Methods that learn gene regulatory networks from CRISPR screens treat genes as nameless nodes. A gene network is a map of which genes switch which other genes on or off, and these methods recover it by knocking genes down one at a time and watching what changes. Yet genes sit at fixed places on chromosomes, and neighboring genes share packaging and local regulatory switches. When one gene is knocked down, or simply fluctuates, its physical neighbors can move with it without either one regulating the other. We propose to measure how much of what current methods and benchmarks count as "causal edges" is really chromosomal neighborhood. We then model that neighborhood effect as a nuisance and remove it before network inference. Comparative synteny mapping, the study of which gene neighborhoods are conserved across species, supplies the one ingredient raw genomic distance cannot: a way to separate neighborhoods kept together by evolution, where coupling is likely real biology, from neighborhoods that are adjacent by accident, where coupling is more likely an artifact.
The obvious way to join these fields is the opposite move: treat genes in conserved syntenic blocks as more likely to regulate each other, and use that as a prior that pushes edges toward neighbors. We argue that prior is probably backwards for the benchmarks people actually use. Nearness creates shared variation, which is the textbook signature of a hidden common cause. A prior that favors neighbors would reward exactly the edges most likely to be confounded. The claim of this note is that genome position belongs in the confounder model, not the edge prior, and that the claim can be tested directly on existing human perturbation data.
Core Claim
In large single-cell CRISPR screens, the shift in a target gene's expression after a regulator is perturbed contains a component that depends only on the two genes' positions on the chromosome. Regulation between distant genes is sparse and, to first order, should not depend on how far apart the two genes sit. So any systematic dependence of perturbation response on genomic distance can be estimated by pooling over all regulator–target pairs and then subtracted. We expect this component to be large enough that a predictor which knows nothing except chromosomal neighbors scores non-trivially on current benchmark metrics. We expect that removing it changes method rankings and shrinks some apparent advantages.
A secondary, falsifiable claim concerns where comparative synteny earns its place. At matched genomic distance, coupling between neighbors whose adjacency is conserved across vertebrates should replicate across cell lines more often than coupling between neighbors whose adjacency is lineage-specific. If that holds, conservation tells us which neighborhood couplings to keep as biology and which to discard as artifact. If it does not hold, the honest conclusion is that distance alone suffices, and we will report that.
Why This Gap Is Real
The causal-discovery side shows that something is systematically off on real data. CausalBench builds benchmarks from roughly 400,000 single-cell CRISPR profiles in K562 and RPE1 cells. It scores methods by a statistical metric, the mean Wasserstein distance (how far a putative target's expression distribution moves when its putative regulator is perturbed), and by overlap with curated interaction databases. It found that methods using interventional data do not beat observational ones. The CausalBench challenge was won by heuristics: supervised LightGBM on correlations (Guanlab), GRNBoost with distribution tests (BetterBoost), raw mean shifts (MeanDifference), and a sparse linear model with root-cause masks (SparseRC). PerturbCausal showed that networks inferred by GIES from real Perturb-seq data were not robustly above chance against TRRUST, a curated transcription-factor database. These results motivate the proposal: the field's ground truth and its metrics are shaky in ways a structured nuisance could partly explain. All of these works treat genes as unordered labels, so none can see a position-dependent signal.
Two methods could absorb such a nuisance in principle but are not given the structure. SPOT handles hidden confounders generically by steering continuous graph optimization with a learned posterior over which variable pairs are connected at all. It reports skeleton F1 of 0.63 versus 0.48 for ABIC on the 11-variable Sachs data, but it is evaluated only up to 100 variables and its posterior is learned from simulated tables. Sample, estimate, aggregate (SEA) runs classical estimators on random five-variable subsets and denoises them with a pretrained transformer. It processes the 622-gene CausalBench graph in 5 seconds, against 5 minutes for the fastest classical baseline and 9 hours for DCDI. Its subsets are drawn uniformly, so it is blind to position. Our proposal supplies both methods with a specific, biologically motivated confounder rather than asking them to discover one. Mechanisms Matter constrains the design. It shows that perturbation predictions collapse across contexts when the underlying interaction matrix rewires, and it offers no way to say which interactions persist. Our cross-cell-line replication test asks exactly that question for the neighborhood component.
On the synteny side, conservation of gene neighborhoods is strong but not absolute, which sets how we use it. Carlton et al. report 73% gene-order conservation in chromosome cores between rodent and human malaria parasites. Jaillon et al. use pufferfish, with its duplicated blocks, to reconstruct a 12-chromosome bony-vertebrate ancestor, and they show that human chromosomes are mosaics of ancestral blocks. That makes fish a deep outgroup for asking which human adjacencies are ancient. Sun et al. find that even within maize about 10% of genes are nonsyntenic between two inbred lines. Conservation therefore has to enter as a graded score, not a hard label. The file's own caveat that positional conservation does not imply functional invariance is precisely the secondary claim we test, rather than assume.
Research Program
The model is deliberately simple. Let $\Delta_{ij}$ be the observed shift in target gene $j$ when gene $i$ is perturbed, measured as the standardized mean shift or the Wasserstein distance. We write
$$\Delta_{ij} = T_{ij} + g(d_{ij}, c_{ij}) + \varepsilon_{ij},$$
The terms are defined as follows. $T_{ij}$ is the trans-regulatory effect we want; it is sparse and assumed independent of position. $d_{ij}$ is the genomic distance between the two genes, set to infinity when they lie on different chromosomes. $c_{ij} \in [0,1]$ is a synteny conservation score: the fraction of a fixed panel of vertebrate genomes (mouse, chicken, a teleost, with pufferfish as the deep outgroup) in which the orthologs of $i$ and $j$ remain within the same syntenic block. $g$ is a smooth function that is monotone in distance, with $g(\infty, \cdot) = 0$. The function $g$ is fitted by robust pooled regression over all perturbation–target pairs. This works because trans effects are rare and do not depend on $d$, so they act as heavy-tailed noise around the distance trend. Pairs on different chromosomes anchor the zero level. We compute synteny blocks with standard block-detection tooling of the kind OrthoClusterDB provides (OrthoClusterDB), using ortholog calls that require neighborhood support rather than best sequence hits alone.
For observational cells, the analogous nuisance is a local shared factor. Genes inside the same block co-vary through shared chromatin state, which behaves like a hidden common cause. We estimate a leave-pair-out local factor for each block and partial it out of the expression matrix before any discovery method sees the data.
The program then has four parts. First, a diagnosis: estimate $g$ on K562 and RPE1 and report how much of the benchmark's statistical signal it carries. Second, a correction: feed residualized shifts and partialled expression into existing discovery methods without changing them, so that any gain is attributable to the nuisance model. Third, the conservation test: compare the replication of cis coupling across cell lines in conserved versus non-conserved neighborhoods. Fourth, a synthetic check on the CAUSALDGP simulator from Mechanisms Matter. There we inject a known position-dependent term and verify that $g$ is recovered. We also verify that the correction leaves trans-edge recovery untouched when the true $g$ is zero.
Baselines
The naive reference point is a neighbor-only predictor. For each perturbed gene it ranks candidate targets purely by genomic proximity, ignoring expression entirely. It is the cheapest possible explanation of the benchmark signal. If it scores above random on CausalBench's Wasserstein metric, that alone shows position contaminates the evaluation. It needs no tuning, and we will run it on the same K562 and RPE1 splits and the same edge budgets as every other method. Any method we call successful must exceed it on trans pairs, meaning pairs on different chromosomes or more than a few megabases apart, not just on the full edge set.
The strongest existing competitors are the CausalBench challenge winners: Guanlab, BetterBoost, MeanDifference and SparseRC (CausalBench challenge). These are the methods that actually top this benchmark, and SEA (SEA) is the fastest learned aggregator on the same 622-gene graph. We found no reported Wasserstein or false-omission-rate figures on Lacuna for the challenge winners; the file's Guanlab K562 number is truncated in the source. The only reported SEA figure on this benchmark is its 5-second runtime. We will therefore rerun every competitor ourselves from released code, on identical data splits, with an identical hyperparameter search budget per method, with and without our residualization. Beating them means two things. On the trans-only Wasserstein metric and on STRING and CORUM overlap restricted to trans pairs, the corrected version of a method must be at least as good as its uncorrected version. And the uncorrected version's apparent margin over the neighbor-only predictor must shrink on the full metric. Mean and standard deviation over five seeds will be reported, because PerturbCausal showed that single-seed advantages fail to replicate.
Two ablations isolate the key ingredient. The first replaces $g(d, c)$ with a distance-only $g(d)$, which tests whether comparative synteny adds anything over raw position. The second replaces the soft residualization with a hard mask that deletes all pairs within a fixed window. This tests whether modeling the neighborhood is better than simply discarding it. We also run the obvious intersection idea as a competitor: SEA with subsets drawn preferentially from within syntenic blocks. Our hypothesis is that this variant inflates confounded edges.
Why not just…?
Why not just mask out nearby gene pairs? A hard window is the strongest simple alternative, and it may capture most of the artifact. Its cost is that it deletes every genuine short-range regulatory relationship and must choose an arbitrary window. The hard-mask ablation settles this directly. If masking matches residualization on trans metrics and loses nothing on curated cis interactions, we will recommend the mask and present the modeling as unnecessary.
Why not just use genomic distance and skip comparative synteny? This is a real threat to the secondary claim, and the distance-only ablation exists to test it. Distance alone should explain the artifact. Conservation is only useful if, at matched distance, conserved adjacencies replicate across K562, RPE1, Jurkat and HepG2 (the multi-line data used in Mechanisms Matter) better than lineage-specific ones. The maize result that 10% of genes are nonsyntenic within a single species (Sun et al.) shows that conservation scores vary enough to test this. If the replication gap is absent, the paper's contribution reduces to the diagnosis and correction, and we will say so.
Why not just use synteny as a prior that favors edges between neighbors, or guide SEA's subset sampling with it? This is the obvious combination, and the research file lists it as an open gap. But if neighbors share hidden causes, favoring them raises false positives on any metric that checks for trans regulation. We do not argue this from principle alone. The synteny-guided SEA arm is run against uniform sampling, and if it wins on trans-restricted metrics our confounder framing is wrong.
Why not just let a generic hidden-confounder method absorb the effect? SPOT and related tools for hidden confounders could in principle learn this structure. In practice they are evaluated at 100 variables or fewer and learn their posterior from simulated tables with no positional input. Giving them our residualized data is a cheap test of whether explicit structure helps where generic machinery has nothing to grab onto.
Decisive Checks
Three results would decide the proposal. First, the fitted $g$ on K562 and RPE1 must show a clear, monotone distance dependence that is absent for pairs on different chromosomes, and the neighbor-only predictor must beat random on the full Wasserstein metric. If neither holds, the premise fails and the project stops after the diagnosis. Second, on CAUSALDGP with an injected position term, residualization must recover the injected $g$ and restore trans-edge F1 to its level with no injected term, while leaving F1 unchanged when nothing is injected. Third, across the four cell lines, replication of cis coupling at matched distance must be higher in conserved than in non-conserved neighborhoods for synteny to stay in the method. As a further check, we will report whether the gap between interventional and observational methods documented in CausalBench widens or narrows after correction. We make no prediction about the direction, but either outcome bears on how far the field can trust that benchmark's statistical evaluation.
References
- CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data
- The CausalBench challenge: A machine learning contest for gene network inference from single-cell perturbation data
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- Scalable Differentiable Causal Discovery in the Presence of Latent Confounders with Skeleton Posterior
- Sample, estimate, aggregate: A recipe for causal discovery foundation models
- Mechanisms Matter: Transportability of Cellular Perturbation Effects
- Genome sequence and comparative analysis of the model rodent malaria parasite Plasmodium yoelii yoelii
- Genome duplication in the teleost fish Tetraodon nigroviridis reveals the early vertebrate proto-karyotype
- Extensive intraspecific gene order and gene structural variations between Mo17 and other maize genomes
- OrthoClusterDB: an online platform for synteny blocks