Paracrine Remodeling of Autophagy Adaptor Complexes Across the In Vitro–In Vivo Gap

Cells selectively dispose of damaged organelles or dangerous protein complexes by tagging them for delivery to the lysosome, the cell's degradation compartment. This disposal is not bulk housekeeping: it is carried out by physical protein complexes — an adaptor protein that simultaneously grips the doomed cargo and the membrane of the disposal vesicle — and these complexes have been mapped almost exclusively with co-immunoprecipitation (Co-IP), a method in which an antibody attached to a bead fishes one protein out of a gently lysed cell along with whatever partners were bound to it. The problem is that every such map in the literature was drawn in isolated cells grown flat on plastic, while cells in a living tissue are continuously bathed in secreted signals from their neighbors — so-called paracrine signaling. Nobody has asked the obvious question: do those neighbor-derived signals change which adaptor complexes physically assemble, and therefore change what the cell degrades?

This question sits exactly on top of a well-documented failure mode in biomedicine: mechanisms discovered in cell culture routinely fail to reproduce in animals or patients, a shortfall the field calls the in vitro–in vivo translation gap. The existing computational attempts to bridge that gap work only on mRNA measurements and ignore protein complexes and degradation entirely. I propose that a specific, measurable, and currently invisible slice of the translation gap is paracrine remodeling of autophagy adaptor complexes — and that we can test this directly by measuring the same physical complexes and the same degradation rates at each rung of a ladder running from 2D monoculture, through paracrine co-culture, into perfused 3D tissue-like matrices, and finally into living animals.

Core Claim

The claim is that paracrine secretomes from stromal and immune neighbors quantitatively remodel the adaptor–cargo protein complexes that route specific cargoes to lysosomal degradation, and that this remodeling — not transcriptional change — accounts for a large share of the discrepancy between in vitro and in vivo measurements of selective autophagy. Operationally, define the adaptor occupancy $\alpha$ as the fraction of a cargo protein that co-precipitates with its adaptor in a Co-IP assay, normalized to total cargo, and let $t_{1/2}$ be the cargo's degradation half-life measured by cycloheximide chase (a technique where new protein synthesis is blocked so the decay of existing protein can be timed). The claim predicts that $\alpha$ and $t_{1/2}$ measured in 2D monoculture fail to predict their in vivo values, that paracrine conditioning of the same cells in co-culture shifts both measurably, and that the paracrine-conditioned values close much of that gap:

$$D = \left| \log t_{1/2}^{\text{2D}} - \log t_{1/2}^{\text{in vivo}} \right| \;>\; D' = \left| \log t_{1/2}^{\text{paracrine}} - \log t_{1/2}^{\text{in vivo}} \right|$$

where $D$ and $D'$ are the in vitro–in vivo discrepancies with and without paracrine conditioning. This comparison is the decisive number the project delivers.

The choice of model system is deliberate. The cleanest fully mapped adaptor cascade in the literature is the CD317–calnexin–RACK1 axis: Jian Cheng et al. used targeted Co-IP to show that the membrane protein CD317 physically complexes with both calnexin (an ER chaperone) and the autophagy receptor RACK1, routing calnexin into autophagosomes during proteasome-inhibitor treatment in hematologic cell lines — and their own text flags that this pathway lacks validation in multicellular microenvironments where stromal paracrine factors modulate drug sensitivity. That is the explicit invitation this proposal accepts. A second, independent cargo — the p62-bridged autophagic clearance of the mutant FLT3-ITD receptor under arsenic trioxide, established by Co-IP in Xiaojian Liu et al. — provides a generality check, since the authors likewise note the in vivo leukemic niche was never tested.

Research Program

The program has four phases, each producing data the next one consumes.

Phase 1 — replicate and instrument the 2D baseline. In bortezomib-treated K562 and NCI-H929 cells, I will reproduce the CD317–calnexin–RACK1 Co-IP and calnexin degradation, then upgrade the readouts to quantitative ones: adaptor occupancy $\alpha$ from densitometry-normalized Co-IP westerns, $t_{1/2}$ from cycloheximide chase, and lysosomal delivery from organelle-targeted tandem fluorescent flux reporters of the mRFP-GFP type already standard in the field. To address the known lysis artifacts of Co-IP — detergent solubilization can force proteins from different compartments into false contact or rip apart weak, transient complexes — I will apply a membrane-permeable reversible cross-linker before lysis so that complexes are frozen in their in-cell state, following the logic that motivated the in situ proximity ligation validations of Ian T. Foster et al., which confirmed sub-40-nanometer proximity but, crucially, could not resolve multi-protein complex composition the way Co-IP can.

Phase 2 — paracrine remodeling in vitro. The same measurements are repeated with the target cells in transwell non-contact co-culture with macrophage and stromal donor populations, the standard apparatus of paracrine analysis. The candidate mediators come from the one reciprocal paracrine circuit fully characterized in the literature — the CXCL12-driven GM-CSF/HB-EGF feedback loop between tumor cells and macrophages mapped by Antonella Rigo et al. — but I will not assume them. Instead, conditioned medium will be fractionated and tested with neutralizing antibodies and receptor blockers, and, in parallel, bait pull-downs followed by mass spectrometry (Co-IP-MS, the discovery mode demonstrated for interactome mapping by Juan Zhao et al. and Ningning Sun et al.) will reveal, unbiased, whether paracrine exposure adds or removes members of the CD317-centered complex — for example whether additional scaffold proteins appear that compete with RACK1. The prediction is a dose- and time-dependent shift in $\alpha$ and $t_{1/2}$ that is reproducible by conditioned medium transfer and reversible by a single neutralizing antibody.

Phase 3 — the translation ladder. The paracrine-conditioned and unconditioned states are then carried up three rungs: first, 3D extracellular-matrix hydrogels under interstitial flow, because the translation-gap literature shows static planar culture misses matrix binding kinetics and fluid shear that govern signaling assemblies (Zhiyuan Xu et al. on shear-dependent vascular invasion; Alex Scott et al. on paracrine effects requiring intact-animal evaluation). Second, an intermediate in vivo rung using the zebrafish delayed-mosaic transgenesis platform of Library transgenesis in zebrafish, which permits pooled testing of transgene-modified cells inside a living organism at throughput no mouse model offers. Third, an orthotopic mouse model with bortezomib treatment, where Co-IP from recovered tumor cells (cross-linked in situ before lysis) and proximity ligation on tissue sections give the in vivo ground truth for $\alpha$, and ex vivo cycloheximide chase on recovered cells gives $t_{1/2}$. Where feasible, primary patient material anchors the clinical end: the CD317 study itself correlated CD317 expression with poor survival at p = 0.0045 in a patient cohort (Jian Cheng et al.), and the NPM1-leukemia study of Glycolytic Enzyme PKM2 shows the scale of feasible aspirate work (30 diagnostic bone marrow samples). A clinical question immediately follows from Yuan‐Tong Liu et al.: their LIMP-2 study showed autophagic GSK3β degradation maintains cancer stemness in vivo yet never identified the bridging adaptor by Co-IP, and explicitly left open whether stromal paracrine cues feed back into this degradation. Phase 2's discovery pipeline applied to that system would fill that hole as a second deliverable.

Phase 4 — a quantitative translation model with physical features. Finally, the measurements become features in a transfer model of the DeepSADR type. DeepSADR showed that drug mechanisms shift between cell lines and patients and built a transfer-learning bridge from in vitro DepMap screens (966 cell lines, 20 drugs, roughly 225,781 response samples) to 555 TCGA patient samples — but its representations are entirely chemical substructures and pathway-level gene expression. I will test whether adding paracrine-conditioned adaptor occupancy and flux features to the same transcriptomic backbone improves prediction of in vivo degradation flux and, downstream, of drug response. This is a direct answer to the omission the field itself documents: none of the current gap-bridging models — DeepSADR, Cell-Level Virtual Screening, VCWorld, SCALE — includes a single protein-complex or organelle-degradation feature, even though post-translational degradation rewires phenotypes without changing mRNA, as the GSK3β and calnexin cases both demonstrate.

Baselines

The first reference point is the "do nothing" baseline that the entire field currently relies on: take the 2D monoculture Co-IP occupancy and cycloheximide-chase half-life as the truth and assume they hold in vivo. This is what every cited autophagy-adaptor study effectively does (Jian Cheng et al., Xiaojian Liu et al.), and no reported result found on Lacuna for how well in vitro adaptor-complex measurements predict in vivo degradation rates — that absence is precisely the gap. It will be run by carrying the Phase 1 numbers forward unmodified and scoring them against the Phase 3 in vivo measurements; its prediction error $D$ is the quantity the proposal must beat.

The strongest existing competitor is the transcriptomics-only transfer model in the DeepSADR family. DeepSADR reports an average clinical response AUC of 0.856 across five drugs, a Sorafenib AUC of 0.957 versus 0.811 for the next best baseline, and, importantly for us, a drop to 0.662 when its adaptive readout module is ablated — establishing both its strength and its sensitivity to representation choices. On the drug-response task I will retrain this architecture on the matched in vitro/in vivo pairs my ladder generates, with the same feature budget, tuning budget, and cross-validation folds as my extended model, so the only difference is the presence of adaptor-occupancy and flux features. On the flux-prediction task (predicting in vivo $t_{1/2}$), the transcriptomics-only model is retrained identically; beating it means lower median log-error on held-out in vivo conditions, with the DeepSADR reported AUC of 0.856 serving as the calibration point for what expression features alone achieve on patient endpoints.

The key ablation isolates the proposal's central ingredient: the paracrine-conditioning step. The full pipeline is run identically but with target cells never exposed to a donor secretome — same 3D matrix, same flow, same in vivo recovery, same Co-IP and chase protocols. If the predictive gain survives without paracrine exposure, the improvement came from 3D mechanics or measurement upgrades, not from paracrine remodeling, and the core claim fails. A second, softer ablation conditions cells with a cocktail of pure recombinant candidate ligands instead of live co-culture; the prediction, grounded in Jie Xia's finding that co-cultured secretomes are non-additive and exceed anything recombinant dosing reproduces, is that this partially but incompletely rescues the shift.

Why not just…?

Why not just dose recombinant cytokines on monoculture? This is the cheapest experiment imaginable and it will be run — as a control, not a substitute. The evidence against it as a stand-in is direct: Jie Xia showed that the co-culture secretome is synergistic, producing cytokine outputs absent from monocultures treated with single recombinant factors, and Antonella Rigo et al. showed the macrophage–tumor signaling there is a reciprocal loop (tumor GM-CSF provoking macrophage HB-EGF, itself primed by CXCL12) that a one-time ligand dose cannot sustain. If, against expectation, the recombinant cocktail in my ablation fully reproduces the co-culture shift in $\alpha$ and $t_{1/2}$, that would be a genuine simplification and I would report it as such — the ablation is designed to settle this rather than argue around it.

Why not just use proximity ligation or computational docking instead of Co-IP? Proximity ligation is attractive because it works directly in tissue, and it will be used as the orthogonal in vivo confirmation, in the spirit of Ian T. Foster et al.. But it reports that two proteins sit within roughly 40 nanometers; it cannot say whether CD317, calnexin, and RACK1 form one ternary degradative intermediate or merely co-localize, which is exactly the composition question on which the mechanism turns — a limitation Jian Cheng et al. faced and resolved only with Co-IP. Computational binding prediction is weaker still: on phenotypic retrieval in Cell-Level Virtual Screening, molecular-fingerprint baselines reached a Hits@1 of only 0.0357 against 0.1250 for cell-level models, and docking cannot tell you whether a predicted interaction actually routes cargo to a lysosome. The proposal's honest weakness — that even cross-linked Co-IP is an averaged, semi-quantitative readout over heterogeneous subpopulations — is mitigated, not solved, by pairing it with single-cell-resolution proximity ligation in tissue.

Why not just measure bulk autophagy markers or run the in silico models? Bulk LC3 westerns and autophagy-gene transcriptomics are the standard shortcuts, and both demonstrably miss the phenomenon at issue. Jian Cheng et al. rescued ER calcium refilling through the selective removal of one specific cargo that bulk autophagy markers could not explain, and Yuan‐Tong Liu et al. showed GSK3β clearance rewires cell fate post-translationally, with no corresponding mRNA change — the exact failure mode that gene-set-based prediction exhibits at scale, with Alain R. Bateman et al. documenting wild instability of transcriptomic signatures across gene-set collections. As for skipping in vivo work entirely: the CellVS-Net authors themselves state their virtual screening network does not yet predict human-subject outcomes (Cell-Level Virtual Screening), and Huatao Li) showed test-tube binding measurements can be wholly undone by in vivo processing. In vivo measurement here is the ground truth the models are eventually trained on, not a luxury. That said, the computational alternative is a real threat to the modeling phase, which is why it is the strongest baseline rather than something dismissed.

Decisive Checks

Three results would decide the proposal's fate. First, paracrine conditioning must shift $\alpha$ and $t_{1/2}$ in vitro in a mediator-attributable way: reproducible by conditioned medium transfer, reversed by a single neutralizing antibody, and absent when donor cells are engineered not to secrete the implicated factor. Second, the ladder must show a monotone story: the 2D baseline error $D$ against in vivo ground truth should be large, the paracrine-conditioned error $D'$ substantially smaller, and the 3D-mechanics-only ablation should tell us how much of the closure belongs to paracrine signals versus physical context. Third, the extended transfer model with occupancy and flux features must beat the transcriptomics-only retrained baseline on held-out in vivo flux prediction, with the DeepSADR-reported average AUC of 0.856 on clinical response as the reference performance level for what expression features alone deliver. If the paracrine shift proves reproducible with recombinant ligands alone, or if occupancy features add nothing over transcriptomics, the proposal still yields a useful negative result — a bound on how much of the translation gap this mechanism can explain — but the core claim would be falsified, and I would say so.

References

  1. CD317 maintains proteostasis and cell survival in response to proteasome inhibitors by targeting calnexin for RACK1-mediated autophagic degradation
  2. Arsenic trioxide induces autophagic degradation of the FLT3-ITD mutated protein in FLT3-ITD acute myeloid leukemia cells
  3. HPF-MPI: A Programming System Supporting Task and Data Parallelism
  4. Macrophages may promote cancer growth via a GM-CSF/HB-EGF paracrine loop that is enhanced by CXCL12
  5. Orbitrap Astral–based proteome and phosphoproteome analysis identifies candidate proteins associated with the phosphatidate phosphatase MoPah1 in Magnaporthe oryzae
  6. Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
  7. FOXC1 Contributes To Microvascular Invasion In Primary Hepatocellular Carcinoma Via Regulating Epithelial-Mesenchymal Transition
  8. Mast cells exert pro-inflammatory effects of relevance to the pathophyisology of tendinopathy
  9. Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes
  10. Glycolytic Enzyme PKM2 Mediates Autophagic Activation to Promote Cell Survival in NPM1-Mutated Leukemia
  11. LIMP-2 enhances cancer stem-like cell properties by promoting autophagy-induced GSK3β degradation in head and neck squamous cell carcinoma
  12. DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
  13. Cell-Level Virtual Screening
  14. VCWorld: A Biological World Model for Virtual Cell Simulation
  15. SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction
  16. Interleukin-1β induces intercellular adhesion molecule-1 expression, thus enhancing the adhesion between mesenchymal stem cells and endothelial progenitor cells via the p38 MAPK signaling pathway
  17. Importance of collection in gene set enrichment analysis of drug response in cancer cell lines