Locat: Joint enrichment and depletion testing identifies localized marker genes in single-cell transcriptomics
Lewis, W. R., Aizenbud, Y., Strino, F., Kluger, Y., Parisi, F.
bioinformatics
Quick verdict
Locat joint-tests expression concentration and depletion for specific cell markers; sound statistics and public code, but validation metrics are self-defined.
Summary
The paper introduces Locat, a statistical framework for marker gene discovery in single-cell transcriptomics that jointly tests two properties: expression concentration within compact regions of a cell embedding and significant depletion elsewhere. For each gene, weighted Gaussian mixture models estimate gene-specific and background densities; concentration is assessed against prevalence-matched randomizations, depletion via a beta-binomial tail test with effective sample size corrections, and both are combined into a localization p-value using a Cauchy combination test with sparsity and sensitivity penalties. On synthetic benchmarks, Locat stays sensitive for genes expressed in 1.6-16% of cells while the likelihood-ratio test loses power early and a permutation variant becomes anti-conservative. Compared with six published methods plus a spectral baseline, Locat's top-100 genes show the highest cell-type specificity in three of four datasets. Applications to murine embryonic dermis (430 of 11,700 localized genes), IFN-beta-stimulated PBMCs (454 stimulated-specific and 339 control-specific differential localization), and a retinoic-acid ESC time course (2,431 genes across days 0/2/4/10) recover known biology. Main limitations: the specificity benchmark rewards exactly the property Locat optimizes, the composite p-value's calibration is empirical rather than analytic, and results depend on the user-supplied embedding.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A marker gene requires joint evidence of concentration within compact embedding regions and significant depletion elsewhere; enrichment-only scoring conflates diffuse and specific expression patterns.
- Locat maintains sensitivity for genes expressed in as little as ~1.6% of cells while correctly losing significance when expression becomes indistinguishable from background structure, unlike the likelihood-ratio test (early power loss) and a permutation variant (anti-conservative).
- Locat's top-100 genes show higher cell-type specificity than six published methods plus a spectral baseline in three of four benchmarked datasets.
- Per-sample localization analysis without batch correction, followed by localized-gene co-embedding, preserves stimulus-specific programs (e.g., IFN-responsive monocytes) that Harmony-based integration obscures.
- In the retinoic-acid ESC time course, localized gene sets turn over substantially between day 4 and day 10 (529 shared of ~1,300-1,400 per timepoint), marking a developmental watershed.
Strengths
Conceptually sharp dual criterion with concrete motivating examples
Section 2.1/Fig. 1 distinguishes concentrated-only genes (H19, Nnat; localization p = 0.089 and 0.082) from concentrated-and-depleted markers (Sox2, Hist1h2bb; localization p = 9.89e-3 and 2.48e-2), making the depletion argument falsifiable rather than rhetorical.
Statistical engineering addresses real confounds
Prevalence-matched null calibration (Eq. 7), Kish effective sample size with 0.6 scaling factor and sqrt(n) cap, beta-binomial overdispersion rho=0.02, and Sidak adjustment across the lambda scan (Methods 4.1.3-4.1.4); parameters held fixed across all datasets (Supp. Table S1).
Self-falsifying simulation design
Actionable multi-sample workflow
Strong reproducibility infrastructure
Weaknesses
Primary benchmark is self-referential
Composite p-value calibration is empirical, not analytic
Biological validation is narrative
Baseline fairness partially open
Embedding dependence acknowledged but narrowly tested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Locat (bioRxiv, 10.64898/2026.04.03.716370) is a methods paper for single-cell RNA-seq marker discovery. Its central claim: enrichment-only scoring conflates a local expression peak with a specific marker; a true localized marker must additionally be significantly depleted from the rest of the cellular manifold. The paper formalizes this as a dual hypothesis test, benchmarks against six published tools (LMD, singleCellHaystack, Hotspot, GSPA, GiniClust, Scanpy rank_genes_groups) plus an in-house spectral baseline (SpectralRH), and applies the method to murine embryonic dermis (GSE255534), IFN-beta-stimulated PBMCs (GSE96583), and a retinoic-acid ESC differentiation time course (GSE227320). Claims are mostly calibrated to the evidence; the abstract even hedges the embedding-preservation result with 'in several cases.'
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the Locat method do in single-cell RNA-seq analysis?
Locat identifies marker genes whose expression is both concentrated within a compact region of a cell embedding (such as PCA space) and significantly depleted elsewhere. It fits weighted Gaussian mixture models to gene-specific and background cell densities, tests concentration against prevalence-matched randomizations and depletion with a beta-binomial tail test, and combines both into a single localization p-value per gene.
How is Locat different from Hotspot or singleCellHaystack?
Hotspot and singleCellHaystack score genes for non-random, locally structured expression (graph autocorrelation or density deviation) but do not explicitly test whether expression is absent outside the enriched region. Locat adds this depletion requirement, which filters out broadly expressed genes that merely peak locally; in the paper's benchmarks its top-100 genes show higher cell-type specificity in three of four datasets.
Is Locat's code and data publicly available?
Yes. The package is on PyPI as locat (version 0.1.5 was used in the paper), with code at github.com/KlugerLab/Locat and analysis notebooks at github.com/KlugerLab/Locat-paper. All datasets are public, including GEO accessions GSE255534 (murine dermis), GSE96583 (IFN-beta PBMCs), and GSE227320 (retinoic-acid ESC time course).
Why do marker genes need to be depleted outside their expression region?
A gene can peak locally yet remain detectably expressed across the whole tissue or manifold; such genes correlate with structure but do not cleanly identify a cell population. Requiring statistical evidence of absence from regions where many other cells reside yields markers whose presence and absence are both informative, which the authors argue is the practical definition of a specific marker.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extraction is truncated mid-paper: part of Section 2.2 (first synthetic scenario) and the middle of Section 2.6 (trajectory-gene analysis) are missing; Figure 3 details are absent.
- Figures could not be inspected; figure-level statements (Fig. 2 confidence intervals, Fig. 4 embeddings) rely on captions and in-text numbers only.
- Supplementary methods and tables (S1.1-S1.8, Tables S1-S5) are referenced but not included, so some parameterizations (prevalence interpolation, penalty coefficients, stability tables) could not be verified directly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.