Specifying Embedding Methods and Preprocessing Steps

The main functionality of scloop is to provide a unified interface for embedding and clustering methods with a common set of preprocessing steps. The embedding_algs argument is a list of strings which specify the embedding methods and preprocessing steps to run. The embedding methods are specified by the name of the method followed by a + and then a comma-separated list of preprocessing steps.

List of Embedding Methods

Baselines:

Name	Description
1d_pca	1-dimensional PCA baseline. Takes each contact matrix or preprocessed matrix and aggregates over rows to produce a 1D vector for each cell. Embed using PCA.
1d_lsi	1-dimensional LSI baseline. Same as 1D PCA baseline, but embed using LSI instead
2d_pca	PCA baseline. Takes each contact matrix or preprocessed matrix and unravels the specified number of strata into a single vector representation for each cell. Embed using PCA
2d_lsi	LSI baseline. Same as PCA baseline but embed using LSI instead
scHiCluster	scHiCluster method for embedding scHi-C based on VCSQRT normalization, convolution, and random-walk imputation prior to PCA embedding. By default this method will run with these preprocessing steps unless otherwise specified.
fastHiCRep	Similarity-based method which relies on the HiCRep stratum-adjusted correlation coefficient (SCC) as a distance metric for MDS embedding
InnerProduct	Generalization of fastHiCRep which ignores distance in correlation computation and simply computes cosine similarities of each strata vector
cisTopic	Convert dataset into a set of discrete locus-pair occurences (bag-of-words representation) and run Latent Dirichlet Allocation for topic modeling

Conventional scRNA-seq/scATAC-seq methods:

Name	Description
scVI	Aggregate each contact matrix into a 1D vector and treat it like a transcription vector. Embed using default scVI model
scVI_2d	Unravel each contact matrix into a 1D vector and embed using default scVI model
peakvi	Aggregate each contact matrix into a 1D vector and treat like a binary peak vector. Embed using PeakVI
peakvi_2d	Unravel each contact matrix into a 1D vector and embed using PeakVI

Deep learning methods:

Name	Description
Higashi	Represent entire dataset as a hypergraph and learn cell node embeddings by training a hypergraph neural network
Fast-Higashi	Higashi model based on tensor decomposition rather than training a hypergraph neural network
3DVI	Train an scVI model on each strata independently and concanenate to obtain final cell embeddings
VaDE	Variational deep embedding model, trains a VAE with Gaussian mixture prior (if number of clusters is specified). Similar to 3DVI but embeds entire matrix instead of independent strata

Biological feature representations:

Name	Description
deTOKI/deDOC	Identify TADs and domain boundaries in each contact matrix and embed using PCA of domain density vectors
InsScore	Identify domain boundaries by computing Insulation Score over sliding window and embed using PCA of domain density vectors
scGAD	Map each cell to a gene score vector based on a set of known loci. Scores represent z-scores from the BandNorm normalization method

List of Preprocessing Steps

Filtering operations:

Name	Description
quantile_<q>	Filter low values based on a quantile cutoff specified by q
min_count_<N>	Filter values with count lower than N

Normalization operations:

Name	Description
vc_sqrt_norm	Vanilla square-root coverage correction. Normalize by the square-root of the row-sums and column sums.
oe_norm	Distance correction. Compute average of each distance strata for expected values and return the observed/expected ratios
kr_norm	KR normalization. Convert the contact matrix into a doubly-stochastic matrix using the KR algorithm

Graph operations:

Name	Description
convolution	Perform simple neighbor averaging with a box filter
random_walk	Perform random-walk imputation on the contact matrices
network_enhance	Compute KNN transition matrix for each contact matrix
google	Compute PageRank transition matrix for each contact matrix