Research Radar — 2026-10-25
Methods & AI
Computational
Probabilistic uncertainty in virtual spatial transcriptomics from histology is largely redundant with the predicted mean
bioRxiv Subject Collection: Bioinformatics Published 2026-10-24 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.23.760875
virtual spatial transcriptomics uncertainty calibration benchmarking
Summary: Across three cancer cohorts, this benchmark asks whether uncertainty in histology-predicted spatial expression identifies errors beyond what the predicted mean already explains. The authors find that calibration can look good while mean-adjusted error informativeness remains weak.
Why it matters: A directly relevant warning against equating calibrated predictive distributions with useful uncertainty for deciding which virtual measurements to trust. The benchmark covers specified likelihood heads and approximate-Bayesian approaches on Visium cancer cohorts, not every spatial predictor or tissue. Author code documents patient-held-out evaluation, but results were not independently reproduced.
Why for Yiru: Relevant to rigorous measurement, representation and inference in single-cell and spatial omics.
Overcoming segmentation bottlenecks in brain tissue for spatial proteomic inference at population scale
bioRxiv Subject Collection: Bioinformatics Published 2026-10-24 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.22.761199
spatial proteomics segmentation downstream inference
Summary: SPInfer adapts Cellpose-SAM to human prefrontal-cortex cell bodies using 5,832 annotations, with reported segmentation F1 rising from 0.37 to 0.84. In 14,765 images from 69 donors, changing the masks changed which Alzheimer-pathology associations survived correction.
Why it matters: Shows why segmentation should be evaluated through downstream population-level inference, not visual plausibility or generic benchmark performance alone. More significant associations after fine-tuning are not automatically more true; annotation selection, donor-level splits and association-model assumptions need scrutiny. The findings are observational and brain-domain-specific.
Why for Yiru: Relevant to rigorous measurement, representation and inference in single-cell and spatial omics.
Averaging in gene perturbation benchmarks hides perturbation-strength variation: Consequences for model evaluation
bioRxiv Subject Collection: Bioinformatics Published 2026-10-23 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.20.760537
computational omics model evaluation
Summary: Using X-Atlas/Orion CRISPRi data, the authors associate depth-corrected guide abundance with response reproducibility in HCT116 and HEK293T. In the evaluated HCT116 retrieval setting, STATE and Cell2Sentence predictions did not reliably retain perturbation identity, preventing a meaningful within-perturbation strength comparison.
Why it matters: A directly actionable perturbation-evaluation audit: label averages can conceal efficacy heterogeneity and failure to retain intervention identity.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Information contracts determine what inverse perturbation benchmarks measure
bioRxiv Subject Collection: Bioinformatics Published 2026-10-23 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.19.760271
computational omics model evaluation
Summary: InverseBench explicitly specifies candidate information and endpoint construction, showing that the same ridge pipeline scores much higher on a transcriptomic proxy than on independently measured K562 growth fitness. Shortlist comparisons do not show a uniform corresponding penalty, emphasizing that endpoint correlation and useful biological decisions are different evaluation targets.
Why it matters: Separates proxy prediction, response-informed prioritization and genuinely unseen intervention ranking; highly relevant to causal claims and benchmark leakage.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
GeoSinkhorn Flow: Geometry-Aware Flow Matching for Perturbation-Conditioned Dynamics in Single cell Data Phenoscapes
bioRxiv Subject Collection: Bioinformatics Published 2026-10-22 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.16.759724
computational omics model evaluation
Summary: GeoSinkhorn Flow combines heat-diffusion geometry, a geometric autoencoder and perturbation-conditioned flow matching to infer transitions between cellular distributions. The authors report held-out-condition tests across organoid drug responses, Perturb-seq and time-resolved CyTOF, with an unbalanced variant for changing abundance.
Why it matters: Geometry-aware conditional transport is directly relevant to perturbation prediction; it models distributions rather than treating condition embeddings as mechanistic trajectories.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
scGSCA: gene-set differential co-expression analysis for single-cell RNA sequencing data
bioRxiv Subject Collection: Bioinformatics Published 2026-10-12 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.05.756579
single-cell RNA-seq differential co-expression mixed models
Summary: scGSCA estimates pathway-level differential co-expression from quantile-residual correlations after fitting gene-wise zero-inflated negative-binomial mixed models. It includes subject-level random effects and an interaction test for phenotype-dependent co-expression differences between cell types.
Why it matters: A useful statistical alternative to testing pathway mean expression alone, with explicit attention to donor dependence and cellular context. Inference depends on count-model and random-effect assumptions, adequate independent subjects and residual-correlation calibration; co-expression changes are not causal pathway regulation.
Why for Yiru: Relevant to rigorous measurement, representation and inference in single-cell and spatial omics.
BiNVaST: niche boundary-preserving and synthetic data-guided variational autoencoders for concordant inference across high-resolution spatial transcriptomic platforms
bioRxiv Subject Collection: Bioinformatics Published 2026-10-21 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.15.759369
spatial transcriptomics cell annotation variational autoencoder
Summary: BiNVaST combines a dual-branch variational autoencoder with platform-adjusted synthetic training data derived from spatial profiles and an annotated single-cell reference. The authors report cell-typing and cross-platform concordance gains in six specimens spanning four tissue types and four spatial platforms.
Why it matters: Directly addresses rare or divergent cancer cell states and annotation across spatial assays while retaining niche boundaries. Reference labels, segmentation and synthetic-to-real transfer can propagate errors; six specimens do not establish universal cross-platform robustness.
Why for Yiru: Relevant to rigorous measurement, representation and inference in single-cell and spatial omics.
On technical biases in imaging spatial transcriptomics
bioRxiv Subject Collection: Bioinformatics Published 2026-10-16 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.16.759529
imaging spatial transcriptomics technical bias measurement models
Summary: This analysis organizes imaging-spatial-transcriptomics biases across sample preparation, acquisition and computational quantification. Public-data examples and a unified count-modeling framework show how multiple technical effects can compound.
Why it matters: Useful for distinguishing technical spatial structure from biology and for designing explicit measurement diagnostics before fitting spatial models. This is a methodological analysis and perspective, not a validated universal correction algorithm; bias components may not be separately identifiable from a single observed tissue.
Why for Yiru: Relevant to rigorous measurement, representation and inference in single-cell and spatial omics.
Reference-free isoform discovery links unannotated RNA processing to disease risk
bioRxiv Subject Collection: Bioinformatics Published 2026-10-23 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.21.760529
computational omics model evaluation
Summary: Torino adapts non-negative matrix factorization to learn RNA coverage patterns and quantify their usage without transcript annotations. Across species, GTEx and 1,193 brain samples, the authors identify tissue-associated patterns, coverage-pattern QTLs and disease-associated intron retention.
Why it matters: High-priority annotation-independent transcriptomics with population QTL and disease-cohort evidence; stronger relevance than generic agent wrappers.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Programmable design of functional RNAs through multimodal generative learning
bioRxiv Subject Collection: Bioinformatics Published 2026-10-15 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.14.759035
computational omics model evaluation
Summary: RNAverse combines sequence, structural-probing information and functional annotations for conditional RNA generation. The authors report experimental tests in snoRNAs, aptazymes, aptamers and untranslated regions, including cell assays and a mouse immune-response experiment.
Why it matters: Strong AI-for-science candidate because generation is connected to measured RNA functions in several experimental settings.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Biomedical discoveries
Biomedicine
Multiomic characterization of microsatellite-stable colorectal cancer reveals distinct biology and targetable vulnerabilities.
Cancer Cell (via PubMed) Published 2026-10-22 Peer-reviewed original biomedical research DOI: 10.1016/j.ccell.2026.09.014
single-cell and spatial biology mechanistic evidence
Summary: An integrated analysis combines about 2,000 colorectal tumor samples, 4,500 molecular profiles and 195,246 cells or nuclei, including newly profiled MSS disease, to resolve regulatory and metabolic programs. In RAS-mutant models, tumoroid and xenograft experiments support testing FAK, PDK1/2 or proteasome inhibition alongside MEK blockade, not established patient benefit.
Why it matters: Large multiomic integration with functional follow-through in patient-derived models, highly relevant to interpreting heterogeneous omics and drug-response hypotheses.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Single-cell profiling of physically interacting immune cells using image-guided sorting
bioRxiv Subject Collection: Immunology Published 2026-10-16 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.09.758094
single-cell and spatial biology mechanistic evidence
Summary: Image-guided sorting separates contacting T-cell–monocyte pairs from coincidental doublets in blood from tuberculosis patients before profiling individual partners. Contacting pairs show distinct metabolic and cytotoxic programs and apparent monocyte-to-T-cell RNA transfer, offering an experimentally grounded complement to inferred cell communication.
Why it matters: Direct physical-contact measurement tackles a core ambiguity in cell-cell inference and doublet filtering.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
SPEAR: Spatial Epitranscriptomic Analysis by Mass Spectrometry Reveals Localized RNA Modification Remodeling During Glioma Progression
bioRxiv Subject Collection: Cancer Biology Published 2026-10-16 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.13.758238
single-cell and spatial biology mechanistic evidence
Summary: SPEAR combines laser microdissection with LC–MS/MS to quantify dozens of RNA modifications from FFPE regions as small as 0.1 square millimetres. Mouse-brain sections and human glioma specimens reveal spatial modification patterns, including lower methylated modifications in higher-grade foci that bulk analysis obscures.
Why it matters: Adds a chemically measured spatial epitranscriptomic layer with FFPE applicability and a clear resolution caveat.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Spatial profiling of Kawasaki disease cardiac inflammation identifies STING as a therapeutic target
bioRxiv Subject Collection: Immunology Published 2026-10-22 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.16.757759
single-cell and spatial biology mechanistic evidence
Summary: A 64-plex imaging study maps immune infiltrates and stromal disruption in Kawasaki-disease autopsy coronary tissue and identifies myeloid STING expression. Genetic and pharmacological STING inhibition reduces vasculitis in a mouse model, supplying a testable mechanistic lead rather than proof of treatment efficacy in children.
Why it matters: Connects spatial pathology to perturbation evidence rather than stopping at a cell atlas.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
A multimodal precision oncology platform integrating genomics,single-cell transcriptomics, and functional drug screening expandstherapeutic options in pediatric brain tumors
bioRxiv Subject Collection: Cancer Biology Published 2026-10-14 bioRxiv preprint; not peer-reviewed DOI: 10.64898/2026.10.12.758619
single-cell and spatial biology mechanistic evidence
Summary: A platform combining whole-genome sequencing, single-cell RNA sequencing and ex vivo drug screening expands therapy nominations in 17 children with malignant CNS tumors. Functional classifications align directionally with outcomes in eight of ten evaluable cases, but combination therapies and radiation prevent attributing those outcomes to the platform.
Why it matters: Direct fit to multimodal drug-response research with appropriately restrained translational claims.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
NF-κB phosphorylation controls enhancer selection to shape the inflammatory gene program.
Immunity (via PubMed) Published 2026-10-23 Peer-reviewed original biomedical research DOI: 10.1016/j.immuni.2026.09.025
single-cell and spatial biology mechanistic evidence
Summary: A RelA S276A mutation separates primed from latent enhancer responses in macrophages: phosphorylated NF-kB is needed to recruit CBP/p300 and open otherwise occluded sites. Mutant mice resist endotoxin shock but become more susceptible to bacterial infection, showing that suppressing an inflammatory program can trade off host defense.
Why it matters: Perturbation-resolved enhancer mechanism is relevant to regulatory-network and causal omics modeling.
Why for Yiru: Useful for critically evaluating measurement, validation and biological interpretation in computational omics.
Methodology Review: D-optimal experimental design: buy complementary information, not more of the same
English · About 10 minutes
Historical thematic reading, not current news. Main and foundation inspected in full text; broader survey uses primary abstract evidence. Original worked example is illustrative; no omics implementation or empirical validation performed.
The intuition
Suppose you can afford four measurements of a system with two unknown properties. A tempting plan is to repeat the easiest measurement four times. That can make one answer precise while leaving the other entirely unknown. Experimental design asks a different question from estimation: before observing the data, which measurements would make the eventual answers informative? The budget is fixed; the object being optimized is the arrangement of information, not the amount of data already collected.
Imagine uncertainty as an ellipse drawn around two parameter estimates. Repeating the same kind of measurement compresses the ellipse mainly in one direction. A complementary measurement compresses its neglected direction. D-optimality uses the volume of this joint uncertainty region as its target. This is an intuition, not a promise that every parameter improves equally. A thin but very long ellipse can have a modest area while still leaving one scientifically important combination poorly determined. [R1] [R2]
The main reading turns this geometry into a finite sensor-selection algorithm: a convex relaxation produces both a candidate selection and a performance bound. The final discrete selection remains a heuristic; the gap need not be small. Read it as a careful separation of what an optimizer certifies from what a rounded experimental plan actually achieves. [R1]
A small worked example
Here is an original toy experiment, not an empirical omics result. Two unknown quantities are \(\theta_1\) and \(\theta_2\). A type-A measurement observes only the first, and a type-B measurement observes only the second. Each observation has independent Gaussian noise of variance one. Four observations are available, and repeated measurements are permitted. Let \(n_A\) and \(n_B\) be their respective counts, with \(n_A+n_B=4\). If each count is positive, averaging within each type gives unbiased estimates whose variances are \(1/n_A\) and \(1/n_B\).
Compare two allocations. Three A measurements and one B measurement give variances one third and one. Two of each give one half and one half. The determinant of the covariance is therefore one third for the first plan and one quarter for the balanced plan. For these diagonal matrices it is simply the product of the two variances. Their uncertainty-ellipse area ratio, balanced divided by unbalanced, is \(\sqrt{3}/2\approx0.866\). Thus balance reduces joint area by about thirteen percent, even though estimation of the first quantity becomes less precise.
Four A measurements are not merely a slightly worse plan: without a prior or another constraint, the second quantity is unidentifiable. Its uncertainty cannot be represented by a finite two-dimensional covariance ellipse. This distinction matters when implementing an optimizer. A singular information matrix should not be treated as an ordinary finite score. Adding a tiny numerical diagonal can make code run, but it also hides whether the data identify every parameter.
Now change the objective. If only the first quantity matters, spending everything on A may be sensible. The balanced answer was optimal for a joint two-parameter volume criterion, not for every possible scientific question. Likewise, if B measurements cost twice as much, the feasible allocations change. Write down the estimand, budget and permitted repetitions before optimizing. This example separates those choices so that a seemingly impressive determinant does not substitute for a scientific objective.
The necessary mathematics
The standard linear model makes the geometry explicit. For measurement index \(i\), write \(y_i=a_i^\top\theta+\epsilon_i\). Here \(y_i\) is the scalar observation, \(\theta\) is the unknown parameter column vector, \(a_i\) is its known sensitivity column vector, and the superscript \(\top\) means transpose. The noise \(\epsilon_i\) has mean zero and common variance \(\sigma^2\), independently across measurements. For a selected index set \(S\), define its information matrix \(M_S\) by the following equation. If it is invertible, the least-squares estimator \(\hat\theta\) has the stated covariance; \(\operatorname{Cov}\) denotes covariance.
\[M_S=\frac{1}{\sigma^2}\sum_{i\in S}a_i a_i^\top,\qquad \operatorname{Cov}(\hat\theta)=M_S^{-1}.\]The inverse notation means matrix inverse, not elementwise reciprocals. Gaussian noise makes the usual fixed-confidence ellipsoid exact in this model with known noise variance. [R1]
In the toy allocation, the sensitivities are the two coordinate directions, so the information matrix is diagonal with entries \(n_A\) and \(n_B\). Its determinant is \(n_A n_B\). Substituting \(n_B=4-n_A\) gives \(n_A(4-n_A)=4-(n_A-2)^2\), maximized at two. This direct calculation explains the balanced solution without relying on an optimization package. It instantiates the determinant criterion rather than proving that balanced allocation is generally optimal. [R2]
For continuous approximate designs, the classical equivalence theorem links determinant maximization to minimization of worst-case prediction variance over the design space. Its conditions and continuous design measures matter. It does not say that every fixed-size integer subset inherits the same optimality certificate. Keep the theorem, the finite allocation and the relaxation distinct: related mathematical objects can answer different operational questions. [R2]
Assumptions and failure modes
The sensitivity vectors are part of the model, not discovered by the determinant. If a nonlinear response is approximated around the wrong parameter value, a locally informative design can be unhelpful for the real system. Before committing the entire budget, compare several plausible parameter settings and ask whether the chosen measurements change radically. This sensitivity analysis is a proposed diagnostic here, not a robustness guarantee supplied by the toy calculation. The broader survey motivates considering models beyond the linear setting; the suggested diagnostic is not a result verified from that abstract. [R3]
Independent measurement noise is also substantive. Technical repeats sharing a calibration error do not provide the information of independent replicates. In the toy example, a common offset affecting every B observation would not vanish merely by averaging more B observations. Similarly, an unmodeled batch effect can be precisely estimated as if it were the desired parameter. More precise model-based estimates can still be wrong. Include nuisance effects and the actual dependence structure before interpreting a design score. [R1]
The integer decision creates a second failure mode. Fractional allocation can ask for part of a sensor or a fraction of an experimental run. Rounding must restore a feasible plan and its information matrix must then be checked again. A good fractional score alone is not the quality of the executed experiment. Large numerical dynamic ranges are another warning: compare log determinants rather than multiplying many eigenvalues directly, and investigate near singularity rather than silently regularizing it away. [R1]
Verified applications
Joshi and Boyd study selecting measurements for parameter estimation and evaluate their approach numerically. This is a verified computational sensor-selection application, not evidence of a deployed biomedical assay or a clinical benefit. Their paper supports the relaxation-based formulation and its explicit lack of a universally small optimality gap. No biological validation is claimed for the present worked example. [R1]
Huan, Jagalur and Marzouk survey classical, Bayesian and decision-theoretic design, including nonlinear and non-Gaussian settings and sequential policies that depend on previous outcomes. This establishes a broader computational context, not a claim that our linear determinant formula works unchanged in all those settings. Only the primary abstract and publication metadata are used for this related reading; its individual application studies have not been independently reassessed here. [R3]
Possible transfers — speculative
A possible omics transfer would allocate a small pilot budget among perturbation conditions whose predicted responses depend on a few explicitly chosen model parameters. This is speculative. A condition with a large response is not automatically useful if it repeats the same parameter sensitivity as conditions already chosen. One could compare a determinant-based plan with random selection and a simple biologically motivated baseline, while reserving independent validation conditions. The cited sensor paper does not validate this proposal.
The experiment would first need credible sensitivities, a count or transformed-data observation model, and donor-level replication. Thousands of cells from one donor should not be substituted for thousands of independent experimental units. Include required controls and minimum donor coverage as constraints, not optional rewards. If the scientific question concerns a single contrast, optimize that contrast rather than all model coefficients indiscriminately. If the model cannot make defensible response predictions, stop at a pilot study instead of presenting a computed design as authoritative.
What to try and what to read next
Try the four-measurement example by enumerating every allocation. Mark singular plans explicitly, calculate both individual variances and joint area, then change either measurement cost or the scientific target. This small exercise reveals what the objective rewards and what it ignores. For a real project, record the feasible experiments, independent unit, noise assumptions, uncertainty target and validation plan before selecting an algorithm.
Read the main sensor-selection paper for the bridge from geometry to a computable plan, the equivalence paper for the precise classical theorem, and the modern survey for questions that outgrow the linear model. The practical lesson is not to maximize a determinant whenever possible. It is to spend scarce measurements on information the current design lacks, while keeping model error, feasibility and the actual scientific decision visible.
Reading list
-
[R1] Primary reading: Sensor Selection via Convex Optimization. Siddharth Joshi and Stephen Boyd (2008). Main reading: study the measurement model and relaxed subset-selection problem before the numerical experiments. The author-hosted PDF is the journal article, not a separate paper.
-
[R2] Related reading: The Equivalence of Two Extremum Problems. J. Kiefer and J. Wolfowitz (1960). Foundational theorem: understand why determinant optimality and worst-case prediction variance coincide for approximate designs under the stated regression assumptions, not arbitrary finite subsets.
-
[R3] Related reading: Optimal experimental design: Formulations and computations. Xun Huan, Jayanth Jagalur and Youssef Marzouk (2024). Broader perspective: see how the objective and computational problem change outside the linear-Gaussian setting. The author manuscript and journal article are one reference; abstract-level evidence is used here.