The reproducibility stack is auditing itself: QC order, perturbation benchmarks, and provenance-tracked design

What changed

Three papers posted to bioRxiv's research-methods-reproducibility feed this week share a single target: the scientific pipeline itself. One reports that the order of quality-control steps in single-cell RNA sequencing can hide the very artifacts QC is meant to remove. Another introduces a benchmark that grades computational models of cellular perturbation on both how much RNA they predict and where it sits in the cell. A third offers an open platform for designing structural colours that tracks the provenance of its training data end to end. None of the three is peer-reviewed — all are preprints — but together they read like a methods section for methods sections: the reproducibility stack learning to check itself. m:…1782 m:…8669 m:…9903

The common thread is infrastructure honesty. Each paper builds something that audits something else in the stack: an ordering comparison that exposes a failure mode in doublet detection, a benchmark that exposes weaknesses in perturbation models, and a design platform whose dataset carries its own lineage. For professionals downstream, the practical point is that these are the quiet papers — the ones that change how tools get run, graded, and trusted. m:…1782 m:…8669 m:…9903

Doublet detection, out of order

In single-cell RNA sequencing, quality control typically removes two kinds of noise: ambient RNA, free-floating transcripts that contaminate each droplet, and doublets, droplets that accidentally captured two cells instead of one. A preprint from Meier and colleagues benchmarks the order in which those two corrections get applied and reports that the standard sequence, ambient RNA correction first, confounds the doublet detection that follows, reducing the share of doublets detected. m:…1782

The mechanism the authors describe is straightforward: decontX and soupX, two ambient RNA correction tools, remove proportionally more ambient RNA from doublets, which masks the doublet signal — so doublets that remain undetected stay in the dataset. Their recommendation is blunt: run doublet detection before ambient RNA correction, and re-examine existing pipelines. CellBender, they note, is the exception to that rule. m:…1782

A report card for virtual cells

Predicting how a cell responds to a perturbation, a drug or a knockout, is moving from counting transcripts to modeling where RNA actually sits inside the cell. In a preprint, the authors introduce SubcellPert as a benchmark for perturbation-response prediction at single-RNA-molecule resolution, built to model how perturbations affect both RNA abundance and intracellular spatial organization. m:…8669

The benchmark's early results are unflattering for current approaches. Gaussian transport achieves the lowest Oracle spatial error on 8 of 11 cellular-context out-of-distribution datasets, which is good on position. But only 19 of 190 Joint adapter fits retain source counts after validation — a count-fidelity failure rate that matters, because models that predict positions while losing counts are grading well on the wrong test. m:…8669

Provenance as a feature

The third preprint, from M. S. Morioka, describes Kawasemi-Colours, an open end-to-end platform for structural-colour design inspired by avian plumage. It pairs a provenance-tracked dataset of keratin-air nanostructures — spongy, channel, and quasi-ordered sphere geometries with their associated colour metadata — with interpretable machine learning and wave optics. The platform supports forward colour prediction and inverse design: give the analytic inverse design tool a target colour and it generates candidate nanostructures. m:…9903

The notable part, from a reproducibility standpoint, is that provenance is built into the platform rather than documented afterwards. In a week when other papers are exposing gaps in QC ordering and model evaluation, a dataset that ships with its own lineage reads as the constructive version of the same instinct: don't just trust the pipeline, instrument it. m:…9903 m:…1782 m:…8669