When biologists pull a protein out of a cell with an antibody, they get back everything stuck to it. This method is co-immunoprecipitation, or Co-IP when coupled to mass spectrometry. Some of those partners touch the bait directly. Many others ride along because they belong to a large complex the bait happens to touch. The standard next step is functional enrichment analysis: ask which pathway categories (Gene Ontology terms, KEGG pathways) are overrepresented in the list. That step treats every co-purified protein as equally informative, so abundant bystander complexes such as the spliceosome or ribosome can dominate the answer.
We propose to score each co-purified protein by how likely it is to make direct physical contact with the bait, using structure-based protein–protein docking and a check on whether the binding site is physically exposed. Enrichment is then computed on these directness weights, with a null distribution that controls for protein abundance. The claim is simple: the pathway story told by a pull-down should come from the proteins the bait actually touches. Structure-based docking, used as a soft weight rather than as decorative confirmation, can recover that story better than flat enrichment or interactome-topology enrichment can.
Core Claim
Co-IP cannot tell direct binding from bridging through a third protein. Its users know this, which is why the MoPah1 study in rice blast fungus had to add recombinant GST pull-down and yeast two-hybrid assays to show that MoPah1 contacts the kinase Pmk1 directly (Orbitrap Astral–based proteome and phosphoproteome analysis…). It is also why the IQGAP3–GLI1 Co-IP in lung cancer left the contact interface explicitly unresolved (IQGAP3 activates Hedgehog signaling…).
Enrichment methods inherit this blindness. Overrepresentation analysis, a hypergeometric test on a flat gene list, ignores structure entirely. Topology-aware methods such as NetGSA use network edges taken from expression or databases, and those edges carry no information about physical contact (A comparative study of topology-based pathway enrichment analysis methods; Network-based pathway enrichment analysis with incomplete network information).
Our specific, testable claim has two parts. First, a directness score built from predicted inter-chain contacts plus binding-site accessibility separates validated direct partners from co-complex bystanders. Second, enrichment weighted by that score, under an abundance-stratified null, changes which pathways are called in a direction that orthogonal direct-binding assays confirm.
The statistic is a weighted pathway score
$$T_P = \sum_{i \in L \cap P} w_i ,$$
where $L$ is the set of co-purified proteins, $P$ is a pathway's member set, and $w_i \in [0,1]$ is the predicted probability that protein $i$ contacts the bait directly. Significance comes from permuting the weights $w_i$ only among proteins in the same abundance bin. A pathway therefore cannot score highly just because its members are abundant, sticky, or structurally well modeled.
Relation to Existing Work
The most visible existing combination of docking and enrichment comes from network pharmacology, and it plays the role of motivation by counterexample. The resveratrol study in premature ovarian insufficiency ran GO and KEGG enrichment over 609 database-derived genes. It then picked three hub genes (ATM, BRCA1, RAD51) and docked the small molecule onto them, reporting energies between −6.1 and −6.9 kcal/mol (Multi-target modulation of the homologous recombination pathway by resveratrol…). The Fufang Banmao study used docking similarly as a final rationale step (Integrating Network Pharmacology And Mendelian Randomization…). In both, docking happens after enrichment, checks a ligand rather than a protein–protein contact, and cannot change the enrichment result. We reverse the order: protein–protein docking feeds into enrichment and can overturn its conclusion.
Two papers supply methods. ProTact predicts inter-chain contacts and binary complex poses with an SE(3)-invariant graph network, meaning one whose outputs do not change when the molecules are rotated or translated, plus physical constraints (Accurate Protein-Protein Interactions Modeling through Physics-informed Geometric Invariant Learning). DiffAffinity predicts how point mutations change binding free energy (TianZhuAI4S/DiffAffinity). We reuse the first to build $w_i$ and the second to design interface-disrupting mutants for prospective tests. Neither tool was built to interpret pull-down lists, and we do not modify their architectures.
HIGH-PPI links residue-level structure graphs to a protein interaction network (Hierarchical graph learning for protein–protein interaction). It predicts whether an edge exists. It does not ask whether an observed co-purification reflects direct contact, and it has no enrichment layer. DeepSyn replaces static annotation with a literature-mined knowledge graph (Deep functional synthesis). Its edges come from text co-occurrence, which is exactly the kind of non-physical association we aim to filter.
D-finder supplies a constraint rather than a method (Computational Prediction and Experimental Verification of New MAP Kinase Docking Sites…). Its peptide predictions bound JNK1 76% of the time (45/59), but roughly 30% failed in full-length proteins because the motif was buried. This is direct evidence that a contact score must include accessibility in the folded protein, and also that it must not penalize motifs in disordered regions. The D-sites in GLI3 sit in exactly such flexible stretches.
Research Program
The pipeline has three stages.
Stage one handles bait–prey contacts. For each co-purified protein we predict a bait–prey complex with ProTact, using predicted structures where no experimental structure exists. From the top-ranked inter-chain contacts we take a contact-confidence score. Because ProTact is binary, multi-protein assemblies are handled by also docking each prey against the bait's top-scoring direct partners. A prey that contacts a direct partner well but contacts the bait poorly is labeled a likely bridged member.
Stage two adds an accessibility term. For the predicted interface residues on the prey, we measure solvent exposure in the monomer structure and treat predicted-disordered regions as exposed. The final weight $w_i$ comes from a small logistic calibrator fit on held-out complexes. Its inputs are contact confidence, accessibility, and the bridging flag.
Stage three runs weighted enrichment with the abundance-stratified permutation null. Abundance bins come from the mass-spectrometry intensities in the dataset itself. This targets the known failure in which Co-IP lists are skewed toward hyper-stable, abundant complexes.
The calibrator is trained only on structural benchmark complexes from DIPS-Plus. It never sees the pull-down datasets used for evaluation, so enrichment results cannot leak into the weights.
Baselines
The naive reference is flat overrepresentation analysis on the full co-purified list. This means a Fisher's exact test against a proteome background, as run in standard pipelines and packaged in funkea. It is what nearly every Co-IP-MS paper does, so a method that cannot beat it on recovering validated direct-partner pathways has no case.
The strongest existing enrichment competitor is NetGSA, run with a condition-specific network estimated from BioGRID and KEGG priors (Network-based pathway enrichment analysis with incomplete network information). On TCGA breast cancer it reported 25 significantly enriched pathways at FDR 0.01, where standard gene set analysis found zero or very few. The topology comparison found NetGSA and DEGraph to be the most powerful of nine topology-based methods (A comparative study…). NetGSA is the strongest competitor because it already uses network structure to sharpen enrichment and tolerates heavily misspecified networks. If interactome topology alone separates signal from bystanders, NetGSA will show it. We run it on the same co-purified lists and intensities, with the same pathway definitions and FDR threshold, and tune its penalty by its own recommended procedure.
For contact prediction itself, the strongest component baseline is ProTact used without our accessibility and bridging terms. Its paper reports a 31.94% relative gain in Precision@10 on DIPS-Plus and 31.63% on CASP13&14 over prior contact predictors. No absolute Precision@10 value was found on Lacuna, so we will measure it ourselves on the same splits.
Two ablations isolate the key ingredients. The first replaces $w_i$ with STRING-style interactome evidence for bait–prey edges, which tests whether structure adds anything beyond database knowledge. The second removes the abundance stratification, which tests whether the null rather than the weights does the work.
Beating the baselines means two things. First, higher recall of the pathway containing the orthogonally validated direct partner at matched FDR. Second, fewer calls of abundance-driven bystander pathways in synthetic tests where ground truth is known.
Why not just…?
Why not just filter contaminants and run ordinary overrepresentation analysis? Contaminant subtraction removes proteins that bind beads or antibodies nonspecifically. It does nothing about genuine co-complex members that are real but indirect, such as the host spliceosome captured by influenza nucleoprotein (Ningning Sun et al.). The comparison study shows that overrepresentation analysis also loses power when the true signal is a handful of proteins (A comparative study…). That is precisely the direct-partner regime. The synthetic spike-in experiment below measures this directly.
Why not just use NetGSA or another topology-aware method on a known interactome? This is the real threat. If database edges already mark which prey are adjacent to the bait, structure adds cost without gain. Database edges are, however, missing exactly for uncurated proteins and non-canonical contacts, the gap the ovarian cancer stem cell analysis ran into (Core signaling pathways in ovarian cancer stem cell…). They also do not distinguish direct from co-complex membership. The STRING-weight ablation and the head-to-head NetGSA comparison settle this. If NetGSA matches us on the MoPah1 and synthetic tests, we will report that structure is not worth its compute for this task.
Why not just trust a structure predictor's confidence score and skip the accessibility term? The D-finder results argue against it. About 30% of motifs that bound as peptides failed in the full protein because of burial (Whisenant et al.), and pairwise contact predictors score interfaces on isolated chains. The D-finder accessibility check below tests whether our term predicts those failures. If it does not, the term is dropped.
Proposed Experiments
We run four experiments, ordered so the cheapest one can kill the idea first.
The first is the D-finder accessibility check. Among the 45 peptides that bound JNK1, we ask whether accessibility computed on predicted full-length structures, with disordered regions counted as exposed, ranks the full-length failures below the successes. This is measured by AUROC against D-finder's original ranking. If accessibility fails to separate them, the structural premise is weak.
The second is calibrated directness on DIPS-Plus and CASP13&14. We measure Precision@10 for contact prediction, and AUROC for separating true binary interfaces from decoys built by pairing members of the same complex that do not touch.
The third is synthetic Co-IP lists. We embed a bait's true direct partners from one pathway inside abundant co-complex members from another pathway, with intensities sampled from the MoPah1 dataset (ProteomeXchange PXD056310). We then measure how often each method ranks the true pathway first at FDR 0.05.
The fourth combines real and prospective tests. On the MoPah1 data, the question is whether Pmk1's MAP kinase pathway rises in rank under weighting relative to the flat analysis's glycerophospholipid and autophagy themes; Pmk1 is the partner validated as direct by GST pull-down and yeast two-hybrid. Prospectively, we dock IQGAP3 against GLI1 and design two interface mutants predicted by DiffAffinity to disrupt binding, plus one predicted-neutral control. We then test whether endogenous Co-IP is lost only for the disrupting mutants.
The decisive outcome is concordance between computational prediction and experiment. Weighted enrichment should promote the pathway of the validated direct partner where flat analysis and NetGSA do not, and the predicted interface mutants should abolish co-purification while the neutral control preserves it. If both hold, structure has earned its place in enrichment. If neither holds, the honest conclusion is that pull-down enrichment should stay structure-free.
References
- Orbitrap Astral–based proteome and phosphoproteome analysis identifies candidate proteins associated with the phosphatidate phosphatase MoPah1 in Magnaporthe oryzae
- IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer
- A comparative study of topology-based pathway enrichment analysis methods
- Network-based pathway enrichment analysis with incomplete network information
- Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation
- Integrating Network Pharmacology And Mendelian Randomization To Explore Potential Targets Of Fufang Banmao Capsule Against Non-Small Cell Lung Cancer
- Accurate Protein-Protein Interactions Modeling through Physics-informed Geometric Invariant Learning
- Hierarchical graph learning for protein–protein interaction
- Deep functional synthesis: a machine learning approach to gene functional enrichment
- Computational Prediction and Experimental Verification of New MAP Kinase Docking Sites and Substrates Including Gli Transcription Factors
- funkea: Functional Enrichment Analysis in Python
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- Core signaling pathways in ovarian cancer stem cell revealed by integrative analysis of multi-marker genomics data