Research Radar — 2026-10-31
Methods & AI
Computational
General prediction of T cell receptor antigen specificity from sequence using AlphaFold 3 | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1126/sciadv.aef5674
TCR specificity protein structure benchmarking
Summary: AlphaFold 3 structural-confidence features discriminate cognate TCR–peptide–MHC combinations from controls, including held-out epitopes. Reported median validation AUROCs are 0.81–0.92 across MHC classes.
Why it matters: A concrete structure-based route to TCR specificity prediction beyond epitope-specific training.
Why for Yiru: Relevant to combining structural predictions with immune-repertoire measurements rather than treating receptor sequence similarity as specificity.
Spatial single-cell mapping of transcriptional differences across genetic backgrounds in mouse brains
Nature Methods Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1038/s41592-026-03289-2
spatial transcriptomics reference mapping imputation
Summary: ATLAS encodes reference-derived transcriptional signatures in tissue and decodes cell types and imputed transcriptomes. The study profiles 40 million cells across 15 mouse brains and reports regional and cell-type composition differences.
Why it matters: Scalable replicated spatial comparison comes with an explicit distinction between measured signatures and inferred transcriptomes.
Why for Yiru: Useful for scaling anatomical cell-state maps while keeping measured panel signals distinct from reference-imputed genes.
IDPForge: deep learning of proteins with global and local regions of disorder
Nature Communications Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1038/s41467-026-79191-4
diffusion models protein ensembles intrinsically disordered proteins
Summary: IDPForge uses a transformer-based diffusion model to generate all-atom structural ensembles for disordered proteins and regions while retaining folded domains, with optional experimental restraints.
Why it matters: Moves protein modeling from a single folded structure toward ensembles that can be compared with solution measurements.
Why for Yiru: Relevant to ensemble rather than single-structure representations of proteins and integration of experimental restraints.
Genomic heterogeneity inflates the performance of variant pathogenicity predictions
Nature Communications Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1038/s41467-026-78623-5
variant prediction benchmark design dataset shift
Summary: A stratified evaluation shows that mixing variant categories with different pathogenicity prevalences can inflate genome-wide prediction scores. The authors evaluate models separately within genomic contexts.
Why it matters: A transferable benchmark lesson: aggregate discrimination can partly reflect category composition rather than within-category prediction.
Why for Yiru: Directly relevant to evaluating biological predictors under class imbalance and heterogeneous genomic contexts.
Digital twins forecast individual immune trajectories
bioRxiv Subject Collection: Bioinformatics Published 2026-10-30 Research preprint (not peer reviewed) DOI: 10.64898/2026.10.23.761347
single-cell models longitudinal prediction digital twins
Summary: ImmuneTwins personalizes a shared single-cell model using repeated observations of each person. Retrospective vaccination and external-cohort evaluations report improved immune-state forecasts as individual histories accumulate.
Why it matters: Tests whether longitudinal personalization adds information beyond a single baseline profile.
Why for Yiru: Relevant to state-dependent response models and the distinction between cells and independent donors.
Advancing morphological profiling with self-supervised learning through comparative insights into representation learning of cellular perturbations
Nature Methods Published 2026-10-29 Peer-reviewed original research DOI: 10.1038/s41592-026-03270-z
self-supervised learning Cell Painting batch effects
Summary: A benchmark trains self-supervised image models on JUMP Cell Painting and evaluates perturbation retrieval and transfer. DINO and DINOv2 outperform the tested masked-reconstruction models and CellProfiler in the reported settings; postprocessing substantially affects results.
Why it matters: Separates representation objective, evaluation task and batch correction in morphological profiling.
Why for Yiru: Useful for comparing masked prediction with teacher-student representation learning.
Geometric inference of cross-species transcriptome correspondence.
Cell Reports Methods (via PubMed) Published 2026-10-28 Peer-reviewed computational methods article; abstract and author repository reviewed, full journal methods not reviewed. DOI: 10.1016/j.crmeth.2026.101643
single-cell omics optimal transport cross-species integration
Summary: Species-OT uses Gromov-Wasserstein optimal transport to align transcriptome-distribution geometry across species, returning probabilistic gene correspondences and a transcriptomic discrepancy. Tests span differentiation bulk RNA-seq in three species and pluripotent single-cell data from six mammals.
Why it matters: Offers an alternative to sequence homology for comparing regulatory roles across species. Geometric alignment is not proof of orthology, conserved mechanism or causal regulation.
Why for Yiru: Directly relevant to cross-species single-cell integration and model-organism translation; useful to compare against homology-constrained baselines.
Quantification method affects replicability of eQTL analysis, colocalization and TWAS
Nature Genetics Published 2026-10-26 Peer-reviewed original research DOI: 10.1038/s41588-026-02827-0
eQTL reproducibility RNA quantification
Summary: Across 48 GTEx tissues and seven phenotypes, transcript annotation and RNA quantification choices alter eQTL discovery, expression prediction, colocalization and TWAS conclusions. Concordance varies with gene characteristics.
Why it matters: Shows that upstream measurement choices propagate into genetic interpretation.
Why for Yiru: Relevant to assessing whether cell-state and genotype effects are robust to preprocessing.
Biomedical discoveries
Biomedicine
Human hair–bearing skin organoids reveal cell type–specific susceptibility and innate immune responses to herpes simplex virus 1 | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1126/sciadv.aef5563
skin organoids spatial transcriptomics experimental infection
Summary: Spatial transcriptomics and imaging of HSV-1-infected human skin organoids reveal compartment-specific infection and inflammatory responses. Cytokine assays support secreted signaling, while interferon-β, but not the tested TNF-α or 4-1BBL conditions, reduces viral replication.
Why it matters: An experimentally tractable organoid setting for separating infection status, cell identity and neighborhood-associated responses.
Why for Yiru: Links controlled infection to spatially heterogeneous epithelial and stromal responses in a human tissue model.
Decoder-FFPE-seq enables sensitive, genome-wide spatial transcriptomics of archival tissues at single-cell resolution | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-10-30 Peer-reviewed research article; newly published journal version DOI: 10.1126/sciadv.ady5085
spatial transcriptomics FFPE measurement technology
Summary: Decoder-FFPE-seq combines probe conversion with dendrimeric coordinate barcoding for archival spatial transcriptomics. Mouse and human tissue benchmarks report improved sensitivity and applications to tumor RNA programs.
Why it matters: A relevant archival-tissue technology to compare with existing spatial platforms.
Why for Yiru: An enabling assay for spatially resolving archived tissue and testing how measurement chemistry changes recovered niches.
Evolution of immune infiltration and immune triad formation across early melanoma progression
bioRxiv Subject Collection: Cancer Biology Published 2026-10-30 Research preprint (not peer reviewed) DOI: 10.64898/2026.10.29.762853
spatial transcriptomics immune niches melanoma
Summary: Spatial profiling of patient-matched early melanoma lesions finds higher T-cell, dendritic-cell and immune-triad densities in invasive lesions. Depth-matched comparisons suggest tissue depth alone does not explain the pattern.
Why it matters: Separates local multicellular arrangements from broad cell-density or clustering summaries.
Why for Yiru: Directly relevant to functional-niche representations and spatial immune context.
Transcriptional and functional profiling reveals a dynamic continuum of beta-cell activation states.
Cell Reports (via PubMed) Published 2026-10-30 Peer-reviewed experimental biology study; indexed author abstract reviewed, full methods and supplements not reviewed. DOI: 10.1016/j.celrep.2026.118233
single-cell omics cell-state continua functional profiling metabolism
Summary: Phospho-S6 measurements and transcriptomic profiling support sequential glucose-dependent beta-cell recruitment, with higher activation thresholds concentrated in the islet core. Single-cell profiles favor a continuous spectrum; high-fat-diet demand shifts activation thresholds and functional programs.
Why it matters: Challenges treating beta-cell states as fixed discrete clusters and connects molecular variation with measured functional thresholds. A marker-based continuum is not itself proof of stable lineage subclasses.
Why for Yiru: Relevant to interpreting continuous cell states and pairing single-cell readouts with functional and spatial assays.
Cancer cell adaptations along a proliferation-robustness axis drive TME pro-tumor phenotypes
bioRxiv Subject Collection: Cancer Biology Published 2026-10-29 Research preprint (not peer reviewed) DOI: 10.64898/2026.10.27.762245
co-culture tumor microenvironment cell-state plasticity
Summary: Long-term co-culture with replenished fibroblasts or macrophages produces distinct cancer-cell adaptations, including increased drug tolerance or stress-associated programs. Adapted cancer cells also alter surrounding-cell states; some changes are reversible.
Why it matters: Uses experimentally controlled interaction history to distinguish adaptation from immediate response.
Why for Yiru: A concrete biological setting for separating intrinsic state, niche exposure and reciprocal effects.
KRASG12D-targeted therapy sensitivity depends on interferon signaling activation in pancreatic cancer.
Cancer Cell (via PubMed) Published 2026-10-29 Peer-reviewed preclinical cancer study; indexed author abstract reviewed, full methods and supplements not reviewed. DOI: 10.1016/j.ccell.2026.10.003
multi-omics perturbation pancreatic cancer drug response
Summary: Multi-omics across 242 patient-derived pancreatic-cancer organoids and matched tumors links intrinsic KRAS-targeted therapy resistance to reduced interferon signaling. Suppression induces resistance in sensitive models, whereas interferon activation increases sensitivity in resistant models.
Why it matters: Moves beyond an expression association with functional perturbations supporting interferon signaling as a response determinant. The MYC relationship remains association-based in the abstract.
Why for Yiru: A useful example of connecting multi-omics stratification to experimentally testable response mechanisms and combination hypotheses.
Methodology Review: Changepoint detection: pay for a boundary only when it earns its place
English · About 10 minutes
Algorithm claims are bounded to inspected primary texts; CBS application evidence is abstract-level. The worked examples are original illustrations, and the spatial transfer is speculative. No new experimental validation or universal runtime guarantee is claimed.
The intuition
An ordered profile can be explained as one noisy level, several stable plateaus, or a smooth trend. Today we ask a deliberately narrower question: where should a piecewise-constant description place its boundaries? Imagine recording a sensor that reads near zero and later near four. Giving every observation its own level fits perfectly, but explains little. Keeping one level is simple, but may conceal a persistent shift. A useful model must make an extra boundary pay for itself by reducing enough lack of fit. This is a modeling exercise, not a claim that real biological transitions are always abrupt.
There are two separate decisions: what counts as a good segmentation, and how to search for it. PELT, short for Pruned Exact Linear Time, accelerates a particular penalized search without changing its optimum. It discards past boundary candidates only when they cannot improve a future solution. Exactness refers to the chosen objective, not to discovering the true biological mechanism. [R1]
A small worked example
Use the invented sequence \(y=(0,0,4,4)\) and measure fit by the sum of squared deviations from each segment mean. Here \(y\) is the ordered observation vector. A single segment has mean two and squared error sixteen. A boundary after position two produces two constant segments and zero error. Suppose each boundary costs \(\beta=3\), where the positive number \(\beta\) is our penalty in the same units as squared error. The one-segment score is sixteen; the two-segment score is three. Better fit justifies the boundary in this toy problem.
Check the alternatives rather than trusting the picture. Splitting after position one or three leaves squared error \(32/3\), so either score is \(32/3+3\). Any solution with two or more boundaries pays at least six, even if its fitting error is zero. Therefore the split after position two is the global optimum. If the penalty becomes twenty, every segmentation containing a boundary costs at least twenty, and the single-segment score of sixteen wins. At penalty sixteen, the two descriptions tie. The data have not changed: the price of complexity has.
A small perturbation makes the exercise less artificial. Replace the observations by \((0,1,4,5)\). The one-segment mean is 2.5, with squared error seventeen. Splitting in the middle gives means 0.5 and 4.5, with total squared error one. At penalty three the score becomes four. These numbers are hand calculations, not performance evidence. Try replacing the second observation by a large outlier: a method rewarded for fitting squared deviations may spend boundaries isolating the outlier rather than finding the transition you intended to study.
Notice what is absent from this calculation. There are no labels saying which plateau represents a cell type, no replicate donors, and no probability assigned to the boundary. Even an unambiguous optimum is only the best answer to this numerical question. If the scientific target is a gradual response, forcing the answer into constant blocks can give an attractive but misleading description. Write down the expected shape before looking at the segmentation, then use the example to test whether your proposed loss rewards that shape.
The necessary mathematics
Let \(n\) be the number of observations, \(i\) an observation index, and \(j\) a segment index. For ordered values \(y_1,\ldots,y_n\), let \(C(a,b)\) be the squared-error cost of positions \(a\) through \(b\), inclusive. Their fitted level is their mean \(\bar y_{a:b}\). With \(m\) internal boundaries \(\tau_1,\ldots,\tau_m\), set \(\tau_0=0\) and \(\tau_{m+1}=n\). The objective \(J\) adds segment costs and a penalty for each boundary. This is the mean-shift specialization of additive segmentation.
\[C(a,b)=\sum_{i=a}^{b}(y_i-\bar y_{a:b})^2,\qquad J=\sum_{j=0}^{m}C(\tau_j+1,\tau_{j+1})+\beta m.\][R4]
Let \(F(t)\) be the minimum penalized cost for the first \(t\) observations. Initialize \(F(0)=-\beta\) so the first segment is not charged as a boundary. The index \(s\) identifies the last boundary before \(t\). Evaluate earlier prefixes once and reuse them.
\[F(t)=\min_{0\leq s\lt t}\left(F(s)+C(s+1,t)+\beta\right).\]Backtracking the minimizing indices gives the boundaries. For our first example and penalty three, \(F(1)=F(2)=0\) and \(F(3)=F(4)=3\). [R1] [R4]
Pruning improves computation rather than changing this statistical objective. PELT uses an inequality that establishes permanent domination of a candidate boundary under a condition on segment costs. Rigaill instead retains costs as functions of the segment parameter and prunes candidates that cannot be optimal anywhere in the relevant parameter space. His paper treats the best solutions across a range of segment counts; that constrained problem should not be confused with fixing a single penalty. [R1] [R2]
Assumptions and failure modes
PELT does not guarantee linear runtime for every sequence. Its linear expected-cost result needs assumptions, including a suitable model for segment lengths and sufficient improvement from real changes. Without effective pruning, its worst-case search remains quadratic in the number of observations. A faster exact solver does not select the penalty or validate the loss for you. [R1]
For this particular toy loss, an abrupt mean shift is the feature being rewarded. A changing variance, correlated noise, a gradual ramp, or an isolated artifact can change the interpretation. Multiplying every measurement by ten multiplies every squared-error cost by one hundred; leaving the penalty unchanged would therefore ask a different question. This follows directly from squaring the rescaled residuals. Similarly, deleting positions changes possible segment lengths. Record units, filtering, minimum-length constraints and coordinate order before comparing two outputs. A constraint-aware implementation must preserve its pruning guarantees; simply modifying a recurrence is not a proof.
The observations must have a defensible order. Sorting unrelated genes by the measured response manufactures apparent runs and then asks the same data to explain them. Genomic coordinates, sampling times, and a prespecified spatial path have different meanings and resolution limits. A boundary between adjacent observations is not a precisely located event between them. If positions are far apart, report that interval rather than a false point estimate. These are interpretation checks proposed for applying the model, not a confidence interval supplied by the optimizer.
Verified applications
Olshen and colleagues developed circular binary segmentation for ordered microarray DNA copy-number measurements. Their publisher abstract reports simulations, cell lines with known alterations, and a breast-cancer cell-line example. This establishes a biological segmentation application, but CBS is a different procedure and does not demonstrate PELT performance. The inspected evidence for that reading is abstract-level. Maidstone and colleagues compare exact pruning algorithms and evaluate copy-number variation data, illustrating why the objective and computational strategy should be assessed separately. No current single-cell benchmark or clinical validity is claimed here. [R3] [R4]
Possible transfers — speculative
A speculative experiment would segment a prespecified one-dimensional spatial transect using a pathway score measured at successive positions. Choose the transect without using that same score, retain coordinates, and compare constant-level segmentation with a smooth baseline. Check stability across independently sampled tissues rather than merely across thousands of neighboring cells. A suspected boundary should also survive reasonable preprocessing choices and correspond to an independently measured feature before receiving a biological label. This proposed transfer is not validated by the array copy-number applications. Irregular spacing, branching tissue structure and uncertain cell measurements could make the one-dimensional model inappropriate.
What to try and what to read next
Start with four points and enumerate every possible segmentation. Reproduce the scores above, increase the penalty, and verify where the preferred model changes. Then compare that enumeration with a dynamic-programming implementation before adding pruning. For a larger experiment, record both the selected boundaries and the choices that created them: loss, penalty, allowable lengths and coordinate order. Read Killick for safe pruning, Rigaill for a contrasting pruning representation, Olshen for an established biological setting, and Maidstone for the distinction between optimization problems and solvers. The useful question is not just where a line jumps, but what evidence made that boundary worth its cost.
Reading list
-
[R1] Primary reading: Optimal detection of changepoints with a linear computational cost. Rebecca Killick, Paul Fearnhead and Idris A. Eckley (2012). Main reading: understand optimal partitioning and the safe-pruning argument before the complexity theorem. Journal online publication is December 21, 2012; the linked arXiv manuscript first appeared January 7, 2011.
-
[R2] Related reading: A pruned dynamic programming algorithm to recover the best segmentations with 1 to Kmax change-points. Guillem Rigaill (2015). Algorithmic contrast: functional pruning for a range of segment counts. This is not the same optimization task as fixing one penalty. The journal appeared in 2015; the preprint began April 6, 2010.
-
[R3] Related reading: Circular binary segmentation for the analysis of array-based DNA copy number data. Adam B. Olshen, E. S. Venkatraman, Robert Lucito and Michael Wigler (2004). Established biological application and a different search strategy: read how ordered microarray intensities become candidate copy-number regions. Abstract-level evidence only in this review; this paper does not validate PELT itself.
-
[R4] Related reading: On optimal multiple changepoint algorithms for large data. Robert Maidstone, Toby Hocking, Guillem Rigaill and Paul Fearnhead (2016). Comparison and implementation perspective: separate the statistical objective from its solver, and compare functional with inequality pruning. Published online February 15, 2016, in the 2017 volume; linked preprint September 5, 2014.