MOON data-to-PKN mapping principles

Scope

This article describes how to decide which measurements should be used as MOON inputs, consistency checks, filters, or annotations when scoring a COSMOS prior knowledge network (PKN).

The central question is:

What biological quantity does this measurement represent, and does that match the type of causal interaction encoded by the PKN?

MOON scores signed mechanistic paths in a PKN. A feature should therefore not be mapped to a node only because the gene, protein, or metabolite identifier matches. The measured quantity must also match the biological state represented by the node and its edges.

This article focuses on MOON workflows. Classic COSMOS/CARNIVAL workflows remain available, but they have heavier solver and setup requirements and are best used when specifically needed.

The basic MOON cases

When no method is specified, a MOON workflow is usually the practical default. Two common entry points are:

• RNA only: infer transcription factor (TF) activities from RNA, then use those TF activities as downstream observations to infer plausible upstream regulators. The upstream layer may be unspecified, or defined by receptors, ligands, drugs, perturbations, or another relevant source layer.
• RNA plus metabolomics: infer TF activities from RNA and use them as upstream activity inputs; use mapped metabolite measurements as downstream inputs. This is the compact workflow illustrated by the COSMOS_basic material.

More advanced datasets require a more explicit mapping step. The rest of this article gives practical rules for that decision.

Direct inputs, footprints, gates, and annotations

Different omics can enter the analysis in different roles:

The important distinction is that these roles are not interchangeable. A total protein abundance value can be useful, but it does not automatically become a protein activity score. RNA can be essential for TF activity inference, but the RNA abundance of a kinase is not usually evidence for kinase activity.

RNA

RNA abundance is best matched to transcriptional regulation.

Preferred uses:

• Infer TF activities from RNA with a TF-target regulon, for example using CollecTRI with decoupleR.
• Use RNA target expression as a TF-target consistency check after MOON scores TFs.
• Use RNA to decide which genes are expressed and can reasonably remain in a context-specific PKN.
• Keep RNA abundance or RNA factor weights as annotations when interpreting scored protein nodes.

Avoid treating RNA abundance of a receptor, kinase, enzyme, or TF as direct protein activity unless the analysis explicitly justifies that assumption.

For TF-target consistency, keep the RNA vector separate from TF activity scores. TF activity is a derived upstream or downstream activity variable; target RNA is the observed transcript-level readout used to check whether TF-target edges have the expected sign.

Total proteomics

Total proteomics measures protein abundance, not protein activity.

Preferred uses:

• Use total protein abundance as evidence that a protein is present or associated with the signal being interpreted.
• Use matched total proteomics to gate transcript-supported TF-target regulation when the question requires a downstream protein-level readout.
• Use total proteomics as an annotation for scored nodes.

A useful pattern is protein-gated RNA for TF-target filtering:

1. RNA asks whether the TF-target transcript changes in a direction compatible with the upstream TF activity and TF-target sign.
2. Matched total proteomics asks whether this transcript-level signal plausibly translates into protein-level downstream support.
3. If RNA is coherent but the matched protein abundance is weak or contradictory, the TF-target edge can be treated as unsupported for a functional downstream readout.

This does not mean that total proteomics proves a TF-target edge. It only makes the retained edge more compatible with the chosen multi-omic assumptions.

Phosphoproteomics

Phosphoproteomics is usually better matched to signaling activity than total proteomics, but it still often needs a footprint step.

Preferred uses:

• Infer kinase or phosphatase activities from phosphosite data and a kinase-substrate or phosphatase-substrate prior.
• Use these inferred activities as MOON inputs when they match the signaling layer being scored.
• Use a measured phosphosite directly only when the PKN explicitly represents that modified site or a node whose state is defined by that modification.

Avoid collapsing phosphoproteomics into total protein abundance semantics. A phosphosite measurement is a modification-specific signal and should be mapped accordingly.

Metabolomics

Metabolite measurements can often be used more directly as downstream inputs because metabolite nodes in the COSMOS PKN represent metabolite states.

Preferred uses:

• Map metabolite identifiers and compartments explicitly, for example with prepare_metab_inputs().
• Use metabolite abundances, differential statistics, or factor weights as downstream inputs when inferring upstream regulators.
• Keep compartment assumptions visible. Mapping the same measured metabolite to multiple compartments is an analysis choice, not a neutral default.

DNA sequencing and genetic lesions

DNA sequencing identifies candidate causal alterations. These are usually perturbations or annotations, not direct activity measurements.

Preferred uses:

• Use high-confidence, directionally interpretable alterations as upstream perturbation candidates. Examples include known activating mutations, homozygous deletions, truncating loss-of-function events, amplifications with expression support, or fusions with known direction.
• Use ambiguous alterations as annotations unless their functional direction is clear enough for the analysis.
• Use RNA-seq to check whether the altered gene is expressed and to infer downstream TF activities or pathway footprints.

For an early stop codon or other likely loss-of-function lesion in node B:

• Disable incoming A -> B edges as explanations for the loss of B activity. The loss is genetically imposed, not necessarily caused by upstream regulation.
• Retain outgoing B -> C edges by default and represent B as a forced negative activity. The inactive state of B should propagate to downstream targets according to the sign of B’s outgoing edges.
• Disable outgoing edges only when the specific edge mechanism requires a molecular function that the lesion destroys and this cannot be represented as lower B activity.

Check allele status, loss of heterozygosity, clonal fraction, RNA expression, domain position, protein stability, and whether the mutation could be dominant-negative or gain-of-function before making a hard rule.

Knockdowns, knockouts, and imposed perturbations

A designed perturbation is an intervention on a node, not an ordinary observed consequence of the network.

For a knockdown, knockout, or other imposed perturbation of node B:

• Disable incoming A -> B edges as explanations for the state of B.
• Treat B as an upstream input or candidate driver with a forced sign when the perturbation is effective, for example B = -1 for a functional knockdown or loss-of-function perturbation.
• Retain outgoing B -> C edges by default so the forced state of B can propagate through the signed PKN.

The reasoning is causal: the perturbation fixes or constrains B, so upstream regulators should not be used to explain the state of B; the downstream side is kept because the analysis asks what follows from that imposed state.

Important caveats:

• Transcript knockdown does not guarantee immediate protein or activity loss.
• RNAi, CRISPR knockout, CRISPR interference, degron systems, and drugs have different kinetics and off-target risks.
• Partial knockdown may require a weaker score, sensitivity analysis, or exclusion of the target as a hard constraint.
• Timepoint matters, especially when RNA decreases before protein turnover or when later readouts include compensation and feedback.

Timepoint alignment across omics

Timepoint matching is a modeling choice. It asks whether measurements are treated as the same network state or as different stages of a causal trajectory.

Two common assumptions are:

• Synchronous alignment: measurements at the same or nearest timepoint are treated as compatible views of a shared quasi-steady state.
• Lagged alignment: an earlier upstream-like layer is paired with a later downstream-like layer, for example early phosphoproteomic signaling followed by later RNA-derived TF activity.

There is no universal rule that a 24 h RNA profile should always be matched to 24 h phosphoproteomics, or always to an earlier phosphoproteomic timepoint. The right choice depends on the perturbation, pathway kinetics, sampling design, and question being asked.

Recommended workflow:

1. Define contrasts within each omic and timepoint, such as treated minus control at the same timepoint.
2. Convert each omic to its PKN-compatible role at that timepoint.
3. Enumerate plausible alignments, including same-time, nearest-time, and biologically motivated lagged alignments.
4. Run MOON separately for alignments that could change the interpretation.
5. Compare sign coherence, recovery of known controls, stability of high-scoring nodes, pathway plausibility, and whether inferred paths respect the expected response order.
6. If alignments disagree and no external evidence resolves the choice, keep the results separate or report them as sensitivity analyses.

Limits to state explicitly:

• Static PKN edges do not encode kinetic delays, feedback loops, or pathway-specific response times.
• Sparse sampling can make nearest-time matching misleading.
• Late timepoints may reflect adaptation rather than the primary mechanism.
• Bulk profiles can average over asynchronous cell states.
• Missing matched controls can make time alignment an additional confounder.

A practical checklist

Before running MOON, write down:

1. What each data type measures: abundance, activity, modification, perturbation, or latent weight.
2. Which PKN edge types are involved in the planned analysis.
3. Which measurements become direct inputs, footprints, consistency checks, functional-readout gates, filters, or annotations.
4. For RNA, how TF activity is derived and how target RNA is kept separately for TF-target coherence.
5. For proteomics, whether the data are total proteomics or phosphoproteomics.
6. For DNA-seq and designed perturbations, which node states are imposed and which incoming edges should stop being used as explanations.
7. For metabolites, which identifiers and compartments are used.
8. For time-course data, whether each omic alignment is synchronous or lagged.
9. What assumptions should be tested by sensitivity analyses.

The final network should be interpreted as a coherent mechanistic hypothesis under these assumptions, not as experimental proof that every retained edge is active in the studied context.

Further implementation notes

The package-facing article is intentionally concise. Coding agents and developers can find more operational detail in agent-docs/moon-data-pkn-mapping-principles.md, including implementation notes for TF-target filtering, protein-gated RNA, and edge-pruning defaults.