Co-immunoprecipitation (Co-IP) is the standard way to find out which proteins physically sit in a complex with a chosen "bait" protein: you grind up millions of cells, use an antibody to pull the bait out of the soup, and identify the proteins that came along for the ride. The catch is the grinding. Because Co-IP needs that large cell input, it returns a single population-averaged answer. If a complex exists only inside a rare 3% subpopulation — cancer stem-like cells, virus-infected macrophages, a transitional state in a regenerating tissue — the signal is diluted toward the detection floor. Worse, if the bait binds partner A in one cell type and partner B in another, the pulldown reports all three together, creating the illusion of a tripartite complex that exists nowhere in the body. This is the same averaging failure that single-cell sequencing exposed for gene expression, but it has never been fixed for physical protein interactions.
The reason it has not been fixed is that the entire deconvolution toolkit — methods that un-mix bulk measurements into per-cell-type signals — operates on abundances, not on interactions. POPPY aligns bulk RNA-seq with single-cell references to recover per-cell-type transcript profiles, and Count Bridges deconvolves bulk counts into integer single-cell profiles, but neither can say anything about which proteins are bound to which (POPPY, Venkat et al.; Count Bridges). On the proteomics side, GraphDEC deconvolves bulk and spatial proteomics into cell-type abundance proportions, and it does this well — but explicitly cannot determine whether co-expressed proteins are physically bound (GraphDEC). Meanwhile, the computational models that do predict cell-type-specific protein interactions — PINNACLE, AWARE, and scPPIN — build them by draping single-cell RNA data over static, bulk-derived reference interactomes, and none has ever checked a single prediction with a physical pulldown (PINNACLE, Li et al.; AWARE; scPPIN, Klimm et al.). PINNACLE's own limitations section concedes this: the underlying interactome "is not measured in a cell type-specific manner," so its cell-type networks may contain systematic false positives and negatives. There is, to our knowledge, no framework — mathematical or experimental — for assigning a Co-IP pulldown's prey proteins to the subpopulations in which they were actually bound. No reported result found on Lacuna for deconvolution of Co-IP pulldown data.
Core Claim
We claim that a pulldown from a heterogeneous mixture can be un-mixed into subpopulation-specific physical interactomes, if you stop treating the pulldown as a single measurement and start treating it as a designed experiment. The recipe has three ingredients. First, measure the mixture itself with single-cell profiling, so the subpopulation proportions and per-subpopulation expression levels are known rather than guessed. Second, run the pulldown not once but as a series at deliberately varied mixture compositions, which turns occupancy — the unknown "was this prey bound in subpopulation k?" — into an identifiable quantity. Third, run lysis-mixing controls, where the subpopulations are lysed separately and the lysates combined afterward, to flag prey associations that form only after membranes are destroyed. The output is, for each prey protein in the pulldown, an assignment of which subpopulations it was physically bound in, with an occupancy estimate and an artifact flag.
Formally, for prey $j$ in mixture ratio $r$, the observed mass-spectrometry count is modeled as
$$y_{rj} = \sum_{k} \pi_{rk}\, e_{jk}\, \theta_{jk} + \varepsilon_{rj},$$
where $\pi_{rk}$ is the measured proportion of subpopulation $k$ in mixture $r$, $e_{jk}$ is the measured expression of prey $j$ in subpopulation $k$, $\theta_{jk} \in [0,1]$ is the unknown occupancy of prey $j$ on the bait within subpopulation $k$, and $\varepsilon$ is noise. The $\pi$ and $e$ terms come from the single-cell profiling; the ratio series makes the $\pi$ matrix vary across $r$ so that the $\theta$ terms are separable rather than confounded. A group-sparsity penalty across subpopulations keeps the estimator from smearing occupancy where none exists, which is precisely the mechanism that should dissolve the phantom tripartite-complex illusion.
This is an integrated experimental approach in the strict sense of Direction 1: the perturbation axis (mixture ratios), the single-cell molecular axis (proportions and expression), the biochemical axis (crosslinked Co-IP followed by nano-LC-MS/MS, following the protocol of Sun et al.'s H7N9 nucleoprotein study, which identified 32 unique peptides for NP and 17 co-purified spliceosome components in bulk HEK293T lysates), and the functional axis (phenotypic assays on the sorted subpopulations) are coupled in one design rather than patched together after the fact (Sun et al.).
Research Program
The program has three linked stages, each pairing an experiment with the part of the model it identifies or tests.
Stage one builds ground truth from mixtures of cell lines whose interactomes we can measure purely. The anchor system is HEK293T cells transiently transfected with influenza H7N9 nucleoprotein, exactly as in Sun et al., mixed at known ratios with untransfected HEK293T cells. This directly attacks the heterogeneity problem Sun et al. acknowledge but could not address: transfection efficiency is inherently uneven, so high-expressing cells, low-expressing cells, and non-transfected cells were all lysed into one slurry. We will define the ground truth interactome of each pure fraction by separate pulldowns on FACS-isolated NP-positive and NP-negative cells (feasible here because both fractions are abundant), then create mixtures at 3:97, 10:90, 50:50, and 90:10, apply the model, and score whether it recovers the known per-fraction interactomes. The same design repeats with a second, unrelated bait to test generality.
Stage two adds the two controls that make the answer trustworthy. The lysis-mixing control mixes lysates from separately lysed pure fractions and compares prey recovery against the intact-mixture pulldown: prey proteins that appear only in the co-lysed sample are flagged as post-lysis artifacts, because compartments and cells that never touched in life were merged in the tube. This is the experimental answer to the Co-IP lysis vulnerability that proximity ligation assays were invented to escape, and it does so at discovery throughput rather than the pairwise, hypothesis-locked throughput of PLA (Foster et al., Tax–Scribble PLA). The second control is a mild pre-lysis crosslinker titration, calibrated so that transient complexes survive dilution without creating new ones; the titration curve itself is diagnostic, since a genuine in-cell interaction should appear at low crosslinker concentration while a post-lysis one should not.
Stage three moves to the biological settings where the masking problem actually decides scientific conclusions. The first is tamoxifen-resistant MCF-7/TAMR cultures, where Lin et al. showed resistance is driven by a rare stem-like subset with a 74.8-fold upregulation of SOX2 and epigenetic silencing across 44 genes — dynamics that bulk profiling conflates with uniform adaptation (Lin et al.). We will pull down a resistance-relevant bait from mixed sensitive and resistant cultures and ask which protein partnerships are exclusive to the SOX2-high subset. The second is A549 lung cancer cultures under sphere-forming conditions, where Li et al. proved by bulk Co-IP that IQGAP3 physically binds GLI1 to activate Hedgehog stemness genes — but could not determine whether that complex is restricted to the small sphere-forming stem-like subpopulation or present everywhere (Li et al.). Both are exactly the "100% occupancy in a 3% subpopulation versus 3% occupancy everywhere" ambiguity that bulk Co-IP cannot resolve. Finally, orthogonal single-cell confirmation: predicted subpopulation-restricted interactions get validated by PLA and split-luciferase assays on the matched samples, giving a physical, per-cell readout of the computational assignments.
On the computational side, the model is deliberately simple — a regularized regression on the equation above, with the option of a neighborhood-aware variant in the spirit of GraphDEC's similarity graphs when subpopulations are related cell states rather than clean cell types. We will use Graphein for protein-network featurization where structural priors on candidate edges help regularize $\theta$ (Graphein).
Proposed Experiments
The decisive experiments, in order. First, the mixture-recovery test: with pure-fraction pulldowns as ground truth, measure prey-to-subpopulation assignment accuracy, interaction-level precision and recall, and the false tripartite-complex rate — the fraction of bait-A-B co-recoveries the naive bulk reading produces that the ground truth says never co-occur in one cell type. Second, the rarity stress test: spike the NP-positive fraction down to 3% and quantify at what fraction size the deconvolved assignment still recovers the known interactions, comparing against bulk Co-IP-MS detection directly. Third, the artifact-filter test: how many lysis-artifact associations the mixing control flags, verified against the pure-fraction ground truth. Fourth, the biological tests in MCF-7/TAMR and A549 spheres, with PLA/BiFC single-cell confirmation rates for predictions restricted to the stem-like subsets. Fifth, the head-to-head against computational contextual interactome prediction: we run PINNACLE's transcript-overlay logic on our own single-cell references and compare its top-ranked cell-type-specific interactions against our pulldown-anchored assignments, judged by PLA confirmation. Metrics are chosen to be reportable in the same spirit as the field's existing numbers: GraphDEC, the strongest abundance deconvolution, reports a Concordance Correlation Coefficient of 80.7%, a 15% improvement over the next best method (CellDART), and 93% average CCC on cross-species datasets with a 26.4% improvement over Tangram (GraphDEC); we will report occupancy-CCC against pure-fraction ground truth in the same format so the numbers are directly comparable in kind.
Baselines
The do-nothing reference point is conventional bulk Co-IP-MS, run exactly as Sun et al. ran it: one pulldown on the whole mixed population, prey list taken at face value as the interactome (Sun et al.). No reported result found on Lacuna for its prey-assignment accuracy on heterogeneous mixtures — which is the point: it has never been scored against subpopulation-resolved truth. We will score it. It is the strongest competitor for discovery throughput and the weakest for assignment; beating it means showing a nonzero false tripartite-complex rate in mixtures where the ground truth contains no such complex, and recovering interactions at spike-in fractions where the bulk read loses them.
The strongest existing physical method is FACS-sort-then-Co-IP: sort subpopulations by surface marker, then run standard pulldowns on each pure fraction. This is the gold standard and we use it as our ground-truth generator, so the comparison is honest by construction — same antibody, same beads, same nano-LC-MS/MS instrument time, same prey database and filtering thresholds. Our claim is deliberately narrower than "beat sorting everywhere." Where a subpopulation can be sorted at 50% abundance with a clean marker, sorting wins and we say so. We win only where sorting structurally fails: below roughly 5% abundance where cell numbers fall under Co-IP's input requirement, and for transitional states that lack a binary surface marker. Damdinsuren et al. showed that highly plastic transitional and differentiation trajectories cannot be cleanly isolated by single-antibody gating, which is precisely the regime we target (Damdinsuren et al.). The decisive number is interaction recovery at the 3% spike-in fraction: if sorted Co-IP and our method both recover it, sorting remains sufficient; if only ours does, the method earns its existence.
The strongest computational competitor is the PINNACLE-style transcript overlay: map single-cell expression onto a static reference interactome and predict cell-type-specific complexes from transcript presence (PINNACLE; the same overlay logic underlies scPPIN and AWARE). We will run PINNACLE's released contextualization on our own scRNA-seq references with identical compute and tuning budget, and compare on physical confirmation: PLA-validated precision of top-ranked subpopulation-specific interactions. PINNACLE's reported strength is on benchmark rankings — for rheumatoid arthritis target nomination it beat random walk in 100% of cell types, GAT in 44.2%, and BIONIC in 11.5% at APR@10 — but those are disease-gene association benchmarks, not physical interaction measurements, and its structural binding-discrimination gains (a 0.011 score gap between binding and non-binding pairs at PD-1/PD-L1 versus 0.003 for context-free models) required 3D priors rather than transcript presence alone. Beating it means a higher PLA-confirmed precision on the same samples, which is a test no contextual model has yet passed.
The key ablation isolates the ratio series. Running the identical pipeline with a single mixture composition instead of four collapses the identifiability of $\theta_{jk}$ — proportions and occupancy become confounded, and the model can no longer distinguish "expressed but unbound" from "bound but rare." A second ablation removes the lysis-mixing control, predicted to admit artifact edges into the subpopulation assignments at a measurable rate. Both ablations use the same data, crosslinking, and MS time, differing only in the design and filtering steps.
Why not just…?
Why not just infer complexes from single-cell RNA co-expression? The strongest evidence against this is within scPPIN itself: the physically important connector proteins EGFR and APP in liver subclusters ranked 1,800th and 373rd in differential expression — statistically invisible to any co-expression screen — yet they were the hidden hubs the module detection needed (scPPIN). Co-expression also cannot see post-translational state, which Sun et al. show matters directly: the histones co-purified with influenza NP carried multi-site lysine acetylation, a modification with no transcriptomic shadow (Sun et al.). Our mixture experiment will quantify the gap: prey proteins recovered by pulldown deconvolution that co-expression ranks as noise.
Why not just FACS-sort and run standard Co-IP? This is the real threat, and we treat it as the primary baseline rather than argue around it. For abundant, marker-positive subpopulations it will work and should be preferred — our ground truth depends on it. It fails on three documented fronts: input mass at rare fractions, marker absence for transitional states (Damdinsuren et al.), and sorting-induced dissociation stress that can disrupt labile complexes before lysis. The 3% spike-in experiment settles it empirically: if sorted Co-IP recovers the known interactions at that fraction, the simpler method suffices and we will report that.
Why not just run PINNACLE or AWARE? Their own authors state the limiting defect: the reference interactome is static and unmeasured at cell-type resolution, so the cell-type networks inherit false positives and negatives by construction (PINNACLE). They also train on healthy atlases (Tabula Sapiens, Human Cell Atlas), so they cannot represent perturbation-driven interactome remodeling — the kind Weber et al. documented when CAR rest epigenetically reprograms exhausted T cells into memory-like fates, or that Lin et al. observed under tamoxifen selection (Weber et al.; Lin et al.). The head-to-head PLA-confirmation experiment is the clean test: prediction from transcript overlay versus measurement from pulldown, on identical samples.
Why not just use PLA everywhere? PLA gives true single-cell spatial resolution without lysis and is our chosen orthogonal validator (Foster et al.). But it requires two pre-defined antibodies against two suspected partners — it cannot discover a 17-protein spliceosome recruitment from an unbiased pulldown, which is exactly what Sun et al.'s Co-IP-MS achieved (Sun et al.). PLA confirms candidates; it does not generate them at network scale. Our design uses discovery where discovery is needed and PLA where physical per-cell confirmation is needed.
Decisive Checks
The proposal stands or falls on three numbers. First, at the 3% spike-in fraction, interaction recovery of the deconvolved pulldown versus both bulk Co-IP-MS and sorted Co-IP-MS — the method must recover known interactions that the bulk read dilutes away, and be honest about whether sorting still reaches them. Second, the false tripartite-complex rate: in designed mixtures where bait-A and bait-B interactions live in different subpopulations, the naive bulk read must show a measurable phantom rate and the deconvolution must suppress it substantially. Third, PLA-confirmed precision of subpopulation-restricted assignments versus the transcript-overlay baseline on matched MCF-7/TAMR and A549 sphere samples. If the first two hold, the field gains the thing it currently lacks — a way to ask, of any pulldown from a mixed tissue, not just "what binds the bait" but "who, among the cells, was actually bound."
References
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Deciphering Cell Type Abundance in Proteomics Data Through Graph Neural Networks
- Contextual AI models for single-cell protein biology
- Deep Contextual Learners for Protein Networks
- Functional module detection through integration of single-cell RNA sequencing data with protein–protein interaction networks
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer
- Graphein - a Python Library for Geometric Deep Learning and Network Analysis on Biomolecular Structures and Interaction Networks
- TGF-β1-induced cell growth arrest and partial differentiation is related to the suppression of Id1 in human hepatoma cells
- Transient rest restores functionality in exhausted CAR-T cells through epigenetic remodeling