Research Radar — 2026-11-18
Methods & AI
Computational
Efficient count-based models improve power and robustness for large-scale single-cell eQTL mapping
Nature Genetics Published 2026-11-10 Peer-reviewed computational methods study; primary publisher evidence reviewed. DOI: 10.1038/s41588-026-02830-5
single-cell eQTL pseudobulk negative binomial calibration
Summary: jaxQTL fits negative-binomial count models to donor-level pseudobulk data. In 982 OneK1K participants across 14 PBMC cell types, it identified more eGenes than linear models and assessed replication in independent cohorts.
Why it matters: It connects count distributions, cell abundance and statistical power. Its speed advantage is relative to single-cell mixed models, not every linear tool; permutation testing remains a bottleneck.
Why for Yiru: Useful for immune single-cell genetic regulation and for comparison with the previously covered SAIGE-QTL, while distinguishing donor-level analysis from continuous cell-state analysis.
Multimodal cell foundation models require synergistic integration beyond alignment
Nature Machine Intelligence Published 2026-11-09 Peer-reviewed computational benchmark and theory study; primary publisher evidence reviewed. DOI: 10.1038/s42256-026-01328-3
multimodal integration spatial transcriptomics foundation models synergy
Summary: The Synergistic Information Score distinguishes redundant from complementary multimodal information. Comparisons of ten fusion methods on spatial transcriptomic tasks indicate greater benefit for complex niche prediction than for tasks already well served by one modality.
Why it matters: Better alignment does not necessarily mean better biological integration. Linear alignment of frozen encoders may capture shared signals without exploiting complementary information; this does not invalidate every fusion architecture.
Why for Yiru: Useful for spatial-omics and cell-foundation-model evaluation, particularly for avoiding comparisons against weak unimodal baselines.
Methodology Review: Gillespie simulation: let the next event choose the clock
English · About 10 minutes
Primary full paper for direct SSA and BioNetS; primary abstracts for next-reaction and tau-leaping contrasts; numerical example checked by hand.
The intuition
A reaction simulation need not ask what happens during each tiny time step. The Gillespie direct method instead samples when the next reaction occurs and which channel fires. It is exact for its specified stochastic reaction model, not automatically for a real cell. [R1]
The useful conceptual shift is from a smooth movie to an event log. A trajectory records intervals of unchanged molecular counts separated by jumps. That representation makes the random clock part of the scientific model rather than an implementation nuisance. It also forces a useful question before coding: are we trying to represent the history of one system, the average of many systems, or the noisy measurements made on those systems? Those are different objects. Today we use a small invented example to keep all three apart.
A small worked example
Consider one molecular species with a count that starts at four. Molecules enter singly at a constant rate of two per minute. Each existing molecule independently disappears at a rate of one half per minute. The numerical choices below are illustrative, not fitted RNA kinetics or experimental evidence. At count four, the birth channel has propensity two and the death channel also has propensity two. Their total is four events per minute. Accordingly, the next event has a mean waiting time of one quarter of a minute, but that is a mean, not a compulsory interval.
For a hand calculation, choose the first uniform number to be exp(−1) and the second to be 0.8. These numbers are deliberately selected to illustrate the transformation, not presented as a random experiment. The waiting time is 1/4 minute. Order birth before death in the channel list. The selection threshold is 0.8 times four, or 3.2. It exceeds the birth interval ending at two, so the event is a death. The count changes from four to three at time 0.25 minute. No intermediate fractional molecule is inserted.
Recalculate after the event. Birth still has propensity two, but death now has propensity 1.5, giving a total of 3.5. Choose the next pair of illustrative uniform numbers as exp(−0.7) and 0.2. The waiting time is 0.7/3.5, or 0.2 minute. The threshold is 0.7, which lies in the birth interval. The count returns to four at time 0.45 minute. Reusing the old total of four would have produced the wrong clock; keeping both channel probabilities equal would also have been wrong after the first event.
Several checks follow without a large simulation. Starting from four, the probability of no event during half a minute is exp(−2), approximately 0.1353. The probability that the next event is a birth is one half. Starting instead from zero, the death propensity is zero, so only births are possible. This last check is especially valuable: a valid implementation should not produce negative counts and then hide them by clipping. At zero the mean time to the next birth is half a minute. The change in waiting time reflects the changed set of available events.
Now compare the event trace with its ensemble mean. For this linear birth–death example, let m denote the expected count. Its rate of change is two minus one half of m. Starting at four therefore gives a constant expected count of four, even though the explicitly constructed trajectory visits three. There is no contradiction: an expectation is not a prediction that each realization stays flat. This closure works because the death propensity is linear. Do not infer from this example that replacing every count by its mean is justified for nonlinear reaction networks. Finally, a measured count could differ again because an instrument captures only part of the molecular population. That measurement layer was not included in the event calculation.
The necessary mathematics
Let \(x\) be the count vector, \(a_j(x)\) the propensity of reaction \(j\), and \(a_0(x)=\sum_j a_j(x)>0\). With propensities constant between events, the joint waiting-time and channel density is
\[p(\tau,j\mid x)=a_j(x)\exp[-a_0(x)\tau].\]For independent uniforms \(u_1,u_2\in(0,1)\), set \(\tau=-\log(u_1)/a_0(x)\) and choose the first channel whose cumulative propensity exceeds \(u_2a_0(x)\). Advance time by \(\tau\) and update \(x\leftarrow x+\nu_j\), where \(\nu_j\) is that channel's stoichiometric count change. Recompute affected propensities and repeat. The exponential survival law explains the clock; normalized propensities explain channel selection. [R1]
Assumptions and failure modes
The basic formulation assumes a well-mixed Markov reaction model. Exact sampling does not validate its rates or reaction list. [R1]
State representation matters. BioNetS distinguishes discrete counts, continuous chemical Langevin approximations and hybrid models. Its numerical examples test solver behavior on selected systems, not universal biological correctness. A continuous option is therefore a modeling choice to justify, not merely a faster setting with an identical guarantee. [R4]
Gillespie’s later tau-leap paper explicitly trades exactness for potentially faster approximate simulation. It reports simple test systems rather than a blanket accuracy guarantee. This is a different lever from optimizing an exact event scheduler: before comparing runtimes, specify whether both methods are meant to sample the same law. [R3]
For our autonomous toy model, zero total propensity would mean no further events. An externally scheduled change could invalidate that conclusion, so a simulator with interventions needs to represent their timing explicitly. Similarly, replacing a time-varying hazard by its value at the previous event requires separate justification. These are boundary checks for a proposed implementation, not claims that the elementary formula covers every extension.
Verified applications
Gibson and Bruck present the exact Next Reaction Method and apply it to a model of the lambda-phage lysis/lysogeny decision circuit. Their author-repository abstract describes a performance comparison against standard and optimized standard methods. That supports a concrete computational application and an exact-algorithm distinction; it does not establish that simulation alone predicts an experimental biological decision. The abstract also describes extensions beyond constant-rate Markov settings, but this column does not reproduce or validate those extensions. [R2]
BioNetS is a second documented application: software that represents biochemical networks with discrete and continuous variables, including hybrid models. Its discrete component uses a Gillespie implementation, while continuous variables use chemical Langevin equations. The authors report numerical accuracy and efficiency tests. This is evidence of a modeling and software workflow, not a claim that the historical download links still work or that the package is the best current implementation. [R4]
Possible transfers — speculative
A possible research use is to build a simulation-based stress test for a proposed burst-detection statistic. First define competing count-generating models, then add the same explicit observation model to each. Vary sampling intervals and capture efficiency separately from kinetic parameters. Ask whether the statistic distinguishes the kinetic alternatives after observation, rather than merely separating latent trajectories that an experiment cannot resolve. This is a proposed transfer, not a result established by the cited papers or a recommendation that such a test identifies biological mechanisms uniquely.
Keep a small exact event simulator as a reference on tractable cases when developing a faster approximation. Compare distributions of a prespecified endpoint, first-passage time or event count across independent ensembles, rather than expecting two trajectories with different random-number consumption to coincide. Sampling uncertainty belongs in that comparison. A visually convincing pair of traces is not an accuracy assessment, and agreement on one mean does not check rare outcomes.
What to try and what to read next
The durable lesson is to specify the event law before choosing the time grid. Reproduce the two-event calculation, check behavior at zero count, then state which observable your experiment actually records. Only after those checks should runtime optimization become the central question. Exactness is a statement about sampling a mathematical model; model adequacy and measurement adequacy remain separate scientific responsibilities.
Reading list
-
[R1] Primary reading: Exact stochastic simulation of coupled chemical reactions. Daniel T. Gillespie (1977). Read the direct-method derivation for why random waiting times and channel probabilities jointly reproduce the specified jump process.
-
[R2] Related reading: Efficient Exact Stochastic Simulation of Chemical Systems with Many Species and Many Channels. Michael A. Gibson and Jehoshua Bruck (2000). Compare an exact implementation improvement with methods that change the target simulation law.
-
[R3] Related reading: Approximate accelerated stochastic simulation of chemically reacting systems. Daniel T. Gillespie (2001). Read as the contrasting approximation strategy, not as an exact replacement for event-by-event simulation.
-
[R4] Related reading: Biochemical Network Stochastic Simulator (BioNetS): software for stochastic modeling of biochemical networks. David Adalsteinsson, David McMillen and Timothy C. Elston (2004). See how one modeling interface can expose distinct mathematical approximations rather than one universal stochastic solver.