Evolutionary Age as a Prior for Finding Direct Contacts in Pulldown Data

When biologists pull a protein out of a cell lysate with an antibody (co-immunoprecipitation, or Co-IP) and identify everything that comes with it by mass spectrometry (affinity purification–mass spectrometry, AP-MS), they get a list of proteins that share a complex with the bait. They do not learn which of those proteins actually touch the bait and which are held on through intermediate subunits. This note proposes a source of information for separating the two that current methods ignore: the evolutionary order in which the subunits of a complex appeared. Complexes are not built all at once. Newer subunits are added onto older ones, and in the symmetric complexes where this has been modeled, the older contacts are larger and stronger than the newer ones. If that history holds generally, a young subunit that co-purifies with ten old ones is probably touching only one or two of them. The age of each subunit then constrains which direct-contact graphs inside the complex are plausible.

Concretely, we would date every subunit in the yeast complexes from the two landmark AP-MS screens. We would then add a likelihood term to the step that turns co-complex lists into direct-contact maps. The term rewards contact graphs that could have grown by sequential attachment of younger proteins onto older ones, and it penalizes dense contact among young peripheral subunits. Predicted contacts are scored against structurally resolved interfaces and high-confidence binary interaction data. The central test is whether real ages help beyond protein abundance, network degree, and within-complex shuffled ages.

Core Claim

Inside a multi-protein complex, the direct-contact graph carries a recoverable imprint of the order in which subunits were added. Using subunit age as a structured prior improves the recovery of direct contacts from AP-MS co-complex data, beyond what topology-only methods and confound-only priors (abundance, degree) achieve.

Three bodies of work motivate the claim. Villar et al. showed with an idealized model of proteins as spheres with sticky surface patches that in dihedral homotetramers the older interface is larger and stronger. They also found that evolution keeps the newer-to-older interface strength ratio below about 0.5 so that assembly proceeds through stable intermediates (Self-Assembly and Evolution of Homomeric Protein Complexes). That work is limited to homomers and never touches pulldown data; our claim extends its logic to heteromeric complexes and tests it on real AP-MS screens. Saeed and Deane found that protein age is the dominant predictor of interaction degree, with a binned correlation near 0.98, consistent with newer proteins attaching to older hubs (Protein protein interactions, evolutionary rate, abundance and age). Their analysis pooled direct and indirect edges into an unweighted graph; we use the same attachment logic at the level of individual contacts inside one complex. Navlakha and Kingsford reversed network growth models and found that duplication-with-divergence fits yeast interactome history better than preferential attachment or forest-fire growth (Network Archaeology: Uncovering Ancient Networks from Present-Day Interactions). Their method infers ages from topology alone. We do the reverse: externally estimated ages constrain the topology, and the duplication model enters as a second growth mechanism our prior must allow, because paralogous subunits of similar age can legitimately contact each other.

Two lines of work supply the problem formulation and the baselines. One is the cut-edge and genetic-algorithm method that infers a direct interaction graph from AP-MS (Predicting direct protein interactions from affinity purification mass spectrometry data.). Its authors generated synthetic test networks with preferential-attachment and duplication models but never used age in the scoring function. The other is bipartite bait–prey clustering for complex detection (Detection of protein complexes from affinity purification/mass spectrometry data). Both treat every bait–prey observation as exchangeable. We keep their machinery and change only the prior over contact graphs.

Research Program

The first component is dating. Each yeast protein receives two independent ages. One is a phylostratum, meaning the deepest branch of the tree of life on which a homolog is found, assigned with the ProteinHistorian-style pipeline used in High GC content causes orphan proteins to be intrinsically disordered. The other is the conservation-based excess-retention score of Saeed and Deane. Where a structural domain assignment exists, we add the SCOP superfamily age on the 0-to-1 scale (1 being the last universal common ancestor) from Exploring Fold Space Preferences of New-born and Ancient Protein Superfamilies. Using three dating schemes matters because of the homology-detection horizon: a fast-diverging ancient protein can be misdated as young, and we want conclusions that survive the choice of scheme.

The second component is the prior. For a complex with subunits $1,\dots,n$ ordered by age, and a candidate direct-contact graph $G$, we define

$$\log P(G) = \sum_{(i,j)\in G} \left[\alpha \, \min(a_i,a_j) - \beta \, |a_i - a_j| \, \mathbb{1}[\text{not paralogs}]\right] - \gamma \sum_{k} \max\!\left(0, d_k^{\text{young}} - 1\right),$$

where $a_i$ is the age of subunit $i$ and $(i,j)\in G$ ranges over contacts. The first term favors contacts whose younger partner is old. The second term penalizes age gaps between non-paralogs, reflecting the finding that structurally bridged folds have a median age difference of 0 (Structural Bridges through Fold Space). The third term penalizes each young subunit $k$ (bottom age tercile within the complex) for every contact beyond one to an older subunit, where $d_k^{\text{young}}$ counts those contacts. The weights $\alpha,\beta,\gamma \ge 0$ are fit on a training split of complexes. Paralogy is called from sequence, so duplicated subunit pairs escape the age-gap penalty, as the duplication model of Navlakha and Kingsford requires. The prior is added to the existing AP-MS likelihood of the direct-interaction method and optimized with the same search procedure, so the only change is this term.

The third component is a secondary, mechanistic prediction about purification itself. If young contacts are weaker, preys attached through young interfaces should be detected less reproducibly. We test this as recovery across the two independent screens, Gavin_2006 and Krogan_2006, after controlling for abundance from PaxDb, the database used in ProteomeVis: a web app for exploration of protein properties from structure to sequence evolution across organisms' proteomes. A small targeted Co-IP follow-up would tie this to wash resistance directly. For six to ten complexes with confident predicted contacts of differing age, we would pull down one subunit under a graded salt series and quantify prey retention by Western blot, following the targeted Co-IP practice in IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer.

Baselines

The naive reference is raw AP-MS: score each bait–prey pair by its co-purification evidence and call the top pairs direct. This is the starting point against which the direct-interaction method reported its gain.

The strongest existing method is that topology-only direct-interaction inference (DIGCOM, the direct interaction graph method of Predicting direct protein interactions from affinity purification mass spectrometry data.). It is the only method in our file that explicitly addresses the direct-versus-indirect problem on the same yeast data. Its authors report that predicted direct interactions achieve about 35% greater overlap with high-confidence yeast two-hybrid networks than top-scoring raw AP-MS interactions on Krogan data. We found no absolute precision or recall figure on Lacuna for this method, so we will rerun it ourselves. Fairness is easy to enforce because our method is that method plus one prior term. Both get the same input matrices, the same genetic-algorithm budget (generations, population size, restarts), and the same held-out evaluation complexes. Beating it means a statistically significant gain in precision at matched recall against structural contacts, evaluated on complexes not used to fit $\alpha,\beta,\gamma$, and replicated on both Gavin_2006 and Krogan_2006.

For complex membership, which we need only as the stage that defines the subunit sets, we use the bipartite Jaccard clustering of Detection of protein complexes from affinity purification/mass spectrometry data. Its authors report about a 20% accuracy improvement over the Markov clustering algorithm and about a 50% F-measure improvement on Krogan_2006 against CYC2008. We also run the analysis with CYC2008 complexes given directly, so that errors in complex detection cannot drive the result.

Three ablations isolate the key ingredient. The first replaces ages with protein abundance. The second replaces ages with interactome degree, which Saeed and Deane show is nearly collinear with age when binned. The third shuffles ages among subunits within each complex, which preserves the age distribution but destroys the ordering. If any of these matches the real-age prior, the evolutionary claim fails.

Why not just…?

Why not just use abundance as the prior? Abundance is the strongest correlate of evolutionary rate, at about −0.49 in yeast according to ProteomeVis: a web app for exploration of protein properties from structure to sequence evolution across organisms' proteomes. It also drives detectability in every high-throughput assay, so an abundance prior could capture most of what looks like an age effect. This is a real threat. The abundance ablation settles it directly. We also fit a combined model and ask whether age retains a significant coefficient when abundance is present. If it does not, we report that the "age signal" in pulldowns is a detection artifact.

Why not just use network degree, since it already tracks age? The binned correlation of about 0.98 in Protein protein interactions, evolutionary rate, abundance and age is between age bins and mean degree. Individual proteins scatter far more widely, and degree measured from AP-MS is contaminated by the very indirect edges we are trying to remove, which makes it circular as a prior. Our prior also uses age differences and paralogy, which degree cannot express. Still, the degree ablation is the test, and we expect it to recover some but not all of the gain.

Why not just predict contacts from structure and ignore AP-MS altogether? For complexes with solved or confidently modeled structures, structural docking would beat any statistical prior. But solved structures cover only about 13% of the yeast proteome (ProteomeVis), and young subunits are disproportionately short and disordered (High GC content causes orphan proteins to be intrinsically disordered). Those are exactly the peripheral attachments where structure is weakest and the age prior should help most. We therefore report performance separately for complexes with and without structural coverage, and we treat structural contacts as ground truth rather than as a competitor where they exist.

Why not just run network archaeology on the AP-MS graph? The method in Network Archaeology infers ages from a given topology. Applied to co-complex data, it would read indirect edges as genuine history and inherit their noise. Its authors also note that it ignores gene loss. Using external sequence-based ages breaks this circularity. As a secondary analysis, we compare archaeology-inferred ages with sequence ages as alternative inputs to our prior.

Proposed Experiments and Decisive Checks

Ground-truth direct contacts come from yeast complexes with experimentally resolved multi-subunit structures, where a contact means atoms of two chains lie within interface distance. A secondary truth set is the high-confidence binary interactions used in the direct-interaction paper. Complexes are split by complex, never by edge, into fitting and evaluation sets. The primary metric is precision–recall area against structural contacts. Secondary metrics are overlap with binary interactions and, for the mechanistic prediction, cross-screen reproducibility of prey detection. We model reproducibility as a function of the age of the inferred attachment interface, with abundance and bait identity as covariates.

The proposal fails cleanly in any of three cases. The first is if the shuffled-age prior performs as well as real ages. The second is if the gain vanishes once abundance is included. The third is if the gain appears under only one of the three dating schemes, which would point to a dating artifact rather than biology. It succeeds if real ages beat DIGCOM at matched search budget on both screens, with the gain concentrated in complexes with young peripheral subunits and weak structural coverage, and if the salt-series Co-IP shows predicted young contacts releasing at lower stringency than predicted old ones in the same complex.