Neural Data Visualization for Scalable and Generalizable Single Cell Analysis

Hyunghoon Cho, Bonnie Berger, Jian PengbioRxiv 2018 · DOI

TL;DR

Researchers have developed net-SNE, a novel neural network-based approach for visualizing single-cell RNA sequencing (scRNA-seq) data. Unlike traditional methods like t-SNE, net-SNE is highly scalable to millions of cells and, crucially, can generalize to new, unseen data, enabling knowledge transfer across experiments.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful technology for understanding cellular heterogeneity, but the sheer volume of data generated (millions of cells) poses significant computational challenges. Existing visualization methods, most notably t-stochastic neighbor embedding (t-SNE), struggle with this scale, requiring days or even weeks to process large datasets. Furthermore, t-SNE visualizations are not inherently generalizable; they cannot easily incorporate new data points without recomputing the entire embedding. This limitation hinders the ability to integrate and compare data from different experiments or to analyze newly acquired cells in the context of existing knowledge.

Methods

The core innovation is neural t-SNE (net-SNE). Instead of directly computing pairwise similarities and embedding them, net-SNE trains a neural network (NN) to learn a mapping function. This function takes a cell's gene expression profile as input and outputs its low-dimensional coordinates for visualization. The NN is trained to optimize the same objective function as t-SNE, aiming to preserve local structure.

Key aspects of the method include:

  • Parametric Embedding: The NN learns a function $y_i = NN(x_i; \theta)$, where $x_i$ is the input expression profile and $y_i$ is the output embedding. This allows for the embedding of any new input $x$, not just those present in the training data.
  • Scalability Enhancements: To handle large datasets, net-SNE leverages optimizations from t-SNE, such as the Barnes-Hut approximation, and introduces novel techniques for stochastic optimization (mini-batch gradient descent) and training with a reference visualization. The latter uses a pre-computed t-SNE embedding as a target for the NN, significantly speeding up training.
  • Generalizability: By learning a mapping function, net-SNE can project new cells onto an existing visualization, even if these cells represent entirely new subtypes or come from different experimental batches.

Datasets

The authors evaluated net-SNE on a variety of datasets:

  • 13 benchmark scRNA-seq datasets: These datasets varied in size (from tens to tens of thousands of cells) and had known cluster labels, allowing for direct comparison of visualization quality with t-SNE.
  • Brain1m dataset: A large-scale dataset of 1.3 million mouse brain cells was used to specifically test scalability and generalizability on massive datasets. This dataset did not have known labels, so objective scores were used for comparison.
  • Six purified blood cell type datasets: These were used to demonstrate net-SNE's ability to map new cell populations onto an existing visualization.

Key Insights and Ideas

  • Redundancy in scRNA-seq data: The paper highlights that as datasets grow, a significant portion of newly sequenced cells are redundant, meaning they are very similar to already observed cells. This redundancy is key to enabling scalable visualization by learning a generalizable mapping.
  • The power of a learned mapping: By training a neural network to learn a mapping function, net-SNE overcomes the static nature of traditional embedding methods like t-SNE. This learned function is the key to its generalizability.
  • Bridging the gap between visualization quality and speed: net-SNE achieves visualization quality comparable to t-SNE while drastically reducing computation time, especially for large datasets.
  • Translational potential: The ability to generalize to new data opens up new workflows for integrating diverse scRNA-seq studies and transferring knowledge across experiments.

Key Results

  • Visualization Quality: net-SNE achieved visualization quality comparable to t-SNE on 13 benchmark datasets, as measured by the adjusted Rand index (ARI) for clustering accuracy.
  • Scalability: For the 1.3 million-cell Brain1m dataset, net-SNE achieved a high-quality visualization in approximately 20 minutes using a bootstrap approach, compared to 13 hours for default t-SNE.
  • Generalizability:
    • In cross-validation experiments, net-SNE accurately positioned held-out cell subtypes as distinct clusters, even when these subtypes were absent from the initial training data. A naive t-SNE extension (t-SNE+k-NN) failed to do this.
  • When projecting new FACS-purified blood cell populations onto a pre-trained net-SNE embedding of whole blood cells, net-SNE precisely mapped these new cells to their corresponding clusters, demonstrating effective cross-dataset knowledge transfer.
  • Speedup with Reference Visualization: Training net-SNE to match a pre-computed t-SNE reference significantly reduced training time, enabling visualization of large datasets in minutes rather than days.

Limitations and Future Work

  • Reference Visualization Dependency: The "Training net-SNE with Reference Visualization" approach requires an initial t-SNE computation, which can still be time-consuming for very large datasets, though it serves as a good starting point.
  • Optimization Challenges: While net-SNE offers significant improvements, training neural networks for complex objectives like t-SNE can still be challenging, and the authors note difficulties with very small mini-batches.
  • Future Directions: The authors suggest exploring the parameters of the trained neural networks to gain insights into which gene expression patterns are driving the visualizations. They also envision net-SNE as a foundation for building comprehensive reference maps of cell types, similar to the Human Cell Atlas, and for learning reduced vector space representations for broader biological analysis.

Summary updated

Abstract

Summary Single-cell RNA sequencing is becoming effective and accessible as emerging technologies push its scale to millions of cells and beyond. Visualizing the landscape of single cell expression has been a fundamental tool in single cell analysis. However, standard methods for visualization, such as t-stochastic neighbor embedding (t-SNE), not only lack scalability to data sets with millions of cells, but also are unable to generalize to new cells, an important ability for transferring knowledge across fast-accumulating data sets. We introduce net-SNE, which trains a neural network to learn a high quality visualization of single cells that newly generalizes to unseen data. While matching the visualization quality of t-SNE on 14 benchmark data sets of varying sizes, from hundreds to 1.3 million cells, net-SNE also effectively positions previously unseen cells, even when an entire subtype is missing from the initial data set or when the new cells are from a different sequencing experiment. Furthermore, given a “reference” visualization, net-SNE can vastly reduce the computational burden of visualizing millions of single cells from multiple days to just a few minutes of runtime. Our work provides a general framework for newly bootstrapping single cell analysis from existing data sets.
BibTeX
@misc{cho2018neural,
  author = {Hyunghoon Cho and Bonnie Berger and Jian Peng},
  title = {{Neural Data Visualization for Scalable and Generalizable Single Cell Analysis}},
  year = {2018},
  doi = {10.1101/289223},
  howpublished = {bioRxiv},
}
Figure