Iterative single-cell multi-omic integration using online learning

Abstract

Integrating large single-cell gene expression, chromatin accessibility and DNA methylation datasets requires general and scalable computational approaches. Here we describe online integrative nonnegative matrix factorization (iNMF), an algorithm for integrating large, diverse, and continually arriving single-cell datasets. Our approach scales to arbitrarily large numbers of cells using fixed memory, iteratively incorporates new datasets as they are generated, and allows many users to simultaneously analyze a single copy of a large dataset by streaming it over the internet. Iterative data addition can also be used to map new data to a reference dataset. Comparisons with previous methods indicate that the improvements in efficiency do not sacrifice dataset alignment and cluster preservation performance. We demonstrate the effectiveness of online iNMF by integrating more than a million cells on a standard laptop, integrating large single-cell RNA-seq and spatial transcriptomic datasets, and iteratively constructing a single-cell multi-omic atlas of the mouse motor cortex.

MultiMAP: dimensionality reduction and integration of multimodal data

“…We assessed and benchmarked the performance of MultiMAP against several popular approaches for integrating single-cell multi-omics, including Seurat [ 10 ], LIGER [ 11 ], iNMF [ 12 ], Conos [ 29 ], and GLUER [ 30 ].…”

“…Most data integration methods project multiple measurements of information into a common low-dimensional representation to assemble multiple modalities into an integrated embedding space. Recently published methods employ different algorithms to project multiple datasets into an embedding space, including canonical correlation analysis (CCA) [ 10 ], nonnegative matrix factorization (NMF) [ 11 , 12 ], or neural network models [ 13 ]. These methods have demonstrated utility, yet suffer from shortcomings, including challenges with scaling and being limited to consideration of features shared across data sets (e.g., the same genes).…”

Dynamic inference of cell developmental complex energy landscape from time series single-cell transcriptomic data

“…Therefore, developing an on-line framework of GraphFP that can cluster and annotate the single-cell time series scRNA-seq data in different batches in a serial fashion should be an interesting topic. The newly developed single-cell data analysis tools such as the on-line integration method online iNMF [ 40 ] and the cell type annotation method scArches based on transfer learning [ 41 ] can be adopted.…”

Effective and scalable single-cell data alignment with non-linear canonical correlation analysis

“…Many single cell data alignment methods have been recently developed. The majority of them, with a few notable exceptions such as the recent iNMF ( 9 ), are targeted towards small-size and medium-size datasets. These existing methods can be summarized into four categories: (i) reference-based methods, such as Scmap-cluster ( 10 ) and scAlign ( 11 ), which align new inquiry datasets based on a well-annotated reference dataset; (ii) clustering-based methods, such as Harmony ( 12 ), DESC ( 13 ), which remove batch effects and align samples in an embedding space by iteratively optimizing a clustering objective function; (iii) matching-based methods, such as MNN ( 14 ) and Scanorama ( 15 ), which apply a mutually nearest neighbors strategy to identify overlapped cells across datasets and (iv) projection-based methods that use a statistical model to project individual cells from different datasets into a lower dimensional space, including Seurat ( 16 , 17 ) that applies canonical correlation analysis for projection, LIGER ( 18 ) that uses latent factors from non-negative matrix factorization for projection, and scVI ( 19 , 20 ) and others ( 21–23 ) that use variational techniques for projection.…”

“…The ZINB-based methods such as scVI may be less efficient in capturing complex expression features for multiple datasets. Although some existing recent methods ( 9 ) can be scaled up to large-size datasets, they still have potential to inaccurately align cells due to complicated parametric models. Therefore, it is in urgent need to develop effective alignment methods that are also computationally efficient.…”