Survival for patients with metastatic or recurrent osteosarcoma has been stuck near 20–30% for three decades, despite surgery and intensive chemotherapy. One reason is that the disease's worst prognostic drivers survive at the protein level, not the gene level: the deubiquitinase USP3 physically strips ubiquitin tags off the receptor tyrosine kinase EPHA2, extending EPHA2's lifetime from roughly six hours to a nearly stable state, which keeps PI3K/AKT signaling constantly on and drives lung metastasis — all without changing how much EPHA2 messenger RNA the tumor makes. Every transcriptomic prognostic model built for osteosarcoma is blind to this kind of event, and no selective USP3 inhibitor exists.
Cells already possess a delivery system that could, in principle, undo this: molecular chaperones. An Hsp70 chaperone is a small ATP-burning clamp that grabs exposed stretches of misfolded or newly made proteins, and its co-chaperone partners decide what happens next — hand the client to a folding platform, to a targeting complex, or to the 26S proteasome, the cell's protein shredder. This sequential hand-off is called client relay, and a central unresolved question in the field is what tips a bound client toward refolding versus destruction. We propose to test whether that decision can be biased pharmacologically in a cancer cell: whether small molecules that accelerate the relay from Hsp70 to the proteasome can override USP3-mediated stabilization and push EPHA2 back into degradation machinery — and whether the state of this relay machinery carries prognostic information that osteosarcoma's existing gene-expression signatures miss.
Core Claim
The claim is two-part and testable. First, in osteosarcoma cells the fraction of EPHA2 routed toward proteasome delivery versus USP3-mediated rescue is a measurable, perturbable property of the chaperone relay machinery, and enhancing relay — through co-chaperone manipulation first, then through designed small molecules — can restore rapid EPHA2 turnover even when USP3 is present. Second, the expression state of the relay machinery is an independent prognostic feature in osteosarcoma, detectable in existing clinical cohorts and complementary to current signatures.
The proposal uses the four input directions in sharply unequal roles. Client relay enhancement is the core mechanism; lead compound optimization is the tool-building discipline that turns a genetic observation into a drug-like molecule; co-immunoprecipitation — pulling a "bait" protein out of cell lysate with an antibody to see what physically comes with it — is both the discovery instrument (coupled to mass spectrometry) and the primary validation readout; osteosarcoma prognosis is the application and the ultimate clinical justification.
The Tension That Makes It Falsifiable
The literature contains an unresolved contradiction that this proposal is designed to settle. Yin et al. showed that the proteasome activator REGγ is an oncogene in osteosarcoma: silencing it arrests the cell cycle and induces apoptosis, implying the tumor needs heightened proteasomal activity (REG γ knockdown suppresses proliferation by inducing apoptosis and cell cycle arrest in osteosarcoma). Yet Li et al. showed osteosarcoma cells also survive by escaping proteasomal degradation of specific clients through USP3 (USP3 promotes osteosarcoma progression via deubiquitinating EPHA2 and activating the PI3K/AKT signaling pathway). Our working hypothesis resolves these: bulk proteasome capacity is not the bottleneck — client routing is. REGγ-dependent tumors degrade what they are fed, while a handful of stabilized oncoproteins are diverted away from the feed. If this is right, global proteasome or broad chaperone inhibition should fail clinically and mechanistically, while client-selective relay enhancement should succeed. If it is wrong, our own perturbation experiments will say so.
Research Program
The program has four linked phases, each grounded in precedents that stop short of this intersection.
Phase 1 — map the endogenous relay interactome. We will run co-immunoprecipitation coupled to nano-LC-MS/MS in the four osteosarcoma lines Li et al. used (143B, HOS, MG63, SAOS-2), using inducible Hsp70 and EPHA2 as baits, with crosslinked antibodies, IgG and empty-vector controls, and peptide-count thresholds exactly as established for interactome mapping by Sun et al. in their influenza nucleoprotein study (Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)) and by Zhao et al. with Orbitrap Astral instrumentation (Orbitrap Astral–based proteome and phosphoproteome analysis identifies candidate proteins associated with the phosphatidate phosphatase MoPah1 in Magnaporthe oryzae). This matters because the field's relay findings rest on yeast model substrates and grafted reporters: Cho et al. proved J-domain proteins accelerate Hsp70-to-Sgt2 transfer using a photocrosslinkable Bos1 substrate in the yeast GET pathway (J-domain proteins promote client relay from Hsp70 during tail-anchored membrane protein targeting), and the degron work of the HSP70-motif study used Ura3-HA-GFP fusion reporters in yeast and HEK293T cells (HSP70-binding motifs function as protein quality control degrons). Nobody has mapped the physical chaperone hand-off complexes around an endogenous, post-translationally stabilized human oncoprotein. We will also stratify cells by the pro-protein-synthesis (PPS-OS) phenotype — the worst-outcome subtype Yi et al. defined by single-cell clustering of GSE152048, marked by RPS27, RPS28, and JTB (Identification of a pro-protein synthesis osteosarcoma subtype for predicting prognosis and treatment) — because heavily translating cells should be the most relay-dependent, and nobody has looked at their physical chaperone complexes.
From these data we define a relay bias ratio for EPHA2:
$$R \;=\; \frac{I(\text{EPHA2}\!\cdot\!\text{BAG6})}{I(\text{EPHA2}\!\cdot\!\text{BAG6}) \;+\; I(\text{EPHA2}\!\cdot\!\text{USP3})}$$
where $I(\cdot)$ is the Co-IP-MS intensity of the corresponding co-precipitating partner. BAG6 is one of the ubiquitin-like-domain co-chaperones that physically bridge Hsp70-bound clients to the 19S proteasome cap, per the co-chaperone routing literature (Co-Chaperones in Targeting and Delivery of Misfolded Proteins to the 26S Proteasome); USP3 is the rescue arm. $R$ is a single number we can track through every perturbation below, and lysis artifacts on fragile hand-off complexes — a known Co-IP limitation (Validating Physical Protein Interactions via Co-Immunoprecipitation) — will be controlled with gentle crosslinking and orthogonal proximity assays.
Phase 2 — shift the fate genetically before doing chemistry. We will overexpress J-domain proteins and BAG-family nucleotide-exchange factors and ask whether $R$ rises and the EPHA2 half-life (measured by cycloheximide chase, following Li et al.'s ~6-hour baseline) falls, with and without USP3 co-expression. Cho et al. give the mechanistic handle: Ydj1 and Sis1 promote client transfer strictly through their J-domains stimulating Hsp70's ATPase activity, and Sis1 does this without aiding initial client capture — meaning relay enhancement is separable from chaperone binding itself (J-domain proteins promote client relay from Hsp70 during tail-anchored membrane protein targeting). ATPase-stimulation-defective J-domain mutants, grounded in the state-dependent J-domain–Hsp70 interaction (Role of the J-domain in the cooperation of Hsp40 with Hsp70) and in the interdomain-coupling BiP mutants (BiP mutants that are unable to interact with endoplasmic reticulum DnaJ proteins provide insights into interdomain interactions in BiP), serve as the mechanistic negative controls. Readouts beyond half-life: EPHA2 ubiquitination smear by Co-IP, EPHA2 association with the 19S particle, p-AKT, and migration/invasion assays, mirroring Li et al.'s endpoints.
Phase 3 — build molecules that do what the genes did. Phase 2 tells us which interface is rate-limiting; Phase 3 targets it. We will run an active-learning loop of the kind Gusev et al. built for the LRRK2 WDR domain — docking-filtered analogs from the ~5.5-billion-compound Enamine REAL library, relative binding free energies by thermodynamic integration, experimental confirmation of predicted binders, which achieved a 23% hit rate (8 confirmed inhibitors out of 35 tested) there (Active Learning-Guided Hit Optimization for the Leucine-Rich Repeat Kinase 2 WDR Domain Based on In Silico Ligand-Binding Affinities) — repointed at a dynamic protein–protein hand-off interface rather than a static pocket, which is exactly where no such pipeline has been applied. Where scaffolds need reshaping for toxicity or pharmacokinetics, we will use ScaffoldGVAE, whose multi-view graph autoencoder kept side-chain integrity with a near-100% hop rate where REINVENT2 forgot side chains below 40% hopping (ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks). Binding hits get confirmed by surface plasmon resonance and differential scanning fluorimetry before any cell assay, because the field's own lesson is that in-silico scoring — ScaffoldGVAE's MM/GBSA energies of −40 to −60 kcal/mol — produces false positives without biophysical cross-validation. We will also automate pharmacophore cluster definition, since the strongest fragment-based workflow, FragmentScout, still needs manual cluster and excluded-volume curation (A New Fragment‐Based Pharmacophore Virtual Screening Workflow Identifies Potent Inhibitors of SARS‐CoV‐2 NSP13 Helicase). Pharmacokinetic failure is a designed-in risk from the start: He et al.'s Ewing sarcoma repurposing screen showed auranofin needed intraperitoneal suspension delivery to reach therapeutic levels at all (In silico and in vitro drug screening identifies new therapeutic approaches for Ewing sarcoma), so solubility and metabolic stability filters run from the first triage round.
Phase 4 — ask whether relay state predicts outcome. Using the Phase 1 map to pick the components that matter, we build a relay-machinery expression signature and test it on the TARGET osteosarcoma cohort against the strongest published signatures. We are honest about the ceiling here: TARGET and the GEO validation cohort ($n = 90$; GTEx $n = 396$ normals) offer only transcriptomics, so we can only test an mRNA proxy of relay state — the same limitation Qiu et al. concede for their polyamine signature, which is retrospective and infers metabolism from mRNA abundance (Identification and verification of a polyamine metabolism-related gene signature for predicting prognosis and immune infiltration in osteosarcoma). The Co-IP work is what tells us which proxy genes are worth including; without it, the signature would be another co-expression exercise.
Baselines
The naive reference point is the untreated cell. In 143B and HOS cells with endogenous USP3, EPHA2 is stabilized — Li et al. report its half-life prolonged from about six hours to a stable state (USP3 promotes osteosarcoma progression via deubiquitinating EPHA2 and activating the PI3K/AKT signaling pathway). Vehicle-treated cells define the floor; any intervention counts as working only if it measurably restores rapid EPHA2 turnover and ubiquitination in the presence of USP3.
The strongest existing pharmacologic competitor is the chaperone-inhibition strategy from sarcoma drug repurposing: ganetespib, the HSP90 inhibitor He et al. identified from a 2,316-drug FDA screen as synergistic with auranofin in Ewing sarcoma xenografts (In silico and in vitro drug screening identifies new therapeutic approaches for Ewing sarcoma). It is the closest thing to a working chaperone-directed agent in pediatric sarcoma, and it attacks the same proteostasis system from the opposite direction — broad inhibition rather than selective routing. We will run it head-to-head with our relay enhancers in the same four osteosarcoma lines, at equipotent anti-proliferative concentrations, with identical assay timing, and with matched vehicle and no-treatment arms. Beating it means degrading EPHA2 at least as effectively while inducing less global proteotoxic stress — quantified as lower compensatory Hsp70 induction and less death in non-malignant osteoblast controls — and degrading the intended client selectively rather than collapsing proteostasis broadly.
For the prognostic arm, the strongest reported baseline on the TARGET cohort is the six-gene PPS-OS signature of Yi et al., with 1-, 3-, and 5-year time-dependent AUCs of 0.779, 0.847, and 0.881 respectively (Identification of a pro-protein synthesis osteosarcoma subtype for predicting prognosis and treatment) — numerically stronger at 5 years than Qiu et al.'s polyamine signature (0.799). We will refit their signature on the same TARGET training set under identical censoring, use their GEO cohort ($n = 90$) for external validation, and grant it the same tuning budget we give ours. Beating it means a higher time-dependent AUC at 5 years, or a statistically significant likelihood-ratio improvement when the relay signature is added on top of theirs.
Two ablations isolate the proposal's key ingredient. The first is an ATPase-stimulation-defective J-domain mutant (and, later, a compound analog that binds Hsp70 without enhancing relay): if the effect on EPHA2 half-life vanishes while Hsp70 binding persists, relay enhancement — not generic chaperone engagement — is the mechanism, exactly as Cho et al. found relay strictly dependent on JDP-stimulated ATPase activity (J-domain proteins promote client relay from Hsp70 during tail-anchored membrane protein targeting). The second is USP3 knockdown alone, which serves as the ceiling for any USP3-directed strategy and tells us whether relay enhancement can match direct rescue of the stabilization arm. For the computational stage, single-round Glide SP docking is the cheap baseline; Doijen et al. showed it is markedly less selective than pharmacophore-based selection (1,275 hits needed from ~26,111 structures versus 1,000 from ~10,969 for the same site), which is why we treat it as a first-pass filter only (A New Fragment‐Based Pharmacophore Virtual Screening Workflow Identifies Potent Inhibitors of SARS‐CoV‐2 NSP13 Helicase).
Why not just…?
Why not just repurpose an HSP90 inhibitor like ganetespib? This is the cheapest path to a clinical-grade compound, and He et al.'s xenograft synergy data make it genuinely attractive (In silico and in vitro drug screening identifies new therapeutic approaches for Ewing sarcoma). But broad ATP-site chaperone inhibition produces generalized proteotoxic stress, not selective oncoprotein degradation, and the same screen showed the pharmacokinetic fragility of repurposing — auranofin could not reach therapeutic systemic levels orally. More telling, the one modern subtype-guided trial in refractory osteosarcoma, Tian et al.'s tinengotinib arm, achieved a median progression-free survival of only 57 days (Targeting high-risk MYC-overexpressed osteosarcoma with an Aurora kinase inhibitor:--results from a pilot umbrella trial). Our head-to-head ganetespib arm settles the question empirically: if broad inhibition matches relay enhancement on EPHA2 degradation at tolerable stress levels, the simpler strategy wins and we will say so.
Why not just inhibit or knock down USP3? USP3 knockdown works beautifully in vitro — Li et al. showed it collapses EPHA2 signaling and lung metastasis — so this is the most serious alternative. Three things temper it. First, Li et al. themselves state that selective small-molecule USP3 inhibitors are lacking, and deubiquitinase active sites are notoriously similar, which is why the target remains undrugged years after its discovery. Second, client stabilization may be redundant across deubiquitinases, while the proteasome relay sits downstream of every stabilizer at once. Third, knockdown is not a deliverable therapy for metastatic sarcoma. We are not arguing around this: USP3 knockdown is built in as a comparator in every Phase 2 assay. If it matches relay enhancement on $R$, half-life, and metastasis endpoints, then relay enhancement's only advantage is druggability — a real but narrower claim, and we would report it that way.
Why not just read risk from mRNA signatures? Because the phenomenon we target is post-translational and therefore invisible to transcriptomics by construction: USP3 changes EPHA2's half-life without touching EPHA2 transcription, and Qiu et al. explicitly acknowledge that their signature rests on retrospective mRNA proxies rather than measured turnover (Identification and verification of a polyamine metabolism-related gene signature for predicting prognosis and immune infiltration in osteosarcoma). The honest version of our counter-argument is also empirical: if the relay signature adds no significant prognostic value over Yi et al.'s 0.881 five-year AUC on the same TARGET data, the prognosis claim fails, and only the therapeutic arm of the proposal stands.
Decisive Checks
Three results decide the proposal's fate. The first and most important: in USP3-co-expressing osteosarcoma cells, does JDP- or BAG-mediated relay enhancement (and later, the designed compound) raise $R$ and shorten EPHA2 half-life back toward six hours or below? A clean yes resolves the REGγ-verses-USP3 paradox in favor of routing and validates the whole mechanism; a no means the fate decision sits upstream of USP3, which redirects the field's attention and is itself a publishable, decision-ruling result. The second: does the best-optimized compound, validated by SPR and Co-IP for binding and routing, reduce lung metastatic nodules in the mouse model Li et al. established, with pharmacokinetics good enough to dosing without suspension workarounds? The third: does the relay-machinery signature beat or significantly augment the reported 5-year AUC of 0.881 on TARGET-OS? No reported result found on Lacuna for a chaperone-relay-directed small molecule, a chaperone Co-IP-MS interactome in osteosarcoma, or a proteostasis-based prognostic model in this disease — which is precisely the space this proposal occupies.
References
- REG γ knockdown suppresses proliferation by inducing apoptosis and cell cycle arrest in osteosarcoma
- USP3 promotes osteosarcoma progression via deubiquitinating EPHA2 and activating the PI3K/AKT signaling pathway
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- Orbitrap Astral–based proteome and phosphoproteome analysis identifies candidate proteins associated with the phosphatidate phosphatase MoPah1 in Magnaporthe oryzae
- J-domain proteins promote client relay from Hsp70 during tail-anchored membrane protein targeting
- HSP70-binding motifs function as protein quality control degrons
- Identification of a pro-protein synthesis osteosarcoma subtype for predicting prognosis and treatment
- Co-Chaperones in Targeting and Delivery of Misfolded Proteins to the 26S Proteasome
- IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer
- Role of the J-domain in the cooperation of Hsp40 with Hsp70
- BiP mutants that are unable to interact with endoplasmic reticulum DnaJ proteins provide insights into interdomain interactions in BiP
- Active Learning-Guided Hit Optimization for the Leucine-Rich Repeat Kinase 2 WDR Domain Based on In Silico Ligand-Binding Affinities
- ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks
- A New Fragment‐Based Pharmacophore Virtual Screening Workflow Identifies Potent Inhibitors of SARS‐CoV‐2 NSP13 Helicase
- In silico and in vitro drug screening identifies new therapeutic approaches for Ewing sarcoma
- Identification and verification of a polyamine metabolism-related gene signature for predicting prognosis and immune infiltration in osteosarcoma
- Targeting high-risk MYC-overexpressed osteosarcoma with an Aurora kinase inhibitor:--results from a pilot umbrella trial