Which Cells Lose Their Proteasomes First?

Every cell runs a disposal machine called the proteasome — a multi-part protein complex that chops up damaged or misfolded proteins. In the aging vertebrate brain, this machine slows down surprisingly early: well before any visible protein clumps appear, and early enough that its decline in young adulthood predicts how long the animal will live Reduced proteasome activity in the aging brain results in ribosome stoichiometry loss and aggregation. When the proteasome falters, protein complexes that must be built in exact subunit ratios fall apart stoichiometrically, and their orphaned parts aggregate. Yet everything we know about this early failure comes from homogenized whole tissue, which averages thousands of dissimilar cell types into one number. Meanwhile, single-cell RNA sequencing of the aged brain shows that different cell lineages age in opposite directions — but it cannot see protein complexes at all, because RNA levels are a poor proxy for assembled, working proteins.

The proposal here is to close both blind spots at once. I will combine single-cell transcriptomics, cell sorting, and co-immunoprecipitation — a method that uses an antibody against one "bait" protein to physically pull down the intact complexes it belongs to, followed by mass spectrometry to identify every "prey" partner — to build the first lineage-resolved map of which specific brain cell types lose proteasome and chaperone-triage complex integrity first, and which physical interactions break. The most vulnerable lineage and its failing interaction then become a validated therapeutic target: perturbed genetically and pharmacologically, with physical target engagement proven by the same co-IP pipeline, and efficacy tested in a short-lived vertebrate aging cohort.

Core Claim

The claim is threefold. First, proteasome-driven stoichiometric collapse in early aging is not uniform across cell types, and the lineages that transcriptionally upregulate ribosomal genes — mature neurons, in the aged-mouse-brain atlas of Single-cell transcriptomics of the aged mouse brain reveals convergent, divergent and unique aging signatures — are doing so futilely, in response to the worst underlying stoichiometric loss, not successfully adapting. Second, the earliest physically measurable lesion is the dissociation of specific assembly and triage complexes — the 19S regulatory cap on the 20S core, and the chaperone-cochaperone delivery assemblies (HSP70 with J-domain proteins, BAG-6, HSP110) that the review Co-Chaperones in Targeting and Delivery of Misfolded Proteins to the 26S Proteasome identifies as the triage machinery whose aging behavior is unmeasured. Third, stabilizing one of these upstream physical interactions is a broader therapeutic strategy than dissolving any single downstream aggregate, because it addresses the root cause that generates orphan subunits across many complex families simultaneously.

Research Program

The program has three linked aims, run in mouse brain with a killifish validation arm.

Aim 1 — a single-cell early-aging time course. Existing brain atlases sample only extremes: Single-cell transcriptomics of the aged mouse brain reveals convergent, divergent and unique aging signatures compared 2–3 month against 21–23 month animals (50,212 cells, 25 populations), so the early window in which proteasome decline begins is unsampled at single-cell resolution. I will run single-nucleus RNA-seq across a 3, 6, 12, 18, and 24 month time course, with standard preprocessing (ambient RNA removal, doublet filtering, PCA, Leiden clustering) and, critically, a pseudotime-ordered proteasome assembly score per cell: the summed expression of the assembly chaperones and 19S cap subunits whose early-adulthood decline predicted mortality in Reduced proteasome activity in the aging brain results in ribosome stoichiometry loss and aggregation. The output is a ranked list of candidate vulnerable lineages and the age at which each first deviates. The bidirectionality warning from Single-cell transcriptomics of the aged mouse brain... — 112 of 451 age-altered pathways move in opposite directions across cell types, and classic markers like Gfap switch lineages — means bulk averaging would rank the wrong cells.

Aim 2 — lineage-sorted Co-IP-MS of assembly and triage complexes. For the top-ranked vulnerable lineages (plus a resilient reference lineage), I will FACS-sort each population from pooled mouse brains and perform targeted co-IP followed by mass spectrometry on endogenous baits: 19S cap subunits (to quantify capped 26S versus uncapped 20S complexes), and HSP70-family triage baits (to quantify delivery-complex composition). Co-IP needs roughly 10^6–10^7 cells per reaction, so populations will be pooled across animals within age groups; this is feasible for abundant lineages and will be honestly reported as the resolution limit for rare ones. Each age-point and lineage also gets native PAGE with in-gel fluorogenic activity assays (Suc-LLVY-AMC for the chymotrypsin-like catalytic rate), the same assay used by Reduced proteasome activity in the aging brain..., plus size-exclusion chromatography–mass spectrometry to measure complex stoichiometry directly. The key output is a per-lineage, per-age "interaction integrity map": which bait–prey edges weaken, and in which order. Co-IP is admittedly vulnerable to lysis artifacts and loses weak, transient interactions — two known limits catalogued for this method — so every edge called as broken in aging will be cross-checked with proximity ligation assays in intact tissue sections, which fluoresce only when two proteins sit within ~40 nm of each other. This is exactly the discovery-then-orthogonal-validation pattern that made NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17 (Co-IP-MS discovery of a NAT10–DDX17 physical interaction, then functional proof) a credible validation pipeline.

Aim 3 — target validation of the earliest failing interaction. Whichever interaction fails first in the most vulnerable lineage becomes the candidate target. Validation follows the same logic as Selective Aurora A-TPX2 Interaction Inhibitors Have In Vivo Efficacy as Targeted Antimitotic Agents, which validated a physical protein–protein interface — not a catalytic pocket — as a druggable target with CAM2602 ($K_D = 19$ nM, no detectable activity across a 97-kinase panel, >1000-fold selectivity over Aurora B). Here the steps are: genetic perturbation (knockdown and overexpression of the failing partner) in primary sorted cells to show the stoichiometry and activity phenotypes depend on that interaction; a small-molecule or stabilizing-perturbation screen for compounds that maintain it, with engagement proven by cellular thermal shift assay and restored Co-IP pull-down; and an in vivo arm in Nothobranchius furzeri, the short-lived fish in which Reduced proteasome activity in the aging brain... established the longitudinal design — a cohort of ~159 animals sampled at 10 and 20 weeks with Cox proportional hazards modeling — extended with functional endpoints (grip strength, rotarod) mirroring the PKM2 study's readouts.

Baselines

The naive reference point is bulk-tissue measurement exactly as performed in Reduced proteasome activity in the aging brain results in ribosome stoichiometry loss and aggregation: whole-brain native PAGE activity assays and bulk SEC-MS. That work reports that 30% of protein complexes show stoichiometry alteration in old versus adult fish (against 11% between young and adult), and that mRNA–protein Pearson correlation falls from 0.48 in young to 0.33 in old brain, with >40% of age-related proteome variance post-transcriptional. These are the numbers my per-lineage measurements must be reconciled against: bulk values are the cell-type-weighted average of my lineage values, and any successful decomposition should reproduce them when re-aggregated. Beating this baseline means showing that the bulk average conceals real structure — specifically, that the spread of per-lineage 26S/20S activity ratios at a given age is large enough that the earliest-declining lineage crosses a defined activity threshold (say, a 25% drop from its own young-adult baseline) at least two time points before the bulk homogenate does.

The strongest competitor on the discovery side is the scPPIN pipeline from Functional module detection through integration of single-cell RNA sequencing data with protein–protein interaction networks, which maps single-cell differential-expression p-values onto BioGRID interaction networks and solves a maximum-weight connected subgraph problem to find active protein modules; it famously surfaced hidden hubs like EGFR (ranked 1800th in differential expression) in liver spheroids. I will run scPPIN on my own Aim 1 dataset — same cells, same compute, default tuning, no extra budget — and treat its module calls as the competing hypothesis set. Beating it means one concrete thing: among the assembly and triage interactions my Co-IP-MS measures as age-fragmented in sorted lineages, scPPIN's static-network modules fail to predict a substantial fraction (pre-registered threshold: at least half), and its predicted module members show no measurable interaction change in the corresponding Co-IP. PIKE-R2P from PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction is the related deep-learning competitor, but it predicts abundance of only 10–13 surface proteins from CITE-seq data, and no reported MSE or Pearson values for its CBMC/PBMC benchmarks were found on Lacuna, so it cannot be used as a quantitative baseline here; it functions only as a design contrast.

On the therapeutic side, the comparator is the downstream-aggregate strategy of PKM2 aggregation drives metabolism reprogramming during aging process, where compounds K35 and K27 (binding confirmed by microscale thermophoresis) dissolve PKM2 aggregates and extend lifespan in aged mice; the exact lifespan-extension percentage is not reported on Lacuna, so the comparison will be run head-to-head in the same killifish cohort rather than against a published figure. Beating it means broader rescue: an upstream intervention should restore stoichiometric balance across multiple complex families (measured by SEC-MS), not just one aggregate species, at matched target-engagement doses.

Finally, one ablation isolates this proposal's key ingredient: rerun the full analysis pipeline with the Co-IP-MS-derived empirical interaction edges replaced by static BioGRID/STRING edges of the kind scPPIN and PIKE-R2P use. If the lineage vulnerability ranking and the chosen target are unchanged, the empirical interactome added nothing and the simpler computational route should have won; that is a real possible outcome and I state it as a falsification condition in advance.

Why not just…?

Why not just do bulk proteomics across ages, as Sacramento et al. already did? Because bulk averaging provably cancels the signal we need. In Single-cell transcriptomics of the aged mouse brain..., 112 of 451 age-altered pathways move in opposite directions across cell types, and ribosomal biogenesis is upregulated in mature neurons but downregulated in neural stem cells — exactly opposite shifts that sum to near-zero in a homogenate. And Mouse aging cell atlas analysis reveals global and cell type-specific aging signatures shows cell types age at radically different rates (immune and stem cells scoring high on a 330-gene Global Aging Gene score, parenchymal cells low). Bulk proteomics found the phenomenon; it structurally cannot localize it.

Why not just run scPPIN or PIKE-R2P on public aging atlases and skip the wet lab? Because these tools anchor everything to mRNA and a static interaction database, and the central quantity here — assembled, functional complexes — is post-transcriptional by the strongest number in the file: mRNA–protein correlation drops from 0.48 to 0.33 with age and >40% of proteome variance is post-transcriptional Reduced proteasome activity in the aging brain.... A network that assumes a 1:1 mRNA-to-protein mapping cannot, even in principle, predict which physical complex dissociates in which lineage. This is precisely what the scPPIN head-to-head baseline tests empirically, so the claim does not rest on argument alone.

Why not just dissolve downstream aggregates, like K35/K27 for PKM2? This is the most serious alternative, because PKM2 aggregation drives metabolism reprogramming during aging process already achieved functional rescue and lifespan extension in vivo. But it treats one species while the upstream lesion generates orphans across 30% of all complexes Reduced proteasome activity in the aging brain..., and newly synthesized wild-type proteins keep falling into the collapsing proteostatic environment Intrinsically aggregation-prone proteins form amyloid-like aggregates and contribute to tissue aging in Caenorhabditis elegans. The killifish head-to-head (upstream stabilizer versus a PKM2-aggregate-dissolving comparator, multi-complex stoichiometry as the primary endpoint) is designed to settle this directly, and if the downstream approach rescues stoichiometry as broadly, it wins.

Why not replace Co-IP entirely with proximity assays (PLA, BiFC)? Because those require a pre-existing hypothesis about the interacting pair and, for BiFC, genetic fusion tags — they cannot discover which interactions break among unknowns. Co-IP-MS is the only unbiased option for mapping an uncharacterized aging interactome, a role it played in the NAT10–DDX17 discovery NAT10 Promotes Tubular Epithelial Cell Senescence.... Its real weaknesses — lysis artifacts, loss of transient interactions, and incompatibility with insoluble aggregates — are handled by design: PLA as orthogonal in-situ confirmation of every "broken edge" call, and SEC-MS plus native PAGE to cover complexes that resist detergent extraction.

Proposed Experiments and Decisive Checks

The decisive experiments, in order. First, the falsification of the compensation hypothesis: if sorted mature neurons with the highest ribosomal transcript upregulation show the worst ribosomal stoichiometry loss and orphan-protein aggregation by SEC-MS, the futile-compensation claim stands; if their stoichiometry is intact, the transcript upregulation is adaptive and Aim 1's vulnerability ranking must be re-derived. Second, the ordering test: the first measurable event in the most vulnerable lineage must be loss of 19S-cap or triage-complex co-IP edges before detectable aggregate deposition, replicating at lineage resolution the temporal precedence established in bulk by Reduced proteasome activity in the aging brain.... Third, the causality test in the killifish cohort: perturbing the chosen interaction must shift proteasome activity, stoichiometry, and Cox-model mortality hazard in the predicted direction, with engagement shown by restored co-IP pull-down — matching the engagement-then-phenotype logic of CAM2602 ($K_D = 19$ nM) Selective Aurora A-TPX2 Interaction Inhibitors... and the MST-verified PKM2 binders. Fourth, the breadth test versus the PKM2-style comparator, with multi-complex stoichiometric rescue as the pre-registered primary endpoint. If all four checks pass, the field gains its first cell-type-resolved, physically validated map of the earliest proteasome lesion in aging — and its first upstream target worth pursuing.