Certify or Abstain: Auditing Cross-Layer Claims in Small Multi-Omics Cohorts with Disjoint Assays

Small cancer studies often measure different molecular layers on different patients. One subset of tumors gets DNA methylation profiling, another gets metabolite profiling, and a third gets gene-expression profiling. The authors then tell a story that runs across layers, for example "vitamin C depletion starves the demethylating enzymes, which causes promoter hypermethylation." No single patient was measured on all the layers that story connects. We propose a tool that takes the layers actually measured together, and on whom, and returns one of three verdicts for each cross-layer claim. The claim is either implied by what was observed, refuted by it, or undetermined. Only the third verdict asks for more matched samples. The tool also recommends which samples to match next so that the most undetermined claims get resolved per assay dollar.

Information theory sets the shape of the tool. It cannot be a complete decision procedure. Undecidability of Network Coding, Conditional Information Inequalities, and Conditional Independence Implication proves that deciding whether a set of conditional-independence statements implies another is undecidable in general, and so is deciding whether a linear conditional information inequality holds. Asking "does what we saw about layer pairs force a conclusion about an unseen layer pair?" is a question of exactly this kind. So any honest auditor must be sound, meaning it never certifies a false implication, and must be allowed to answer "undetermined." The open question is empirical: on real cohort sizes and layer structures, how often does a sound-but-incomplete auditor actually reach a verdict, and do the verdicts survive when the missing matched data are revealed?

Core Claim

Cross-layer conclusions drawn from disjoint-subset cohorts can be sorted, with calibrated finite-sample error, into certified, refuted, and undetermined. Today's standard practice is to impute the missing layers or chain pairwise correlations. That practice turns many undetermined claims into confident ones, because the imputation model assumes the very cross-layer link being claimed.

The new pieces are three. The first is the three-valued audit, which carries confidence intervals from sample sizes in the tens through to the verdict. The second is a direct test of whether information-theoretic bounds can certify claims that classical Gaussian covariance-completion bounds cannot. The third is a design rule that picks which additional sample-matched assays to buy.

Why this intersection, and what it is not

The motivating case is Comprehensive metabolomic and epigenomic characterization of microsatellite stable BRAF-mutated colorectal cancer. That study profiled 39 tumors with three assays: whole-genome bisulfite sequencing, RNA-seq, and mass-spectrometry metabolomics. The metabolomics arm had only 24 tumors per group. The assays were run on largely non-overlapping subsets, yet the paper links ascorbate depletion to a 96.4% hypermethylation bias.

The repository side faces the same problem at scale. The LINCS consortium (Comprehensive Proteomics Metadata and Integrative Web Portals Facilitate Sharing and Integration of LINCS Multiomics Data) unified metadata across assays. It states that quantitative harmonization between assay types remains unsolved, and its data production ended in 2020, so new matched measurements now cost money nobody budgeted. These two papers motivate the work. They supply no method.

Li's undecidability results constrain the design: they rule out a complete auditor. The partially fixed-size construction in The Undecidability of Network Coding with some Fixed-Size Messages and Edges strengthens this, since even small fixed alphabets do not restore decidability in general. We borrow one tool from network information theory: computing bounds on unknown information quantities by linear programming over Shannon-type inequalities. We restrict it to a handful of discretized variables, where it is computable.

We do not claim that epigenetic regulation is a network code. We also do not use the topological-brittleness results on 3-source networks (On the Multiple-Unicast Capacity of 3-Source, 3-Terminal Directed Acyclic Networks). The analogy to patient heterogeneity is evocative but untestable with the data we have.

Research Program

The input is a co-measurement graph. Its nodes are molecular variables; an example is the promoter methylation of a gene, its expression, the transporter SLC23A1, and ascorbate. Each set of variables that was measured on the same samples carries the sample count for that set. A cross-layer claim is a statement about a variable pair that was never co-measured, such as "the mutual information between ascorbate level and promoter methylation is at least $\tau$" or "the correlation is negative."

Two bound engines. The auditor computes the range of values the unseen quantity can take, consistent with the observed marginals. It does this with one of two interchangeable engines.

The Gaussian engine finds the range of the missing correlation over all positive-semidefinite completions of the partially observed correlation matrix. For three variables this range is

$$r_{13} \in r_{12}r_{23} \pm \sqrt{(1-r_{12}^2)(1-r_{23}^2)},$$

where $r_{ij}$ is the correlation between variables $i$ and $j$, and only $r_{12}$ and $r_{23}$ were observed.

The information engine discretizes each variable into a few states. For methylation the states are unmethylated, partially methylated, and methylated, which matches the biology. It then solves a linear program over the entropies of all variable subsets. The program enforces the Shannon inequalities, with the observed entropies entering as interval constraints, and its output bounds the unseen mutual information.

From bounds to verdicts. At $n\approx 24$, plug-in entropy and correlation estimates are biased and noisy. We therefore build bootstrap confidence sets for every observed quantity. A claim is certified only if every completion consistent with every point in the confidence set satisfies it. It is refuted only if no such completion does. Every other case is undetermined. This gives coverage guarantees that are conservative by construction.

Design rule. Given a budget of $B$ additional matched assays, the allocator greedily chooses which variable sets to co-measure, and on how many samples. Each choice maximizes the expected shrinkage of the undetermined bounds, estimated by simulating new data from the current confidence set.

Repository artifact. The co-measurement graph and the verdicts are stored as structured metadata next to the data, as a claim certificate. A later reader can then see which published cross-layer links were ever identifiable. This extends the data-selection and model-export governance steps described in Biomedical data repositories require governance for artificial intelligence/machine learning applications at every step. That paper covers consent and access but not inferential validity.

Baselines

Naive reference: correlation chaining. If A correlates with B and B correlates with C, the claim A→C is accepted. This is how many papers combine disjoint layers in their narrative. We run it on the same masked cohorts. Any rule with a lower false-certification rate beats it.

Strongest existing practice: impute, then correlate. The missing layers are filled in from the shared layer with multiple imputation or low-rank matrix completion, and the cross-layer correlation is then tested on the completed data. It gets the same cohorts and the same hyperparameter search budget as our tuning of the discretization and bootstrap settings, with imputation rank and regularization chosen by held-out reconstruction error on the observed entries.

It is the strongest competitor because it is what an analyst would actually do, and it always produces an answer. Beating it means a lower false-certification rate at matched or higher certified-claim yield, judged against ground truth from fully matched data. No reported result was found on Lacuna for cross-layer claim certification on disjoint-subset cohorts, by imputation or any other method.

Gaussian covariance completion alone. This is our own pipeline with only the Gaussian engine. It is the ablation that isolates the key ingredient, the nonparametric information engine. If it certifies as many correct claims with equal error, the information-theoretic machinery adds nothing, and we will report that.

Oracle. The full matched-data estimate defines the ground truth and the upper bound on yield.

Why not just…?

Why not just impute the missing layers and run the analysis on complete data? Imputing layer C from layer B implicitly assumes that C is independent of A once B is known, or else carries a correlation structure learned from the few samples that have both B and C. In either case, the imputed A–C relationship is manufactured by the model rather than observed. Our masked-TCGA experiment measures this directly: we count how often imputation declares a significant A–C link that the matched data contradict. This is the main empirical bet. If imputation's false-certification rate turns out comparable to ours at higher yield, the proposal loses its central argument.

Why not just use classical bounds on a missing correlation and skip information theory? This is a real threat, and it is why the Gaussian engine is built in as an ablation rather than ignored. Covariance completion is cheap and well understood. The information engine should matter only where relationships are nonlinear or variables are genuinely discrete, which methylation states plausibly are. The ablation settles which regime the data are in. If Gaussian completion wins or ties, it becomes the default engine. The three-valued protocol and the design rule stand either way.

Why not just collect matched samples for every patient? That is the ideal, but tissue availability is exactly what forced the Taira design. The infrastructure papers also document that funding for continued assays is short-lived (The Computing Research Repository on grant dependence; LINCS on post-2020 maintenance). The design experiment below asks how much matching is enough. It compares our allocator against spreading the same budget uniformly at random.

Proposed Experiments

Masked ground truth from TCGA. TCGA, as mined pan-cancer in CLIC6's role in cancer: from broad analysis to breast cancer validation, has matched methylation, expression, and protein data for many tumors. We take tumor types with matched layers and draw sub-cohorts of 8 to 50 patients. These sizes span the 8 breast-cancer tissue pairs in the CLIC6 validation and the 50 colorectal pairs in MicroRNA-143 targets DNA methyltransferases 3A in colorectal cancer. We then hide layers to reproduce disjoint designs, including the exact three-layer, partial-overlap structure of the Taira study.

Claims are drawn from documented silencing axes, such as CLIC6 promoter methylation versus expression, and from random gene–metabolite and gene–protein pairs. The primary metrics are three: the false-certification rate on implied and refuted verdicts, the yield (the fraction of claims that receive a verdict), and the empirical coverage of the bound intervals. Each is reported against cohort size.

A sanity check uses a known matched-data effect. Ng et al. report a miR-143–DNMT3A protein correlation of $r=-0.71$, measured on a subset of their 50 pairs. When the axis is fully observed, the auditor should certify a negative link. When the shared layer is removed, it should abstain.

Cross-assay transfer on LINCS. LINCS profiles the same cell lines and perturbations with several proteomic and chromatin assays. Many conditions are therefore matched across assay types, and we can mask assay types per condition. This tests the auditor on the exact harmonization gap LINCS reports. The evaluation is verdict accuracy on claims linking one assay's readout to another's.

Design value. We give a budget of $B$ extra matched assays, from 2 to 20, to the allocator and, as the comparison, to uniform random matching. We then measure the drop in undetermined claims and in false certifications.

Decisive outcomes. The proposal succeeds if three things hold at $n\le 30$. First, imputation's false-certification rate is materially higher than the auditor's. Second, the auditor keeps nominal coverage. Third, the allocator resolves more claims per assay than random matching. It fails cleanly in two ways. If the yield is near zero at realistic sizes, the honest conclusion is that small disjoint cohorts cannot support cross-layer claims at all, which is still worth stating. If the Gaussian ablation matches the information engine, the conclusion is that the simpler engine should be used.