Drugs that halt cancer cells at a specific division checkpoint—G1, S, or G2/M—often work beautifully in a plastic dish and then disappoint in patients. Two recent findings suggest why, and they point at the same blind spot. First, the machine-learning models we use to translate cell-culture drug responses into clinical predictions treat the cell cycle as a static list of RNA abundances: DeepSADR, the strongest such system, clusters 1,426 genes into 13 pathways (including "cell cycle regulation") and learns statistical associations between drug fragments and expression modules, but never asks whether the physical protein machines that actually execute the arrest—SKP2 gripping p27, CDK1 bound to Cyclin B1—actually form or dissolve under drug pressure (DeepSADR). Second, transcript-only "virtual cell" models have now been caught systematically hallucinating: they inflate predicted perturbation effect sizes by roughly 23-fold and fail to recover causal gene-gene relationships, and post-hoc calibration fixes the magnitudes while leaving the causal wiring wrong (Virtual Cell Models Inflate Perturbation Effect Sizes).
The core problem is that arrest is executed by physical protein complexes, and RNA abundance does not tell you whether those complexes assemble. In the curcumol study that anchors this proposal, p27 protein was stabilized not because its mRNA changed but because curcumol physically pried SKP2 away from p27, blocking ubiquitination—something only co-immunoprecipitation (Co-IP, an assay where an antibody against a "bait" protein fishes out whatever proteins are physically bound to it) could show (Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells). Worse, the physical wiring itself is context-dependent: the CKAP2–HIF-1α complex forms under hypoxia in cultured endothelial cells but its temporal behavior shifts between early and late stages of the living mouse retina (The association between CKAP2 and HIF-1α in retinal neovascularization). So the wiring that a drug must break in a patient may differ from the wiring in the dish that informed the model.
Core Claim
I propose that co-IP-verified physical checkpoint complexes, measured across microenvironments, can serve as inductive anchors that simultaneously fix two documented failures: the causal hallucination of transcript-only virtual cell models, and the in vitro-to-in vivo collapse of drug response prediction for cell-cycle-arresting therapeutics. Concretely, we will build a transfer-learning model in the DeepSADR lineage whose middle layer is not a static gene list but a set of measured, context-dependent physical edges—SKP2–p27, CDK1–Cyclin B1, CDC25B, USP7–PLK1—supplied by our own Co-IP mass-spectrometry (Co-IP-MS, where the pulled-down complex is digested and every bound protein is identified by mass spectrometry) experiments, and whose output includes the intermediate phenotype that current frameworks ignore: which arrest gate (G0/G1, S, G2/M) a cell population lands in, not just a viability number or a RECIST category.
This is deliberately different from the two closest works. The curcumol study already combines RNA-seq, Co-IP, and a xenograft, but it is a single-cell-line mechanism paper with no model, no transfer learning, and no cross-context generalization claim (Curcumol study). DeepSADR already does pathway-level transfer learning but its "cell cycle" module is an unordered expression set with no physical interaction data anywhere in the pipeline (DeepSADR). Nobody, to my knowledge from this literature, has injected Co-IP-verified physical complexes as constraints into in vitro-to-in vivo transfer models, or tested whether those anchors correct the 23-fold causal distortion.
Research Program
The program has three components, one wet-lab and two computational, designed so that each constrains the others.
First, we generate a matched physical-plus-transcriptomic perturbation atlas. Across a panel of checkpoint drugs spanning both gates—CDK4/6 inhibition in the palbociclib/dalpiciclib lineage for G1 (Palbociclib decade review; Dalpiciclib phase 1), curcumol as the SKP2–p27 disruptor (Curcumol study), and fenofibrate, the CDC25B degrader WG-391D, and USP7 disruption of PLK1 for G2/M (Fenofibrate radiosensitivity; WG-391D in ovarian cancer; USP7 sustains PLK1)—we profile, in several cancer cell lines, three matched readouts under each drug: RNA-seq, Co-IP-MS with SKP2, CDK1, PLK1, and CDC25B as baits, and flow-cytometry cell-cycle gate distributions. Critically, we repeat the triplet across microenvironments: standard 2D culture, 3D matrix culture, and hypoxia, plus xenograft tissue where feasible. The CKAP2–HIF-1α result motivates this: physical complexes themselves rewire with oxygen and tissue context, so a static PPI database would bake in exactly the artifact we are trying to remove (CKAP2–HIF-1α study). Where Co-IP leaves scaffolding-versus-direct-binding ambiguous, we cross-check with proximity ligation assay, which lights up two proteins only when they sit within nanometers of each other in intact cells, following the precedent of the IQGAP3–GLI1 work (Ian T. Foster et al.).
Second, we build the model. Starting from DeepSADR's architecture—drug fragments from BRICS decomposition on one side, pathway modules on the other, a Set Transformer readout learning permutation-invariant interaction-set representations—we replace the static cell-cycle expression module with a physical-anchor layer: the Co-IP-MS edge weights measured in each context, so the model's representation of "the G1 checkpoint under hypoxia plus curcumol" is literally the measured occupancy of the SKP2–p27 complex in that condition. The readout becomes multi-task: predict the cell-cycle gate distribution and the physical-complex dissolution profile as intermediate outputs, then the drug response. The hypothesis is that forcing the model to explain the measured physical edges prevents it from inventing statistical dependencies to fit viability labels.
Third, we run the two decisive test beds: effect-size fidelity and causal recovery on the Perturb-seq benchmarks where virtual cell models were shown to fail, and clinical transfer on the DeepSADR/GANDALF evaluation stack.
Proposed Experiments
The effect-size test asks whether physical anchors tame the 23-fold inflation. Using the Replogle K562 and RPE1 CRISPRi datasets and the Norman 2019 CRISPRa dataset—the exact benchmarks used by the PerturbCausal evaluation—we train our architecture with and without the physical-anchor layer and measure predicted-versus-actual perturbation effect-size distributions and causal recovery F1, comparing against the sparse linear ElasticNet comparator that no deep virtual cell model managed to beat (Virtual Cell Models Inflate). Prediction: the anchored model closes most of the inflation gap; the unanchored ablation will not, and quantile-matching calibration on the anchored model will not be needed for magnitude—distinguishing our mechanism from the calibration-only patch that the PerturbCausal authors showed fails to restore causal structure. For causal-network scale, we additionally evaluate on the CausalBench suite of over 200,000 interventional single-cell data points, scored by mean Wasserstein distance and false omission rate, where supervised winners like Guanlab's LightGBM approach already scale well with interventional data (CausalBench challenge); our claim there is narrower—that physical anchors improve causal edge recall on cell-cycle genes specifically.
The transfer test asks whether physical anchors improve clinical prediction for checkpoint drugs. We run on the DeepSADR evaluation stack: pretraining on the 966 DepMap cell lines with 225,781 labeled response samples, fine-tuning on the 555 TCGA patient samples across the five clinical drugs, and reporting per-drug AUC. The comparison bar is steep: DeepSADR reports AUC 0.957 on Sorafenib versus 0.811 for the next best baseline, and 0.856 average AUC that collapses to 0.662 when its adaptive readout is ablated (DeepSADR). We also evaluate on the retrieval benchmarks from Cell-Level Virtual Screening, where gene-network representations already beat raw expression: CellVS-Net achieves Hits@1 of 0.1250 on DDR-Bench against 0.0357 for molecular fingerprints and 0.0179 for random (Cell-Level Virtual Screening); our question is whether physical edges beat statistical network edges on cell-cycle drug retrieval, using the same DTR-Bench logic of matching a drug to the knockdown of its target.
The prospective wet-lab check is the one that no computational baseline can co-opt: before running the mouse, the model predicts which checkpoint complexes will persist under tumor-like hypoxia and which drug–context combinations will fail to arrest. We then test those predictions by Co-IP on xenograft tissue, following the curcumol study's demonstration that Co-IP-level mechanisms can be interrogated in vivo and that arrest can be rescued by overexpressing the target (Curcumol study). Model-predicted complex dissolution rates that match measured ones, and predicted arrest failures that actually fail, are the real proof.
Baselines
The do-nothing reference point on the retrieval side is random retrieval, which scores Hits@1 of 0.0179 on DDR-Bench as reported by the Cell-Level Virtual Screening paper; anything near it means the representation carries no signal. On the clinical side the naive reference is direct global alignment of cell-line and patient transcriptomes, which the DeepSADR authors report fails because the biological mechanism itself shifts between culture and tumor.
The strongest existing competitor is DeepSADR, because it is the current best system on exactly our evaluation: 966 cell lines and 225,781 pretraining samples, fine-tuned on the same 555 TCGA patients and five drugs, reporting per-drug AUC up to 0.957 on Sorafenib and 0.856 on average (DeepSADR). We will run it unmodified on identical data splits, identical drug sets, and an identical tuning budget; our model gets the same transcriptomic inputs plus only the Co-IP-MS edge weights. Beating it means average AUC above 0.856 with no per-drug degradation, and the more informative test is on the checkpoint-mechanism drugs among the five where physical anchors should matter most. GANDALF, which wins best AUROC on four of five drugs and beats DruID and PANCDR by up to 10.96% (GANDALF), and CellVS-Net, which cut gene-network prediction error by 12–47% on LINCS L1000 perturbations (Cell-Level Virtual Screening), round out the deep-learning competition; CellVS-Net is the key rival for the claim that networks beat snapshots, since it uses statistical co-variation networks where we use measured physical ones.
The most dangerous simple baseline is ElasticNet, the sparse linear comparator that, on the PerturbCausal benchmark, no deep virtual cell model significantly outperformed on causal recovery F1 (Virtual Cell Models Inflate). We run it two ways: on raw expression, replicating the published comparison, and—crucially—on our Co-IP edge features, to test whether the physical data alone carries the gain regardless of model class. The key ablation isolates our ingredient: the identical architecture with physical edges replaced by statistical co-expression edges of matched density, trained on matched data. If the ablation performs as well as the anchored model, the win belongs to network structure in general, not to physical measurement, and we will report it that way.
Why not just…?
Why not just do RNA-seq and pathway enrichment, as thousands of studies already do? The file gives three concrete reasons this falls short for this question. The p27 mechanism is invisible to transcriptomics: p27 was stabilized through physical SKP2 binding and ubiquitination, demonstrable only by Co-IP and protein chase, while its mRNA said nothing (Curcumol study). Transcript-only virtual cell models inflate perturbation effect sizes by roughly 23-fold and cannot recover causal wiring even after magnitude calibration (Virtual Cell Models Inflate). And the physical wiring changes with context in ways expression does not track, as the CKAP2–HIF-1α complex showed across in vitro hypoxia and in vivo retinal stages (CKAP2–HIF-1α study). RNA-seq remains one of our two main readouts; the proposal is about what it cannot supply.
Why not just use molecular fingerprints or docking, since the drugs are small molecules? The Cell-Level Virtual Screening paper already quantified this: molecular fingerprints scored Hits@1 of 0.0357 on DDR-Bench, barely above the random baseline of 0.0179, while cell-level network representations reached 0.1250—molecular structure alone misses the cellular network adaptation that determines whether a checkpoint drug works (Cell-Level Virtual Screening). Our question is specifically about the translation gap, which is a system-level phenomenon, so molecule-level models are answering a different question.
Why not just use ElasticNet on everything, since it beat the deep virtual cell models? This is the real threat, and I will not argue around it. It is true that on the Perturb-seq benchmarks no deep model significantly beat ElasticNet on causal recovery F1 (Virtual Cell Models Inflate). But two things limit it as a full answer here: ElasticNet cannot perform the cross-domain transfer from cell lines to patients that is the point of the proposal, and the CausalBench results show that supervised non-linear and distributional methods scale better as interventional data grows (CausalBench challenge). Still, we take the honest path: ElasticNet run on our Co-IP features is an explicit baseline, and if a sparse linear model on physical edges matches our deep anchored model, the finding stands—physical measurement, not architecture, is the contribution. That would still be a result nobody has reported.
Why not just do standard domain adaptation between cell lines and patients, without the pathway or interaction decomposition? DeepSADR's own ablations answer this: replacing its adaptive set-based readout with ordinary pooling collapsed average AUC from 0.856 to 0.662, and simple global expression alignment failed outright due to mechanism shift between culture and tumor (DeepSADR). GANDALF likewise degraded when cross-attention and multi-task learning were removed (GANDALF). The decomposition is doing measurable work; our change is what goes inside the decomposition.
Decisive Checks
Three results would settle the proposal, and each has a failure mode we pre-commit to. If the physical-anchor model reduces the 23-fold effect-size inflation toward measured distributions while the co-expression ablation does not, the anchor hypothesis wins; if both reduce it equally, physical measurement is unnecessary and statistical network structure suffices—a publishable negative. If the anchored model beats DeepSADR's 0.856 average AUC and its 0.957 Sorafenib ceiling on the checkpoint drugs specifically, physical anchors improve clinical translation; if gains appear only on non-checkpoint drugs, our mechanism story is wrong. And if prospective Co-IP on xenograft tissue matches model-predicted complex dissolution and correctly forecasts which drug–context combinations fail to arrest, the model has captured physical wiring rather than fitting labels. Conversely, if the anchors help prediction but the wet-lab forecasts miss, the model is exploiting statistical shortcut features of the Co-IP data, and we would report that as the leading caveat. No reported result found on Lacuna for any existing model incorporating Co-IP-verified physical interactions into in vitro-to-in vivo transfer learning, which is precisely the space this proposal occupies.
References
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- Curcumol Induces G1 Phase Arrest in SK-Hep-1 Cells by Targeting SKP2-Mediated p27 Degradation
- The association between CKAP2 and HIF-1<b><i>α</i></b> in retinal neovascularization under hypoxic conditions
- 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
- A phase 1 study of dalpiciclib, a cyclin-dependent kinase 4/6 inhibitor in Chinese patients with advanced breast cancer
- Fenofibrate Increases Radiosensitivity in Head and Neck Squamous Cell Carcinoma via Inducing G2/M Arrest and Apoptosis
- Inhibition of CDC25B With WG-391D Impedes the Tumorigenesis of Ovarian Cancer
- USP7 is a novel Deubiquitinase sustaining PLK1 protein stability and regulating chromosome alignment in mitosis
- HPF-MPI: A Programming System Supporting Task and Data Parallelism
- The CausalBench challenge: A machine learning contest for gene network inference from single-cell perturbation data
- Cell-Level Virtual Screening
- GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment