Doctors can sample blood easily, but the organs they care about, such as the brain or liver, usually cannot be sampled at all. Blood-based molecular markers, including DNA methylation (chemical tags on DNA that switch genes on or off in a cell-type-specific way), often fail to reflect what is happening inside those organs. Studies keep finding that blood mostly records systemic immune noise rather than organ pathology. Internet engineers solved a similar problem twenty years ago: measure a few end-to-end paths, then use the shared structure of the network to infer the delays on every unmeasured path. We propose to treat the body the same way. Internal-organ molecular states are the unmeasured links, blood features are the measured paths, and a low-rank statistical predictor chooses which few blood features to measure and how to combine them to reconstruct an organ's profile.
The main question is diagnostic as well as predictive. It is already known that much of the blood–brain similarity comes from shared genetic control, so an organ state can partly be predicted from genotype alone, with no blood draw. The useful and unanswered question is how much of an internal tissue becomes observable from blood beyond what the genome already predicts, and which small set of blood features carries that surplus. Answering it means extending the network-monitoring method in the direction its authors flagged as missing: noise that is correlated across measurements.
Core Claim
A Statistical Framework for Efficient Monitoring of End-to-End Network Properties observed that routing matrices have low effective rank. It combined a Best Linear Predictor (the minimum-variance linear estimate of unmeasured quantities given measured ones) with QR factorization with column pivoting, a greedy linear-algebra routine that picks the most informative, least redundant rows. With this, it reconstructed delays on all 110 Abilene paths with under 1% error using 10% of the paths needed for exact algebraic recovery. The method assumes uncorrelated noise across links. In tissue data that assumption is plainly false: co-methylation modules and linkage disequilibrium make the residuals densely correlated (Characterization of cross-tissue genetic-epigenetic effects and their patterns in schizophrenia).
We claim three things. First, a cross-tissue version of this predictor that conditions on genotype and models correlated residual noise will give a quantitative, per-gene map of what blood can and cannot reveal about each internal tissue. Second, a panel of a few hundred blood features chosen by noise-aware pivoting will match dense penalized regression on all blood features. Third, noise-aware selection will beat the original diagonal-noise selection at every panel size. The network-scalability literature supplies the method and its known failure mode. Tissue divergence supplies the application and the scientific question.
Research Program
Let $y_c$ be the central-tissue profile (for example, brain expression across $G$ genes), $y_p$ the peripheral blood profile, and $g$ the individual's genotype at local regulatory variants. We model
$$ y_p = A_p z + B_p g + \epsilon_p, \qquad y_c = A_c z + B_c g + \epsilon_c . $$
Here $z$ is a $k$-dimensional latent state shared across tissues, with $k$ much smaller than $G$; it plays the role of the physical links that many network paths share. $A_p$ and $A_c$ are the loadings of each tissue on that state. $B_p$ and $B_c$ are genetic effects in each tissue. $\epsilon_p$ and $\epsilon_c$ are residuals with a full, structured covariance $\Sigma_\epsilon$, estimated as low-rank plus block-diagonal, with blocks given by co-expression or co-methylation modules.
For a chosen subset $S$ of $m$ blood features, the predictor is the conditional mean $\hat y_c = \mathbb{E}[y_c \mid y_{p,S}, g]$. The quantity we report is the incremental observability of gene $j$:
$$ \Delta_j(m) = R^2_j(\text{genotype} + \text{blood panel } S_m) - R^2_j(\text{genotype only}), $$
measured in held-out individuals. This is the share of a tissue's variation that only a blood draw can reveal.
Panel selection reuses Chua et al.'s pivoted QR, with one change. It operates on the blood loadings after whitening by $\Sigma_\epsilon^{-1/2}$. Two blood features whose noise is correlated then no longer count as independent evidence. As a second selector we adapt the weighted-$\ell_1$ idea from Sparsity-Promoting Optimal Wide-Area Control of Power Networks. That paper penalized the number of nonzero feedback links and cut 460 possible gains to 37 while staying within 1.5882% of the centralized controller. We apply the same kind of group penalty to the weights on blood features.
Scale is where the network literature pays off. Chua et al. worked with 110 paths. We work with tens of thousands of features per tissue and many tissues at once. The low-rank structure keeps both fitting and selection linear in the number of features.
The primary testbed is GTEx v8. The file describes it as bulk RNA-seq plus whole-genome sequencing across 47 post-mortem tissues, used by Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations. Donors contribute several tissues, including whole blood, so the peripheral–central pairing is within individuals. This is expression, not methylation, and we say so plainly. The research file contains no paired blood–brain methylation cohort. The Lin et al. brain, blood and saliva samples come from different people.
The methylation arm therefore uses a two-sample design. Using Lin et al.'s shared cis-methylation QTLs (genetic variants that shift methylation at nearby CpGs), we estimate the genetically mediated blood–brain covariance. Lin et al. report that 45–73% of these QTLs are shared between tissue pairs, with 85–93% direction consistency. We then ask how much of brain methylation is recoverable through that genetic route alone. This yields a lower bound on methylation observability. The full non-genetic term needs paired samples, which we flag as the natural follow-up.
Baselines
The naive reference predicts each held-out donor's tissue profile from age, sex and technical covariates only. Every other method must beat it, and it anchors the $R^2$ scale.
The strongest genotype-only competitor is MTWAS (Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations). It splits genetic effects into a cross-tissue component and a tissue-specific component, fitted by EBIC-penalized sparse regression, and it reports an average expression-prediction $R^2$ improvement of 47.4% over PrediXcan across the 47 GTEx tissues. It is the strongest competitor because it answers the same question with no blood measurement at all. If it already explains everything blood explains, a blood panel is pointless. We run it on the same donor splits, with its published tuning procedure, and it defines the genotype-only term in $\Delta_j$. The file reports only the relative improvement, so for absolute per-tissue brain $R^2$ there is no reported result found on Lacuna for MTWAS brain-region $R^2$; we reproduce it ourselves.
The strongest blood-using competitor is dense penalized regression from all blood features plus genotype to each central gene. We use elastic net and a Bayesian shrinkage variant in the spirit of the BayesR+ EpiScores of Blood-based epigenome-wide analyses of cognitive abilities, which explained about 4% of variance in cognition in external cohorts, a different target but the same peripheral-surrogate logic. It gets the same folds and the same hyperparameter search budget as our method.
We beat it if a panel of $m \le 500$ blood features reaches its held-out $R^2$ within one standard error. We clearly beat it if we exceed it on low-sample tissues, where the shared low-rank structure should borrow strength.
Two ablations isolate our key ingredient. The first is Chua et al.'s original diagonal-noise predictor with unwhitened pivoted QR. The second is our full predictor with randomly chosen panels of equal size. The claim about correlated noise holds only if whitened selection beats diagonal selection at matched $m$ across a range of panel sizes.
Why not just…?
Why not just rank blood features by their correlation with the target tissue and keep the top ones? Univariate filters pick redundant features that track the same underlying factor. Decoding menopause-induced tissue fibrosis using pan-tissue network inference found that single-gene differential expression barely overlapped across organs, while network-level structure revealed shared hubs. We include top-$m$ correlation filtering as a baseline at every panel size. The prediction is that it saturates early, because many of its picks carry the same information.
Why not just predict the organ from the genome and skip blood entirely? This is the real threat, and we do not argue around it. We make it the yardstick. Lin et al.'s high sharing and direction consistency of methylation QTLs suggest much blood–brain concordance is genetic. The null associations of blood clocks with amyloid burden in Comprehensive analysis of epigenetic clocks reveals associations between disproportionate biological ageing and hippocampal volume are consistent with blood adding little beyond that. If $\Delta_j$ is near zero for most brain genes, that is a publishable negative result: blood panels for those genes are redundant with genotyping. The per-gene map would then say which minority of genes are exceptions.
Why not just fit a large nonlinear model on everything? The steering ceiling in Sparse Autoencoders Find Causal, Lineage-Specific Context Features in Chromatin Foundation Models closed only 11.2% of the cell-line-to-tissue gap and suggests divergence is nonlinear. With a few hundred donors, however, a dense nonlinear model is likely to overfit. We test this directly with a gradient-boosted and an MLP variant on the selected panel and on all features. If nonlinearity on a small panel beats our linear predictor, that is evidence the missing variance is nonlinear rather than unobservable, and we will adopt it.
Proposed Experiments
Experiment one maps incremental observability. For each GTEx tissue, using five-fold donor-level cross-validation, we compute the distribution of $\Delta_j(m)$ for $m$ from 10 to 2,000. We expect the curves to flatten at a few hundred features, if the low-rank premise holds as it did for Abilene, and brain regions to show lower $\Delta$ than other organs.
Experiment two tests correlated noise. We compare whitened and diagonal selection at matched $m$ and report paired differences in held-out $R^2$.
Experiment three is the decisive falsification control. We shuffle blood profiles across donors while keeping each brain sample's own genotype. Incremental observability must collapse to zero. Any residual gain would expose leakage or batch structure.
Experiment four handles confounding. Post-mortem collection conditions shared by all of a donor's tissues could create spurious blood–brain covariance. We therefore recompute $\Delta_j$ after regressing out donor-level collection and death-circumstance covariates. We count only the covariance that survives as observability.
Experiment five, the methylation arm, estimates genetically mediated blood-to-brain methylation predictability from the Lin et al. cohorts. It checks whether the genes with high observability in GTEx expression also sit near shared methylation QTLs. Agreement would support the methylation extension, and disagreement would show that the expression results do not transfer.
References
- A statistical framework for efficient monitoring of end-to-end network properties
- Characterization of cross-tissue genetic-epigenetic effects and their patterns in schizophrenia
- Sparsity-Promoting Optimal Wide-Area Control of Power Networks
- Partitioning and aggregating cross-tissue and tissue-specific genetic effects to identify gene-trait associations
- Blood-based epigenome-wide analyses of cognitive abilities
- Decoding menopause-induced tissue fibrosis using pan-tissue network inference
- Comprehensive analysis of epigenetic clocks reveals associations between disproportionate biological ageing and hippocampal volume
- Sparse Autoencoders Find Causal, Lineage-Specific Context Features in Chromatin Foundation Models