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 generate one type of data from the other, enabling richer multi-omic analyses even when only one modality is experimentally available, and offers a powerful tool for data exploration and hypothesis generation.
What Problem Does This Paper Address?
Single-cell technologies have revolutionized biology by allowing us to study cellular states at an unprecedented resolution. However, most standard single-cell experiments capture only one type of molecular information per cell, such as gene expression (scRNA-seq) or chromatin accessibility (scATAC-seq). While technologies exist to jointly profile multiple modalities within the same cell, they are often complex, expensive, and can introduce noise. This limitation means that crucial information about how different layers of genomic regulation interact within individual cells is often lost. The authors tackle this problem by developing a method to computationally "translate" between these modalities, effectively creating paired multi-omic data from single-modality measurements.
Methods
The core of BABEL is a deep learning architecture composed of four modular neural networks: two encoders and two decoders. The encoders take either scRNA-seq or scATAC-seq data and project it into a shared, low-dimensional "latent space" that represents the abstract cellular state. The decoders then take points from this latent space and reconstruct the corresponding scRNA-seq or scATAC-seq profiles.
What's novel here is BABEL's interoperable design: the encoders and decoders are trained to work together seamlessly. This means an ATAC encoder's output can be used by an RNA decoder, and vice versa, enabling cross-modality translation. For ATAC data, BABEL incorporates an "chromosome-aware" design, pruning inter-chromosomal connections to focus on more biologically relevant intra-chromosomal interactions, which significantly reduces model complexity. The model is trained using a joint loss function that encourages accurate reconstruction of both input modalities and accurate translation between them.
Datasets
BABEL was trained and evaluated on several datasets:
- Human PBMC, colon adenocarcinoma (COLO-320DM and COLO-320HSR), and GM12878 cells: These were used for training and testing BABEL's ability to translate between scATAC-seq and scRNA-seq.
- Mouse cerebral cortex data (SNARE-seq): Used to demonstrate BABEL's generalizability across species and experimental protocols.
- Patient-derived basal cell carcinoma (BCC) samples: Used to showcase BABEL's application on challenging, unpaired clinical data.
- Human bone marrow and PBMC CITE-seq data: Used to demonstrate BABEL's extension to predict protein epitope profiles.
Key Insights and Ideas
- Bridging the multi-omic gap: BABEL offers a powerful way to overcome the experimental limitations of joint multi-omic profiling, making it accessible to a wider range of researchers.
- "Information bottleneck" for generalization: The use of a constrained latent space acts as an "information bottleneck," forcing the model to learn the most important biological variation and improving its ability to generalize to unseen cell types and contexts.
- Chromosome-aware ATAC processing: By focusing on intra-chromosomal interactions, BABEL's ATAC processing is more efficient and potentially more biologically relevant.
- Interoperability is key: The interoperable design of BABEL's encoder-decoder modules is crucial for learning a shared, aligned latent representation without needing complex adversarial methods or manifold alignment.
- Pseudo-bulk correlation as a validation metric: The authors demonstrate that pseudo-bulk correlation can be a useful metric for evaluating the quality of BABEL's predictions, especially when paired ground truth data is unavailable.
Key Results
- Accurate cross-modality translation: BABEL achieves strong performance in translating between scATAC-seq and scRNA-seq. On a held-out test cluster, it achieved a Pearson's correlation of 0.63 for ATAC-to-RNA inference and an AUROC of 0.92 for RNA-to-ATAC inference.
- Generalization to new contexts: BABEL generalizes well to cell types and tissues not seen during training. For example, it accurately predicted scRNA-seq from scATAC-seq data of basal cell carcinoma (BCC) samples, enabling fine-grained cell state classification.
- Improved performance over existing methods: BABEL's inferred gene expression from ATAC data showed significantly higher correlation (0.63) compared to gene activity scores from tools like ArchR (0.20) and MAESTRO (0.18) on a test cluster.
- Cross-species applicability: BABEL demonstrated similar performance on mouse data as on human data, indicating its generalizability across species.
- Extension to protein prediction: BABEL can be extended to predict protein epitope profiles from scRNA-seq data, and even from scATAC-seq data by first inferring RNA profiles.
Limitations and Future Work
- Performance dependency on training data similarity: BABEL performs best when predicting on cells similar to those in its training set. Samples that deviate significantly from the training data may show poorer predictive performance.
- Extending to new modalities: While BABEL can be extended to new modalities (like protein epitopes), this requires additional jointly profiled data for training new encoder/decoder pairs.
- Future directions: The authors suggest that BABEL can be integrated with other analysis workflows, such as improving multi-omic integration techniques by substituting BABEL's more accurate RNA inferences for gene activity scores. They also envision BABEL as a powerful tool for single-cell regulome analysis in an increasingly multi-omic world, enabling more efficient data exploration and hypothesis generation.
Summary updated