Evaluating the false-positive rate of single-cell copy-number and malignant-cell classifiers

Single-cell studies of cancer partition cells into malignant and non-malignant compartments using copy-number inference or supervised classifiers, and every downstream analysis inherits the errors of that partition. These methods are rarely applied to tissue that contains no tumor, so their false-positive rate is largely unknown. We assembled 14 tumor-free single-cell RNA-seq datasets (36,528 cells; 10 healthy organs and 4 datasets from inflammatory-disease cohorts) and ran three widely used callers, inferCNV (as infercnvpy), CopyKAT and ikarus, under published decision rules. In this panel every positive call is false by construction. infercnvpy and CopyKAT made false calls in every dataset, and ikarus in 12 of 14. A twocriterion rule adapted from Tirosh and colleagues called 0.5–19.3% of cells malignant, and a forced two-group split called 4.3–56.2% (up to 78.0% without a reference). CopyKAT called 3.0–63.0% of cells aneuploid (above 20% in 9 of 14 datasets) and ikarus 0–100% (above 50% in 4 of 14); pooled over the cells each scored, both assigned about 31% to the malignant class. In-silico positive controls, made by imposing arm-level dosage changes on real cells, showed per-cell discrimination near chance for single events (AUC 0.53), rising to 0.85 at 18 events. At a false-call rate matched on the same diploid cells used to measure it, ranking cells by the continuous copy-number score detected 2.2–5.3 times as many altered cells as the published rules in all eight estimable rule comparisons, and up to 10.3 times as many as an external classifier. Because the threshold was set on those diploid cells, these factors are upper bounds. The false calls share a pseudo-karyotype that recurs across tissues (median Spearman 𝜌 = 0.54 over 91 dataset pairs) and tracks local gene density (𝜌 = −0.53; 𝑃 < 10−4 against a within-chromosome circular-shift null that preserves genomic autocorrelation): gene-dense regions look deleted and genepoor regions look gained. Replacing the decision step with a cell-type-conditional empirical null and Benjamini–Hochberg control at 𝛼 = 5% cut the per-cell false-call rate to a median of 0.03% (maximum 0.43%) in 13 of 14 datasets: 28 false calls among 33,536 cells. The extreme-tail 𝑝-values were not exactly calibrated, however: 7 of those 13 datasets still received at least one call, more than expected under Benjamini–Hochberg control whether it is applied per dataset or per donor. In the remaining dataset, psoriatic skin from a separate study, a null drawn from Tabula Sapiens tissues called 26.8% of cells. In a separate donor-held-out evaluation, a cross-tissue null called 27.1% of psoriatic cells and a null drawn from held-out donors of the same study called 0.03%. The sensitivity of the calibrated rule was not evaluated. We recommend that malignant-cell calls be accompanied by a false-positive rate measured on tumor-free cells processed like the query.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24324080
Primary Topic
Single-cell and spatial transcriptomics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Evaluating the false-positive rate of single-cell copy-number and malignant-cell classifiers

Preonath Chondrow Dev, Md. Jubayer Hossain
5 links
Zenodo (CERN European Organization for Nuclear Research)
Single-cell and spatial transcriptomics
preprint

Evaluating the false-positive rate of single-cell copy-number and malignant-cell classifiers

Preonath Chondrow Dev, Md. Jubayer Hossain
preprint en

Abstract

Single-cell studies of cancer partition cells into malignant and non-malignant compartments using copy-number inference or supervised classifiers, and every downstream analysis inherits the errors of that partition. These methods are rarely applied to tissue that contains no tumor, so their false-positive rate is largely unknown. We assembled 14 tumor-free single-cell RNA-seq datasets (36,528 cells; 10 healthy organs and 4 datasets from inflammatory-disease cohorts) and ran three widely used callers, inferCNV (as infercnvpy), CopyKAT and ikarus, under published decision rules. In this panel every positive call is false by construction. infercnvpy and CopyKAT made false calls in every dataset, and ikarus in 12 of 14. A twocriterion rule adapted from Tirosh and colleagues called 0.5–19.3% of cells malignant, and a forced two-group split called 4.3–56.2% (up to 78.0% without a reference). CopyKAT called 3.0–63.0% of cells aneuploid (above 20% in 9 of 14 datasets) and ikarus 0–100% (above 50% in 4 of 14); pooled over the cells each scored, both assigned about 31% to the malignant class. In-silico positive controls, made by imposing arm-level dosage changes on real cells, showed per-cell discrimination near chance for single events (AUC 0.53), rising to 0.85 at 18 events. At a false-call rate matched on the same diploid cells used to measure it, ranking cells by the continuous copy-number score detected 2.2–5.3 times as many altered cells as the published rules in all eight estimable rule comparisons, and up to 10.3 times as many as an external classifier. Because the threshold was set on those diploid cells, these factors are upper bounds. The false calls share a pseudo-karyotype that recurs across tissues (median Spearman 𝜌 = 0.54 over 91 dataset pairs) and tracks local gene density (𝜌 = −0.53; 𝑃 < 10−4 against a within-chromosome circular-shift null that preserves genomic autocorrelation): gene-dense regions look deleted and genepoor regions look gained. Replacing the decision step with a cell-type-conditional empirical null and Benjamini–Hochberg control at 𝛼 = 5% cut the per-cell false-call rate to a median of 0.03% (maximum 0.43%) in 13 of 14 datasets: 28 false calls among 33,536 cells. The extreme-tail 𝑝-values were not exactly calibrated, however: 7 of those 13 datasets still received at least one call, more than expected under Benjamini–Hochberg control whether it is applied per dataset or per donor. In the remaining dataset, psoriatic skin from a separate study, a null drawn from Tabula Sapiens tissues called 26.8% of cells. In a separate donor-held-out evaluation, a cross-tissue null called 27.1% of psoriatic cells and a null drawn from held-out donors of the same study called 0.03%. The sensitivity of the calibrated rule was not evaluated. We recommend that malignant-cell calls be accompanied by a false-positive rate measured on tumor-free cells processed like the query.

Zenodo (CERN European Organization for Nuclear Research)
Child Health Research Foundation (BD)
Good health and well-being
Single-cell and spatial transcriptomics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon