Research Radar — 2026-11-21
Methods & AI
Computational
A multimodal foundation model for molecular and morphologic data in oncology
Nature Cancer Published 2026-11-20 Peer-reviewed methods article; earlier full preprint used for detailed evidence DOI: 10.1038/s43018-026-01252-w
multimodal learning computational pathology representation learning
Summary: THREADS aligns whole-slide morphology with molecular profiles to learn reusable oncology representations.
Why it matters: The biologically informed training target is the central design choice, not simply a larger image encoder.
Why for Yiru: Relevant to choosing molecular targets for spatial representation learning, while remaining distinct from perturbation-effect prediction.
Optimizing phenotype scale improves genetic analyses in large-scale biobanks
Nature Genetics Published 2026-11-19 Peer-reviewed statistical methods article DOI: 10.1038/s41588-026-02826-1
quantitative genetics phenotype scaling model validation
Summary: SIQReg learns phenotype transformations that reduce scale-induced genetic heterogeneity while improving some additive analyses.
Why it matters: Apparent nonadditivity can change with measurement scale; removing it is not proof that the underlying biology is irrelevant.
Why for Yiru: A useful stress test for FEAST parameterizations and intrinsic/extrinsic response claims, rather than a ready-made single-cell model.
Biomedical discoveries
Biomedicine
Posttranscriptional regulation by TIA1 and TIAL1 controls the transcriptional program enforcing T cell quiescence | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-11-18 Peer-reviewed mechanistic immunology study DOI: 10.1126/sciadv.aeh4461
T-cell quiescence posttranscriptional regulation conditional perturbation
Summary: TIA1 and TIAL1 jointly maintain peripheral naïve T-cell quiescence. Conditional genetics and RNA-binding assays link their loss to lower expression of quiescence-enforcing transcription factors, aberrant activation and impaired survival.
Why it matters: Shows why RNA abundance alone can miss a regulatory mechanism: RNA-binding proteins can support transcriptional identity through target translation.
Why for Yiru: Useful mechanistic background for CD4 T-cell perturbation modeling and for distinguishing direct regulatory effects from secondary state shifts.
PIC-NIC reveals that DNA mechanical strain impacts transcription factor recognition
Nature Structural & Molecular Biology Published 2026-11-18 Peer-reviewed primary research DOI: 10.1038/s41594-026-01919-z
transcription factors DNA mechanics regulatory priors
Summary: PIC-NIC experimentally separates DNA backbone continuity from sequence identity, mapping nick-sensitive TF recognition across 15 human factors in eight structural families. Structural, binding-kinetic and simulation analyses show that local DNA mechanics can modify affinity and specificity.
Why it matters: It exposes a regulatory variable that sequence-only motif models cannot fully capture.
Why for Yiru: Relevant to mechanistic priors for regulatory-response models and to distinguishing sequence change from physical-context perturbation.
TransA-MERFISH reveals divergent post-retinal direction-selective circuits
Nature Published 2026-11-18 Peer-reviewed primary research DOI: 10.1038/s41586-026-11235-7
spatial transcriptomics circuit tracing perturbational validation
Summary: TransA-MERFISH combines anterograde trans-neuronal labeling with a 140-gene spatial panel to connect neuronal identities to retinal-input cell types. Orthogonal tracing, circuit physiology and retinal perturbations distinguish inherited and independently computed motion selectivity in superior colliculus.
Why it matters: It pairs molecular taxonomy with experimentally tested connectivity rather than inferring interactions from proximity alone.
Why for Yiru: A strong example for validating spatial interaction hypotheses with independent tracing and perturbation readouts, even though the system is neural rather than immune tissue.
Enhancer-targeted CRISPR-A rescues haploinsufficiency and mutant phenotypes in organoid models of autism
Nature Neuroscience Published 2026-11-13 Peer-reviewed primary research DOI: 10.1038/s41593-026-02504-9
CRISPR activation organoids gene regulation
Summary: Enhancer-targeted CRISPR activation raises CHD8 or SCN2A expression and rescues selected molecular, developmental and electrophysiological phenotypes in gene-edited and patient-derived neuronal or cortical-organoid models.
Why it matters: It tests a regulatory intervention that uses retained wild-type alleles and developmental enhancer context rather than assuming more gene expression is always better.
Why for Yiru: Useful evidence for context- and time-dependent perturbation-response modeling, with organoids providing an intermediate experimental domain.
A single-cell lung cancer atlas reveals cellular and molecular drivers of metastasis and clinical outcome
Nature Cancer Published 2026-11-12 Peer-reviewed primary research DOI: 10.1038/s43018-026-01260-w
single-cell atlas tumor microenvironment external validation
Summary: LuCaS+ profiles over 2.5 million nuclei from 454 patients with lung cancer, separating a 364-patient discovery cohort from 90 external validation patients. Tumor-state and microenvironment composition associate with relapse and survival, supplemented by spatial profiling.
Why it matters: It provides a large clinically annotated resource for testing tissue-state and outcome models across centers.
Why for Yiru: Useful for evaluating niche and cell-state representations against donor-level outcomes without treating cell count as independent patient replication.
STAT3 controls IL-6–induced epigenomic remodeling and restrains a latent STAT1-associated interferon program in human macrophages | Science Advances
AAAS: Science Advances: Table of Contents Published 2026-11-11 Peer-reviewed research article DOI: 10.1126/sciadv.aek8053
macrophages cytokine signaling epigenomics
Summary: IL-6-driven macrophage chromatin programs depend on STAT3 recruitment. Removing STAT3 reveals a delayed STAT1-associated interferon program, partly sensitive to STING and IFNAR interventions.
Why it matters: Illustrates why an intracellular perturbation can change the response to the same extracellular cue.
Why for Yiru: Read as a context-dependent regulatory example, not a validated intervention model or proof of disease causality.
Methodology Review: Markov-chain lumpability: when a smaller state space keeps the same predictions
English · About 10 minutes
Original theorem chapter and three full primary papers read; toy calculations checked, no benchmark or raw-data reproduction.
The intuition
Imagine a technician who records whether a machine is in room A or room B, but cannot see its setting inside room A. Suppose there are two settings. If either setting gives the same chance of entering B at the next observation, knowing the hidden setting adds nothing to that particular prediction. But if one setting nearly always leaves and the other almost never does, the label A hides consequential information. The history of room visits may then indicate which setting the machine is in, even though the current room does not. A smaller state description can create apparent memory in a system that originally had none.
Strong lumpability gives an exact test for when this loss of detail is harmless for predictions about rooms. It does not ask whether two settings look alike, or whether they move among the hidden settings in the same way. It asks whether their total transition probabilities into each observed room agree. This distinction matters whenever a graph or dynamic model is summarized by clusters. A cluster can be visually coherent and internally sticky while hiding two very different ways to leave. Conversely, hidden states can behave very differently inside a block and still make identical predictions about the next block. [R1]
A small worked example
Consider an invented discrete-time chain on hidden states \(a_1,a_2,b\). Record both \(a_1\) and \(a_2\) as room A, and \(b\) as room B. In that order for both rows and columns, let the transition matrix be \(P=\begin{pmatrix}1/2&1/4&1/4\\1/8&5/8&1/4\\1/2&0&1/2\end{pmatrix}\). Its first row means that from the first setting in A, the next observation stays there half the time, moves to the other setting one quarter of the time, and enters B one quarter of the time. The second setting has a different internal pattern but the same one-quarter chance of entering B.
At room level the matrix is therefore \(Q=\begin{pmatrix}3/4&1/4\\1/2&1/2\end{pmatrix}\). Starting in A, there are two ways to end up in B after two observations: remain in A first and then leave, or enter B first and then stay there. Their probabilities sum to \((3/4)(1/4)+(1/4)(1/2)=5/16\). This is the same whether the hidden start is the first or second setting. The aggregate model preserves this prediction, but not the probability of ending at a specific setting within A.
Now replace the transitions by a deterministic cycle: the first A setting goes to the second, the second goes to B, and B returns to the first. In a stationary run, a present A following B must be the first setting, so the next room is A with certainty. A present A following A must be the second setting, so the next room is B with certainty. The current room is identical in the two cases; the previous room changes the prediction. Averaging over both settings would give a one-step escape probability of one half, but it would erase the pattern that reveals the hidden phase.
The necessary mathematics
Let \(X_t\) denote the microstate at integer time \(t\), and let \(P_{xy}=\Pr(X_{t+1}=y\mid X_t=x)\) be the fixed transition probability from microstate \(x\) to microstate \(y\). Partition the microstates into disjoint blocks \(C_1,\ldots,C_k\), where \(k\) is the number of blocks, and record only the block of the current state. For any starting block \(C_i\) and target block \(C_j\), the row-sum criterion requires that \(\sum_{y\in C_j}P_{xy}=\sum_{y\in C_j}P_{x^{\prime}y}\) whenever \(x,x^{\prime}\in C_i\). The common value is the macro-transition entry \(Q_{ij}\). The sums compare whole blocks, not the entries within them. [R1]
The proof is short. Conditional on any observed history ending in block \(C_i\), the microstate is a mixture over that block. At the microstate level, the Markov assumption says that the future depends on the current microstate, not its past. Averaging a constant next-block probability over the mixture gives the same constant, regardless of the weights. Conversely, initialize the chain separately at two states in the same block. If a single macro-transition matrix works for every initial distribution, its first row of predictions cannot differ across those two starts. The resulting exact model therefore preserves every finite sequence probability of observed block visits for a given initial block distribution. [R1]
Assumptions and failure modes
The statement assumes a finite, time-homogeneous Markov chain, a fixed partition, and the chosen discrete observation interval. A model estimated at a different lag or under a different intervention has a different transition matrix to check. Failure of the criterion means there is no common transition matrix valid for every microstate initialization. It does not rule out Markov behavior for particular initial distributions; that weaker possibility has a separate definition. Nor does passing the criterion certify an empirically fitted microstate model. Exact lumpability of a wrong model remains exact only inside that model. [R1]
Sampling matters when the matrix is fitted rather than known. Two rarely observed states may have very similar empirical transition rows just because neither has been observed leaving. Conversely, small noisy row differences need not reflect a meaningful failure at the forecast horizon. Compare direct estimates of multi-step block transition probabilities with iterated one-step predictions on held-out trajectories when possible. A discrepancy is a warning; close agreement at a few lags is not a proof that the projection is Markov under all initializations. The test also cannot identify the mechanism of a transition. [R1] [R3]
Verified applications
In molecular simulation, Bowman builds a hierarchy of coarser Markov state models by comparing microstates with finite-sampling uncertainty in mind. The article analyzes molecular dynamics-derived models of alanine dipeptide and villin, and compares candidate partitions with other methods using Bayes factors. That is a demonstrated setting in which dynamic, rather than purely geometric, similarity is useful. It is not a demonstration that the resulting molecular macrostates satisfy exact lumpability, or that the reported comparison transfers to other data types. [R3]
Network studies make another distinction concrete. E and colleagues choose an approximation objective for a random walk and study when it recovers lumpable partitions. Piccardi uses stationary-weighted, one-step block retention to score network communities. One-step retention is meaningful for the specified walk and weighting even when the observed block sequence is not itself Markov. In both cases the outputs are properties of the chosen model of movement on a graph, not evidence that nodes physically move or that communities represent biological mechanisms. [R2] [R4]
Possible transfers — speculative
A possible omics exercise is to ask whether a coarse cell-state label makes a good forecast for a genuinely longitudinal or lineage-tracked dataset. At one defined interval and condition, estimate the distribution of next coarse labels separately from each finer starting state. Compare those distributions across fine states within a coarse label, preserve uncertainty for rare transitions, and evaluate held-out biological replicates. Destructive snapshots, pseudotime, or a neighborhood graph alone do not supply observed transitions. One can still audit lumpability of a learned graph walk, but that result would concern the walk, not measured cell-state kinetics.
What to try and what to read next
Start with two questions: what detail am I dropping, and what future event must remain predictable? Add each transition row over the proposed destination blocks and compare rows within each starting block. If they match, use the smaller chain for predictions about blocks. If they differ, retain the relevant hidden distinction or make an explicitly approximate forecast and test it. Reproducing both toy chains by hand makes the key distinction clear: a good state label is one that keeps the information needed for the prediction, not merely one that makes the diagram smaller. [R1]
Reading list
-
[R1] Primary reading: Finite Markov Chains. John G. Kemeny and J. Laurie Snell (1976). Read the all-initial-distributions definition and the block row-sum criterion in sections 6.3.1–6.3.2. Section 6.4 distinguishes weak lumpability.
-
[R2] Related reading: Optimal partition and effective dynamics of complex networks. Weinan E, Tiejun Li and Eric Vanden-Eijnden (2008). Read for the distinction between a partition selected for prediction under a specified metric and an exact state aggregation.
-
[R3] Related reading: Improved coarse-graining of Markov state models via explicit consideration of statistical uncertainty. Gregory R. Bowman (2012). Read the toy transition graph and Table I for how finite sampling affects a merging decision.
-
[R4] Related reading: Finding and Testing Network Communities by Lumped Markov Chains. Carlo Piccardi (2011). Read the Methods and the persistence-probability examples. A useful one-step network diagnostic is not a certificate of exact macrostate dynamics.