Arrested but Exposed: Mapping How CDK4/6 Inhibition Remodels the GPX4 Interactome

When a cancer drug halts the cell cycle, the stalled cells do not simply freeze. They stay metabolically active, remodel their protein complexes, and become vulnerable to some kinds of death while resisting others. This proposal asks a question that sits at a genuine gap: when CDK4/6 inhibitors — a clinically approved class of breast cancer drugs — trap cells in the first growth phase of the cell cycle, what happens to the physical protein complexes that protect cell membranes from iron-driven oxidative destruction? Ferroptosis is that form of death: iron catalyzes lipid peroxidation — essentially rusting of fat-rich membranes — unless the enzyme GPX4 and its supporting proteins neutralize the peroxidized lipids. If arrest weakens GPX4's bodyguards, then a safe, oral, approved pill could prime tumors for ferroptosis, sidestepping the dire pharmacokinetics of direct GPX4 inhibitors. If arrest instead strengthens them, we need to know that before combining these drugs in patients.

The question is unanswerable with gene expression alone, because the interactions that set GPX4 stability are gated by chemical modifications on proteins — ubiquitin tags and fatty-acid attachments — that do not track mRNA levels. We propose to answer it with co-immunoprecipitation (Co-IP), an assay that uses an antibody against one chosen "bait" protein to physically pull it, and anything bound to it, out of a cell lysate, followed by mass spectrometry to identify the bound partners. The answer will then be tested for generalizability: first across a panel of cell lines with different genetic backgrounds, then in mouse tumors, and finally computationally, by asking whether models that predict how a cell-line finding transfers to patient tumors improve when fed context-specific physical interactions instead of static database networks.

Core Claim

Pharmacological CDK4/6 arrest is not death-neutral. We hypothesize that palbociclib-induced arrest remodels the physical complexes governing GPX4 stability — specifically the USP5–GPX4 deubiquitination complex and the p53–USP11/MDM2 switch described below — and that the direction of this remodeling (loss versus reinforcement of GPX4 protection) determines whether arrested cells become ferroptosis-primed. Further, we claim this remodeling is heterogeneous: the minority of cells that escape arrest through non-genetic plasticity may retain GPX4 protection, and this subpopulation — not the bulk — dictates whether CDK4/6 plus ferroptosis-inducing combinations work. Only context-specific Co-IP, run as a time course during arrest, can resolve this, and only multi-model testing can tell us whether the answer holds beyond a single cell line.

Three findings motivate this claim without settling it. A proteomic study found that seven days of palbociclib in triple-negative breast cancer cells downregulates the DNA-synthesis enzymes RRM1 and RRM2 while leaving central carbon metabolism running, and explicitly did not measure any ferroptosis markers or validate protein complexes (Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition). Separately, genetic knockdown of RRM1 — not pharmacological arrest — was shown to destabilize p53, lower p21 and GPX4, and roughly double lipid peroxidation when combined with radiation, about a ~2-fold increase in lipid peroxidation compared to radiation alone in HCT116 cells (Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis). And a breast cancer study used Co-IP followed by mass spectrometry to discover that the enzyme Fatty Acid Synthase promotes an attachment of a fatty acid chain to the protein USP5 (S-palmitoylation), which is required for USP5 to bind GPX4 and protect it from destruction; when FASN was removed, the interaction vanished (FASN inhibits ferroptosis in breast cancer via USP5 palmitoylation-dependent regulation of GPX4 deubiquitination).

None of these three papers did what we propose. The palbociclib proteomics paper stopped at the inventory, never measuring lipid damage or pulling down complexes. The RRM1 paper used genetic silencing in two cell lines (HCT116, HeLa) and never tested a pharmacological cell cycle inhibitor. The FASN paper never varied cell cycle phase at all. Our proposal connects the causal chain end to end: approved drug → arrest state → physical GPX4-complex remodeling → ferroptosis sensitivity → generalizability across models.

Research Program

The program has three linked stages.

Stage one is a time-course interactome map during arrest. We will treat a panel of cell lines spanning genetic backgrounds and ferroptosis baselines — T47D (ER+ breast, where fractional palbociclib resistance was characterized), MDA-MB-231 (triple-negative breast, the line used in the palbociclib proteomics study), HCT116 (colon, the RRM1 study line), and a clear-cell carcinoma line, since clear-cell histology confers intrinsic GPX4-inhibitor sensitivity through PUFA-enriched lipids (A GPX4-dependent cancer cell state underlies the clear-cell morphology and confers sensitivity to ferroptosis) — with palbociclib at concentrations matching the 1 µM, 7-day protocol of the proteomics study. At 1, 3, and 7 days we will run endogenous-bait Co-IP for GPX4, USP5, p53, USP11, and MDM2, with mass spectrometry readout for GPX4 pulldowns to catch unanticipated partners, mirroring the discovery strategy that found USP5 in the FASN study. In parallel we will quantify GPX4 ubiquitination (ubiquitin pulldown of GPX4), USP5 palmitoylation by acyl-biotin exchange as in the FASN paper, and lipid peroxidation with BODIPY-C11 and MDA assays, plus the ferroptosis-specific rescue controls (lipid-radical-trapping antioxidant and iron chelation) that distinguish ferroptosis from other death modes.

Stage two interrogates the resistant minority. The fractional-resistance study showed that a subpopulation of T47D cells escapes palbociclib arrest through non-genetic plasticity, with premature accumulation of cyclin B1 in G0/G1 and elevated CDK2 and E2F1, and its authors explicitly hypothesized noncanonical cyclin B1 wiring but never pulled down the complexes to find cyclin B1's partners in the arrested state (Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells). We will sort arrested versus escaping cells by live-cell cycle and cyclin B1 reporters and run the same Co-IP panel on each fraction, asking whether escapers preserve USP5–GPX4 binding that the bulk loses — and whether they are correspondingly ferroptosis-resistant. This directly tests whether combination therapy fails not because arrest does not prime ferroptosis, but because the primed cells die and the unprimed ones escape.

Stage three tests generalizability in two layers. First, in vivo: the FASN study established that orlistat (a FASN inhibitor) plus anti-PD-1 immunotherapy suppresses 4T1 mammary tumors in BALB/c mice with increased CD8+ T cell infiltration (FASN inhibits ferroptosis in breast cancer via USP5 palmitoylation-dependent regulation of GPX4 deubiquitination); we will extend this backbone with palbociclib arms and measure whether in-vitro synergy (palbociclib plus a GPX4-pathway perturbation) survives the microenvironment. Second, computationally: we will ask whether Co-IP-derived, condition-specific interaction edges improve transfer of drug-response predictions from cell lines to patient tumors, on benchmarks described below.

Baselines

The naive reference point is palbociclib monotherapy: arrest alone, with the working expectation drawn from the proteomics study that cells remain metabolically active and no ferroptosis was measured (Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition). Any claim that arrest primes ferroptosis must show lipid-peroxidation and viability effects clearly above this monotherapy arm at matched drug exposure.

The strongest experimental competitor is direct GPX4 inhibition with RSL3 or ML210, the small molecules shown to be selectively lethal in clear-cell carcinoma lines (A GPX4-dependent cancer cell state underlies the clear-cell morphology and confers sensitivity to ferroptosis). Our combination arms (palbociclib plus RSL3, with Bliss and Loewe synergy scoring) must beat RSL3 alone at matched RSL3 dose to justify adding the cell cycle drug; if the combination is merely additive or antagonistic, the priming claim fails. We will run both arms in the same cell lines, same passage windows, same assay batch, with equal replicate counts.

The strongest computational competitor for the transfer question is DeepSADR, a deep transfer learning system that decomposes drugs into chemical subsequences, maps them against 13 functional pathways, and uses a Set-Transformer readout to transfer response predictions from 966 DepMap cell lines (225,781 samples) to 555 TCGA patient samples across 5 drugs (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction); on its benchmark it reports a clinical AUC of 0.957 for sorafenib versus 0.811 for the next best baseline, and ablating its Adaptive Readout module dropped average AUC from 0.856 to 0.662, which we take as the scale of what module-level design choices are worth. We will retrain DeepSADR's architecture on our data with identical input dimensions, tuning budget, and cross-validation folds, changing only the interaction-edge source, so that any delta is attributable to the edges and nothing else.

For in-vitro combination prediction specifically, the benchmark competitor is the Extreme Randomized Trees approach evaluated on O'Neil's high-throughput combination screen (583 drug combinations across 39 cancer cell lines), which reports a Pearson correlation of 0.738 between predicted and observed synergy and an F1 of 0.954 for the synergism-versus-antagonism classification (In silico drug combination discovery for personalized cancer therapy); beating it means achieving higher Pearson correlation on the same cell-line subset under the same folds.

The key ablation isolates our ingredient: context-specific edges. We will compare three edge sets inside the same architecture — static STRING protein-interaction edges in the style of AGMI, which achieved RMSE 0.7943 and $R^2$ of 0.9184 on 80,758 cell-line/drug pairs from GDSC but used static database interactions and never transferred to patients (AGMI: Attention-Guided Multi-omics Integration for Drug Response Prediction with Graph Neural Networks); palbociclib-condition Co-IP edges from stage one; and a shuffled-edge control. If Co-IP edges do not beat STRING edges on transfer performance, the computational layer of this proposal is falsified, though the biological stages would stand on their own.

Why not just…?

Why not just use static protein-interaction databases instead of running Co-IP? Because the decisive interactions in this system are condition-dependent in ways databases cannot represent. In the FASN study, USP5 bound GPX4 only when USP5 was palmitoylated, and FASN knockdown abolished the interaction (FASN inhibits ferroptosis in breast cancer via USP5 palmitoylation-dependent regulation of GPX4 deubiquitination); in the RRM1 study, depleting RRM1 switched p53's binding from the deubiquitinase USP11 to the ubiquitin ligase MDM2 — a competitive shift visible only by physical pulldown (Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis). Our STRING-edge ablation makes this testable rather than rhetorical: if static edges predict transfer as well as Co-IP edges, the reviewer wins.

Why not just give direct GPX4 inhibitors and skip the cell cycle entirely? This is the strongest objection, and we answer with the cited limitation: direct GPX4 inhibitors RSL3, ML210, and ML162 are potent in vitro but suffer poor pharmacokinetics and bioavailability in vivo (A GPX4-dependent cancer cell state underlies the clear-cell morphology and confers sensitivity to ferroptosis), whereas CDK4/6 inhibitors are orally bioavailable and approved, with palbociclib a decade into clinical use (A Decade After Approval of the First CDK4/6 Inhibitor: A Look Back at Palbociclib's Journey from Discovery to Approval and What's Next in CDK Inhibition in Breast Cancer) and dalpiciclib validated in a phase 1 trial in advanced breast cancer (A phase 1 study of dalpiciclib, a cyclin-dependent kinase 4/6 inhibitor in Chinese patients with advanced breast cancer). The honest caveat is that RSL3 remains the more direct ferroptosis trigger; our experiments will settle the trade-off by measuring whether palbociclib lowers the RSL3 dose needed for equivalent lipid damage (a dose-shift endpoint), which is the clinically meaningful form of "priming."

Why not just assume arrest protects cells from death and drop the combination idea? The evidence points the other way: prolonged palbociclib arrest leaves cells metabolically active with upregulated central carbon metabolism while downregulating RRM1 and RRM2 (Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition), and RRM1 loss degrades GPX4 and doubles lipid peroxidation under DNA damage (Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis). Still, this objection could be right in specific backgrounds — arrest may shield some lines — which is exactly why the multi-line panel with per-line Co-IP readouts is the core design rather than a single-line case study, the limitation the curcumol authors themselves flagged for SK-Hep-1 (Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells by Targeting SKP2-Mediated p27 Degradation).

Why not just use standard transcriptomic transfer learning (ComBat, PRECISE, or plain deep networks)? Linear batch-correction methods were directly outperformed by non-linear transfer in the TRANSACT comparison across GDSC, PDX, TCGA, and HMF cohorts (Predicting clinical drug response from model systems by non-linear subspace-based transfer learning); no reported result found on Lacuna for a numerical TRANSACT summary statistic, so we treat its superiority as qualitative there. More decisively, DeepSADR's own ablation showed that global expression alignment without sub-pathway-aware readout collapsed average AUC from 0.856 to 0.662 (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). Our claim is one step further — that even sub-pathway models miss PTM-gated binding — and it is falsifiable by the edge ablation above.

Proposed Experiments and Decisive Checks

The first decisive check is the interactome delta. If arrest primes ferroptosis, GPX4 Co-IP pulldowns during the palbociclib time course should show reduced USP5 recovery, reduced USP5 palmitoylation, increased GPX4 ubiquitination, and a p53 binding shift toward MDM2, temporally preceding the rise in lipid peroxidation. Temporal precedence is essential: complex dissolution must come before, not after, membrane damage, or it is a consequence rather than a cause.

The second check is the dose-shift and synergy endpoint. Across the four-line panel, we will quantify whether palbociclib pre-treatment shifts the RSL3 lethality curve, with Bliss and Loewe scores computed on matched-plate data, and ferroptosis specificity confirmed by rescue with radical-trapping antioxidants and iron chelation. For the computational extension we will train on the O'Neil combination benchmark subset with genomic profiles (31 cell lines, 16,575 samples; the ERT baseline reports Pearson 0.738 and F1 0.954 (In silico drug combination discovery for personalized cancer therapy)), then evaluate transfer on the DeepSADR DepMap-to-TCGA setup, reporting AUC against DeepSADR's 0.957 sorafenib figure and against our retrained STRING-edge variant. Clinical plausibility will be checked by asking whether palbociclib-treated breast tumors in the TCGA-BRCA cohort of 1,069 patients from the FASN study show expression signatures matching the sensitized interactome state; no reported result found on Lacuna for a numerical FASN–TCGA correlation, so this remains a hypothesis-generating, not confirmatory, analysis.

The third check is the escaper question: sorted fractionally resistant cells either retain USP5–GPX4 binding and ferroptosis resistance (in which case combination therapy needs a senolytic or escape-targeting partner, consistent with the metabolically active senescence-like state noted after CDK4/6 inhibition (Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition)), or they lose it too (in which case arrest is uniformly sensitizing). Finally, the 4T1/BALB/c mouse experiment on the orlistat-plus-anti-PD-1 backbone established in the FASN study determines whether any in-vitro synergy survives an immune-competent microenvironment — the translation gap that motivates the whole program (FASN inhibits ferroptosis in breast cancer via USP5 palmitoylation-dependent regulation of GPX4 deubiquitination).

The proposal fails cleanly in defined ways: if arrest-time-course Co-IPs show no GPX4-complex change, the priming hypothesis is dead; if complexes change but combination synergy never exceeds RSL3 alone, the pharmacology is not actionable; and if Co-IP edges do not improve transfer over STRING edges, the computational layer adds cost without signal. Each failure mode is informative and none of them is currently resolved in the literature we could find on Lacuna.