Skip to main content
Back to timeline
Nature MethodsSource publication:

Echidna puts bulk whole-genome sequencing and single-cell transcriptomes in one Bayesian model, beating four CNA tools on clonal reconstruction

Synopsis

Copy-number alterations and transcriptional plasticity in tumors have largely been analyzed apart: methods that infer copy number from single-cell RNA sequencing rely on hidden Markov models whose sequential constraint makes it hard to model how copy number relates to expression. Echidna places bulk whole-genome sequencing gene dosage and single-cell RNA sequencing in one Bayesian hierarchical model and infers copy number jointly across samples and time points. On matched endometrial cancer single-nucleus DNA and RNA data, its clonal reconstruction beats inferCNV, CopyKAT, Numbat and XClone on purity, AMI and ARI (P < 1 × 10−10).

Source-provided article image: Echidna: a Bayesian framework for integrative inference of copy number alterations impacting phenotypic plasticity
Nature Methods

Interpretation

Echidna lets bulk whole-genome sequencing and single-cell RNA sequencing from the same specimen jointly constrain copy-number inference, and extends that inference across multiple samples and treatment time points. Earlier single-cell copy-number methods inferred copy-number states along the chromosome with hidden Markov models, and that sequential constraint limited modeling of copy number together with downstream expression effects, as well as clonal evolution across time points and tissues. On endometrial cancer tumor specimens 2 and 5, with single-nucleus DNA sequencing as ground truth, Echidna's clonal reconstruction beat inferCNV, CopyKAT, Numbat and XClone on purity, adjusted mutual information (AMI) and adjusted rand index (ARI), with P < 1 × 10−10.

The model treats each phenotypic cluster's gene dosage as a latent variable and the bulk-sequenced gene dosage as a weighted sum of cluster dosages, with cluster proportions at that time point as weights. Single-cell transcript counts are modeled as Gamma-Poisson to absorb overdispersion, and the transcriptional rate is a latent parameter following a Gamma distribution scaled by gene dosage, linking copy number and expression in one generative chain. After fitting, gene dosage is smoothed along the genome with a one-dimensional Gaussian kernel, cluster-specific neutral values are learned with a Gaussian mixture model, and a Gaussian hidden Markov model assigns discrete amplified, deleted or neutral states; cluster covariance is parameterized by a Cholesky decomposition with an LKJ prior to stabilize estimates of clonal dependence.

Perspective

The demonstrated scope is tumor samples in which copy-number alterations are a hallmark: endometrial cancer, glioblastoma, pancreatic ductal adenocarcinoma, non-small cell lung cancer brain metastasis and one melanoma patient treated with anti-PD-1 therapy, with single-cell RNA sequencing and bulk whole-genome sequencing as inputs.

The model does not currently account for transcriptional changes from epigenetic regulation; the neutral state is taken as the most frequent Gaussian mixture component, an assumption that may need extra supervision in multiploidy; and clone reconstruction degrades progressively as bulk whole-genome sequencing noise rises.