Research Radar — 2026-11-17

Generated 2026-11-17T05:53:18.123084+08:00 approved RSS sources Curated daily research digest; 19/20 primary RSS feeds and 8/8 PubMed-indexed journal feeds available. Indexing may lag publication; coverage is not exhaustive.

Methods & AI

Computational

2 selected
Computational #1 Use the decision tree to formulate the intended biological question. This is a methods review, not evidence that one algorithm universally wins.

Principles and assumptions underlying domain detection in spatial transcriptomics

Nature Genetics Published 2026-11-16 Peer-reviewed methodological review DOI: 10.1038/s41588-026-02812-7

Authors: Pratibha Panwar; Carissa Chen; Sichang Ding; Farhan Ameen; Maya Shen; Daniel Kim; Rojashree Jayakumar; Maximilian Woollard; Jeyson Chuah; Xiaoqi Liang; Jean Yang; Ellis Patrick; Pengyi Yang; Stephanie C. Hicks; Shila Ghazanfar

spatial transcriptomics domain detection method selection

Summary: This review organizes 47 spatial-omics clustering methods by their assumptions and applicability, surveying 108 variables across 26 groups and providing a research-question-driven decision tree.

Why it matters: Different clustering objectives need not identify the same biological structure.

Why for Yiru: A useful planning reference before selecting a spatial-domain or niche-analysis pipeline.

Computational #2 Read for evaluation design, not autonomous-discovery claims. Retain expert review; this evidence does not demonstrate a general closed-loop wet-lab scientist.

An open agentic AI co-scientist for clinical and biomedical discovery

Nature Medicine Published 2026-11-16 Peer-reviewed accepted early article; linked medRxiv preprint DOI: 10.1038/s41591-026-04808-6

Authors: Kaleigh F. Roberts; Zachary B. Abrams; Luca Cappelletti; Mahdi Moqri; Nicholas Heugel; J. Harry Caufield; Mathieu Bourdenx; Yan Li; Jineta Banerjee; Luca Foschini; Diego Galeano; Nomi L. Harris; Melody Li; Kejun Ying; Justin A. Melendez; Nicolas R. Barthélemy; James G. Bollinger; Yingxin He; Vitaliy Ovod; Tammie L. S. Benzinger; Shaney Flores; Brian A. Gordon; Taylor Pedersen; Jinsheng Yu; Brendan P. Lucey; Erik S. Musiek; Sarah A. Naylor; Adegoke A. Ojewole; Mukta Phatak; Donald L. Elbert; Sarah Biber; Eric C. Landsness; Christopher J. Mungall; Randall J. Bateman; Justin T. Reese

agentic AI biomedical data analysis validation design

Summary: OpenScientist is an open-source agentic analysis system evaluated in five biomedical case studies. Reported checks include hidden-cohort survival prediction, external-cohort hypothesis testing and a blinded comparison with human analysis.

Why it matters: Data-grounded evaluation is more informative than fluent reports alone, although a small case-study set does not establish general reliability.

Why for Yiru: Relevant to auditable AI-assisted biomedical analysis and the boundary between generating hypotheses and validating them.

Biomedical discoveries

Biomedicine

1 selected
Biomedicine #1 Prioritize the assay design and off-target controls. The full protocol is not yet validated in primary cells; do not equate promoter insertion tests with complete assay validation.

Joint single-cell capture of Cas9 edits and transcriptomes reveals on- and off-target effects on gene expression

Nature Methods Published 2026-11-16 Peer-reviewed research article; earlier bioRxiv version exists DOI: 10.1038/s41592-026-03319-z

Authors: Michael H. Lorenzini; Brad Balderson; Shanna N. Lavalle; Karthyayani Sajeev; Aaron J. Ho; Madelyn Light; Olalekan H. Usman; Gavin Kurgan; Graham McVicker

CRISPR perturbation single-cell sequencing off-target effects

Summary: Superb-seq jointly measures Cas9 edits and transcriptomes using T7 in situ transcription and combinatorial single-cell RNA sequencing. The authors demonstrate on- and off-target expression effects across three cell lines.

Why it matters: Capturing the edit itself can reveal heterogeneity that a guide identifier alone misses.

Why for Yiru: Directly relevant to perturbation-to-transcriptome experimental design and interpretation.

Methodology Review: Permutation tests: what are you allowed to shuffle?

English · About 10 minutes

Theory and related methods inspected in full-text sources, with claims limited to cited sections. Toy calculations are original and checked by exhaustive enumeration; no biological dataset or software benchmark rerun.

The intuition

Imagine that an analysis reports a large difference between two groups. Before celebrating, ask a different question: how surprising would this statistic look among assignments that the null hypothesis treats as interchangeable? A permutation test constructs that comparison by rearranging observed data rather than inventing a parametric noise distribution. The hard part is not writing a shuffle command. It is deciding which rearrangements preserve the scientific null.

Hemerik and Goeman formulate validity through invariance under a finite transformation group. The group contains the unchanged data, inverses and compositions. Their results distinguish exhaustive testing from random sampling of transformations and show why arbitrary appealing subsets can fail. Exact here does not mean that every nominal significance threshold has exactly the same rejection probability: discreteness and ties can make a valid nonrandomized test conservative. [R1]

A small worked example

Here is an original four-unit example, with a deliberately strong null: all four independent measurements have the same distribution. Observed control values are 1 and 2; treatment values are 4 and 5. Use the absolute difference of group means as the statistic. The observed means are 1.5 and 4.5, so the statistic equals 3. This is an arithmetic illustration, not an experiment with four real patients.

Keep the four values fixed and list every choice of two treatment units. In order, the choices are (1, 2), (1, 4), (1, 5), (2, 4), (2, 5), and (4, 5). The complementary pair is the control group. Their absolute mean differences are 3, 1, 0, 0, 1, and 3. Two of the six equally weighted assignments are at least as extreme as the observed one, giving a two-sided permutation p-value of 2/6, or one third.

This calculation is a useful antidote to visual overconfidence. The groups are completely separated, yet there are only six possible assignments. In this particular example the strongest two-sided result is still one third. More cells measured within each of these four experimental units would not create additional independently assigned units. If you replace each value by a hundred identical copies and shuffle those copies independently, you have changed the experiment being represented, not merely improved numerical precision.

For a prespecified directional question, treatment minus control gives the six signed differences −3, −1, 0, 0, 1, and 3. Only one reaches the observed value of 3, so the corresponding one-sided result is 1/6. Choosing that direction after noticing which group is higher would not be the same procedure. Save the direction and statistic before looking at the result; the two calculations answer different questions.

Now imagine instead that the units are paired: the pairs are (1, 4) and (2, 5), with one control and one treatment measurement inside each pair. Under a null that permits swapping labels within each pair, there are four swap patterns, not six arbitrary group assignments. For the same absolute mean-difference statistic, these produce 3, 0, 0, and 3, hence 2/4. The changed answer comes from the allowed transformations. Pairing is not an optional decoration on the figure.

A final hand check: add 10 to every value. Both means shift by 10, so all differences and both reference lists remain unchanged. Multiply every value by a positive constant of 2: the statistics double, but their ordering and the exceedance counts stay unchanged. These invariances make good unit tests. They do not establish validity for a different design; they only check that this implementation matches its stated statistic.

The necessary mathematics

Let X denote the data, G a finite valid transformation group, g one transformation, and T the chosen statistic. Let M be the number of transformations in G. The exhaustive upper-tail value is \(p=\frac{1}{M}\sum_{g\in G}\mathbf{1}(T(gX)\geq T(X))\). The indicator is one when its condition holds, zero otherwise. The identity contributes at least one count. Equal-sized equivalence classes allow the six assignments above to replace all within-group reorderings. [R1]

For B independent uniform draws from the full group, with replacement, retain the original observation separately. If b sampled statistics are at least as large as the observed statistic, use \(p_{\mathrm{MC}}=(b+1)/(B+1)\). Under the required invariance and sampling conditions this gives a valid conservative p-value; it is not simply an unbiased estimate of the exhaustive value. The plus one does not repair an invalid shuffle scheme. [R1] [R2]

For example, if B = 999 and b = 0, this formula gives 0.001, not zero. Repeating the sampling can change b. State the sampling rule, number of draws, seed and tie handling so another person can reconstruct the calculation. Phipson and Smyth emphasize that understated permutation p-values can have serious consequences when many hypotheses are tested. [R2]

Assumptions and failure modes

Exchangeability is a statement about the joint distribution under the null, not merely equality of means. Winkler and colleagues explain why nuisance effects can prevent direct exchange of observations and why structured designs require more care. Their GLM framework includes approximate methods; its success in studied scenarios is not a universal proof for every residual-shuffling recipe. [R3]

Use the paired example as a diagnostic question for your own data: which entries must move together, and which must never cross a boundary? Write these restrictions before computing a result. If treatment and batch are perfectly aligned, unrestricted relabeling may create combinations the study never observed. A shuffling script cannot manufacture the missing comparison. Investigate the design rather than treating more permutations as a remedy.

The same numerical statistic can be embedded in several different tests. Specify whether the target is a difference in distributions, a directional mean contrast under stronger assumptions, or an effect justified by randomized treatment assignment. This review develops the invariance-based route. It does not claim that every randomization test must be derived from the same group argument, or that a small p-value proves a causal mechanism.

Verified applications

Winkler and colleagues provide a concrete neuroimaging application: a general-linear-model permutation framework, detailed simulations and the randomise algorithm, including examples of common imaging designs. The inspected paper discusses nuisance variables and certain dependent-data structures. This verifies a methodological application, not an independent rerun of their experiments and not validation on single-cell data. [R3]

Possible transfers — speculative

A speculative omics exercise is to compare a fixed pathway summary across independently sampled donors. Start with one summary per donor, retain design strata, and write a defensible null before deciding which donor labels can move. Reproduce the four-unit arithmetic first, then simulate data under your proposed null to inspect false positives. These simulations diagnose the implemented design; they cannot prove that real donors satisfy it.

For spatial data, a second proposed exercise is to compare unrestricted cell-label shuffling with transformations that preserve a specified spatial structure. State exactly which scientific hypothesis each procedure tests. Do not choose whichever produces the smallest p-value. If no plausible transformation preserves the nuisance structure while breaking the effect of interest, stop and reconsider the analysis. This is a suggested validation exercise, not a ready-made spatial inference method.

The deliverable from a first implementation should be an auditable list of transformations and statistics, not just a significance label. Recover all six assignments and all four paired swaps without omissions. Confirm the observed assignment is counted and ties use the stated rule. Then separate effect size, sampling uncertainty and test calibration in your report. A result becomes interpretable when you can explain both the observed contrast and the alternative arrangements against which it was judged.

Reading list

Page Last Updated: