BGPT: Paper Review: Identifying tumor cells at the single-cell level using machine learning.

Explore by Goal

Quick Explanation Copied

Paper in one line: ikarus builds a cross-cancer tumor-vs-normal gene signature (162 genes) and classifies single cells via gene-set scoring + logistic regression + label propagation on a gene-set-derived cell-cell network, with optional CNV “proofreading” to reduce false positives.

Evidence source:

Long Explanation

Review (skeptical, evidence-based): Identifying tumor cells at the single-cell level using machine learning

DOI: 10.1186/s13059-022-02683-1 | Year: 2022

1) What the paper claims (minimal, testable)

Builds a cross-cancer tumor gene signature of 162 genes and a complementary normal gene signature, derived by differential expression across multiple expertly annotated single-cell RNA-seq datasets and taking intersections/unions to form robust enriched/depleted sets.
Classifies each cell as tumor vs normal using gene-set scoring (AUCell/ssGSEA), a logistic regression trained on tumor/normal signature scores, then refines labels via network propagation on a cell-cell graph derived from tumor/normal gene sets.
Optional CNV “proofreading” (via inferCNV calls and an additional logistic model) aims to reduce false positives while keeping balanced accuracy approximately constant.
Across multiple cancer types and platforms, they report high cross-dataset performance (average balanced accuracy reported around ~0.98) and outperform several baseline/classification tools in their comparisons.

2) Visual overview of the pipeline (from the paper)

Step A — Signature discovery: differential expression per dataset → select top genes (top ~300 used internally) → aggregate via intersection (tumor enriched) / union (tumor depleted) logic across datasets → cross-validation to choose signature/model.

Step B — Cell scoring + prediction: AUCell/ssGSEA gene-set scores → logistic regression (tumor/normal) → network propagation on a gene-set-derived cell-cell graph with decreasing stringency threshold.

Optional CNV proofreading: inferCNV-derived CNVs → train CNV-based logistic regression using pseudo-labels from ikarus → flip discordant labels to reduce false positives.

3) Reported signature structure & robustness (visualized)

The paper reports that among the 162 tumor-signature genes, only one robust co-expression module appears consistently across datasets: a 34-gene cell-cycle module; the remaining genes show largely near-zero pairwise co-expression (low modularity).

4) Ablation & error-sensitivity (what would falsify the model)

Robustness result reported: removing 20–80% of genes shows accuracy decreases; removing up to 40% yields a ~12% drop (from 99% to 87% median), and removing 80% drives accuracy toward random (~50%).

Single-gene ablation reported: 160/162 genes show no observable change in accuracy on test datasets; two genes (SAA1 and FGB) are particularly sensitive on the Lambrechts dataset.

5) CNV proofreading effect (visual)

The manuscript reports that adding the CNV “proofreading” step decreases the false positive rate from ~4% to ~1% while maintaining the average balanced accuracy.

6) What’s biologically plausible vs. what remains uncertain

Plausible (mechanistic intuition constrained by data)

Low co-expression modularity suggests the signature may be capturing distributed tumor programs (or tumor-vs-normal transcriptional shifts) rather than a single pathway. However, “low correlation” does not establish causality or specific biology.
Cell-cycle enrichment is directly supported by the reported 34-gene module and GO analysis centered on cell-cycle/mitosis terms; but the paper also reports that removing that module did not materially change classification accuracy, which argues the signature is not only “cell cycle.”

Uncertain / could be dataset-label artifacts

Reliance on expert labels: signature derivation and evaluation depend on the availability and accuracy of expert cell-type annotations in the training/test datasets; misannotation could propagate into gene signatures and inflate apparent generalization. The authors themselves note that occasional erroneous labeling can contribute to observed false positives, and they also use a healthy blood test as a proxy to estimate false-positive behavior.
Domain shift outside tested modalities: their strongest evidence is in epithelial tumors and neuroblastoma; reduced performance is reported for synovial sarcoma. That implies the approach may be sensitive to tumor biology and/or annotation differences that are not captured by the current signature.
CNV proofreading is not universally guaranteed: CNV-based discrimination is reported to not generalize well across cancer types in their random-forest experiments, yet CNV helps as a proofreading step. This pattern could reflect dataset composition and how pseudo-label flipping interacts with label noise.
AUCell/ssGSEA scoring assumptions: gene-set scoring depends on having sufficient gene detection coverage and on preprocessing choices; the authors explicitly discuss that AUCell can behave oddly in some bulk sequencing contexts and suggest alternative scoring algorithms.

7) Counterfactuals: what would most strongly falsify the central thesis?

Out-of-sample failure across independently labeled cohorts: If ikarus trained/tested as described does not separate tumor vs normal in independently collected data with different annotation protocols and at least comparable label quality, the claimed generalization is weakened. (This is implied by the pipeline’s dependence on annotation and signature derivation.)
Signature contamination hypothesis: If the tumor gene set primarily captures dataset-specific technical artifacts (e.g., batch/processing/dissociation effects) rather than tumor biology, performance will likely deteriorate when technical variables change sharply. The paper discusses dissociation artifacts and distribution-shift issues in the introduction, and uses gene-set scoring designed to be more rank-based.
CNV proofreading miscalibration: If CNV proofreading reduces false positives only in the specific evaluation regime but increases false negatives or degrades performance on other cancers/modalities, then the claimed improvement would not be reliable. The authors report a tradeoff as they flip discordant labels.

8) Reproducibility & implementation notes (what you can try to verify)

The paper states the ikarus code is available as a Python package on GitHub and provides a separate repository for reproducing figures.
They use dataset-specific preprocessing expectations (normalized to total reads and log2-transformed with pseudocount; avoid scaling; require complete gene sets and AUCell gene coverage). This is critical for faithful reproduction.
Bias check: comparisons to other methods are reportedly done with default hyperparameters/tutorial settings and without hyperparameter optimization, which can under- or over-estimate relative performance depending on method sensitivity.

Author reviews (bespoke)

Feedback:

Updated: April 28, 2026