Neuroscientists can now read every gene a single nerve cell expresses and, at the same time, watch where that cell sends its long-range axon. Platforms like Projection-TAGs deliver retrograde viruses carrying RNA barcodes into distant brain regions; axon terminals pick them up, transport them back to the cell body, and the barcode then labels that neuron with its projection identity alongside a full single-nucleus transcriptome (Projection-TAGs). The tempting inference is that genes whose expression tracks a projection target are the genes that build that projection. Two well-documented failures break this inference. First, mRNA is a poor proxy for the physical machinery of wiring: across mammalian tissues, transcript levels predict actual protein abundance only 40–60% of the time, leaving roughly half the variance unexplained (Adhikari) — and axon guidance and synaptic adhesion are executed by proteins, not transcripts. Second, a gene that correlates with a projection target is usually not a driver but a passenger: it may sit downstream of the same transcription factors, or share a common cause, and the correlation survives even though the gene does nothing for the wiring. Finally, even the short lists that survive statistical filtering are almost never causally tested, because perturbation experiments scale orders of magnitude more slowly than sequencing (the functional validation gap).
This proposal builds a single pipeline that confronts all three failures at once, and states up front what it is not: it is not another transcriptome-to-connectome predictor, and not another RNA-to-protein imputation tool. It is a ranking and triage method. We construct a joint partial-correlation graph over three observed layers — transcript counts, predicted protein abundances, and projection-target labels — and rank genes by their partial correlation with a target conditioned on everything else in the graph, evaluated at the protein layer where wiring actually happens. The top-ranked genes then go to a small, deliberately chosen perturbation experiment rather than a genome-wide screen. The core claim is specific: conditioning on the full network at the protein layer will separate direct wiring genes from co-expression passengers, and its short lists will have a higher causal hit rate than lists from differential expression or raw co-expression.
Core Claim
The mechanism is a Gaussian graphical model — a graph in which an edge means two variables are dependent even after conditioning on all others, estimated through the inverse covariance (precision) matrix. We extend the multi-layer shrinkage estimator of DRAGON, which assigns each omic layer its own regularization strength, from its original two bulk layers (methylation and RNA in 765 TCGA breast tumors) to three single-cell layers: RNA counts, protein abundances predicted by a generative translation model, and one-hot projection-target labels. For a protein-coding gene $g$ and target $t$, the score is the partial correlation read off the joint precision matrix $\Theta$:
$$\rho_{gt \cdot \text{rest}} = -\frac{\Theta_{gt}}{\sqrt{\Theta_{gg}\,\Theta_{tt}}}$$
Genes are ranked by $|\rho_{gt \cdot \text{rest}}|$. Because $\Theta$ mixes layers, a transcript can only reach a target label through its protein feature, so the score automatically penalizes genes whose mRNA is abundant but translationally repressed or rapidly degraded — exactly the false positives that models like SynaptoGen, which feed raw transcript vectors $X$ into a protein-interaction compatibility matrix $O$ (Boccato et al.), must accept. And because the score conditions on every measured variable, it removes the indirect gene–target associations that inflate pairwise analyses, the same correction that made plain partial correlation beat sophisticated causal inference algorithms on physical synaptic recovery in the 2014 Connectomics Challenge (Sutera et al.) and that reallocated spurious connectivity hubs in whole-brain benchmarks (Liu et al.).
Research Program
The program has three components. The first is the translation layer: a Zero-Inflated Negative Binomial variational encoder that absorbs count noise and dropout into distributional parameters, following scChord's design (Zhang et al.), but with one structural change. scChord's authors report that its fidelity on Correlation Matrix Distance — a metric comparing whole predicted protein–protein covariance matrices — degrades as target panels grow toward hundreds of proteins, because unconstrained flow matching has no prior on the joint covariance. Our encoder is trained with a precision-matrix penalty on the protein manifold, which is precisely the structure the second component needs. It is trained and evaluated on paired CITE-seq benchmarks (161,000+ cells and 228 proteins in GSE164378; ~8,600 cells and 13 proteins in GSE100866) before being applied to neuronal datasets.
The second component is the joint inference layer. Here the honest obstacle is that single-cell counts are zero-inflated and discrete, while precision estimation assumes approximate Gaussianity — a tension documented for discretized conditional-independence testing in the file's methodological literature. Our approach is to run the precision estimation not on raw counts but on the continuous latent residuals from the translation encoder, which are Gaussian by construction, with the projection labels appended as a third block. Layer-specific shrinkage parameters are estimated as in DRAGON, and edges are filtered against DRAGON's analytical null distribution. The output is a ranked list per projection target, with the protein-layer partial correlation as the score.
The third component is validation triage. Rather than proposing a genome-wide screen, we take the top and bottom deciles of the ranked list — top decile predicted drivers, bottom decile predicted passengers with equal marginal correlation to the target — and perturb both in situ, using Perturb-FISH-style lentiviral guide amplification and multiplexed imaging (Perturb-FISH), which retains the spatial and tissue context that dissociated Perturb-seq destroys, and whose perturbation profiles matched dissociated Perturb-seq gold standards at up to 92% correlation. The decisive quantity is the difference in hit rate between the two deciles. The primary organism for the full loop is C. elegans, because it has both a complete single-cell transcriptomic atlas (70,296 profiles resolving 128 neuron types, with 95.9% annotation coverage and markers enriched up to 24-fold in bulk validation; Taylor et al.) and an invariant, fully mapped 302-neuron connectome — a wiring ground truth that no mammalian system offers. Mouse cortex Projection-TAGs data provides the mammalian projection-target layer.
Baselines
The do-nothing reference point is raw Pearson co-expression between each gene and each projection label. This is what most marker-gene analyses effectively report, and the evidence that it distorts network structure is quantitative: across 326 subjects of the HCP S1200 cohort, structure–function coupling ranged from $R^2 = 0$ to $0.25$ purely as a function of the chosen interaction statistic, and precision-based measures achieved a fingerprinting identifiability index above 2.1 versus roughly 1.5 for Pearson correlation, with Pearson systematically manufacturing hubs out of shared redundancy (Liu et al.). We will run it on identical inputs, genes, and holdout splits, with no tuning budget, and beating it means higher target-prediction AUROC and a higher perturbation hit rate in the validation arm — if conditioning does not improve on Pearson here, the proposal is wrong in its central mechanism.
The strongest existing method for the wiring question is SynaptoGen, which derives synaptic weights from transcriptomic profiles through a learned compatibility matrix and achieved up to $r = 0.92$ correlation with empirical C. elegans co-expression data, plus 100% success on MountainCar-v0 and 47% on LunarLander-v2 with pre-wired feedforward networks (Boccato et al.). It is the direct competitor because it is the only existing model that maps single-cell transcriptomes onto a physical connectome. No reported result was found on Lacuna for SynaptoGen's edge-recovery AUROC against the actual connectome — its 0.92 is correlation with co-expression, not wiring recovery — so we will retrain it ourselves on the same CeNGEN profiles and holdout neurons, with matched feature sets and equal compute, and compare AUROC for chemical-synapse adjacency prediction. The strongest method for the translation question is scChord itself, which achieved the lowest RMSE and highest Pearson correlation across CITE-seq benchmarks against totalVI, sciPENN, and scTranslator while training on 100,000 cells in about 27 minutes with a 229 MB GPU footprint (Zhang et al.) — exact PCC/RMSE table values were no reported result found on Lacuna, so we reproduce its numbers under our own rerun with the same tuning budget. The key ablation isolates the protein layer: a single-layer RNA-only version of our joint graph, with identical shrinkage machinery and conditioning. DRAGON showed that layer-stratified shrinkage improved edge-recovery AUC by roughly 0.14 over single-penalty graphical models on bulk methylation-plus-RNA data (Shutta et al.), and our ablation tests whether the analogous gain appears when the second layer is predicted protein rather than methylation. A second ablation swaps partial correlation for marginal correlation at the protein layer, isolating the conditioning ingredient from the translation ingredient.
Why not just…?
Why not just run SynaptoGen on better single-cell data? Because its misspecification is structural, not a data-quality problem: it assumes the molecular features meeting at a synaptic cleft are deterministic linear functions of transcript abundance, while the proteomic synthesis in the file shows mRNA fails to predict protein 40–60% of the time (Adhikari), and neuronal mRNA half-lives of hours coexist with synaptic protein half-lives of weeks — a neuron can silence a gene whose protein product is still wiring its axon. More data cannot fix a wrong layer. Experiment E2 settles this directly: the protein-swap ablation replaces SynaptoGen's transcript inputs with our predicted protein abundances, and if connectome-recovery AUROC does not move, we would conclude the discordance correction is not the operative ingredient and report that honestly.
Why not just impute proteins with scChord and then correlate with targets? This is the most serious alternative, and we treat it as a genuine threat rather than arguing around it: it reuses an existing tool and adds almost nothing. Three things suggest it will fall short. It propagates per-cell imputation error without any mechanism to correct it; it scores genes by marginal correlation, which Liu et al. show manufactures spurious associations and which Sutera et al. show loses to conditioning even on clean calcium data; and the ~67% barcode dropout in Projection-TAGs single-nucleus data punishes two-stage pipelines whose first stage never saw the barcode (Yang et al.). Our joint estimation instead shares shrinkage strength across all three layers and lets the precision matrix integrate dropout through the encoder's latent space. But we will run exactly this pipeline — scChord followed by Pearson, and scChord followed by partial correlation — as a first-class baseline in E2 and E3 with matched compute. If scChord-then-partial-correlation matches the joint model, the contribution collapses to a glue pipeline and we will say so; if the joint model wins only at high dropout and large protein panels (where scChord's own covariance fidelity degrades), that is itself the result.
Why not just take differentially expressed marker genes for each projection target? Tasic et al. showed that excitatory projection neurons carry fine-grained, region-specific transcriptomic signatures aligned with their axon targets (Tasic et al.), and Goldberg et al. showed regional expression predicts macro-scale connectivity (Goldberg et al.), so markers plainly exist. But the functional validation gap exists precisely because statistical association cannot separate the causal driver from linked passengers and reactive feedback elements, and transcript families (cell adhesion molecules in particular) are rife with paralog redundancy that makes single knockouts uninformative when the wrong family member is chosen. The E4 hit-rate experiment is designed to decide this: differential-expression ranking and partial-correlation ranking receive the same perturbation budget, and the comparison is the fraction of candidates whose loss actually alters wiring.
Proposed Experiments and Decisive Checks
Experiment E1 validates the translation layer on GSE164378 and GSE100866 against scChord, totalVI, sciPENN, and scTranslator under matched compute, with per-protein PCC and RMSE plus Correlation Matrix Distance on the full 228-protein panel — the metric on which scChord itself reports degradation at scale. Experiment E2 is the wiring test: predict chemical-synapse adjacency in the C. elegans connectome from CeNGEN profiles, reporting AUROC and F1 for our joint model, SynaptoGen retrained on the same profiles, Pearson co-expression, and the scChord-then-correlate pipelines. Experiment E3 moves to the mouse Projection-TAGs cortical dataset with its seven downstream targets, treating the ~30% of collateralized neurons as a multi-label prediction problem and reporting per-target and multi-label AUROC under leave-one-target-out splits, with differential-expression markers as the competing ranking. Experiment E4 is the causal triage: in situ perturbation of top-decile versus matched-bottom-decile candidates in C. elegans with wiring readouts, anchored by a Perturb-seq concordance check of the in-situ protocol (Perturb-FISH reported up to 92% correlation with Perturb-seq gold standards). The decisive checks are, in order: does conditioning beat Pearson on target prediction at equal inputs (E3); does the protein layer beat the RNA-only ablation on connectome recovery (E2); and does the top decile beat the bottom decile under perturbation (E4). The first failure anywhere in that chain refutes the corresponding claim, and the pipeline is built so that each claim can fail independently.
References
- Projection-TAGs enable multiplex projection tracing and multi-modal profiling of projection neurons
- A high-stringency blueprint of the human proteome
- DRAGON: Determining Regulatory Associations using Graphical models on multi-Omic Networks
- Optimizing Genetically-Driven Synaptogenesis
- Simple Connectome Inference from Partial Correlation Statistics in Calcium Imaging
- Benchmarking methods for mapping functional connectivity in the brain
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Simultaneous CRISPR screening and spatial transcriptomics reveal intracellular, intercellular, and functional transcriptional circuits
- Molecular topography of an entire nervous system
- Shared and distinct transcriptomic cell types across neocortical areas
- Gene Expression in the Rodent Brain is Associated with Its Regional Connectivity