Pull-down experiments find which proteins stick together. In co-immunoprecipitation, and in its large-scale form, affinity purification with mass spectrometry (AP-MS), an antibody fishes out one "bait" protein together with whatever "prey" proteins came along with it. The result is a noisy two-sided graph linking baits to preys. Some preys are sticky contaminants. Others arrived only because they were attached to something else that was attached to the bait. A tempting way to clean such graphs is to ask whether the two genes' RNA levels rise and fall together across conditions, as RNA sequencing measures. The literature disagrees about whether this works. Methods for dynamic interaction networks and expression-refined clustering treat co-expression as a reliable filter, while Semantic integration to identify overlapping functional modules in protein interaction networks found co-expression to be a poor edge weight, and A high-stringency blueprint of the human proteome reports that mRNA tracks protein abundance only 40% to 60% of the time.
We propose that both sides are right, and that the local shape of the pull-down graph around an edge tells you which side applies. Edges inside tightly reciprocal, clique-like blocks look like stable, fixed-ratio complexes whose subunits should be co-transcribed. Edges on tree-like bridges, or edges touching promiscuous hub preys, look like transient or post-translationally controlled contacts, where co-expression should carry little signal. The proposal is a weighting scheme that runs directly on the raw bait–prey graph. It lets co-expression count only as much as it has earned within each topological context. That per-context trust is calibrated without curated labels, using edges the experiment itself confirms twice: the prey pulled down by the bait and the bait pulled down by the prey.
Core Claim
The claim has two parts. The empirical part is that the predictive value of transcript co-expression for true co-complex membership varies sharply with the edge's topological context in the bait–prey graph. It should be high in reciprocal dense cores and near zero on cut-edges and hub-prey edges. The methods part is that a weight which scales the expression evidence by this context-specific reliability will beat two alternatives. The first uses topology alone, as in Detection of protein complexes from affinity purification/mass spectrometry data. The second adds co-expression uniformly, as in the time-course filtering of Detecting complexes from edge-weighted PPI networks via genes expression analysis. The reliability can also be estimated from the pull-down data itself rather than from a gold-standard complex list, which matters for organisms and cell states that have no curated complexes.
To make the weight concrete, write it for a bait–prey edge $(b,p)$ as a sum of log-odds terms:
$$w(b,p) = T(b,p) - \sigma(p) + \lambda_{k(b,p)}(\rho_{bp})$$
The first term, $T(b,p)$, is a bipartite neighborhood-overlap score: the overlap between the prey sets of $b$ and $p$ when $p$ was also used as a bait, and otherwise the overlap of the baits that retrieved each. The second term, $\sigma(p)$, penalizes preys that appear across many unrelated pull-downs, in the spirit of the specificity penalty in Noise reduction in protein-protein interaction graphs by the implementation of a novel weighting scheme. In the third term, $\rho_{bp}$ is the expression correlation of the two genes, $k(b,p)$ is the edge's topological context class, and $\lambda_k$ is a monotone log-likelihood-ratio curve fitted separately for each class.
The context classes are reciprocal core, quasi-clique member, cut-edge or degree-2 bridge, and hub-prey edge. The bridge and quasi-clique categories are borrowed directly from the decomposition in Predicting direct protein interactions from affinity purification mass spectrometry data..
Research Program
The first stage is a measurement study that resolves the contradiction between the two views of co-expression. On the yeast Gavin_2006 screen (1993 baits, 2671 preys, 19157 bait–prey relationships) and the Krogan_2006 screen (2233 baits, 5219 preys, 40623 relationships), we assign every edge to a context class. For each class, we then measure how well $\rho_{bp}$ separates edges between members of the same CYC2008 complex from other edges, and how well it separates direct interactions in the Yu yeast two-hybrid set from indirect ones. Expression comes from the 36-point yeast metabolic cycle time course (GSE3431) used by Zhang et al. That dataset is microarray rather than RNA-seq. Because the RNA-seq question is the target, we will also assemble a multi-condition yeast RNA-seq compendium from public data. The research file lists no specific such dataset, so its composition must be fixed and reported before any evaluation. If the per-class discrimination curves do not differ, the rest of the proposal fails, and we would report that as the finding.
The second stage fits the curves $\lambda_k$ without gold labels. Reciprocal edges, meaning pairs that retrieve each other when each is used as bait, serve as internal positives. Edges attached to high-$\sigma$ sticky preys, and seen only once, serve as internal negatives. Within each class we fit an isotonic log-likelihood ratio of $\rho$. The obvious worry is that reciprocal edges are over-represented in stable cores. That leaves the bridge and hub classes with few internal positives, and those curves are then shrunk toward zero, which is the conservative choice. The decisive check for this stage is to compare the self-calibrated curves against curves fitted on CYC2008 labels.
The third stage feeds the weighted bipartite graph to two downstream tasks. For complex detection, we reuse the seed-and-expand clustering of Gao et al. unchanged, so that only the weights differ. For separating direct from indirect contacts, we replace the uniform edge probabilities that the direct-interaction deconvolution paper itself names as unrealistic with our weights, and keep its cut-edge, quasi-clique, and genetic-algorithm steps.
The fourth stage tests transfer to human data. We weight BioPlex AP-MS edges using expression correlation computed across conditions and score against GO cellular-component complexes. Mapping cell structure across scales by fusing protein images and interactions fused BioPlex with imaging to map cell structure and listed dynamic transcriptomics as missing; we supply that ingredient at the level of individual edges rather than at the level of cell maps.
Baselines
The naive reference point is the unweighted graph clustered with the Markov Cluster algorithm (MCL), a random-walk clustering method. Gao et al. report roughly 20% higher accuracy over MCL on both yeast screens and about 50% higher F-measure on Krogan_2006. The file gives no absolute F-measure for either screen, so we record: no reported absolute result found on Lacuna for Gao et al. on Gavin_2006 or Krogan_2006. We will re-run MCL ourselves.
The strongest topology-only competitor is Gao et al.'s bipartite shared-prey Jaccard weighting with seed-and-expand clustering. We run it with the same clustering code, the same cluster-size filters, and the same parameter grid as our method. Every method gets an identical tuning budget, selected on a held-out half of CYC2008 and reported on the other half and on GO-CC.
The strongest expression-aware competitor is the Zhang et al. pipeline: topological Jaccard weighting, MCL, then refinement by expression synchrony. It reports precision 0.6004 and F-measure 0.5528 against CYC2008 on the DIP network. We cannot compare to those numbers directly because DIP is a curated union network, not a raw pull-down graph. We therefore re-run the pipeline on the AP-MS projections with the same expression data.
For separating direct from indirect contacts, the competitor is the deconvolution method of Predicting direct protein interactions from affinity purification mass spectrometry data. It reports about 35% better overlap with Yu two-hybrid interactions than the original Krogan scores. We run it with its uniform probabilities and then with our weights substituted.
Three ablations isolate the key ingredient. The first uses a single uniform $\lambda$ across all classes, which is expression added without regard to context. The second drops $\sigma$. The third replaces the self-calibrated curves with label-fitted ones. Beating the baselines means a gain over both the Gao and uniform-$\lambda$ variants on held-out CYC2008 F-measure and on Yu overlap, consistent across both yeast screens, with a paired bootstrap over complexes whose confidence interval excludes zero.
Why not just…?
Why not just weight every edge by expression correlation? Cho et al. found co-expression a weak module predictor, and the transcript–protein gap makes a uniform expression weight likely to hurt precisely on transient contacts. The uniform-$\lambda$ ablation tests exactly this. If it matches the context-gated version, our central claim is wrong, and the experiment will show it.
Why not just train a classifier on topology plus co-expression features? This is the real threat. A gradient-boosted model with interaction terms, trained on CYC2008, could learn the gating on its own, much as the 43-feature network of Protein complex detection with semi-supervised learning in protein interaction networks learned feature combinations. We include this model with the same features and tuning budget. We expect it to match or beat us when labels are plentiful. Our claim therefore rests on the low-label regime. We will retrain it with a shrinking fraction of labeled complexes and test on BioPlex, where transfer from yeast labels is weak. If it still wins with 10% of labels, the self-calibration contribution does not hold, and the note reduces to the measurement study.
Why not just use Gene Ontology similarity instead of expression? Gene Ontology weighting works well where annotations exist. However, 54% of the systems MuSIC resolved were undocumented, and annotations cannot change with cell condition. We will report a Gene Ontology-weighted variant and score it separately on complexes with few annotations.
Why not just threshold spectral counts? The direct-interaction deconvolution paper shows that indirect passengers can be abundant, so abundance alone cannot separate them. Our $\sigma$ term already covers the useful part, the stickiness of preys across pull-downs, and its ablation measures that contribution.
Proposed Experiments
The decisive experiment is the per-class discrimination analysis on Krogan_2006, measured as area under the ROC curve for co-expression against CYC2008 co-membership within each context class. The prediction is a clear gap between reciprocal-core and hub or bridge edges, replicated on Gavin_2006 and across both expression sources. The second check compares self-calibrated with label-fitted $\lambda_k$ curves. The two should agree closely in the dense classes; if they diverge badly, the label-free story fails. The third check is held-out complex F-measure and Yu two-hybrid overlap against the baselines above. The fourth is BioPlex transfer scored on GO cellular-component complexes, together with the low-label curve against the supervised classifier. Together these results say whether expression should get a vote on a pull-down edge, and in which parts of the graph.
References
- Semantic integration to identify overlapping functional modules in protein interaction networks
- A high-stringency blueprint of the human proteome
- Detection of protein complexes from affinity purification/mass spectrometry data
- Detecting complexes from edge-weighted PPI networks via genes expression analysis
- Noise reduction in protein-protein interaction graphs by the implementation of a novel weighting scheme
- Predicting direct protein interactions from affinity purification mass spectrometry data.
- Mapping cell structure across scales by fusing protein images and interactions
- Protein complex detection with semi-supervised learning in protein interaction networks