Research Radar — 2026-10-24
Methods & AI
Computational
Benchmarking segmentation methods for spatial transcriptomics
Nature Methods Published 2026-10-23 Peer-reviewed methods research and benchmarking DOI: 10.1038/s41592-026-03272-x
spatial transcriptomics segmentation benchmarking
Summary: This Nature Methods analysis benchmarks 11 segmentation approaches across 15 spatial-transcriptomics datasets, five technologies and 12 human or mouse tissue types, with biological quality metrics and a reusable Nextflow workflow.
Why it matters: Directly actionable for spatial analysis: segmentation changes transcript purity, cell identity and ligand-receptor inference. The study emphasizes trade-offs rather than a universally best method. Author repository documentation describes an earlier smaller benchmark; use published-paper counts rather than mixing versions.
Why for Yiru: Useful for checking measurement and inference assumptions in single-cell and spatial omics analyses.
SpacePath enables pathway and molecular-family analysis beyond annotation ambiguity in spatial metabolomics and lipidomics
bioRxiv Subject Collection: Bioinformatics Published 2026-10-22 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.16.759546
spatial metabolomics annotation uncertainty pathway analysis
Summary: SpacePath retains ambiguous ion-to-compound mappings while connecting spatial metabolomics and lipidomics signals to pathways and molecular families. The authors report broader feature assignment across public datasets and a glioblastoma xenograft example.
Why it matters: A useful alternative to discarding ambiguous mass-spectrometry features. Greater assignment coverage does not mean greater chemical-identification accuracy; only a small subset received tandem-MS confirmation. Xenograft associations remain exploratory.
Why for Yiru: Useful for checking measurement and inference assumptions in single-cell and spatial omics analyses.
Decoding Cell-level Circadian Phase and Rhythm-state Remodeling in Single-cell Transcriptomes With scRhythmBayes
bioRxiv Subject Collection: Bioinformatics Published 2026-10-22 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.18.759931
single-cell Bayesian inference circadian dynamics
Summary: scRhythmBayes combines seed-gene likelihoods and circular priors to infer cell-level circadian phase and posterior rhythm-state changes.
Why it matters: Adds uncertainty-aware temporal structure to heterogeneous single-cell analyses beyond static cell markers.
Why for Yiru: Relevant to perturbation, uncertainty and interpretable single-cell inference; Phase is latent rather than observed; diseased-tissue phase differences can reflect composition or sampling, and stress-test boundaries matter.
LoCoFlow: Local-Context Flow Matching for Single-Cell Dynamics and Fate Inference
bioRxiv Subject Collection: Bioinformatics Published 2026-10-21 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.15.758594
single-cell dynamics flow matching
Summary: LoCoFlow uses local patches, hierarchical transport and stochastic conditional flows to infer alternative cell-fate trajectories from snapshots.
Why it matters: Explicitly models local geometry and fate uncertainty, with lineage-traced evaluation rather than endpoint distance alone.
Why for Yiru: Relevant to perturbation, uncertainty and interpretable single-cell inference; Lineage-traced benchmarks support prediction but do not validate counterfactual expression edits; local geometry and temporal sampling remain assumptions.
Optimal transport reveals unresolved cells in single cell multi-omics
bioRxiv Subject Collection: Bioinformatics Published 2026-10-20 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.15.759426
single-cell multi-omics optimal transport uncertainty
Summary: A soft optimal-transport coupling makes uncertain cross-modality cell correspondences explicit. Across 78 dataset pairs, the authors report improved label transfer and use ambiguity scores to flag unreliable matches.
Why it matters: Important uncertainty-aware complement to integration benchmarks. Gains in integration mixing alone do not establish biological preservation. Flagged cells include rare and transitional populations, so automatic exclusion can bias biological conclusions; examine them rather than treating ambiguity as technical failure.
Why for Yiru: Useful for checking measurement and inference assumptions in single-cell and spatial omics analyses.
Auditing Confidence-Calibrated Benchmarking of Single-Cell Perturbation Models
bioRxiv Subject Collection: Bioinformatics Published 2026-10-19 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.12.758706
perturbation benchmarking uncertainty
Summary: An audit of a constructed perturbation benchmark uncovers target-cell reconstruction leakage, split failures and disagreement between coverage and useful uncertainty ranking.
Why it matters: Provides actionable null checks and separates marginal calibration from decision-relevant performance.
Why for Yiru: Relevant to perturbation, uncertainty and interpretable single-cell inference; The authors audit their own harness; failure modes are not evidence that all original scVI or CPA analyses are invalid.
U-Chrom: a unified framework for single-cell and spatial 3D genome data
bioRxiv Subject Collection: Bioinformatics Published 2026-10-19 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.18.759996
3D genomics single-cell spatial omics
Summary: U-Chrom provides a common cell-and-genomic-bin data model for single-cell Hi-C, spatial Hi-C and chromatin tracing, alongside analysis tools and a 22-dataset atlas.
Why it matters: Practical infrastructure for comparing three-dimensional genome assays. Shared indexing does not turn contacts into directly measured coordinates or create a joint probabilistic model; reconstructed structures remain inference. The atlas explicitly documents missing contact maps and compartment tracks in some datasets.
Why for Yiru: Useful for checking measurement and inference assumptions in single-cell and spatial omics analyses.
DCAF: Unifying Data Validity, Evaluation Validity and Predictive Modeling for Single-cell Perturbation Simulation
bioRxiv Subject Collection: Bioinformatics Published 2026-10-19 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.13.758713
single-cell perturbation optimal transport
Summary: DCAF combines OT-derived cell couplings, response-focused evaluation and a predictive model for single-cell perturbation simulation.
Why it matters: Addresses data assumptions, metrics and models together rather than reporting only aggregate accuracy.
Why for Yiru: Relevant to perturbation, uncertainty and interpretable single-cell inference; Optimal transport supplies inferred couplings, not observed pre/post measurements of the same cell; benchmark and biological claims remain preprint evidence.
ALBIS: simulation of multi-resolution and multi-dimensional spatial transcriptomics data
bioRxiv Subject Collection: Bioinformatics Published 2026-10-13 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.06.757026
spatial transcriptomics simulation benchmarking
Summary: ALBIS simulates a three-dimensional tissue with transcript locations, then sections and aggregates the same realization into cell-, bin- and spot-resolution observations with known simulation labels.
Why it matters: Useful controlled comparisons across resolution and slice alignment. Simulation truth is not biological truth: current software uses spherical tissues with central and wedge-shaped domains, so apparent robustness may not transfer to irregular real tissues.
Why for Yiru: Useful for checking measurement and inference assumptions in single-cell and spatial omics analyses.
Perturb-Seqr: Signature Search Engine Integrating Connectivity Mapping Resources
bioRxiv Subject Collection: Bioinformatics Published 2026-10-11 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.06.757137
perturbation resource single-cell
Summary: Perturb-Seqr integrates 16 drug/gene connectivity maps into a searchable compendium of expression signatures, supporting mimicker and reverser queries.
Why it matters: A directly usable cross-resource entry point for perturbation hypotheses with context-specific metadata.
Why for Yiru: Relevant to perturbation, uncertainty and interpretable single-cell inference; Cancer-heavy cell lines, varying doses and exposure times, platform-specific backgrounds and gene-set truncation limit transportability; signature reversal is not treatment efficacy.
Biomedical discoveries
Biomedicine
High-throughput droplet-based enzymatic DNA methylome profiling enables single-cell resolution of epigenetic heterogeneity | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-10-23 Peer-reviewed research DOI: 10.1126/sciadv.aef2844
single-cell methylomics epigenetics experimental methods
Summary: scEM-seq couples droplet barcode ligation to post-emulsion enzymatic conversion to profile single-cell methylomes at scale, with immune-cell and prostate-tumor demonstrations.
Why it matters: Extends scalable single-cell measurement beyond RNA and accessibility to DNA methylation.
Why for Yiru: Useful for designing multimodal state measurements and evaluating the coverage-throughput trade-off.
PhagoID: Quantitative in situ proximity labeling defines the proteomes of bead- and microbe-containing phagosome lumina across species, maturation, and immune activation | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-10-23 Journal research with related preprint evidence DOI: 10.1126/sciadv.aee3876
spatial proteomics proximity labeling CRISPR screens
Summary: PhagoID profiles the phagosome lumen by proximity-labeling proteomics and integrates the measurements with 25 genome-scale CRISPR screens of phagosomal phenotypes.
Why it matters: Connects subcellular protein localization with uptake, acidification and antimicrobial phenotypes.
Why for Yiru: A useful measurement-plus-perturbation design for separating molecular presence from functional contribution.
Early hematopoietic fate decisions are independent of bone marrow niche
bioRxiv Subject Collection: Immunology Published 2026-10-23 Preprint, not peer reviewed DOI: 10.64898/2026.10.20.760390
spatial transcriptomics hematopoiesis cell fate
Summary: MERFISH in young and aged mouse femoral marrow finds anatomically distinct niches but little association between niche identity and early stem-cell lineage priming.
Why it matters: Challenges the assumption that a spatially distinct niche necessarily specifies a distinct early fate.
Why for Yiru: Directly relevant to interpreting spatial neighborhoods and avoiding causal claims from colocalization.
Spatial confinement promotes germinal center diversity and tempers immunodominance
Nature Immunology Published 2026-10-21 Peer-reviewed research DOI: 10.1038/s41590-026-02698-x
spatial immunology clonal diversity germinal centers
Summary: Reducing germinal-center confinement in mice increases exchange of B-cell clones between centers and favors dominant clones at the expense of repertoire diversity.
Why it matters: Connects tissue-scale compartmentalization with the diversity of an evolving cellular population.
Why for Yiru: Combines spatial BCR sequencing, lineage tracing and perturbation to test a spatial mechanism rather than only describe neighborhoods.
A single-cell atlas of human neutrophils reveals neutrophil heterogeneity across health and disease
Nature Immunology Published 2026-10-21 Peer-reviewed research DOI: 10.1038/s41590-026-02671-8
single-cell atlas neutrophils cross-study integration
Summary: SCAHN integrates approximately one million neutrophils across 781 samples and 43 datasets, then relates neutrophil-state signatures to independent bulk cohorts.
Why it matters: Provides a harmonized reference for a technically difficult and often underrepresented immune lineage.
Why for Yiru: Relevant to annotation transfer, atlas integration and separating technical from disease-associated cell states.
Behavior-coupled single-cell transcriptomics reveals trafficking protein-dependent regulation of tumor cell migration
bioRxiv Subject Collection: Cancer Biology Published 2026-10-16 Preprint, not peer reviewed DOI: 10.64898/2026.10.15.759353
behavior-coupled sequencing live-cell imaging perturbation
Summary: BC-seq links live-cell migration behavior to single-cell transcriptomes in HPV-negative oral cancer, nominating cholesterol-trafficking mechanisms tested by migration perturbations.
Why it matters: Measures a dynamic phenotype before sequencing rather than inferring behavior from a static expression state.
Why for Yiru: A useful design for connecting single-cell representations with experimentally observed behavior and perturbation.
Cross-disciplinary watchlist
Other Fields
Efficient equivariant framework for universal charge density prediction
Nature Machine Intelligence Published 2026-10-23 Peer-reviewed AI-for-science research; journal version of a 2025 preprint DOI: 10.1038/s42256-026-01326-5
AI for science equivariant graph networks electronic structure
Summary: EdenGNN predicts charge density and PAW augmentation occupancies using equivariant graph networks, including a universal model for cross-element non-magnetic materials. The new journal publication has a linked 2025 preprint.
Why it matters: The surrogate predicts quantities required by downstream electronic-structure calculations rather than an isolated convenient learning target.
Why for Yiru: A cross-field AI-for-science example of structural constraints and downstream physical validation, not a validated biological application.
Methodology Review: Model-X knockoffs: choosing variables by competing against synthetic decoys
English · About 10 minutes
All four author manuscripts inspected in full-text form. Journal dates and preprint identities linked. Approximate learned knockoffs are not represented as exact finite-sample guarantees; biological applications are association analyses, not causal or clinical validation.
The intuition
Suppose a predictive model selects twenty measurements from thousands. Good prediction does not tell you how many selected measurements are genuinely informative after accounting for the others. Correlated proxies may trade places, and a flexible model can exploit chance patterns. The abstract problem is conditional variable selection with an error budget: return a set of features while controlling the expected fraction of false discoveries. That is different from ranking features by predictive importance, and different from proving that changing a selected feature changes the outcome.
Knockoffs turn this into a competition. Give each real feature a synthetic counterpart that resembles its dependence on the other features but carries no additional information about the outcome once the real features are known. Fit the real and synthetic variables together, using a procedure that treats each pair symmetrically. A real feature must beat its own decoy. Decoy victories provide an internal reference for chance victories by null features. The synthetic data are carefully coupled controls, not independently shuffled columns. [R1]
The method does not require the predictive model to be scientifically correct. Its exact Model-X guarantee instead requires a correct joint distribution for the features and a valid knockoff construction. This moves the modelling burden rather than removing it. The distinction matters especially when the response is complex but the feature-generating process is comparatively tractable. [R1]
A small worked example
Here is a threshold calculation, not a fitted biological analysis. Let seven feature pairs have signed competition scores \(W=(6,5,4,3,2,-1,0)\), where the entry at position \(j\) is \(W_j\). Positive means the original wins; negative means the decoy wins. Set the target false discovery rate to \(q=0.2\). For any positive candidate threshold \(t\), count original wins at least as large as \(t\) and decoy wins at least as large in magnitude. Ignore the zero score.
At \(t=1\) there are five original wins and one decoy win. The knockoff-plus estimate is \((1+1)/5=0.4\), above the target. At \(t=2\) there are still five original wins but no qualifying decoy wins, giving \((1+0)/5=0.2\). Therefore the smallest qualifying threshold is two, and features one through five are selected. At threshold three the ratio rises to \(1/4=0.25\): these ratios need not decrease as the threshold increases. [R2]
Try two perturbations by hand. Reducing the target to \(q=0.1\) selects nothing, because no candidate threshold attains that ratio. Alternatively, replacing the score minus one with minus two makes the ratio at threshold two equal to 0.4; higher thresholds also fail at the original target. Neither calculation reveals the actual number of false discoveries, because the toy supplies no true null labels. The numerator is a calibration device, not a measurement of which selected features are false.
The necessary mathematics
Let \(X=(X_1,\ldots,X_p)\) be a random vector of \(p\) measured features and \(Y\) the response; the dataset contains \(n\) independent observations of this pair. Index \(j\) identifies a feature. The null hypothesis for feature \(j\) is conditional independence of \(Y\) and \(X_j\) given \(X_{-j}\), where \(X_{-j}\) denotes all other measured features. A knockoff vector \(\tilde X\) must be independent of \(Y\) conditional on \(X\), and exchanging any subset of original–knockoff pairs must leave their joint distribution unchanged. [R1]
Exchanging an original–decoy pair must reverse its signed score \(W_j\) while leaving the other pairs’ scores unchanged. For example, use \(W_j=Z_j-\tilde Z_j\), where \(Z_j\) and \(\tilde Z_j\) are importance scores from a joint fitting procedure that respects those swaps. Arbitrary importance routines do not automatically satisfy this property. Define \(N_+(t)\) as the number of scores at least \(t\), and \(N_-(t)\) as the number at most \(-t\). Check only positive, observed absolute score values. [R1] [R2]
The knockoff-plus rule chooses the smallest checked \(t\) satisfying the following inequality; if none qualifies, select nothing. Otherwise select every feature with \(W_j\geq t\). The extra one is essential for the usual FDR guarantee, not an optional numerical stabilizer.
\[\frac{1+N_-(t)}{\max(N_+(t),1)}\leq q\]Write \(R\) for the number selected and \(V\) for the number of selected true nulls. Under the theorem assumptions, \(\mathrm{FDR}=\mathbb{E}[V/\max(R,1)]\leq q\), where \(\mathbb{E}\) means expectation over the data and procedure randomness. This is not a per-feature posterior probability or a bound on every realized dataset. [R1] [R2]
Assumptions and failure modes
The hardest assumption is often feature-distribution knowledge. A generator fitted from finite data is generally approximate. Robustness theory links FDR inflation to discrepancies in relevant conditional feature distributions; a visually convincing synthetic dataset or a small generic training loss is not an exact validity certificate. Deep Knockoffs supplies a practical learned construction and empirical diagnostics, but the word deep does not replace the exchangeability condition. Keep the ideal theorem, the implemented generator and observed diagnostics as three separate claims. [R3] [R4]
Conditional importance can be scientifically narrower than expected. Two nearly interchangeable predictors can each add little once the other is present. Weak discoveries may therefore reflect redundancy, not biological irrelevance. Conversely, selecting a feature does not remove confounding by unmeasured causes. A large correlated set can make both generation and discrimination difficult. Before interpreting names on a selected list, specify the measured conditioning set and what omissions or measurement error might change that interpretation.
The independent observational unit also needs an explicit definition. Many cells from one donor are not automatically independent donors. This is a design caution for any proposed omics implementation, not a guarantee furnished by the cited papers. Likewise, response-guided filtering or repeated reruns followed by choosing the most attractive list can change the procedure being calibrated. Freeze the analysis plan, and do not mistake selection stability across random seeds for proof of FDR control.
Verified applications
The Model-X paper studies genetic association using a Crohn disease case-control dataset. This establishes a concrete application to correlated genetic predictors, rather than a biological connection invented from the mathematics. Read its empirical results alongside its feature-distribution construction and modelling choices; the theorem is conditional on those assumptions. The selected associations are not experimental evidence that intervening on each variant would alter disease risk. [R1]
Deep Knockoffs examines HIV mutation features and drug resistance. Its paper distinguishes experiments with simulated responses, where false discoveries can be counted, from analysis of observed resistance measurements. The latter reports selected mutations and repeated knockoff realizations, but does not provide complete ground-truth null labels. Those are different evidence types: benchmark calibration under a simulation cannot by itself establish clinical validity or exact real-data FDR. [R4]
Possible transfers — speculative
A possible omics transfer is donor-level selection of predefined pathway summaries associated with a response after conditioning on the other summaries. This is a proposed experiment, not an application validated by these four papers. Start with a small, auditable feature distribution, hold the response out of generator construction, and use independent units. Test the pipeline on simulated null and signal settings resembling the study, then report sensitivity to plausible feature-distribution misspecification rather than announcing theorem-level protection. An initial feasibility study may correctly conclude that the feature model is not trustworthy enough.
What to try and what to read next
Implement the seven-score threshold before training anything. Enumerate every positive absolute score, reproduce the two perturbations, and verify that no qualifying threshold produces an empty-selection error. Then explain which assumption would justify treating decoy wins as a reference for null original wins. If that explanation is missing, a more elaborate predictor will not repair the inference. The useful mental model is controlled competition, not synthetic augmentation.
Read the Model-X construction and theorem first, then the fixed-X paper to see what changed. Use the robustness paper as an assumption audit, and Deep Knockoffs as an implementation case study. Keep three questions on separate lines in your notes: Is the generator valid? Is the score powerful? What scientific statement does the conditional null answer? A successful workflow needs all three, and prediction accuracy answers only part of the second.
Reading list
-
[R1] Primary reading: Panning for Gold: Model-X Knockoffs for High-dimensional Controlled Variable Selection. Emmanuel Candès, Yingying Fan, Lucas Janson and Jinchi Lv (2018). Main reading: sections 3.1–3.3 introduce the conditional null, exchangeability and the knockoff-plus theorem; the genetics case study shows the intended inferential target. Journal version DOI 10.1111/rssb.12265. First arXiv version 2016-10-07.
-
[R2] Related reading: Controlling the false discovery rate via knockoffs. Rina Foygel Barber and Emmanuel J. Candès (2015). Foundation: compare fixed-design Gaussian regression with Model-X, and read the distinction between the original and plus thresholds. Journal version DOI 10.1214/15-aos1337. First arXiv version 2014-04-22.
-
[R3] Related reading: Robust inference with knockoffs. Rina Foygel Barber, Emmanuel J. Candès and Richard J. Samworth (2020). Assumption audit: the robustness bounds explain why estimating the feature distribution changes the guarantee. Journal version DOI 10.1214/19-aos1852. First arXiv version 2018-01-11.
-
[R4] Related reading: Deep Knockoffs. Yaniv Romano, Matteo Sesia and Emmanuel J. Candès (2019). Implementation and application: inspect the learned generator and HIV drug-resistance analysis, separating simulations from observational results. Journal version DOI 10.1080/01621459.2019.1660174. First arXiv version 2018-11-16. First journal online publication is 2019-10-17; the volume 115 issue is dated 2020.