Paper Review — Verify Claims Fast

Concise critique

The review identifies three primary transcriptomic readouts to detect malignant cells in single-cell RNA sequencing data: cell-of-origin marker expression, inter-patient tumor heterogeneity, and inferred copy-number alterations, and argues for integrative, multi-feature classification rather than any single criterion alonePrimary review conclusion

Detailed review and critique of Identification of malignant cells in single-cell transcriptomics data

What the review covers

• Catalog of transcriptomic readouts usable with scRNA-seq (COO markers, CNAs, SNAs, fusions, proliferation, metabolic and immune-evasion signatures) with practical caveats and examplesFeature catalog
• Technical description and comparative discussion of CNA inference tools (InferCNV, CopyKAT, SCEVAN, CaSpER, Numbat) and their reference requirements and limitsCNA tools overview
• Survey of machine-learning pipelines (van Galen random forest approach, scATOMIC, Ikarus, scMalignantFinder) including strengths and dependence on labelled training dataML methods

Strengths

• Comprehensive synthesis of measurable transcriptomic features relevant to malignancy and how each can be used in practiceComprehensive scope
• Balanced discussion of technical limitations (dropout, reference selection for CNA, low coverage for SNA/fusion calling) with practical recommendationsTechnical caveats
• Useful, actionable guidance: combine COO markers, interpatient heterogeneity metrics, CNAs and secondary features (proliferation, immune evasion, fusions) rather than rely on a single signalRecommendation to integrate features

Limitations and blindspots

1. Transcriptome only: the review acknowledges that RNA readouts may be insufficient in cancers driven mainly by non-CNA mechanisms (gene translocations, point mutations, epigenetics) and recommends multiomic integrationNeed for multiomics
2. Dependence on good references and controls: CNA callers and signature scoring require appropriate reference diploid cells or matched normal tissue; absent these, false positives and false negatives increaseReference dependence
3. Benchmarking and generalization gaps: the review notes that ML tools require well-annotated, cancer-type representative training sets; performance across low-CNA cancers or lineage plasticity contexts remains uncertainML training data caveat

Concrete recommendations for practice

• Start by identifying the cell-of-origin compartment with canonical markers, then subset and run CNA inference vs confident diploid references; cross-check with inter-patient cluster mixing metrics (LISI/entropy) and proliferation/fusion/SNA evidence where availablePractical pipeline
• When possible, combine orthogonal modalities (matched WES/WGS for CNAs, high coverage scRNA or Smart-seq protocols for SNAs/fusions, spatial transcriptomics for context) to raise confidenceMulti-modal validation
• Use cluster-level decisions rather than per-cell binary calls whenever data sparsity or dropout is substantial; CNAs and many signatures are more robust at the cluster levelCluster-level inference

Critical appraisal and open questions

The review is timely and useful, but a few additions would increase utility for practitioners:

• Quantitative benchmarking table: readers would benefit from a standardized performance table (sensitivity/specificity/precision) of CNA callers and ML classifiers across representative tumor types and sequencing platforms (10X vs Smart-seq). The review summarizes tool behavior but does not provide a cross-tool numeric benchmarkNeed for benchmarks
• Guidance on minimal experimental design: explicit rules of thumb (minimum number of diploid reference cells, recommended sequencing depth for reliable SNA/fusion calling) would make the recommendations more prescriptive; the review notes limits but does not quantify thresholdsDesign thresholds missing
• Attention to tumor microenvironment induced transcriptional convergence: the review mentions metabolic and hypoxia-driven changes that can blur differences between malignant and neighboring normal cells, but more discussion on how to disentangle microenvironmental from malignant-intrinsic signatures would be helpfulMicroenvironment confounding

Specific edits that would strengthen the manuscript

1. Add a short benchmarking matrix (tool vs feature vs platform) with suggested default parameters and typical failure modesTool guidance
2. Provide a checklist for experimentalists (sample types, sequencing protocol, minimum controls) and a decision tree for analysts to choose between signature, CNA, SNA, or integrated approaches depending on tumor contextContext dependent decisions

If you want, I can: generate a reproducible analysis notebook implementing the recommended pipeline (Scanpy + CopyKAT/Numbat + logistic classifier + per-cluster consensus), produce a benchmarking matrix template, or draft parameter recommendations for popular platforms such as 10X and Smart-seq.