiv},
}

Loading related content...

Figure