Research Radar — 2026-10-14

Generated 2026-10-14T05:59:52.355666+08:00 approved RSS sources Curated daily research digest; 20/20 primary RSS feeds and 8/8 PubMed-indexed journal feeds available. Indexing may lag publication; coverage is not exhaustive.

Methods & AI

Computational

2 selected
Computational #1 Read the tokenization idea first; inspect split leakage, module ablations and simple baselines before trusting the reported rankings.

BioMod: Biological Module Tokenization for Single-Cell Foundation Modeling

bioRxiv Subject Collection: Bioinformatics Published 2026-10-13 Preprint; not peer reviewed; primary abstract-level assessment DOI: 10.64898/2026.10.12.758490

Authors: Meng, L., Zhu, J., Han, B., Zang, Y., Li, W., Zong, K., Zhong, J., Yang, Y., Liu, J., Jia, J., Sun, R., Yu, G., Han, J., Chen, Y., Zhang, L., Xu, J.

single-cell foundation models module tokenization representation learning

Summary: BioMod replaces individual gene tokens with 660 overlapping pathway- and data-derived modules. The authors pretrain a 50-million-parameter model on about 73 million human cells and report benefits across annotation, integration and response prediction.

Why it matters: Biological modules offer a compact intermediate representation between genes and whole cells.

Why for Yiru: A useful design comparison for a brain representation model, although mouse transfer and preservation of rare gene-level effects remain untested here.

Computational #2 Audit shared-control leakage before comparing models; consider the gate as an empirical safeguard, not guaranteed improvement or a replacement for animal-level validation.

A split-control criterion for few-shot adaptation in single-cell perturbation prediction

bioRxiv Subject Collection: Bioinformatics Published 2026-10-13 Preprint; not peer reviewed; primary abstract and software-documentation assessment DOI: 10.64898/2026.10.05.756906

Authors: Liang, W.-L., Yang, Y.-B., Zhang, Q.-Y., Feng, J.-C., Wang, Q.-F., Zou, D., Tang, J.-L., Liu, Y., Wang, S.-Y., Duan, J., Bai, Y., Li, Y.-F., Sun, W.-Y., Cao, Y.-F., He, R.-R.

perturbation prediction control splitting evaluation bias few-shot adaptation

Summary: SafeDelta uses disjoint control estimates to judge whether a few-shot correction should be released or the shared predictor retained. The preprint reports that independent-control scoring changed 40 of 52 configuration–dataset ranks; its contribution is an adaptation criterion, not the first discovery of shared-control bias.

Why it matters: An apparent adaptation gain can inherit reference noise rather than reflect transferable perturbation signal.

Why for Yiru: Relevant to sparse mouse perturbation benchmarks: independent experimental units still matter, even when cell-level controls can be split.

Methodology Review: Split conformal prediction: an error bar built from held-out mistakes

English · About 10 minutes

Targeted primary full-text sections inspected for all four readings, including the accessible tutorial preprint and its monograph alias. Numerical exercise independently verified; no biological deployment or benchmark reproduction claimed.

The intuition

A prediction model gives an answer; a scientist also needs to know how much room to leave around it. Imagine forecasting a continuous assay score for a new specimen. Two algorithms can make similar average errors yet fail very differently on difficult specimens. Today’s tool is split conformal prediction: a calibration layer that turns a fixed predictor into a set of plausible future outcomes. Its central question is practical: how large should the error allowance be to achieve a chosen long-run coverage level? [R1]

The intuition is to reserve labeled examples that the model did not train on, measure how surprising their outcomes were, and use their ranked surprises as a ruler. The ruler does not require the neural network’s own confidence estimates to be correct. It does require calibration examples and the future example to be interchangeable in the relevant probabilistic sense. A poorly predictive model can therefore produce valid but unhelpfully wide intervals. Validity and usefulness are separate properties. [R1]

A small worked example

Here is an original numerical exercise using nine independent calibration specimens. A predictor has already been fitted elsewhere. Its absolute errors on these nine specimens, in increasing order, are 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8 and 1.2 score units. We want 80% marginal coverage. Include the future specimen when counting possible ranks: there will be ten positions. Eighty percent of ten is eight, so choose the eighth calibration error, 0.8, as the allowance. The finite-sample rank rule comes from the cited split-conformal construction. [R2]

If the frozen model predicts 2.4 for the new specimen, the resulting interval is [1.6, 3.2]. An observed outcome of 3.0 would be covered; 3.3 would not. These are invented numbers, not a reported biological result. Notice what has not happened: we did not change the point prediction, assume normally distributed noise, or infer the probability of one particular specimen’s outcome. We used held-out errors to choose the width of a repeated-sampling procedure.

Repeat the exercise at 90% coverage. The required rank becomes nine, so the allowance grows to 1.2 and the interval becomes [1.2, 3.6]. At 95%, the required rank is ten, beyond the nine available finite errors. The conservative construction then uses an infinite threshold and returns the entire real line. This deliberately small example exposes the resolution limit of calibration data. Nine specimens cannot support an informative version of this particular 95% split-conformal rule without additional assumptions or a different design. [R4]

One more diagnostic: keep the 80% target but replace the largest error, 1.2, by 100. The eighth error remains 0.8, so this example’s interval stays unchanged. Replace the eighth error, 0.8, by 100 instead: after sorting, the eighth becomes 1.2. This is a rank-based calculation, not a standard deviation calculation. It also shows why a single displayed interval conceals sensitivity near the selected order statistic. Try these substitutions by hand before applying the tool to thousands of predictions.

The necessary mathematics

Let \(f\) be the predictor fitted without calibration outcomes; \(x_i\) and \(y_i\) are calibration specimen \(i\)’s input and observed response. With \(n\) calibration specimens, use absolute-error scores \(s_i=\lvert y_i-f(x_i)\rvert\). Let \(\alpha\) be the desired miscoverage probability, between zero and one. Choose \(k\) as the smallest integer at least \((n+1)(1-\alpha)\), and let \(q\) be the \(k\)th smallest score; if \(k=n+1\), set \(q=\infty\). The prediction set for input \(x\) is the interval below. [R2]

\[C(x)=[f(x)-q, f(x)+q].\]

Here \(C\) denotes the set-valued prediction rule. For exchangeable calibration and future scores, the rank argument gives the following marginal coverage statement. \(X_{\mathrm{new}}\) and \(Y_{\mathrm{new}}\) denote the new input and response; \(\Pr\) is probability over calibration and the new observation, conditional on an independently fitted predictor. [R2] [R4]

\[\Pr(Y_{\mathrm{new}}\in C(X_{\mathrm{new}}))\ge 1-\alpha.\]

Without ties, all future-score ranks have equal probability. Ties make this inclusive-threshold construction conservative. The theorem is marginal: it averages over new inputs, rather than promising the target coverage separately at each input or within every biological subgroup. [R4]

Constant-width intervals are only the simplest option. Conformalized quantile regression first learns lower and upper conditional quantile functions, then calibrates their errors on held-out observations. The correction adjusts both endpoints, while the learned endpoints allow width to vary with the input. This addresses heterogeneous noise, but adaptive width alone does not create exact conditional coverage for every subgroup. Treat this as the next reading once the rank calculation feels clear. [R3]

Assumptions and failure modes

Exchangeability is the central boundary. A new laboratory, age range or treatment regime can make future examples unlike the calibration set. Weighted conformal prediction under covariate shift addresses a specific case: the input distribution changes while the response distribution conditional on input remains the same. Its ideal guarantee uses the appropriate known density ratio; estimating that ratio adds practical uncertainty. It is not a general correction for arbitrary biological domain shift. [R4]

For an omics study, write the intended prediction unit on paper before splitting data. Is it another cell from already represented animals, or a new animal? Those are different targets. In the latter case, shuffling cells from the same animals across training and calibration does not establish the required new-animal guarantee. Similarly, do not select the most favorable score after repeatedly inspecting the same calibration outcomes. These are design cautions for the proposed transfer here; the four readings do not certify a particular multi-animal experiment.

Also keep the object of uncertainty explicit. An interval for a future observed response is not automatically a confidence interval for a population mean, a causal effect, or an unobserved molecular state. Measurement noise remains part of the prediction target unless the design models something else. In a proposed workflow, record the response transformation and units next to the interval: a bound for a log-transformed assay score cannot simply be read as a bound in raw count units. This is an interpretation checklist, not an additional conformal theorem.

Verified applications

Romano and colleagues evaluated conformalized quantile regression on eleven regression benchmarks, including medical expenditure, protein tertiary-structure properties, bike sharing and other tabular problems. Their paper reports improved interval efficiency across most tested datasets and discusses conservative random-forest results on two Facebook datasets with tied scores. This is verified methodological benchmark evidence from the full paper, not a claim that conformal prediction has already solved single-cell perturbation uncertainty or that one implementation always wins. [R3]

Possible transfers — speculative

A deliberately bounded omics pilot would predict one prespecified specimen-level pathway score from a fixed feature representation. This is a speculative experiment proposed here. Split independent specimens into model-fitting, calibration and final evaluation groups, keeping repeated measurements together. Freeze the target, predictor and calibration rule before looking at the final evaluation outcomes. Report both coverage and interval width, because a trivially huge interval can cover nearly everything while helping nobody.

Before scaling that pilot, ask whether the calibration specimens span the deployment population’s ages, tissues and assay conditions. Evaluate coverage descriptively within prespecified groups where sample sizes permit, without presenting those diagnostics as a new guarantee. If a future perturbation class is absent from calibration, label that extrapolation explicitly. For high-dimensional expression profiles, separately designed per-gene intervals would not by themselves promise simultaneous coverage of the whole profile. Start with a single interpretable outcome so these distinctions remain visible.

The useful habit is to attach a prediction target, calibration unit and coverage statement to every uncertainty plot. For today’s ten-minute exercise, reproduce the nine-score calculation at 80%, 90% and 95%, then explain why the last answer becomes uninformative. Read the main tutorial for orientation, the regression paper for the exact construction, the quantile-regression paper for variable widths, and the covariate-shift paper for the limits of portability. Stop before deploying if you cannot say what makes the future observation comparable to the calibration observations.

Reading list

  • [R1] Primary reading: A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. Anastasios N. Angelopoulos and Stephen Bates (2021). Preprint; not peer reviewed. Main tutorial: read the opening construction and regression examples. The arXiv manuscript began July 15, 2021 and was revised December 7, 2022; the differently titled 2023 monograph, Conformal Prediction: A Gentle Introduction, has DOI 10.1561/2200000101.

  • [R2] Related reading: Distribution-Free Predictive Inference For Regression. Jing Lei, Max G’Sell, Alessandro Rinaldo, Ryan J. Tibshirani and Larry Wasserman (2018). Technical foundation: inspect Algorithm 2 and Theorem 2.2. First preprint April 14, 2016, revised March 8, 2017; accepted author version posted April 3, 2017; final article published online June 8, 2018 in JASA 113, 1094–1111.

  • [R3] Related reading: Conformalized Quantile Regression. Yaniv Romano, Evan Patterson and Emmanuel J. Candès (2019). Adaptive intervals: read sections 3–4, then the benchmark limitations. First preprint May 8, 2019; also NeurIPS 2019.

  • [R4] Related reading: Conformal Prediction Under Covariate Shift. Ryan J. Tibshirani, Rina Foygel Barber, Emmanuel J. Candès and Aaditya Ramdas (2019). Boundary conditions: see why distribution shift requires extra assumptions. First preprint April 12, 2019, revised July 6, 2020; also NeurIPS 2019.

Page Last Updated: