Research Radar — 2026-11-16

Generated 2026-11-16T05:56:43.774405+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

1 selected
Computational #1 Read for measurement and alignment design. Treat virtual-cell modeling as motivation, not a validated predictive result.

Mapping dynamics of subcellular RNA and protein localization and organization using optimal transport

Nature Communications Published 2026-11-15 Peer-reviewed accepted early research article DOI: 10.1038/s41467-026-79450-4

Authors: Jun Zhu; Hongchao Li; Yafang Li; Yiming Li; Guibin Wang; Yimeng Zhang; Xianju Li; Yang Li; Zirui Wang; Sijun Wang; Tianze Ling; Yun Yang; Min Li; Fuchu He; Xuechen Zhu; Xiaowen Wang; Cheng Chang

spatial multiomics optimal transport RNA-protein localization

Summary: SLOT combines axial cryo-sectioning and multi-omics with optimal transport across six Xenopus developmental stages. The authors report stronger, more dynamic protein polarity than RNA polarity.

Why it matters: Spatial RNA organization alone may miss substantial changes in protein deployment.

Why for Yiru: A relevant example of transport-based comparison of subcellular multiomic distributions.

Biomedical discoveries

Biomedicine

1 selected
Biomedicine #1 Read for regulatory-state interpretation; do not generalize mouse mechanism claims into demonstrated human treatment effects.

Chromatin priming of skeletal stem and progenitor cells governs cell fate during bone repair

Nature Communications Published 2026-11-15 Peer-reviewed accepted early research article DOI: 10.1038/s41467-026-79160-x

Authors: Maria Ethel; Cassandre Goachet; Simon Perrin; Pauline Constant; Salma Lakhlifi; Norine Longé; Fanny Coulpier; Nicolas Narboux-Nême; Céline Colnot

single-cell multiomics chromatin priming cell-fate regulation

Summary: Mouse and human stromal atlases, together with mouse single-nucleus RNA and ATAC profiling, connect skeletal progenitor chromatin states with repair fates. The authors report Sox9-linked cartilage and DLX5-linked bone programs.

Why it matters: Similar transcriptomes can conceal distinct regulatory potential.

Why for Yiru: Useful for relating multimodal cell states to experimentally studied fate programs.

Methodology Review: Sparse Jacobian coloring: asking several derivative questions at once

English · About 10 minutes

Main survey and ColPack author manuscripts inspected; Coleman–Moré and the bicoloring preprint inspected at abstract level. No independent software benchmarks or biological validation; omics transfers are proposals.

The intuition

Suppose a simulator has many inputs and outputs, and you want its local sensitivity matrix. Changing one input at a time sounds unavoidable. But imagine that two inputs affect completely different outputs. You can change both together and still identify their separate effects. The useful prior information is not the size of each effect; it is which effects are impossible. Today we examine a computational compression technique, not a method for discovering biological causal edges.

A Jacobian stores first derivatives of outputs with respect to inputs. In column coloring, connect two input columns when their structural nonzeros share an output row. Give connected columns different colors. Columns with the same color can share a directional derivative because their contributions never collide in one row. The main survey explains this through column-intersection graphs and an equivalent bipartite representation. The graph describes conflicts in derivative recovery, not similarity between cells. [R1]

A small worked example

Consider an original four-input example. Let the input vector have coordinates a, b, c and d. Define three outputs as a squared plus b, b squared plus c, and c squared plus d. Evaluate at a = b = c = d = 1. The derivative rows are (2, 1, 0, 0), (0, 2, 1, 0), and (0, 0, 2, 1). These numbers follow by differentiating the three polynomials directly. The example is a calculation aid, not a fitted biochemical model.

Inputs a and c can share red: a affects only output one, whereas c affects outputs two and three. Inputs b and d can share blue: b affects outputs one and two, whereas d affects only output three. The conflicting pairs are a with b, b with c, and c with d. Thus two colors suffice. Perturbing all four inputs together would not suffice, because each row would mix two unknown derivatives.

The red direction is (1, 0, 1, 0). Multiplying the Jacobian by it gives (2, 1, 2). The blue direction is (0, 1, 0, 1), and its product is (1, 2, 1). Recover output one by placing red at column a and blue at column b. Recover output two by placing blue at b and red at c. Recover output three by placing red at c and blue at d. All six nonzeros are recovered from two products rather than four individual column products.

Check the finite-step version too. Increase both red inputs by a small positive step h, subtract the original outputs, and divide by h. The resulting vector is (2 + h, 1, 2 + h). Blue similarly gives (1, 2 + h, 1). The extra h comes from differentiating a square with a forward difference. Coloring did not remove this approximation error. In contrast, exact directional derivatives of these polynomials give the earlier vectors. Distinguish compression error, which is absent here, from the error in how each compressed derivative is computed.

The necessary mathematics

Let F map n real inputs to m real outputs, and define the Jacobian entry \(J_{ij}=\partial F_i/\partial x_j\). Here i indexes outputs and j indexes inputs. Let c(j) be the color of column j. The seed matrix S has n rows and one column per color; its entry is one when the input has that color and zero otherwise. Compressed differentiation produces \(B=JS\). For every structural nonzero, direct recovery gives \(J_{ij}=B_{i,c(j)}\), provided no other structurally nonzero column in row i has color c(j). This condition explains the graph rule. [R1]

Coleman and Moré formulate the finite-difference evaluation problem through graph coloring and report efficient grouping algorithms. Their work concerns how sparsity reduces function evaluations, not a guarantee that arbitrary observed zero values are permanent zeros. In the example above, a forward difference uses a baseline evaluation plus one perturbed evaluation per color. A fair timing comparison must account for the baseline and use the same difference rule on both sides. [R2]

Assumptions and failure modes

A crucial distinction in our example is structural versus accidental zero. At a = 0, the derivative of the first output with respect to a happens to vanish. That does not make a irrelevant: at a = 1 it is two. If you build the pattern from that one numerical evaluation, you might incorrectly omit a. As a diagnostic exercise, repeat the recovery at zero and at one using the original structural pattern. It should remain valid at both points. A small coefficient also does not justify silently deleting a dependency.

Compression is not guaranteed to be dramatic. Montoison and colleagues point out that a dense row can force column-only coloring to use one color per input; a dense column similarly obstructs row-only coloring. Their preprint develops bicoloring strategies and software to address such cases. Its abstract is the evidence inspected here, so no speedup factor or benchmark superiority is asserted. Treat this as a reason to inspect the dependency pattern before choosing an algorithm. [R4]

In a practical test, time the complete workflow rather than just the directional products. ColPack separates pattern determination, coloring, compressed derivative evaluation and recovery. Its greedy orderings seek useful colorings; they do not promise the minimum color count for every matrix. Reusing a coloring can help when the same structure is evaluated repeatedly, but the reuse is valid only while that structural assumption remains justified. [R3]

Verified applications

ColPack is an established scientific-computing application of this idea. The author manuscript describes software for coloring and derivative recovery, interfaces with ADOL-C and ADIC2, and evaluations using real-world and synthetic graphs. These are verified software and benchmark reports, not an independent reproduction here. They support the claim that compression is implementable in derivative pipelines, but not that an arbitrary omics analysis will become faster. No inspected reading establishes biological prediction or causal discovery from the coloring alone. [R3]

Possible transfers — speculative

A speculative omics transfer would be sensitivity analysis of a deliberately sparse reaction or spatial-interaction simulator. First list which state variables each local update can depend on. Then test compressed and uncompressed derivatives on a tiny instance before increasing the number of cells. Preserve the same numerical precision, tolerances and state values in both calculations. If global normalization makes every output depend on every input, the apparent local sparsity may disappear. That is a modeling fact to measure, not an inconvenience to hide.

For a first experiment, add one extra dependency to the toy model: let output one also contain c. The previous red grouping now combines derivatives with respect to a and c in that output. Recolor the conflict graph or retain an additional direction; do not decode the old product as if nothing changed. This proposed stress test makes a useful acceptance criterion: a pipeline must detect invalid structure, not merely produce plausible looking derivative values. It also separates computational correctness from whether the simulator matches experimental observations.

The core question is not whether many derivatives are small. It is whether known dependencies allow several derivative questions to be asked without mixing their answers. Start with a hand-verifiable function, keep its structural pattern explicit, and compare every recovered entry with an independent calculation. Then evaluate total runtime and memory on the intended model. Stop if pattern assumptions cannot be defended or compression provides no practical saving. A faster calculation is valuable only after you know it computes the same object.

Reading list

Page Last Updated: