Inferring Copy Number Variation (inferCNV)

Inferring Copy Number Variation with inferCNV

Introduction

In oncology and cancer single-cell RNA-seq, distinguishing malignant tumor cells from normal, healthy cells in the microenvironment is the most critical first step. Because cancer is fundamentally driven by genomic instability, malignant cells often have massive chromosomal amplifications or deletions.

inferCNV is a powerful R package that uses single-cell RNA expression as a proxy for DNA copy number variation (CNV). By comparing the expression of genes across the genome in a suspect cell against a known "normal" reference cell, it can identify these large-scale chromosomal alterations.

1. Preparing Data for inferCNV

You need three inputs for inferCNV: 1. A raw count matrix. 2. A metadata file mapping each cell to its annotation (e.g., "Tumor", "Normal_T_cell", "Normal_B_cell"). 3. A gene ordering file mapping each gene to its physical position on the chromosome.

library(infercnv)

# Assuming you have extracted counts from your Seurat object
raw_counts <- GetAssayData(seurat_obj, slot = "counts")

# Create the inferCNV object
infercnv_obj <- CreateInfercnvObject(
  raw_counts_matrix = raw_counts,
  annotations_file = "cell_annotations.txt",
  delim = "\t",
  gene_order_file = "gene_ordering_hg38.txt",
  ref_group_names = c("Normal_T_cell", "Normal_B_cell") # Define the healthy reference!
)

2. Running the inferCNV Pipeline

inferCNV applies a series of smoothing steps, moving averages, and hidden Markov models (HMM) to denoise the RNA data and uncover the true DNA copy number signal.

# Run the core inferCNV algorithm
infercnv_obj <- infercnv::run(
  infercnv_obj,
  cutoff = 0.1, # 0.1 for 10x Genomics, 1 for Smart-seq2
  out_dir = "infercnv_output/",
  cluster_by_groups = TRUE,
  denoise = TRUE,
  HMM = TRUE,
  num_threads = 8
)

Key Considerations

• Cutoff: Use 0.1 for sparse data like 10x Genomics, and 1.0 for full-length methods like Smart-seq2.
• Denoising: Always use denoise = TRUE to remove background noise using the residual distributions.
• HMM: The Hidden Markov Model predicts the actual integer copy number (e.g., 2 copies, 3 copies, or complete deletion).

3. Interpreting the Output

inferCNV automatically generates a heatmap in your output directory (infercnv_output/infercnv.webp).

• Rows are individual cells.
• Columns are genes, ordered strictly by their physical location from Chromosome 1 to Chromosome X/Y.
• Colors: Red indicates chromosomal amplification (e.g., Trisomy). Blue indicates chromosomal deletion.

Large, coordinated shifts across chromosome arms can be consistent with copy-number alteration and may help prioritize clusters for follow-up. They do not, by themselves, prove that cells are malignant: expression programs, technical effects, an unsuitable reference, or subclonal structure can all influence the pattern. A neutral-looking reference is useful, but it is not a guarantee of a diploid genome.

Treat expression-derived CNV inference as a screening and hypothesis-generation step. Before reporting a malignant population, validate the result with orthogonal DNA-based evidence when possible, such as matched bulk or single-cell DNA sequencing, shallow whole-genome sequencing, FISH, or clinically reviewed genomic data. Interpret the result together with pathology, sample context, QC, and canonical lineage markers.

Matched Python and R CNV-inference workflow

Define normal-reference cells before inference and interpret broad expression-derived CNV patterns cautiously. These methods do not replace DNA-based copy-number validation.

import infercnvpy as cnv

cnv.tl.infercnv(
    adata,
    reference_key="cell_type",
    reference_cat=["B cell", "T cell"],
    window_size=100,
    key_added="cnv",
)
cnv_matrix = adata.obsm["X_cnv"]

library(infercnv)

infercnv_obj <- CreateInfercnvObject(
  raw_counts_matrix = counts_matrix,
  annotations_file = "cell_annotations.tsv",
  delim = "\t",
  gene_order_file = "gene_order.tsv",
  ref_group_names = c("B cell", "T cell")
)
infercnv_obj <- infercnv::run(
  infercnv_obj,
  cutoff = 0.1,
  out_dir = "infercnv_output",
  cluster_by_groups = TRUE
)