scBasset: Sequence-based modeling of single cell ATAC-seq using convolutional neural networks

Abstract

Single cell ATAC-seq (scATAC) shows great promise for studying cellular heterogeneity in epigenetic landscapes, but there remain significant challenges in the analysis of scATAC data due to the inherent high dimensionality and sparsity. Here we introduce scBasset, a sequence-based convolutional neural network method to model scATAC data. We show that by leveraging the DNA sequence information underlying accessibility peaks and the expressiveness of a neural network model, scBasset achieves state-of-the-art performance across a variety of tasks on scATAC and single cell multiome datasets, including cell type identification, scATAC profile denoising, data integration across assays, and transcription factor activity inference.

Best practices for single-cell histone modification analysis

“…The neighbor score varies between 0 when both representations disagree completely to 1 when both representations are identical (see Methods and Figure 1B). This evaluation has been previously used in [18, 19] and is currently the standard for evaluating modality alignment tasks in recent community benchmarks such as https://openproblems.bio/.…”

“…Overall, looking at the available datasets and prior biological knowledge, it appears reasonable and the best we can have without proper labels. We can also note that this approach has already been successfully used in evaluating scATAC-seq pipelines in [19, 18].…”

“…To start filling this gap, we perform in this paper a large-scale computational study to evaluate the impact and best choices for the number of cells, coverage per cell, cell selection, matrix construction algorithm, feature selection and dimension reduction algorithm. To quantify the impact of each of these factors, we use two single-cell multi-omics datasets where, in addition to scHPTM, a second modality is measured for each cell (gene expression or cell surface proteins); we then assess how well the cell-to-cell similarity observed with scHPTM data analysis agrees with the one inferred from the co-assay (RNA or protein) [18, 19]. The analysis of more than 10.000 computational experiments allows us to clarify the impact of various experimental choices and data processing factors for scHTPM data, and to suggest practical guidelines.…”

Scalable sequence-informed embedding of single-cell ATAC-seq data with CellSpace

“…Following their default variable peak selection, peaks accessible in fewer than 5% of cells were removed, leaving 38,502 peaks for training. scBasset was trained for 45 epochs, at which point the validation loss and AUC had leveled and the correlation between the intercept and the library size was 0.99 5 .…”

“…This approach can indeed group cells by their cell type identity but introduces bias through a priori motif choice and therefore may miss latent structure in the data. Recently, scBasset 5 used a multi-task neural network approach to learn a sequence model for chromatin accessible peaks that passes through a low-dimensional bottleneck layer, together with cell-specific model vectors that predict whether a peak-given its bottleneck representation-will be accessible in the cell. This approach yields a low dimensional representation of cells via the model vectors and can assign TF accessibility scores to cells by passing a sequence with a planted motif through the model.…”