Research Radar — 2026-10-29
Methods & AI
Computational
PepPr: De Novo peptide sequencing with peptide priors
bioRxiv Subject Collection: Bioinformatics Published 2026-10-28 Research preprint; not peer reviewed DOI: 10.64898/2026.10.27.762126
proteomics sequence inference
Summary: PepPr incorporates a reference-proteome language-model prior directly into de novo peptide sequencing. Reported benchmarks improve reference and variant peptide recovery; antibody examples retain hypervariable-region recall while improving variable–constant boundary recovery. These are sequencing benchmarks, not a demonstration of new experimentally validated protein function.
Why it matters: The method tests whether prior biological knowledge can help noisy inverse problems without restricting every prediction to a database entry.
Why for Yiru: Relevant to probabilistic omics modeling and the trade-off between regularization and discovery of unexpected variants.
From {triangleup}{triangleup}G Prediction to Mutation Prioritization with Conformal Prediction
bioRxiv Subject Collection: Bioinformatics Published 2026-10-28 Research preprint; not peer reviewed DOI: 10.64898/2026.10.27.762132
conformal prediction experimental prioritization
Summary: ProbJanusDDG turns predicted mutation effects into selective recommendations using protein-weighted conformal intervals. At a nominal 90% level, reported mutation-level coverage is 91.6% or 92.5%, but coverage averaged equally across proteins is about 87%. Higher sign accuracy comes with fewer directional calls. No prospective mutation campaign is demonstrated.
Why it matters: Uncertainty is useful only if its calibration unit matches the decision being made.
Why for Yiru: An enabling method for AI-guided experiment prioritization, with a concrete warning about pooled calibration across proteins.
Fixing the Measurement Model Before the Selection Heuristics: Optimiser Repairs for Living-Cell Self-Driving Laboratories
bioRxiv Subject Collection: Bioinformatics Published 2026-10-28 Research preprint; not peer reviewed DOI: 10.64898/2026.10.27.762131
self-driving laboratories measurement error Bayesian optimization
Summary: This simulation study argues that living-cell optimization can benefit more from fixing the measurement model than from changing selection heuristics. It examines robust summaries, noise shrinkage, drift and day effects across simulated campaigns, with independent reimplementation checks. Gains under contamination can come with clean-data penalties; no real-cell closed-loop experiment is reported.
Why it matters: Batch effects and measurement error can dominate an optimizer even when its acquisition rule is sophisticated.
Why for Yiru: Directly relevant to the new self-driving-laboratory interest, but as a computational stress test rather than demonstrated autonomous biology.
Spatialsnake: a flexible workflow for spatial transcriptomics analysis across platforms
bioRxiv Subject Collection: Bioinformatics Published 2026-10-28 Research preprint; not peer reviewed DOI: 10.64898/2026.10.27.762082
spatial omics reproducible workflows
Summary: Spatialsnake packages spatial-transcriptomics analysis in a Snakemake workflow built around SpatialData. Demonstrations span mouse Visium and colorectal-cancer Visium HD, including roughly 1.5 million bins. The contribution is interoperable, reproducible processing across platforms rather than a new causal or cell-state inference method.
Why it matters: A common data representation can reduce bespoke glue code between spatial assays and downstream tools.
Why for Yiru: Useful for comparing spatial pipelines and documenting reproducible preprocessing; bins must not be described as individual cells.
Building virtual cell dynamics that evolve with biological measurements
bioRxiv Subject Collection: Bioinformatics Published 2026-10-27 Research preprint; not peer reviewed DOI: 10.64898/2026.10.23.761350
virtual cells dynamical systems AI for science
Summary: CellCompass builds and revises virtual cell-dynamics models as measurements change. Evaluation includes a simulated dynamics benchmark and retrospective analyses of four real studies, spanning metabolic RNA, lineage barcodes and protein measurements. Reported improvements concern model identification or prediction within these settings, not prospective experimental validation of agent-generated hypotheses.
Why it matters: Different measurements constrain different dynamical ambiguities; model revision is as important as fitting a larger predictor.
Why for Yiru: Relevant to virtual cells, causal-model assumptions and the design of informative biological measurements.
Cell-level integration of gene expression and B cell receptor data captures complementary B cell information
bioRxiv Subject Collection: Bioinformatics Published 2026-10-22 Preprint, not peer-reviewed DOI: 10.64898/2026.10.16.759689
single-cell multimodal integration B-cell receptors
Summary: BCR language-model embeddings capture clonality but can distort cell-type relationships against independent surface-protein measurements. Concatenating expression with isotype, somatic-hypermutation frequency and CDRH3 length retained complementary receptor information while preserving expression-associated structure.
Why it matters: Independent protein measurements make this a useful caution against equating richer sequence embeddings with better cell-state representations.
Why for Yiru: Directly relevant to single-cell multimodal integration and to testing what a representation actually preserves.
Modulator Resolves Transcript-, Allele-, and Condition-associated RNA Modification Differences from Direct RNA Sequencing
bioRxiv Subject Collection: Bioinformatics Published 2026-10-16 Preprint, not peer-reviewed DOI: 10.64898/2026.10.15.759385
direct RNA sequencing RNA modifications measurement models
Summary: Modulator groups direct-RNA reads by shared transcript ends and splice junctions to separate transcript-, allele- and condition-associated modification differences despite truncated reads. In a matched validation subset, 96% of m6A changes measured by GLORI agreed in direction.
Why it matters: Separating transcript composition from modification changes addresses a major interpretation problem in direct-RNA comparisons.
Why for Yiru: Useful for omics measurement models and separating compositional shifts from molecular-state changes.
Biomedical discoveries
Biomedicine
Oncogenic PIK3CA mutations cause lineage-selective niche corruption
bioRxiv Subject Collection: Cancer Biology Published 2026-10-28 Research preprint; not peer reviewed DOI: 10.64898/2026.10.26.761693
single-cell perturbation vascular biology niche signaling
Summary: Using patient material and engineered iPSC-derived vascular models, this preprint reports that oncogenic PIK3CA produces different consequences across vascular lineages despite similar AKT activation. Venous-state changes and HGF signaling alter the surrounding niche. Intervention experiments support prevention in the modeled setting, but not reversal of established disease or clinical efficacy.
Why it matters: The same mutation can act through lineage-specific cell–cell interactions rather than a uniform intracellular response.
Why for Yiru: A useful example of pairing temporal single-cell measurements with engineered perturbations to study niche causality.
Tumour assembloids recapitulate tumour–CAF–T cell crosstalk in renal cancer
Nature Published 2026-10-28 Peer-reviewed original research DOI: 10.1038/s41586-026-11208-w
assembloids tumor microenvironment perturbation
Summary: Renal-cancer assembloids combine tumor cells, cancer-associated fibroblasts and T cells to dissect microenvironmental crosstalk. The study links tumor WNT signaling to fibroblast activin A and CD8 T-cell state, with engineered and humanized mouse experiments supporting the mechanism. This is a mechanistic model system, not a clinically validated predictor for choosing patient therapy.
Why it matters: Mix-and-match multicellular systems can distinguish which compartment supplies a functionally important signal.
Why for Yiru: Especially relevant to organoid experimental design, microenvironment perturbations and interpretable biological validation.
Antigen recognition licenses therapeutic CD8+ T cells to organize a multicellular myeloid interaction network in the tumor microenvironment
bioRxiv Subject Collection: Immunology Published 2026-10-20 Preprint, not peer-reviewed DOI: 10.64898/2026.10.14.759124
single-cell cell-cell communication genetic perturbation
Summary: In vivo interaction recording, single-cell profiling and antigen-presentation perturbations identify an antigen-licensed network between therapeutic CD8 T cells and myeloid cells. Myeloid engagement did not require endogenous myeloid MHC-I, whereas inflammatory macrophage remodeling required tumor-cell peptide–MHC-I.
Why it matters: The experimental design separates contact-network formation from downstream state remodeling instead of inferring communication from expression alone.
Why for Yiru: Strong fit for perturbation-grounded cell–cell communication and causal interpretation of single-cell states.
Droplet-based organoid screens enable high-throughput identification of cancer vulnerabilities
bioRxiv Subject Collection: Cancer Biology Published 2026-10-19 Preprint, not peer-reviewed DOI: 10.64898/2026.10.13.744538
organoids CRISPR perturbation experimental biology
Summary: A droplet organoid platform couples high-throughput compound screening, live imaging and pooled CRISPR perturbations. The authors report a screen of more than 5,000 clinical compounds and a UXS1-linked gastric-to-hepatic cell-state transition in stomach-cancer models.
Why it matters: Combining perturbations with time-resolved organoid phenotypes can expose dependencies not apparent from genomic profiling alone.
Why for Yiru: Substantive experimental-biology infrastructure relevant to future closed-loop experiment design, without claiming that the reported platform is AI-driven or autonomous.
Methodology Review: Persistent homology: which holes survive a change of scale?
English · About 10 minutes
Primary tutorial full text inspected; foundations journal abstract only; stability theorem checked in the 2005 conference manuscript alongside 2006 online journal metadata. No biological benchmark reproduced, and proposed omics uses are speculative.
The intuition
A scatterplot can suggest a ring, several islands, or a filled cloud. But a single plotting scale can be misleading. Connect nearby observations using a very small distance and almost everything is isolated. Increase the distance enough and almost everything is connected. Persistent homology asks which topological features appear and disappear while that distance changes. Its output records intervals of survival rather than insisting that one neighborhood radius is the uniquely correct choice. [R1]
The abstract problem is geometric: summarize connectivity and holes in a finite set of observations. A hole here is a property of a constructed space, not automatically a missing cell type, a regulatory feedback loop, or a physical cavity. Imagine a photograph of people standing around a fountain. Their positions may outline a loop, but those positions alone cannot tell you why the people stood there. Keep that separation between a measured shape and its explanation throughout the analysis.
A small worked example
Take four points at the corners of a unit square: \((0,0)\), \((1,0)\), \((1,1)\) and \((0,1)\). Call the distance threshold \(r\). Our rule is explicit: include an edge whenever the Euclidean distance between its endpoints is at most \(r\), and fill every triangle whose three edges are present. More generally, include every higher-dimensional simplex whose edges are present. This is the Vietoris–Rips convention used in this example; the threshold is an edge length, not a ball radius.
For \(0\leq r\lt1\) there are four isolated vertices and no loop. At \(r=1\) all four sides enter together. The square becomes one connected component with one unfilled loop. No triangle is present because each possible triangle needs a diagonal. At \(r=\sqrt{2}\) both diagonals appear, together with all triangles and the three-dimensional simplex on the four vertices. The loop is filled and disappears. Thus the one-dimensional barcode has one interval \([1,\sqrt{2})\), with lifetime \(\sqrt{2}-1\approx0.414\). The interval includes its birth but excludes its death under our convention.
The zero-dimensional barcode tracks components separately. All four components are born at zero; three terminate when the sides connect them at one, while one component persists indefinitely in ordinary, unreduced homology. Simultaneous merges can make the choice of representative component arbitrary, but they do not change these interval endpoints. The square example is an original hand calculation for this review, not a result from an omics dataset. Its purpose is to make the filtration rule and the difference between components and loops concrete.
Now change the units so every coordinate is multiplied by ten. The loop still has the same qualitative shape, but its birth and death distances both become ten times larger. Its raw lifetime also increases tenfold. Nothing biological has become ten times more convincing. This simple modification shows why a barcode must travel with its distance definition and preprocessing. For another exercise, add the square’s center as a fifth point and determine when triangles enter. You should no longer expect the original loop interval to survive unchanged.
The necessary mathematics
Let \(K_r\) denote the simplicial complex at threshold \(r\). Increasing the threshold gives nested complexes: \(K_r\subseteq K_s\) whenever \(r\leq s\), where \(s\) is a second threshold. Fix a coefficient field, for example arithmetic modulo two. The vector space \(H_k(K_r)\) records independent homology classes in dimension \(k\); \(k=0\) concerns components and \(k=1\) concerns loops. Its dimension \(\beta_k(r)=\dim H_k(K_r)\) is the Betti number. The inclusion maps between complexes also induce maps between these vector spaces. Persistent homology retains those maps, not just a separate count at each threshold. For finite one-parameter filtrations over a field, interval summaries have a rigorous algebraic basis. [R1] [R2]
A separate theorem concerns sublevel sets of continuous tame functions \(f\) and \(g\) on the same triangulable space. Tameness imposes finite-dimensional homology and finitely many homological changes. Writing \(D(f)\) and \(D(g)\) for their persistence diagrams, the theorem states:
\[d_B(D(f),D(g))\leq\lVert f-g\rVert_\infty.\]Here \(d_B\) is the bottleneck distance: the smallest worst matching displacement between diagram points, allowing matches to the diagonal. The norm on the right is the largest absolute difference between function values. It bounds how far the summaries move under this specified perturbation; it does not certify a biological interpretation. [R3]
Assumptions and failure modes
Distance is a scientific choice. If one gene’s numerical scale dominates the representation, nearby cells may mean little more than similar values of that gene. If an embedding erases a feature before the topology is computed, persistence cannot recover it. Conversely, a convincing ring in a two-dimensional visualization need not be a ring in the original representation. For an exploratory analysis, compare a prespecified expression representation with the visualization rather than silently treating the latter as ground truth.
Sampling also matters. A missing region can arise because a state was not sampled, because quality control removed it, or because the system avoids it. A long interval alone does not distinguish these possibilities and is not a p-value. Donor imbalance, batch structure, and correlated observations remain experimental-design problems. Decide in advance how biological replicates support inference. Running the same geometry calculation on more cells from one donor does not create more independent donors.
Computation has a further practical boundary: even a modest number of vertices can generate many higher-dimensional simplices. Start with a small, explicitly bounded problem and report any subsampling or maximum dimension. These choices affect what you can conclude. Failure to compute a feature because the construction was truncated should not be reported as evidence that the feature is absent.
Verified applications
The inspected stability paper develops uses in recovering the homology of sampled sets and comparing geometric shapes. These are established mathematical applications under stated assumptions. Otter and colleagues also benchmark persistent-homology implementations on synthetic and empirical inputs. Their study is useful for understanding the computational pipeline, but its historical benchmark does not identify the best software in 2026. This review has not reproduced those calculations or validated an omics classifier. [R1] [R3]
Possible transfers — speculative
A possible omics experiment would ask whether a predefined cell-state representation contains a reproducible cyclic organization across independently collected donors. This is a speculative transfer, not a claim that the cited papers establish a cell-cycle detector. Choose features without using the desired topological answer, freeze normalization, and compare equally sized donor-specific samples. Inspect whether a putative cycle remains when known technical covariates are controlled. Use independent biological information only afterward to test an interpretation.
A second speculative use is as an audit of a learned representation. Compare a carefully chosen reference geometry with the geometry after compression, while keeping samples and distance conventions explicit. If a loop appears only after aggressive compression, that is a question to investigate, not an automatic discovery. If a feature disappears, inspect which observations and neighborhoods changed before blaming either representation. The useful outcome may simply be a clearer account of what information the embedding preserved or discarded.
What to try and what to read next
Start by drawing the four-point square and writing down exactly which vertices, edges, and filled triangles exist at each threshold. Check both component and loop intervals. Then change one assumption at a time: units, one point’s position, or the addition of a central point. You should be able to explain every altered endpoint before processing thousands of observations. Read the main tutorial for the pipeline, the algebra paper for why interval summaries exist, and the stability paper for the precise scope of robustness. The most useful question is not merely whether a hole appears, but which choices and evidence make that hole scientifically interpretable.
Reading list
-
[R1] Primary reading: A roadmap for the computation of persistent homology. Otter, Porter, Tillmann, Grindrod and Harrington (2017). Main tutorial: read the filtration and barcode sections first, then the computational pipeline. Historical software benchmarks describe 2017, not a current software ranking.
-
[R2] Related reading: Computing Persistent Homology. Zomorodian and Carlsson (2004). Algebraic foundations: understand why a finite one-parameter filtration over a field admits interval summaries. Conference DOI 10.1145/997817.997870 is an earlier version, not another reading.
-
[R3] Related reading: Stability of Persistence Diagrams. Cohen-Steiner, Edelsbrunner and Harer (2006). Guarantees and their boundary: study the stability theorem, then ask whether your preprocessing creates the kind of bounded perturbation it assumes. The 2005 conference DOI is deduplicated with the journal article.