When a compound works in cells but nobody knows what it binds, the usual next step is to measure which genes change their messenger RNA (mRNA) levels and look for the culprit among them. That step almost never works directly. Across more than 500 compounds in the Connectivity Map (CMap), a public library of drug-treated gene-expression profiles, about 97% of known target proteins showed no mRNA change after treatment (Drug target prioritization by perturbed gene expression and network information). The reason is that binding changes what a protein does, not how much of its mRNA is made. The effect then travels through a chain of protein-to-protein signaling steps that are also largely invisible to mRNA measurement, and it only becomes readable once it reaches transcription factors (the proteins that switch genes on and off). The mismatch between mRNA and protein is broad: mRNA explains only 40–60% of protein abundance (A high-stringency blueprint of the human proteome), and phosphorylation states are not captured by transcript data at all (Proteomics, pathway array and signaling network-based medicine in cancer).
We propose to build this "visibility" structure directly into network-based target inference. The protein interaction network is split into a hidden signaling layer, where perturbations spread without leaving an mRNA trace, and an observable regulatory layer, where transcription factors write the perturbation into gene expression. A candidate target is scored by how well a perturbation starting at that node, spread through the hidden layer and read out through the regulatory layer, reproduces the compound's observed expression change. Each score is then compared against degree-matched random nodes and against the same node's scores across all other compounds. The claim is that this pairing does two things. It recovers silent targets better than proximity-to-changed-genes methods. It also separates "specific hubs," which matter for one compound's mechanism, from "generic hubs," which rank highly for every compound because they are well studied.
Core Claim and Relation to Prior Work
The closest prior work is Local Radiality (LR) from Işık et al. (Drug target prioritization…). LR ranks candidates by how close they sit, in shortest-path distance on STRING, to the genes whose expression changed. It rests on their "proximity hypothesis": true targets are on average 2.9 steps from changed genes, versus 3.6 for random proteins. LR treats all edges alike and treats every changed gene as direct evidence. It therefore rewards a node for being physically near changed genes even when no transcriptional route links the two. Our model instead requires the effect to pass through a transcription factor before it can be seen, which is the mechanistic reading of why targets are silent. Işık et al. also motivate our second ingredient. They show that degree and betweenness centrality suffer from popularity bias: targets found only by LR have on average 94 neighbors, versus 248 for targets that every method finds.
The Steiner-network line of work supplies the idea that unmeasured nodes can be inferred from topology. Huang et al. (Linking Proteomic and Transcriptional Data…) connected tyrosine phosphoproteomics to transcriptional output through a prize-collecting Steiner tree, an optimization that finds the cheapest subnetwork linking high-scoring measured proteins. It surfaced unmeasured hubs such as CREBBP, β-catenin and p300, which were validated with inhibitors. Tunçbağ et al. (Network Modeling Identifies Patient-specific Pathways in Glioblastoma) used the related Steiner forest method on eight xenograft lines. It recovered MEK1, which is serine-phosphorylated and so invisible to their tyrosine assay, and Western blots confirmed MEK1 activity in 7 of 8 lines. The multi-omics mode-of-action study (A Multi-Omics Interpretable Machine Learning Model Reveals Modes of Action of Small Molecules) extends Steiner forests to four omic layers. It shows that two antihistamines with nearly identical transcript signatures have divergent mechanisms. It returns pathways rather than binding targets, and it needs dense multi-omics for each compound.
All three constrain our design in the same way. Hidden hubs were recoverable only because protein-level measurements existed. Our proposal asks whether an explicit visibility model can recover part of that hidden-node information from transcripts alone, and it tests this against those papers' protein-derived hubs.
The generative translator scChord (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation) plays a limiting role. It predicts single-cell surface-protein distributions from RNA. Its authors report that correlation structure degrades as panels grow to hundreds of proteins, and they call for pathway-informed graph priors. Its panels (13 and 228 surface proteins) do not include the kinases and adaptors where silent targets live. For that reason we do not use RNA-to-protein imputation as the discordance mechanism.
Research Program
Let $A_s$ be the weighted signaling adjacency, built from STRING or iRefIndex physical and functional edges. Let $R$ be a transcription-factor-to-gene matrix, built either from motif occurrences in accessible chromatin, following Huang et al.'s regression approach, or from curated regulatory edges. For candidate protein $c$, let $p_c$ be the stationary vector of a random walk with restart from $c$ on $A_s$, which describes how a perturbation at $c$ spreads over the hidden layer. The predicted transcript response is
$$\hat{d}_c = R^\top (M \, p_c),$$
where $M$ is a diagonal mask that keeps only transcription-factor entries. This enforces that signal becomes visible only through transcription factors. If $d$ is the compound's observed vector of absolute expression changes, the raw score $S(c) = \mathrm{corr}(\hat{d}_c, d)$ is converted to
$$Z(c) = \frac{S(c) - \mu_{\mathrm{deg}}(c)}{\sigma_{\mathrm{deg}}(c)} + \lambda \, \frac{S(c) - \bar{S}_{\mathrm{cmpd}}(c)}{\sigma_{\mathrm{cmpd}}(c)}.$$
Here $\mu_{\mathrm{deg}}$ and $\sigma_{\mathrm{deg}}$ are the mean and standard deviation of $S$ over random seeds with matching degree. $\bar{S}_{\mathrm{cmpd}}(c)$ and $\sigma_{\mathrm{cmpd}}(c)$ are the mean and standard deviation of candidate $c$'s scores across all CMap compounds. $\lambda$ is a single weight tuned on held-out compounds. The first term asks whether $c$ explains the data better than a node of its size would. The second asks whether $c$ explains this compound better than it explains compounds in general. A node that scores high on raw $S$ but low on the second term is a generic hub. A node high on both is a specific, candidate-target hub. The same decomposition applies to disease hub-gene analysis, with disease-versus-control expression changes in place of compound profiles.
The build has four parts: reimplementing LR and kernel diffusion on the identical network, building the two-layer graph, computing null distributions (precomputable, because random-walk vectors do not depend on the compound), and a held-out split of compounds for tuning $\lambda$ and the restart probability.
Baselines
The naive reference point is fold-change ranking: candidates ordered by their own mRNA change. Işık et al. already show this fails because targets are silent. It is included to anchor the scale, and beating it counts for nothing.
The second baseline is degree centrality on the changed-gene subnetwork, the standard hub-gene workflow seen in studies such as Identification of candidate target genes of pituitary adenomas based on the DNA microarray and the knowledge-graph sieve of Wang et al. (A novel approach for target deconvolution from phenotype-based screening using knowledge graph). We will report its recall and, just as importantly, the average degree of the targets it retrieves.
The strongest existing method is LR. On CMap with STRING, Işık et al. report 22% recall of known targets in the top 1st percentile of candidates, versus 17% for kernel diffusion. We will first reproduce LR on our exact network and compound set. Our comparison uses our reproduced number, with 22% as the sanity check.
Two further baselines represent the strongest cheap fixes. The first is LR z-scored across compounds, which applies our cross-compound normalization to LR and so tests whether popularity correction alone explains any gain. The second is a Steiner forest run on transcript changes only, with the Steiner nodes used as candidates. This asks whether off-the-shelf hidden-node inference already does what we propose.
All methods share the same network version, compound set, known-target annotations and held-out split. Each gets the same budget: a grid over at most two hyperparameters, tuned on the same held-out compounds. Beating the field means higher top-1% recall than the better of cross-compound-normalized LR and transcript-only Steiner forest, with the gain outside a 95% bootstrap interval over compounds. Retrieved targets must also have a median degree no higher than LR's.
The key ablation replaces $R^\top M$ with the identity, so every node is directly visible. Everything else, including both normalizations, stays the same. If this single-layer version matches the full model, the visibility structure contributes nothing and the claim fails.
Why not just…?
Why not just rank genes by differential expression? Because the target's own mRNA is unchanged in about 97% of CMap cases (Işık et al.). This baseline is expected to sit near chance, and it is kept only as a floor.
Why not just pick high-degree hubs among changed genes? Degree rewards whichever proteins have been studied most, independent of the compound. That is the 248-versus-94-neighbor gap Işık et al. report. It also favors broadly connected nodes whose perturbation looks like non-specific toxicity, which is the ambiguity raised by viriditoxin's effect on mitochondrial metabolism (The mycotoxin viriditoxin induces leukemia- and lymphoma-specific apoptosis by targeting mitochondrial metabolism). Our per-compound normalization is designed to demote exactly these nodes, and the retrieved-degree statistic will show whether it does.
Why not just z-score LR across compounds? This is the real threat. Much of LR's remaining error may be popularity bias, and a cross-compound correction on LR could capture most of the gain without any two-layer model. We therefore run it as a first-class baseline. If it matches the full model, and the single-layer ablation matches too, we will report that the useful ingredient is normalization, not visibility, and narrow the contribution accordingly.
Why not just run a Steiner forest on the transcript changes? Steiner methods found MEK1 and CREBBP because phosphoproteomic prizes pinned the hidden layer in place (Tunçbağ et al.; Huang et al.). With transcript-only prizes, the optimization tends to connect changed genes through short, dense paths, which again favors hubs. The transcript-only Steiner baseline tests this directly.
Why not just impute protein from RNA with a translator like scChord, or predict binding from chemistry? Available translators cover surface-protein panels and lose correlation structure at scale (scChord), so they cannot see intracellular signaling states. Chemogenomic predictors such as KGE_NFM (A unified drug–target interaction prediction framework based on knowledge graph and recommendation system) report a 19% AUPR gain over TriModel in the protein cold-start setting on Yamanishi_08. However, they use no cellular response, and they cannot explain divergent mechanisms among structurally similar compounds (multi-omics MoA study). We treat them as complementary and will test a late fusion of KGE_NFM probabilities with $Z(c)$ on compounds present in both Yamanishi_08 and CMap.
Decisive Experiments
The primary test is top-1% and top-5% recall of annotated targets on CMap with STRING, under a compound-level held-out split. It is stratified by target degree, so that gains among low-degree targets are visible separately.
The discordance test uses the glioblastoma datasets from Huang et al. and Tunçbağ et al., assuming their transcript data can be obtained. We run our model using only their expression data and ask whether the protein-derived Steiner hubs (CREBBP, CTNNB1, EP300, MEK1) rank higher than under LR and transcript-only Steiner forest. Success means ranking these hubs at least as high as the stronger of those two baselines without any phosphoproteomic input. Failure would show that the visibility model cannot substitute for protein measurement.
For the hub-specificity test, we compute the generic-versus-specific decomposition across CMap. We then check whether nodes flagged as generic (high raw score, low cross-compound term) coincide with the high-degree set Işık et al. identify, and whether known targets fall in the specific quadrant more often than degree-matched controls.
For qualitative follow-up, the top specific candidates for well-profiled but unexplained compounds, such as erianin (Erianin induces apoptosis and pyroptosis…), are passed to docking and binding assays in the manner of Wang et al.'s USP7 confirmation. These cases serve as illustrations, not as evidence for the claim.
References
- Drug target prioritization by perturbed gene expression and network information
- A high-stringency blueprint of the human proteome
- Proteomics, pathway array and signaling network-based medicine in cancer
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- Network Modeling Identifies Patient-specific Pathways in Glioblastoma
- A Multi-Omics Interpretable Machine Learning Model Reveals Modes of Action of Small Molecules
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Identification of candidate target genes of pituitary adenomas based on the DNA microarray
- A novel approach for target deconvolution from phenotype-based screening using knowledge graph
- The mycotoxin viriditoxin induces leukemia- and lymphoma-specific apoptosis by targeting mitochondrial metabolism
- A unified drug–target interaction prediction framework based on knowledge graph and recommendation system
- Erianin induces apoptosis and pyroptosis via MAPK/ERK and PI3K/Akt pathways and synergizes with anlotinib in anaplastic thyroid carcinoma