Deconvolving Bulk Pull-Downs to Map Which Cells Run Active Mitophagy

When a tissue is stressed — by high glucose, a toxic drug, or hypoxia — its cells dispose of damaged mitochondria through mitophagy: a quality-control pathway in which the kinase PINK1 accumulates on the surface of a failing mitochondrion, recruits the ubiquitin ligase Parkin, and tags the organelle for lysosomal degradation. Whether this pathway is "up" or "down" in a disease is currently read almost exclusively from bulk measurements: Western blots of PINK1, Parkin, and lipidated LC3 on whole-tissue homogenates. The literature itself documents that these readings contradict each other. Qiaoru Wu and colleagues, studying diabetic kidney disease, note that different groups report both mitophagy inhibition and excessive mitophagy activation in essentially the same models, and attribute the disagreement to glucose dose and duration Qing-Re-Xiao-Zheng-Yi-Qi formula relieves kidney damage and activates mitophagy in diabetic kidney disease. We think a more basic confound is being missed: a kidney, or a lung, contains dozens of cell types whose mitophagy machinery may move in opposite directions, and every standard assay — including co-immunoprecipitation, the biochemical gold standard for physical protein interactions — averages them into one number.

Co-immunoprecipitation (Co-IP) uses an antibody against a "bait" protein to physically pull the bait and its bound partners out of a lysate; coupled to mass spectrometry (Co-IP-MS) it gives an unbiased inventory of the protein complexes a bait participates in. But lysis destroys compartmentalization and, when performed on whole tissue, aggregates proteins from every resident cell type, so a detected interaction cannot be attributed to any one lineage. This is exactly the "averaging effect" that single-cell transcriptomics exposed for RNA a decade ago Genome 2011 — yet no existing method resolves it for physical protein complexes, and mitophagy is a worst case because it is governed post-translationally (protein stabilization, phosphorylation, ubiquitination), which RNA levels do not predict. The proposal here is to build the first cell-type-resolved physical interaction map of the mitophagy machinery and use it as the anchor for a computational method — interactome deconvolution — that reads bulk Co-IP-MS data from heterogeneous tissue and reports which cell types harbor active PINK1–Parkin complexes.

Core Claim

The claim is that the apparent contradictions in bulk mitophagy measurements are partly composition artifacts, and that they can be decomposed. Concretely: a bulk Co-IP-MS pull-down of a mitophagy bait from diseased tissue is a mixture over cell types, and the mixture weights and per-cell-type prey profiles can be estimated from integer spectral counts if (and only if) one has a measured, cell-type-specific reference interactome. We will generate that reference ourselves — Co-IP-MS on fluorescence-sorted cell fractions — because it does not exist: the closest computational systems, PINNACLE and Shusi, prune static, context-free protein-interaction databases with single-cell mRNA and openly concede that their cell-type-specific edges were never physically measured in any one cell type Contextual AI models for single-cell protein biology, Systematic discovery of single-cell protein networks in cancer with Shusi. Our method extends the integer-count mixture machinery of Count Bridges — a generative model that treats sequencing counts as discrete Poisson-like quantities rather than continuous real numbers Count Bridges enable Modeling and Deconvolving Transcriptomic Data — from transcript counts to prey spectral counts.

Formally, the observed integer prey-count vector $y$ from a bulk pull-down of one bait is modeled as

$$y = \sum_{k=1}^{K} \pi_k \, x_k, \qquad x_k \sim \text{Poisson}(\lambda_k),$$

where $x_k$ is the prey-count profile that would be measured if only cell type $k$ were present, $\lambda_k$ its rate vector learned from the sorted-fraction reference, and $\pi_k$ the fraction of bait complexes contributed by type $k$, estimated by expectation-maximization exactly as Count Bridges does for bulk RNA mixtures. The scientific payoff is a decomposition test: when bulk mitophagy readouts change between disease stages, does the change live in $\pi$ (composition — different cells present) or in $\lambda_k$ (per-cell activity — the same cells running more or less mitophagy)? These two explanations imply entirely different therapies and have never been separable with existing assays.

Research Program

The program has three parts: build the reference, build the model, and run the decomposition test on a documented contradiction.

First, the reference atlas. We will use mouse kidney, because the file gives us two well-characterized injury systems in this organ: streptozotocin-induced diabetic kidney disease with unilateral nephrectomy, where QRXZYQF restored PINK1/Parkin mitophagy measured by bulk Westerns in glomeruli and MPC5 podocytes Qing-Re-Xiao-Zheng-Yi-Qi formula relieves kidney damage and activates mitophagy in diabetic kidney disease, and cisplatin acute kidney injury, where the NAT10–DDX17 physical interaction was established by Co-IP on whole cortex and HK-2 monocultures, with the authors themselves stating that lentiviral knockdown affected all cortex cell types and calling for cell-type-specific resolution NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17. We will dissociate kidneys from control, cisplatin-injured, and diabetic animals, sort podocytes, proximal tubular epithelial cells, endothelial cells, and infiltrating immune cells by fluorescence-activated cell sorting, and perform Parkin-bait and p62/SQSTM1-bait Co-IP-MS separately on each fraction under mild non-ionic detergent conditions, following the unbiased Co-IP-MS discovery protocol demonstrated for influenza nucleoprotein Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9). In parallel we will sequence single-cell RNA from the same kidneys so that every fraction has both a physical interactome and a transcriptome. This yields the first measured cell-type-specific edges for an organellar quality-control module — something neither PINNACLE's BioGRID/HuRI-derived networks nor Shusi's STRING-E-derived networks contain.

Second, the model. Interactome deconvolution is fitted on the sorted-fraction reference and applied to bulk Co-IP-MS runs of the same baits on unsorted homogenates. We will additionally borrow the ontology-aware alignment idea from POPPY, which improved bulk-to-single-cell mapping by incorporating the Cell Ontology tree and Earth Mover's Distance between cell-type distributions Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, to regularize $\pi$ across related cell states and to attribute bulk changes to specific subsets, the same cross-attention attribution role POPPY played for immunotherapy cohorts.

Third, the decomposition test on diabetic kidney disease. If the documented mitophagy contradiction is a composition effect, deconvolving bulk Co-IP-MS from early and late diabetic kidneys should reveal, for example, podocyte fractions losing mitophagy-complex assembly while tubular fractions mount compensatory mitophagy — opposite signs invisible in the bulk average, in the same way rare stem-like cells drove tamoxifen resistance behind stable bulk expression 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. The model also lets us probe the alternative disposal route documented by Su and colleagues — shedding damaged mitochondria in migrasomes rather than degrading them in lysosomes Migrasomes as intercellular messengers: potential in the pathological mechanism, diagnosis and treatment of clinical diseases — because immune-cell fractions with high migrasome activity should show a distinct prey signature (extracellular-vesicle machinery rather than LC3-tethering adaptors).

Proposed Experiments

The wet-lab benchmark is the sorted-fraction atlas: three injury arms (control, cisplatin, streptozotocin-diabetic), four sorted fractions per kidney, two baits (Parkin, p62), three biological replicates — roughly 72 Co-IP-MS runs, plus matched single-cell RNA-seq. The computational evaluation uses two held-out designs: leave-one-cell-type-out (can the model recover a fraction's interactome when that fraction is excluded from training?) and leave-one-injury-out (train on cisplatin, test on diabetic, and vice versa). The primary metric is a validation score in the sense Shusi uses — the fraction of predicted cell-type-specific bait–prey edges recovered as enriched prey in the measured Co-IP-MS of that fraction — supplemented by Poisson deviance on held-out prey counts. Orthogonal spatial validation will use proximity ligation assays, which light up as fluorescent puncta when two proteins sit within 40 nm inside intact tissue Validating Physical Protein Interactions via Co-Immunoprecipitation: Parkin–p62 puncta co-stained with cell-type markers should rank cell types in the same order the model estimates for $\pi_k$.

Baselines

The do-nothing reference point is standard bulk Co-IP-MS on whole-kidney homogenate — exactly what the NAT10–DDX17 study did on cortex and what the H7N9 study did on HEK293T lysate NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17, Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9). We will run this pipeline on the same kidneys, same baits, same mass-spectrometry time and search settings, so the only difference is that the homogenate is never sorted. This baseline is the strongest competitor for the question "is there active mitophagy machinery anywhere in this tissue," and the comparison that matters is per-cell-type attribution, where the bulk pull-down by construction returns no answer; the relevant prior is Shusi's finding that context-free interactomes achieve less than 1% consistency against cell-type-resolved interaction landscapes Systematic discovery of single-cell protein networks in cancer with Shusi.

The strongest existing method is Shusi itself: a language-model-enhanced variational graph auto-encoder that combines protein sequences, textual functional descriptions, single-cell RNA-seq across 23 cancers, and the STRING-E interactome to predict single-cell-specific PPI rewiring; on its own benchmark it achieved a 60.67% validation score against independent AP-MS data, with 70.54% precision where graph baselines ranged from 0.03% to 18.58% Systematic discovery of single-cell protein networks in cancer with Shusi. We will retrain Shusi on our matched scRNA-seq plus STRING-E, give it the same compute and tuning budget as our model, and ask it to predict which cell types carry active Parkin and p62 complexes. Beating it means a materially higher validation score against our measured sorted-fraction interactomes — our expectation is that it will struggle precisely because its AP-MS training signal comes from bulk cell-line lysates, but this is an empirical question the comparison will settle. We will run PINNACLE, the geometric graph neural network that contextualized 394,760 protein representations across 156 cell types by pruning a global interactome with Tabula Sapiens expression, as a second computational baseline Contextual AI models for single-cell protein biology; it reported beating random walk in 100% of cell types, GAT in 44.2%, and BIONIC in 11.5% at APR@10 on therapeutic target nomination, but its edges are inferred, not measured, which is the axis we differ on.

The key ablation isolates the proposal's central ingredient: the same mixture model, fitted with reference profiles $\lambda_k$ taken from a static bulk Co-IP reference (one unsorted pull-down per injury arm) instead of the sorted-fraction atlas. If the sorted-fraction reference is what carries the signal, this ablation should collapse toward the context-free baseline's performance; if it performs as well, the expensive sorting step is unnecessary and the claim must be weakened accordingly.

Why not just…?

Why not just run bulk Co-IP or AP-MS on the whole tissue, as everyone does? Because it cannot say which cells carry the complex, and the field's own leading papers say so: the NAT10–DDX17 authors explicitly acknowledge their whole-cortex knockdown hit immune and stromal cells indiscriminately and call for cell-type-specific models NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17, and Shusi quantified that context-free interactomes achieve under 1% consistency with cell-type-resolved landscapes Systematic discovery of single-cell protein networks in cancer with Shusi. The decisive experiment is the leave-one-cell-type-out test: the bulk pull-down, run on identical tissue, simply has no output on that axis.

Why not just do single-cell RNA sequencing and read out mitophagy gene expression per cluster? This is the most serious alternative, and we run it ourselves as an input — but it measures the wrong quantity. PINK1 is continuously transcribed and translated in healthy cells and degraded on import; it accumulates on the outer mitochondrial membrane only when membrane potential collapses, a purely post-translational gate Mitophagy Regulation by Kangxian Yixin Granule in a Mouse Model of Dilated Cardiomyopathy. PIKE-R2P demonstrates the general transcript-to-protein gap, predicting protein abundance from single-cell RNA only with the help of a protein-interaction prior, and then only for a handful of surface markers PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction. And co-expression cannot distinguish direct binding from indirect cascades — the very distinction Co-IP exists to make, as shown when IQGAP3 was confirmed to bind GLI1 directly rather than act through a kinase chain Validating Physical Protein Interactions via Co-Immunoprecipitation. Our experiments include the direct check: correlation between per-cell-type PINK1/PRKN mRNA and measured PINK1-complex assembly in the matched fractions, which we expect to be weak.

Why not just use proximity ligation assays in tissue sections and skip the computational work? PLA gives real per-cell-type interaction readouts, and we use it as orthogonal validation. But it requires pre-specified candidate pairs and validated antibody duos — it is a confirmation assay, not a discovery engine — and it does not give the genome-wide transcriptome of the cell harboring the puncta Validating Physical Protein Interactions via Co-Immunoprecipitation. It also cannot answer the retrospective question we care about: explaining bulk Co-IP-MS data already collected across disease models, which only a mixture model fitted to a measured reference can do.

Why not just use off-the-shelf bulk deconvolution tools like CIBERSORTx, MuSiC, or SCDC? Three reasons grounded in the file: they treat counts as continuous reals, violating the integer geometry that Count Bridges exploited to outperform them on bulk RNA deconvolution Count Bridges enable Modeling and Deconvolving Transcriptomic Data; they need a single-cell reference atlas that, for physical mitophagy complexes, does not exist until we build it; and their output is cell-type proportions, not the per-cell-type active-complex profiles $\lambda_k$ that the scientific question turns on.

Decisive Checks

The premise is falsifiable at the first step: if the sorted-fraction Co-IP-MS shows essentially the same Parkin and p62 interactome in every cell type and injury arm, then cell-type resolution adds nothing to mitophagy measurement and the proposal dies early — cheaply and informatively. If the fractions do differ, the decisive downstream result is the decomposition of the diabetic-kidney contradiction: a finding that bulk mitophagy "changes" are driven by $\pi$ (composition) rather than $\lambda_k$ (per-cell activity) would explain why different DKD protocols reported opposite directions Qing-Re-Xiao-Zheng-Yi-Qi formula relieves kidney damage and activates mitophagy in diabetic kidney disease, and would argue that mitophagy trials need cell-type-targeted delivery rather than global activators or inhibitors — a concern already raised for mtROS therapeutics in lung disease Role of mitochondrial stress and the NLRP3 inflammasome in lung diseases. The benchmark to beat is Shusi's 60.67% validation score against physical AP-MS data; no reported result found on Lacuna for any method's cell-type-specific validation against sorted-primary-tissue Co-IP-MS, because no such reference data exists — which is the point.