BABEL enables cross-modality translation between multiomic profiles at single-cell resolution

Kevin Wu, Kathryn E. Yost, Howard Y. Chang, James ZouProceedings of the National Academy of Sciences 2021 · DOI

TL;DR

This paper introduces BABEL, a deep learning model that can translate between single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) data. This allows researchers to computationally infer one modality from the other, enabling richer analyses even when only one type of data is available, and opening doors for new hypotheses and data exploration.

What Problem Does This Paper Address?

Single-cell biology has seen incredible advancements, allowing us to measure various molecular features (like gene expression, chromatin accessibility, protein levels) within individual cells. However, simultaneously measuring multiple of these "modalities" from the same cell is technically challenging, expensive, and can introduce noise. This means researchers often have datasets with only one type of molecular information per cell. This creates a gap: how can we leverage the power of multiomic insights when we only have single-modality data? The authors aim to bridge this gap by developing a method to computationally "translate" between common single-cell modalities, specifically scRNA-seq and scATAC-seq.

Methods

The core of BABEL is a deep learning architecture based on an interoperable neural network. It consists of two "encoders" that project scRNA-seq and scATAC-seq data into a shared, low-dimensional "latent space" representing cellular state. Two corresponding "decoders" then take points from this latent space and predict either the scRNA-seq or scATAC-seq profile.

What's novel here is the "interoperable" design: the encoders and decoders are trained together such that any encoder can feed into any decoder. This forces the latent space to capture a unified representation of cellular state that is modality-agnostic. For scATAC-seq, they leverage the intrachromosomal nature of chromatin accessibility to reduce model complexity. For scRNA-seq, they use a negative binomial loss to predict denoised expression, and for scATAC-seq, a binary cross-entropy loss for peak activity prediction.

Datasets

The authors trained and evaluated BABEL on several datasets:

  • Human Multiomic Data: Jointly profiled scATAC-seq and scRNA-seq data from peripheral blood mononuclear cells (PBMCs), colon adenocarcinoma (COLO-320DM and COLO-320HSR), and lymphoblastoid GM12878 cells. These were used for training, validation, and testing, with specific splits designed to challenge the model's generalization capabilities.
  • Mouse Multiomic Data: Jointly profiled scATAC-seq and scRNA-seq data from adult mouse cerebral cortex (using SNARE-seq) and mouse skin (using SHARE-seq). These were used to test BABEL's cross-species generalization.
  • Unpaired PBMC scATAC-seq Data: Used to test BABEL's ability to impute RNA expression from ATAC data when the data was not jointly profiled.
  • Basal Cell Carcinoma (BCC) Data: Patient-derived scATAC-seq data from BCC tumors, used to demonstrate BABEL's application on challenging clinical samples. This was compared against existing scRNA-seq data from the same patients.
  • CITE-seq Data: Jointly profiled RNA and protein epitope measurements from human bone marrow cells, used to demonstrate BABEL's extensibility to predict protein profiles.

Key Insights and Ideas

  • Bridging the Data Gap: BABEL offers a powerful way to overcome the limitations of single-modality datasets, making multiomic analyses more accessible.
  • Interoperable Latent Space: The core innovation is creating a shared latent space that truly integrates information from different modalities, allowing for flexible translation without explicit alignment steps.
  • Generalization is Key: The model demonstrates impressive generalization to cell types and biological contexts not seen during training, a crucial feature for real-world applications.
  • Beyond Gene Activity Scores: BABEL's imputed RNA expression can provide more nuanced and specific insights than traditional gene activity scores derived solely from ATAC data, particularly for genes with complex regulatory landscapes.
  • Extensibility: The architecture is designed to be modular, allowing for the addition of new decoders to predict other modalities, such as protein epitopes.

Key Results

  • High Accuracy: BABEL achieved strong performance in translating between scATAC-seq and scRNA-seq. For example, on a held-out test cluster, it achieved a Pearson's correlation of 0.62 for ATAC-to-RNA and an AUROC of 0.92 for RNA-to-ATAC.
  • Superior to Baselines: BABEL significantly outperformed k-Nearest Neighbors (KNN) models for these translation tasks.
  • Cross-Species Generalization: The model performed well on mouse data, showing Pearson's correlations around 0.53-0.55 for ATAC-to-RNA.
  • Clinical Sample Application: When applied to patient-derived BCC scATAC-seq data, BABEL's imputed RNA profiles showed good concordance (Pearson's r = 0.70) with empirical pseudobulk scRNA-seq data, and provided more specific gene expression predictions than traditional gene activity scores.
  • Protein Epitope Prediction: BABEL's framework was extended to predict protein epitope profiles from scATAC-seq data, demonstrating its versatility.

Limitations and Future Work

  • Training Data Dependence: BABEL performs best when predicting on cell types similar to those seen during training. For entirely novel cell types (e.g., malignant BCC cells for which there were no basal epithelial cells in the training data), performance can be weaker.
  • Mutual Information Constraint: The accuracy of BABEL's predictions is inherently limited by the mutual information between the modalities. Changes in chromatin accessibility don't always directly correlate with transcriptomic changes.
  • Future Work: The authors suggest that BABEL can be combined with other single-cell analysis workflows to improve multiomic integration. They also highlight the potential for transfer learning by using BABEL's pretrained components for new modalities. The modular design allows for easy extension to predict other data types, and the authors envision it as a valuable tool for hypothesis generation and exploratory analysis in increasingly multiomic research landscapes.

Summary updated

Abstract

Simultaneous profiling of multiomic modalities within a single cell is a grand challenge for single-cell biology. While there have been impressive technical innovations demonstrating feasibility-for example, generating paired measurements of single-cell transcriptome (single-cell RNA sequencing [scRNA-seq]) and chromatin accessibility (single-cell assay for transposase-accessible chromatin using sequencing [scATAC-seq])-widespread application of joint profiling is challenging due to its experimental complexity, noise, and cost. Here, we introduce BABEL, a deep learning method that translates between the transcriptome and chromatin profiles of a single cell. Leveraging an interoperable neural network model, BABEL can predict single-cell expression directly from a cell's scATAC-seq and vice versa after training on relevant data. This makes it possible to computationally synthesize paired multiomic measurements when only one modality is experimentally available. Across several paired single-cell ATAC and gene expression datasets in human and mouse, we validate that BABEL accurately translates between these modalities for individual cells. BABEL also generalizes well to cell types within new biological contexts not seen during training. Starting from scATAC-seq of patient-derived basal cell carcinoma (BCC), BABEL generated single-cell expression that enabled fine-grained classification of complex cell states, despite having never seen BCC data. These predictions are comparable to analyses of experimental BCC scRNA-seq data for diverse cell types related to BABEL's training data. We further show that BABEL can incorporate additional single-cell data modalities, such as protein epitope profiling, thus enabling translation across chromatin, RNA, and protein. BABEL offers a powerful approach for data exploration and hypothesis generation.
BibTeX
@inproceedings{wu2021babel,
  author = {Kevin Wu and Kathryn E. Yost and Howard Y. Chang and James Zou},
  title = {{BABEL enables cross-modality translation between multiomic profiles at single-cell resolution}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2021},
  doi = {10.1073/pnas.2023070118},
}
Figure