Metrics

This package contains all the metrics used for benchmarking scRNA-seq data integration performance. The metrics can be classified into biological conservation and batch removal metrics. For a detailed description of the metrics implemented in this package, please see our publication.

Biological Conservation Metrics

Biological conservation metrics quantify either the integrity of cluster-based metrics based on clustering results of the integration output, or the difference in the feature spaces of integrated and unintegrated data. Each metric is scaled to a value ranging from 0 to 1 by default, where larger scores represent better conservation of the biological aspect that the metric addresses.

hvg_overlap(adata_pre, adata_post, batch[, ...])	Highly variable gene overlap
silhouette(adata, group_key, embed[, ...])	Average silhouette width (ASW)
isolated_labels(adata, label_key, batch_key, ...)	Isolated label score
nmi(adata, group1, group2[, method, nmi_dir])	Normalized mutual information
ari(adata, group1, group2[, implementation])	Adjusted Rand Index
cell_cycle(adata_pre, adata_post, batch_key)	Cell cycle conservation score
trajectory_conservation(adata_pre, ...[, ...])	Trajectory conservation score
clisi_graph(adata, batch_key, label_key[, ...])	Cell-type LISI (cLISI) score

Batch Correction Metrics

Batch correction metrics values are scaled by default between 0 and 1, in which larger scores represent better batch removal.

graph_connectivity(adata, label_key)	Graph Connectivity
silhouette_batch(adata, batch_key, ...[, ...])	Batch ASW
pcr_comparison(adata_pre, adata_post, covariate)	Principal component regression score
kBET(adata, batch_key, label_key[, scaled, ...])	kBET score
ilisi_graph(adata, batch_key[, k0, type_, ...])	Integration LISI (iLISI) score

Metrics Wrapper Functions

For convenience, scib provides wrapper functions that, given integrated and unintegrated adata objects, apply multiple metrics and return all the results in a pandas.Dataframe. The main function is metrics(), that provides all the parameters for the different metrics.

scib.metrics.metrics(adata, adata_int, ari=True, nmi=True)

The remaining functions call the metrics() for

Furthermore, metrics() is wrapped by convenience functions with preconfigured subsets of metrics based on expected computation time:

• metrics_fast() only computes metrics that require little preprocessing

• metrics_slim() includes all functions of metrics_fast() and adds clustering-based metrics

• metrics_all() includes all metrics

metrics(adata, adata_int, batch_key, label_key)	Master metrics function
metrics_fast(adata, adata_int, batch_key, ...)	Only metrics with minimal preprocessing and runtime
metrics_slim(adata, adata_int, batch_key, ...)	All metrics apart from kBET and LISI scores
metrics_all(adata, adata_int, batch_key, ...)	All metrics

Auxiliary Functions

lisi_graph(adata, batch_key, label_key, **kwargs)	cLISI and iLISI scores
pcr(adata, covariate[, embed, n_comps, ...])	Principal component regression for anndata object
pc_regression(data, covariate[, pca_var, ...])	Principal component regression