TL;DR
BASILISC introduces a Bayesian Siamese network framework to match single-cell data across different molecular layers (like RNA and chromatin) while providing a crucial "confidence score." This allows researchers to filter out unreliable matches and detect new cell types the model hasn't seen before, significantly improving the reliability of multimodal data integration.
What Problem Does This Paper Address?
In single-cell biology, researchers often use different assays to measure different molecular layers of a cell—for example, gene expression (RNA) versus chromatin accessibility (DNA structure). Most datasets are "unimodal," meaning they only capture one of these layers. To get a holistic view of cellular function, we need to "align" these different datasets, a task known as diagonal integration.
While existing machine learning models can predict which cells from one assay correspond to cells in another, they typically provide a "point estimate"—a single best guess without any measure of confidence. Because genomic data is notoriously noisy and sparse, these models often make mistakes. The field lacks a principled way to quantify the uncertainty of these cross-modal matches, which is essential for filtering out bad data and identifying novel cell types that weren't present in the training data.
Methods
The authors present BASILISC (Bayesian Alignment of Siamese-Integrated Latents for Inferring Single-cell Correspondence). The core of the framework is a probabilistic Siamese neural network.
The model is instantiated at two levels of granularity:
- Cell-type level: This model determines if two cells (one from each modality) belong to the same cell type. It uses "frozen" embeddings from a pre-trained variational autoencoder (Liam) as inputs.
- Cell level: This model attempts to find the exact partner for a query cell in the other modality, trained end-to-end from raw molecular counts.
To make the models "Bayesian," the authors place a posterior distribution over the network weights. They compare three common approximation methods: Mean-Field Variational Inference (MFVI), Monte Carlo (MC) Dropout, and Deep Ensembles. Instead of a single prediction, the model generates a distribution of scores over multiple "draws" (stochastic forward passes). The standard deviation of these draws serves as the uncertainty score.
Datasets
The framework was evaluated using the NeurIPS 2021 multimodal single-cell benchmark, which includes:
- Multiome: Simultaneous measurement of gene expression (GEX) and chromatin accessibility (ATAC) in the same nucleus.
- CITE-seq: Simultaneous measurement of gene expression and surface protein abundance (ADT) in the same cell.
These "paired" datasets are unique because they provide a ground-truth correspondence that is usually missing in real-world unimodal integration tasks.
Key Insights and Ideas
- Biological Lineage as a Prior: The model learns a similarity space that reflects biology. When presented with a cell type it has never seen before (Out-of-Distribution), the model doesn't just fail randomly; it matches the cell to its closest biological relative (e.g., a progenitor cell) but flags the match with high uncertainty.
- Filtering "Harmful" Errors: At the cell level, not all errors are equal. Matching a cell to a different cell of the same type is often harmless for downstream analysis. Matching it to a completely different cell type is "harmful." The Bayesian uncertainty score is specifically better at flagging these harmful, cross-type errors than standard deterministic scores.
- Uncertainty as a Quality Gate: The uncertainty score can be used to "abstain" from making a prediction. By discarding the 20% of matches the model is least sure about, the researchers could boost matching accuracy significantly.
Key Results
- Superior Accuracy: At the cell level, a deep ensemble of BASILISC matchers achieved a matching accuracy of 0.062 on the benchmark test set, outperforming the original NeurIPS competition winner (0.058).
- Effective Filtering: At the cell-type level, discarding the least certain fifth of pairs raised classification accuracy from 0.93 to 0.98.
- Novelty Detection: The model successfully identified unseen cell types. In "Leave-One-Cell-Type-Out" (LOCTO) experiments, the uncertainty score (psd) separated unseen types from seen ones with an AUROC of 0.627.
- Error Flagging: The Bayesian posterior flagged wrong matches that crossed cell-type boundaries with an AUROC of 0.68, compared to 0.64 for a deterministic version of the same model.
Limitations and Future Work
- Cell-Level Novelty Detection: While the model excels at the cell-type level, detecting unseen types at the individual cell level remains challenging. The authors found that the size of the cell population in a batch heavily influences confidence, sometimes leading to overconfidence in small, unseen groups.
- Scope: The study focused on human bone marrow data. The authors suggest that future work should test the framework across more diverse tissue types and developmental trajectories.
- Architecture: The authors noted that placing the Bayesian posterior on the encoders (the "legs" of the Siamese network) rather than just the merged head might further improve the model's ability to detect novel biological states.
Summary updated