Every year, hundreds of drug candidates kill cancer cells reliably in a plastic dish and then fail in a living tumor. Two failure modes dominate. First, delivery: a compound that is potent in vitro never reaches its target in vivo because it is cleared, metabolized, or physically blocked by tumor tissue. Second, cell-state diversity: a drug that halts the cell division cycle in most cells leaves a fraction unarrested, and that fraction regrows the tumor. Carrier-free drug delivery — building nanoparticles out of the drugs themselves, with no inert shell, by chemically tethering a water-hating drug to a water-loving one so they self-assemble — attacks the first failure mode, because such assemblies carry almost no dead weight and can hold two drugs at a fixed ratio. But nothing in the current toolkit tells you, before you synthesize anything, whether a given carrier-free assembly will keep its two drugs together in the bloodstream of a given patient's tumor, or whether the particular molecular background of that tumor will let cells slip around the cell-cycle checkpoint the drugs are blocking.
This proposal builds a predictive model for exactly that question. It is computational, grounded in an existing in vivo benchmark, and paired with one prospective animal experiment to test the part no existing dataset covers. The core idea: replace the generic "tumor grows, drug shrinks it" equation used by current in vivo response models with a structured model that tracks cells moving between proliferating, arrested, and dying compartments, where the arrest and escape rates depend on the drug formulation's physical integrity and on the tumor's molecular subtype. Escape from arrest then emerges from cell-to-cell variability in those rates — modeling the non-genetic "fractional resistance" recently demonstrated for the CDK4/6 inhibitor palbociclib — rather than being an unexplained residual.
Core Claim
The claim is that in vivo tumor response to cell-cycle-targeting drugs can be predicted more accurately, and — more importantly — predicted earlier and per-subtype, if the dynamics model has three things current models lack: a cell-cycle compartment structure with heterogeneous per-cell transition rates, a drug-concentration term that reflects whether a carrier-free assembly stays intact or collapses when diluted in plasma, and conditioning of all rate parameters on the tumor's molecular subtype. The scaffold is the Graph Neural Network plus Neural Ordinary Differential Equation architecture of Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction, which predicts continuous-time tumor volume trajectories in patient-derived xenografts (PDXs — human tumors grown in mice) and reported fitting accuracy of $R^2 = 0.96$ versus $R^2 = 0.71$ for the classical empirical tumor-growth-inhibition model. But that model treats every treatment as a naked small molecule, uses a black-box growth function with no notion of the cell cycle, and averages over tumor heterogeneity. We keep its graph-based molecular encoder and change the dynamics underneath it.
Concretely, the tumor cell population $N(t)$ is split into proliferating cells $N_P$, arrested cells $N_A$, and dying cells $N_D$, with $N = N_P + N_A + N_D$:
$$\frac{d}{dt}\begin{pmatrix}N_P\\N_A\\N_D\end{pmatrix}=\begin{pmatrix}r-\kappa_{\mathrm{arr}}(c,\theta)&\kappa_{\mathrm{esc}}(\theta)&0\\\kappa_{\mathrm{arr}}(c,\theta)&-\kappa_{\mathrm{esc}}(\theta)-\delta&0\\0&\delta&-\lambda\end{pmatrix}\begin{pmatrix}N_P\\N_A\\N_D\end{pmatrix}$$
Here $r$ is the untreated proliferation rate, $\kappa_{\mathrm{arr}}$ the rate at which drug pushes cells into arrest, $\kappa_{\mathrm{esc}}$ the rate at which arrested cells re-enter the cycle, $\delta$ the death rate of arrested cells, and $\lambda$ the clearance rate of dead cells. The per-cell state $\theta$ — which sets how strongly a given cell responds to the drug — is drawn from a distribution, not a single value. Fractional resistance is then not an error term but a model output: the fraction of $\theta$ values for which $\kappa_{\mathrm{esc}}$ stays high under drug. This is a direct computational counterpart to the finding in Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells, which used single-cell imaging of 14 cell-cycle regulators to show that palbociclib-resistant cells bypass the G1 checkpoint without any mutation, via premature CDK2/E2F1 activity and cyclin B1 appearing in the wrong phase. That paper leaves open whether any delivery or combination strategy can close those escape routes in vivo; we build the model that answers that question before the chemistry is done.
The formulation term enters through $c(t)$, the intratumoral concentration time course. For a free two-drug cocktail, each drug has its own clearance curve, so the ratio at the tumor drifts arbitrarily — the core reason unformulated combinations underperform, as argued in Predicting In Vivo Efficacy. For a carrier-free co-assembly, two regimes exist: while plasma concentration stays above the assembly's critical aggregation concentration (CAC — the dilution threshold below which the nanostructure falls apart, as characterized for the amphiphilic irinotecan–melampomagnolide B conjugate in Amphiphilic irinotecan–melampomagnolide B conjugate nanoparticles for cancer chemotherapy), both drugs are delivered at their built-in fixed stoichiometry, the same principle exploited by the cisplatin–olaparib supramolecular nanostructure of Enhanced cisplatin chemotherapy sensitivity by self-assembled nanoparticles with Olaparib and the drug-inhibitor conjugates of Stimuli-responsive nanodrug self-assembled from amphiphilic drug-inhibitor conjugate for overcoming multidrug resistance in cancer treatment. Once dilution pushes below CAC, the assembly dumps its cargo and each drug clears independently. The model takes measured CAC and release kinetics as inputs and computes which regime a dosing schedule actually operates in — turning a physicochemical property into a predicted efficacy difference.
Research Program
Three components are built. The first is the encoder: a graph neural network that consumes the tumor's molecular profile and the treatment's drug-target graph, in the style of the bipartite graph convolutions over TissueNexus, DGIdb, and DisGeNET used by Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction, but with two added input channels: assembly descriptors (CAC, hydrophobic-hydrophilic balance, linkage lability, measured release half-life) and stoichiometric ratio for co-assemblies. The encoder outputs the ODE rate parameters $r$, the $\theta$-distribution, $\delta$, and $\lambda$ per tumor-treatment pair, and the ODE is integrated continuously — the Neural-ODE trick — so irregularly sampled tumor volumes are handled natively. We deliberately avoid raw-SMILES-only encoders: the symbol-shuffling experiments in Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network showed that a high-scoring convolutional network ($R^2 = 0.826$, Pearson $R_p = 0.909$ on GDSC v6.0) was riding statistical shortcuts and collapsed on truly novel drugs, which is exactly the regime a formulation screener lives in.
The second component is subtype conditioning. Tumors in the PDX cohort will be clustered into molecular subtypes using the network-assisted and rank-based stratification methods validated on TCGA breast cancer — A network-assisted co-clustering algorithm to discover cancer subtypes based on gene expression (which achieved survival separation at $p = 0.044$ on 547 TCGA breast cancer samples where consensus clustering failed at $p = 0.831$) and Simultaneous discovery of cancer subtypes and subtype features by molecular data integration (which split 719 TCGA breast cancer patients into 13 subtypes, resolving "Basal-like" into distinct prognostic groups). The hypothesis to test is that the fitted escape-rate distributions cluster by subtype — that molecular heterogeneity, not just average sensitivity, determines fractional resistance. This directly addresses the tension in the literature between static genomic subtyping, which attributes response variation to fixed mutations, and the single-cell proteomic view of Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells, which attributes it to protein-state variability: our model holds both, since the $\theta$-distribution's shape is conditioned on the genomic subtype while remaining continuous within it.
The third component is prospective and small. We will synthesize one carrier-free co-assembly pairing a G1-arresting CDK4/6 inhibitor with an agent hitting the documented escape route (CDK2/cyclin B1 axis), using the amphiphilic conjugate chemistry of Amphiphilic irinotecan–melampomagnolide B conjugate nanoparticles for cancer chemotherapy or the coordination-driven co-assembly of Enhanced cisplatin chemotherapy sensitivity by self-assembled nanoparticles with Olaparib, measure its CAC and release kinetics in plasma-mimetic buffer, and run it in a PDX mouse arm against the free two-drug cocktail. The model makes the call beforehand: it predicts the full tumor-volume trajectory for both arms in two chosen subtypes — one where it predicts the co-assembly wins (divergent free-drug clearance dominates) and one where it predicts no difference. This design follows the precedent of A Novel Squalenoylated Temozolomide Nanoparticle with Long Circulating Properties Reverses Drug Resistance in Glioblastoma, which showed a carrier-free assembly can be evaluated for in vivo circulation and tumor penetration, but adds the cell-cycle and subtype readout that paper lacks.
Proposed Experiments
The retrospective backbone is the Gao et al. PDX collection as used by Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction: 3,470 PDX experiments covering 191 unique human tumors and 59 treatments, with longitudinal tumor volumes. We evaluate longitudinal forecasting exactly as that paper did — conditioning on the first 7, 14, 21, or 28 days of data and predicting the remaining trajectory — under leave-tumor-out splits, so the model must generalize to unseen molecular backgrounds. Primary metrics are forecasting $R^2$ at each horizon, responder versus non-responder classification, and predicted time-to-regrowth. We additionally score the model's per-tumor escape fraction (the mass of $\theta$ above the resistance threshold) against observed regrowth, which no existing metric captures. Cell-cycle-targeting treatments within the 59 are identified via drug-target mapping (DGIdb, as in the scaffold paper); if no CDK4/6 inhibitor is present, we say so and rely on the broader antimitotic set. In vitro pretraining uses GDSC v6.0 (the same resource used by Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network and DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer), with cell-line-derived parameters transferred to PDX tumors via the shared-subspace alignment of Predicting clinical drug response from model systems by non-linear subspace-based transfer learning, which we reuse as a component rather than compete with. An honest caveat, stated up front: the PDX benchmark contains only nominal small molecules, so the formulation layer cannot be validated retrospectively. It is validated by the prospective mouse experiment, and everything formulation-blind is validated on PDX.
Baselines
The "do nothing" reference point is in vitro potency ranking: use GDSC IC50 values from the nearest matched cell lines to rank the 59 treatments and ask whether that ranking predicts PDX response order. Predicting In Vivo Efficacy documents why this fails — static cytotoxicity assays cannot forecast systemic distribution, half-life, or dose-limiting toxicity — and no reported result found on Lacuna for any IC50-to-PDX transfer accuracy, so we will measure it ourselves as the floor. The strongest existing competitor is the Graph Neural Network + Neural-ODE model of Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction, which reported $R^2 = 0.96$ for trajectory fitting versus $R^2 = 0.71$ for the classical empirical tumor-growth-inhibition model on the same PDX data. That $R^2 = 0.96$ is fitting accuracy on observed trajectories; the meaningful contest is forecasting, so we will re-run their architecture ourselves on identical leave-tumor-out splits and horizons, with the same data, the same graph inputs, and the same hyperparameter search budget as ours — the only difference being the ODE right-hand side and the added input channels. Beating it means higher forecasting $R^2$ at the 14- and 21-day horizons on unseen tumors, better responder/non-responder classification, and materially better calibrated time-to-regrew predictions; if our structured model merely matches its accuracy while also outputting an interpretable escape fraction and accepting formulation inputs it cannot, that is a partial win we will report as such. The third baseline is the empirical tumor-growth-inhibition model, the field standard, run with the same data. The key ablation removes the cell-cycle compartment structure — collapsing $N_P$, $N_A$, $N_D$ into one homogeneous population with a single drug effect — while keeping the encoders and training budget identical; this isolates whether the heterogeneous compartment structure, the proposal's central ingredient, earns its place, since any gain must come from that structure and not from the extra inputs or tuning. A second ablation zeroes the formulation channel to confirm its contribution is confined to the prospective setting.
Why not just…?
Why not just use clinically validated carrier-based nanoparticles? Carriers are proven, and for many drugs they are the right answer. But the inert carrier accounts for 80–95% of nanoparticle mass (Carrier Free Self Assembly of Active Pharmaceutical Ingredients), which caps drug loading and forces high excipient dosing with its own toxicity and clearance burden. More to the point, this proposal is not anti-carrier: the formulation channel takes carrier-based descriptors as inputs too, so the model can tell you when a carrier is unnecessary or counterproductive. The translational critique of Bench-to-bedside translation of podophyllotoxin-based nanomedicines for cancer treatment: utopias and reality? applies to carrier and carrier-free formulations alike — which is exactly why a formulation-aware predictive model is needed rather than more formulation-by-intuition.
Why not just give a free cocktail of the two inhibitors? This is the most serious alternative, and if the model predicts the free cocktail performs identically in a given subtype, the honest conclusion is to use the cocktail. The counter-evidence is that free drugs clear at different rates and do not concentrate synchronously at the target (Predicting In Vivo Efficacy), while co-assemblies enforce fixed stoichiometric co-delivery (Amphiphilic irinotecan–melampomagnolide B conjugate nanoparticles for cancer chemotherapy, Enhanced cisplatin chemotherapy sensitivity by self-assembled nanoparticles with Olaparib). The prospective experiment is designed to settle this directly: both arms are run in both selected subtypes, and the model's pre-registered prediction of where the co-assembly does and does not beat the free cocktail is the decisive read. If the free cocktail matches everywhere, the model — and the carrier-free premise in this indication — is wrong, and we will say so.
Why not just use TRANSACT or standard batch correction for the translation problem? Predicting clinical drug response from model systems by non-linear subspace-based transfer learning already showed linear and batch-correction approaches (Elastic Net, ComBat+deep learning, PRECISE) underperform non-linear subspace transfer, and no reported result found on Lacuna for TRANSACT's exact TCGA or HMF error values. But TRANSACT predicts static response from bulk transcriptomics, assumes monotonic response functions between dish and patient, and represents drugs only as naked small molecules. Non-genetic cell-cycle plasticity violates monotonicity by construction — the same tumor contains cells that respond and cells that bypass — and formulation physics is invisible to it. We reuse its subspace alignment as a transfer component; we do not compete with it on its own static task.
Why not just scale up the black-box model on more data? The symbol-shuffle experiment in Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network is the cautionary tale: a large convolutional network hitting $R^2 = 0.826$ on GDSC relied on statistical regularities, not biochemistry, and degraded sharply on unseen drugs. A formulation screener must generalize to chemical entities and assemblies never seen in training — the blind-drug regime — where the compounding evidence from that paper and from the knowledge-guided designs of DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer says pure capacity does not carry you. Structure is the bet; the compartment ablation is the test of that bet.
Decisive Checks
Three outcomes would decide the proposal. On the PDX benchmark, the structured model must beat the re-run GNN-Neural-ODE baseline on leave-tumor-out forecasting at the 14- and 21-day horizons, not merely on fitting; a tie there with a win only on interpretability is a weaker result. The escape fraction fitted per tumor must correlate with observed regrowth and cluster by molecular subtype, which would be the first quantitative link between the non-genetic plasticity mechanism of Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells and in vivo dynamics. And the prospective mouse experiment must confirm at least one of the two pre-registered predictions about where the carrier-free co-assembly does and does not outperform the free cocktail — the only place the formulation layer meets ground truth, and the reason this proposal is one experiment, not one more model.
References
- Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction
- Cell cycle plasticity underlies fractional resistance to palbociclib in ER+/HER2- breast tumor cells
- Amphiphilic irinotecan–melampomagnolide B conjugate nanoparticles for cancer chemotherapy
- Enhanced cisplatin chemotherapy sensitivity by self-assembled nanoparticles with Olaparib
- Stimuli-responsive nanodrug self-assembled from amphiphilic drug-inhibitor conjugate for overcoming multidrug resistance in cancer treatment
- Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network
- A network-assisted co-clustering algorithm to discover cancer subtypes based on gene expression
- Simultaneous discovery of cancer subtypes and subtype features by molecular data integration
- A Novel Squalenoylated Temozolomide Nanoparticle with Long Circulating Properties Reverses Drug Resistance in Glioblastoma
- DRPreter: Interpretable Anticancer Drug Response Prediction Using Knowledge-Guided Graph Neural Networks and Transformer
- Predicting clinical drug response from model systems by non-linear subspace-based transfer learning
- Bench-to-bedside translation of podophyllotoxin-based nanomedicines for cancer treatment: utopias and reality?