Research Radar — 2026-11-06
Methods & AI
Computational
RPI-Zero: Zero-shot RNA-Protein Interaction Prediction
bioRxiv Subject Collection: Bioinformatics Published 2026-11-05 Original research preprint; not peer-reviewed DOI: 10.64898/2026.11.04.764286
RNA–protein interactions benchmark leakage sequence foundation models
Summary: RPI-Zero predicts RNA–protein binding from jointly encoded sequences. Its benchmark separates RNA and protein similarity clusters and matches negatives by transcript length and binding degree, testing transfer to unseen RNAs, proteins and pairs.
Why it matters: Similarity leakage and easy negative pairs can make interaction prediction look more general than it is. The benchmark design is as valuable as the reported model gains.
Why for Yiru: A concrete template for leakage-resistant evaluation of molecular foundation models and RNA-binding hypotheses.
Benchmarking Statistical Methods for Differential Composition Analysis in Population-Scale Single-Cell Data
bioRxiv Subject Collection: Bioinformatics Published 2026-11-05 Original research preprint; not peer-reviewed DOI: 10.64898/2026.11.04.764099
single-cell cohorts differential composition false discovery control
Summary: A benchmark compares 12 differential cell-composition methods across three population-scale single-cell datasets. Permutations and data-informed simulations reveal inconsistent discoveries; LOCOM has the best reported false-discovery/power balance within these tests.
Why it matters: Apparent disease-associated cell shifts can be artifacts of compositionality, zeros, dispersion and covariates rather than robust cohort effects.
Why for Yiru: Directly useful when choosing statistical tests for donor-level single-cell cohort comparisons.
Harnessing catalytic knowledge to guide enzyme discovery and evolution with AlphaEnzyme
bioRxiv Subject Collection: Bioinformatics Published 2026-11-05 Original research preprint; not peer-reviewed DOI: 10.64898/2026.11.04.764416
enzyme foundation models prospective validation protein engineering
Summary: AlphaEnzyme pairs enzyme structure with reaction chemistry to retrieve catalysts and predict kinetics and mutation effects. Time-separated benchmarks and prospective transaminase prioritization connect computational ranking to a wet-lab assay.
Why it matters: A common enzyme–reaction representation may link starting-enzyme selection with experimental optimization, rather than optimizing unrelated prediction tasks.
Why for Yiru: Relevant to biomolecular foundation models and low-budget, prospective experimental design.
Improving phosphoproteomic kinase activity inference by contextual refinement of Kinase Library relationships
bioRxiv Subject Collection: Bioinformatics Published 2026-11-05 Original research preprint; not peer-reviewed DOI: 10.64898/2026.11.04.764166
phosphoproteomics kinase activity inference perturbation benchmarks
Summary: Using 222 perturbation phosphoproteomic samples, the authors find that broadly expanding motif-based kinase substrates increases coverage while worsening inference. Filtering these relationships with protein-association evidence improves the reported accuracy–coverage balance.
Why it matters: Sequence-compatible phosphorylation sites are not automatically intracellular kinase targets. More prior edges can dilute biological specificity and correlate inferred activities spuriously.
Why for Yiru: Directly relevant to interpreting perturbation phosphoproteomics and selecting regulatory priors for causal hypotheses.
Where single-cell foundation models help: a benchmark of annotation, perturbation and regulatory inference against simple baselines
bioRxiv Subject Collection: Bioinformatics Published 2026-11-04 bioRxiv research preprint; not peer reviewed; abstract and official version/date metadata reviewed DOI: 10.64898/2026.11.03.763968
single-cell foundation models Perturb-seq benchmarking expression-matched controls regulatory inference
Summary: This preprint benchmarks five single-cell foundation models, supervised perturbation methods and task-matched simple baselines across annotation, perturbation and regulatory inference. Reported gains are task-dependent: cross-species annotation improves, while within-species annotation is comparable to principal components. In organoids, expression-matched controls attribute only 12–35% of predicted knockout effects to a gene-specific component.
Why it matters: The useful contribution is an evaluation design that separates model-specific biological information from target-expression shortcuts. Directional agreement with a perturbation is a weaker claim than accurate effect sizes or causal reconstruction, and the benchmark helps keep those claims distinct.
Why for Yiru: Directly relevant to omics and perturbation-model evaluation: reuse the expression-matched control, task-specific baselines and clone/donor-aware replication checks before interpreting virtual-cell knockout predictions.
Autonomous generalizability auditing of scientific foundation models
bioRxiv Subject Collection: Bioinformatics Published 2026-11-04 bioRxiv research preprint; not peer reviewed; abstract and official version/date metadata reviewed DOI: 10.64898/2026.11.03.763979
scientific AI agents generalization auditing perturbation models controlled benchmarking
Summary: AutoSPECTRA is a coding-agent protocol for auditing scientific foundation-model generalization. Informed by a review of 38 studies, it uses measurable generalization axes, prespecified analyses, controlled comparisons and independent confirmation. Four model audits examine protein relatedness and composition, perturbation training coverage, spatial resolution and variant context.
Why it matters: The protocol makes an agent test competing explanations instead of merely generating plausible narratives around benchmark averages. Its scientific contribution is an auditable experimental-analysis process; the four examples do not establish general reliability of autonomous scientific agents.
Why for Yiru: Useful for auditing omics and causal-AI workflows: stratify STATE-like perturbation evaluations by training coverage, compare matched controls and require a separate confirmation analysis before interpreting out-of-distribution gains.
Biomedical discoveries
Biomedicine
Positional replication fork instability drives PARP inhibitor cytotoxicity
Nature Cell Biology Published 2026-11-05 Original experimental research; peer-reviewed journal article DOI: 10.1038/s41556-026-02126-0
CRISPR screening replication stress single-molecule imaging
Summary: CRISPR screening and single-molecule imaging link ESCO2 loss to G-quadruplex stress at pericentromeric replication sites. PARP inhibition disrupts the associated TIMELESS–DDX11 protective response, increasing damage in cancer cell models.
Why it matters: The location and structure of replication lesions may matter alongside their total abundance when interpreting synthetic lethality.
Why for Yiru: An experimental example connecting genetic perturbation with spatially resolved molecular phenotypes rather than inferring mechanism from a response score alone.
Polycomb Marked Promoter Methylation Gains in Aging and Cancer are Independent of Replication
bioRxiv Subject Collection: Cancer Biology Published 2026-11-03 Original research preprint; not peer-reviewed DOI: 10.64898/2026.10.30.763235
clonal epigenomics organoid perturbation selection versus induction
Summary: Low-input methylomes track single-stem-cell-derived colon organoid clones with inducible BRAF V600E. The study reports clonal selection without accelerated Polycomb-promoter methylation, challenging a simple model in which oncogenic signaling directly creates those marks.
Why it matters: Longitudinal clonal experiments can distinguish induction of an epigenetic state from selection of cells already bearing it, a distinction cross-sectional tumor atlases often cannot resolve.
Why for Yiru: Useful for designing perturbation-plus-omics experiments and interpreting cancer-associated methylation trajectories.
Methodology Review: Anderson acceleration: learning a faster fixed-point step from recent mistakes
English · About 10 minutes
Walker-Ni, Evans et al and Saad were inspected in author full text; Toth-Kelley only at publisher-abstract level. The omics transfer is a proposed experiment, not an established result. Saad journal date is publisher-verified 1 July 2025 despite inconsistent arXiv journal-year metadata.
The intuition
Many computational procedures repeat a simple operation until their answer stops changing. The operation may be cheap and reliable, yet painfully slow near its destination. The abstract problem is to find a state that an update leaves unchanged. Instead of replacing the whole solver, ask whether its recent mistakes reveal a better next move. Anderson acceleration keeps a short memory of states and their update residuals, then mixes them using a small least-squares calculation. This is a numerical solver technique, not a new statistical model or a guarantee of biological accuracy. [R4]
Imagine repeatedly walking toward a landmark but covering only one fifth of the remaining distance each time. Your last two positions already reveal something about the distance still missing. Extrapolation uses that information to predict a farther step. In several dimensions, different coordinates can approach their limits at different rates, so a single guessed stride is inadequate. A small collection of residual vectors supplies several directions of evidence. The important distinction is between making those stored residuals cancel and proving that the newly proposed state actually has a small residual. For a nonlinear update, those are different questions. This walking analogy is explanatory, not an empirical result.
A small worked example
Consider the scalar update \(g(x)=0.8x+1\), starting from \(x_0=0\). Here \(x\) is a real number, \(g\) is the update function, and the subscript counts iterations. Its fixed point is \(x_*=5\) because applying the update to five returns five. Ordinary iteration gives \(x_1=1\) and then \(x_2=1.8\). Define the residual as the proposed change, \(r(x)=g(x)-x\). The two stored residuals are therefore \(r(x_0)=1\) and \(r(x_1)=0.8\). These values are exact in this deliberately constructed arithmetic example.
Choose two weights, \(a_0\) and \(a_1\), that add to one and cancel those residuals: \(a_0+a_1=1\) and \(a_0+0.8a_1=0\). Solving gives \(a_0=-4\) and \(a_1=5\). Mix the two update outputs with these weights: \(x_2=-4g(0)+5g(1)=-4+9=5\). The accelerated proposal lands exactly at the fixed point, whereas the ordinary second iterate is still 3.2 away. The negative weight is essential: this is extrapolation, not a weighted average constrained to stay between old states. The calculation illustrates the undamped algorithm, not an observed performance benchmark. [R1]
Why does exact cancellation work here? For an affine update, taking an affine combination of states and then evaluating its residual equals taking the same combination of their residuals. The weights sum to one, which preserves the constant term. Our residual combination is zero, so the extrapolated state is a root. A nonlinear function does not generally preserve this identity. Also change the coefficient 0.8 to 0.99: the corresponding weights become negative 99 and positive 100. The arithmetic still works exactly, but now small errors in stored numbers can be greatly amplified. Fast extrapolation and numerical fragility can arise together.
The necessary mathematics
Let \(x\) be a vector with \(n\) real coordinates and let \(g\) map such states to states. The target satisfies \(x_*=g(x_*)\). At iteration \(k\), retain indices in a set \(I_k\) containing the current state and at most \(m\) earlier states; \(m\) is the chosen memory length. For each retained index \(i\) define \(r_i=g(x_i)-x_i\). Write \(\lVert v\rVert_2\) for the Euclidean length of a vector \(v\). Choose weights \(\alpha_i\) by the constrained least-squares problem below, then form the undamped proposal. Every sum ranges over the retained indices.
\[\min_{\alpha}\left\lVert\sum_{i\in I_k}\alpha_i r_i\right\rVert_2^2\quad\text{subject to}\quad\sum_{i\in I_k}\alpha_i=1\] \[x_{k+1}=\sum_{i\in I_k}\alpha_i g(x_i)\]The symbol \(\alpha\) denotes the collection of weights. No positivity constraint appears. The squared norm and the unsquared norm have the same minimizers. [R1]
A useful independent diagnostic comes from contraction. Suppose a constant \(q\) with \(0\le q<1\) bounds the update according to \(\lVert g(x)-g(y)\rVert_2\le q\lVert x-y\rVert_2\) for all relevant states \(x\) and \(y\). The triangle inequality gives an error bound: \(\lVert x-x_*\rVert_2\le\lVert g(x)-x\rVert_2/(1-q)\). Indeed, insert and subtract \(g(x)\) and move the contracted error term to the left. This elementary derivation explains why a residual tolerance becomes weaker when \(q\) approaches one. The cited convergence analysis supplies the broader conditional framework for Anderson iterations; it does not make a small residual a universal error certificate without assumptions. [R2]
Assumptions and failure modes
The convergence theory is conditional. Toth and Kelley analyze local convergence for contractive mappings with bounded mixing coefficients. A good underlying fixed-point map matters; least squares does not automatically rescue an arbitrary unstable iteration. Evans and colleagues explain a different boundary: an improvement in the leading, linear error term comes with higher-order terms. Consequently, accelerating a linearly converging iteration is not the same proposition as accelerating an already quadratically convergent method. Their result should not be read as a promise that every accelerated step decreases the true residual. [R2] [R3]
Short memories can become nearly redundant as the solver approaches its limit. Saad discusses numerical implementation and conditioning, which matter because a least-squares problem can be small yet unreliable. As a practical safeguard proposed here, evaluate the actual residual at each candidate, check domain constraints, and reject unacceptable candidates in favor of the base update. Keep a rejection log. If an update must preserve positive parameters or a probability simplex, negative mixing weights deserve explicit checks. Changing batches, random seeds or approximation accuracy also changes what the residual history means. Restarting memory after such a change is an experimental precaution, not a theorem asserted here. [R4]
Verified applications
Walker and Ni demonstrate acceleration of expectation-maximization for a three-component normal mixture with known variances and mixing proportions, including difficult cases with poorly separated components. Their paper also studies nonnegative matrix factorization and domain-decomposition examples. These establish uses in statistical estimation and numerical computation, not a tested single-cell analysis workflow. The examples are illustrative, and the paper does not present a comprehensive timing comparison. Evans and colleagues provide additional numerical examples involving nonlinear flow equations. These are evidence that the technique travels across computational settings, while leaving the suitability of a particular biological model unresolved. [R1] [R3]
Possible transfers — speculative
A possible omics experiment would wrap a deterministic latent-variable fitting routine whose full-data update can be exposed as a function. This is a speculative transfer, not a validated omics application of the cited papers. Freeze the dataset, initialization, stopping rule and objective. Compare the base solver with a short-memory accelerated variant, measuring wall-clock time, update evaluations, final objective, residual and feasibility violations. Include rejected proposals in the computational cost. First use simulated data with a known solution or an independently converged reference; only then inspect a biological dataset.
Do not treat a faster fit as a better biological answer. Two solvers can converge to the same biased model, and different initializations can lead to different local solutions. Check whether downstream cell assignments, uncertainty estimates or parameter summaries change only because one run stopped too early. If the original procedure uses stochastic minibatches, first establish a deterministic baseline rather than carrying over contraction claims. A worthwhile negative result would be that least-squares overhead outweighs saved updates, or that constraint violations make acceleration unattractive. Either result clarifies where the computational bottleneck actually lies.
What to try and what to read next
Start with the scalar example and verify both the ordinary and accelerated second steps. Then add a tiny perturbation to one stored residual and observe how the large-weight case reacts. This separates ideal algebra from finite-precision behavior before a large dataset hides the distinction. Read Walker and Ni for the algorithm, Toth and Kelley for conditional convergence, Evans and colleagues for the limitation around quadratic convergence, and Saad for the broader implementation landscape. The reusable idea is modest: a solver can learn from its recent residuals, but every proposed shortcut still needs a trustworthy stopping test and respect for the original problem.
Reading list
-
[R1] Primary reading: Anderson Acceleration for Fixed-Point Iterations. Homer F. Walker and Peng Ni (2011). Main reading: begin with algorithm 1.1, then the implementation cautions and expectation-maximization example; the full paper exceeds ten minutes.
-
[R2] Related reading: Convergence Analysis for Anderson Acceleration. Alex Toth and C. T. Kelley (2015). Theory reading: distinguish assumptions sufficient for convergence from a universal speedup claim. Evidence here is limited to the publisher abstract.
-
[R3] Related reading: A Proof That Anderson Acceleration Improves the Convergence Rate in Linearly Converging Fixed-Point Methods (But Not in Those Converging Quadratically). Claire Evans, Sara Pollock, Leo G. Rebholz and Mengying Xiao (2020). Boundary reading: understand why improving a linearly convergent solver does not automatically improve a quadratically convergent one. Author manuscript first posted 19 October 2018; version 2 dated 20 February 2019; journal publication 20 February 2020.
-
[R4] Related reading: Acceleration methods for fixed-point iterations. Yousef Saad (2025). Context reading: a broad numerical-analysis survey connecting extrapolation, self-consistent-field calculations and Anderson variants. Author manuscript deposited 15 July 2025, after journal publication; these are one reading.