Using Neural Networks to Improve Single Cell RNA-Seq Data Analysis

Chieh Hubert Lin, Siddhartha Jain, Hannah Kim, Ziv Bar‐JosephbioRxiv 2017 · DOI

TL;DR

This paper introduces a novel approach using deep neural networks (NNs) to analyze single-cell RNA sequencing (scRNA-seq) data, improving both cell clustering and the ability to identify cell types from large databases. By learning a more robust, lower-dimensional representation of gene expression, their method outperforms traditional techniques like PCA, offering better accuracy and efficiency for scRNA-seq data analysis.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful technology for understanding cellular heterogeneity, but it presents significant computational challenges. These include dealing with noisy and incomplete data (due to "dropouts"), accurately grouping cells into distinct types or states, and efficiently searching large collections of cell profiles to identify unknown cell types. Existing methods, often relying on Principal Component Analysis (PCA) for dimensionality reduction, struggle with the scale and complexity of scRNA-seq data, and may not fully capture the nuanced relationships between genes that define cell identity.

Methods

The core of this work is the application of deep neural networks (NNs) to scRNA-seq data. The authors explore various NN architectures, including fully connected networks and those incorporating prior biological knowledge (protein-protein and protein-DNA interactions). These NNs are trained to predict cell types from gene expression data, learning a compressed, lower-dimensional representation of the input. This learned representation is then used for downstream tasks like clustering and cell retrieval.

Key methodological aspects include:

  • Supervised Learning: NNs are trained to classify cells into known types, learning to extract relevant features.
  • Dimensionality Reduction: The hidden layers of the NN learn a compact representation of the high-dimensional gene expression data.
  • Biologically Informed Architectures: Some NNs incorporate prior knowledge of gene interactions (PPI and PDI) to guide the learning process, potentially leading to more interpretable and efficient models.
  • Unsupervised Pre-training: Denoising autoencoders (DAEs) are used to pre-train NN weights on unlabeled data, which can improve performance and initialization for supervised tasks.
  • Evaluation Metrics: Adjusted Random Index (ARI) is used for clustering performance, and Mean Average Precision (MAP) is used for cell retrieval accuracy.

Datasets

The study utilizes a substantial collection of scRNA-seq datasets:

  • Training/Evaluation: 3 datasets comprising 16 cell types and 402 single cells were used for initial training, testing, and evaluation of the NN models.
  • Retrieval Analysis: 31 datasets, totaling over 17,000 single-cell profiles, were used to build a comprehensive database for cell retrieval.
  • Gene Selection: A subset of 9437 genes present across training datasets and PPI/PDI data was used for training. For the larger retrieval analysis, the same set of genes was used.

The authors also mention using publicly available data from the Gene Expression Omnibus (GEO).

Key Insights and Ideas

  • NNs Excel at Capturing Complex Relationships: Deep neural networks, with their ability to learn non-linear relationships, are more effective than linear methods like PCA at representing the intricate patterns in scRNA-seq data that define cell identity.
  • Biologically Informed Architectures Offer Advantages: Incorporating prior biological knowledge (PPI and PDI) into NN architectures can lead to models with fewer parameters, potentially reducing overfitting and improving performance, especially for retrieval tasks.
  • Unsupervised Pre-training is Beneficial: Using denoising autoencoders for unsupervised pre-training can serve as a valuable initialization strategy for supervised NN models, enhancing their performance.
  • Reduced Dimensions Dramatically Improve Efficiency: Using the NN-learned low-dimensional representations significantly reduces storage requirements and query run times for large-scale cell retrieval, making it more practical.
  • NNs Improve Both Clustering and Retrieval: The proposed NN-based approach demonstrates superior performance in both unsupervised cell clustering and in accurately identifying cell types from a large database compared to existing methods.

Key Results

  • Clustering Performance: The NN-based dimensionality reduction significantly outperformed PCA, SIMLR, SNN-Cliq, and using original expression values, achieving higher Adjusted Random Index (ARI) scores. The simplest NN architecture (1 fully connected layer with 796 hidden nodes) performed best among the tested NN models for clustering.
  • Retrieval Performance: Methods relying on reduced dimensions (NNs and PCA) performed much better than using raw expression values. NN methods showed an average improvement of ~40% over raw data for retrieval. Specifically, a 2-layer NN incorporating PPI/TF information and pre-training achieved the best retrieval performance, outperforming PCA by a significant margin.
  • Efficiency Gains: For a database of 17,000 cells, the NN-based retrieval reduced query time from 25 minutes to under 5 minutes, while also drastically cutting storage requirements.
  • Functional Interpretation: Analysis of highly weighted hidden layer nodes revealed that NNs can capture biologically relevant gene groups and pathways associated with specific cell types, offering some interpretability.

Limitations and Future Work

  • Limitations:
    • The paper notes that while the NN models achieve high accuracy on training data, this is partly due to the large number of parameters, which can be computationally demanding.
  • While some interpretability was achieved through analyzing node weights, NNs are still largely considered "black boxes."
  • Future Work:
    • Exploring deeper NN architectures.
  • Integrating additional types of prior biological knowledge.
  • Developing an automated tool to download and process new scRNA-seq data for the retrieval application.
  • Addressing the challenge of automatically assigning cell types from published data, given the variability in how such information is encoded.

Summary updated

Abstract

Abstract While only recently developed, the ability to profile expression data in single cells (scRNA-Seq) has already led to several important studies and findings. However, this technology has also raised several new computational challenges including questions related to handling the noisy and sometimes incomplete data, how to identify unique group of cells in such experiments and how to determine the state or function of specific cells based on their expression profile. To address these issues we develop and test a method based on neural networks (NN) for the analysis and retrieval of single cell RNA-Seq data. We tested various NN architectures, some biologically motivated, and used these to obtain a reduced dimension representation of the single cell expression data. We show that the NN method improves upon prior methods in both, the ability to correctly group cells in experiments not used in the training and the ability to correctly infer cell type or state by querying a database of tens of thousands of single cell profiles. Such database queries (which can be performed using our web server) will enable researchers to better characterize cells when analyzing heterogeneous scRNA-Seq samples. Supporting website: http://sb.cs.cmu.edu/scnn/ Password for accessing the retrieval task webserver: scRNA-Seq
BibTeX
@misc{lin2017using,
  author = {Chieh Hubert Lin and Siddhartha Jain and Hannah Kim and Ziv Bar‐Joseph},
  title = {{Using Neural Networks to Improve Single Cell RNA-Seq Data Analysis}},
  year = {2017},
  doi = {10.1101/129759},
  howpublished = {bioRxiv},
}
Figure