Drugs as Control Ports: Port-Hamiltonian Gene-Network Dynamics for Receptor-Negative Cancers

Aggressive tumors that lack the molecular handles of precision oncology — triple-negative breast cancer, driver-gene-negative lung adenocarcinoma, double-negative prostate cancer — are treated with broad, toxic interventions in part because we lack models that say what a given compound will do to a specific tumor's internal signaling network over time. In parallel, a completely separate community studying non-conservative force modeling has built neural dynamical systems that respect the physics of dissipative systems: they split motion into a reversible core, an energy-losing friction term, and an explicit channel for external control inputs, so the model cannot invent energy or explode when simulated past its training data. These two threads have barely touched. The few existing connections either fix the dissipation by hand and ignore gene networks (Contact Wasserstein Geodesics for Non-Conservative Schrödinger Bridges), or model gene networks with unconstrained dynamics and no notion of a drug as a control force (A scalable gene network model of regulatory dynamics in single cells).

This proposal builds a gene-level dynamical model in the port-Hamiltonian form — an architecture borrowed from robot dynamics learning — and represents each drug as a control input entering through a learned port matrix shaped by the drug's known targets. The concrete payoff is a class of receptor-negative-cancer questions that no current model can answer: whether CDK4/6 inhibition pushes cells into a deep or shallow arrest basin (i.e., whether the senescence-like state is permanent or reversible, an explicitly unresolved question in Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition), and how dose and schedule of multi-target combinations reshape the relaxation landscape before any wet-lab experiment is run.

Core Claim

Cell-state dynamics under therapy can be written as a dissipative control system in gene expression space. Let $x \in \mathbb{R}^d$ be the expression of $d$ genes wired by a prior gene regulatory network (GRN). The model learns four components: a scalar potential $H_\theta(x)$ (an energy-like landscape over cell states), a skew-symmetric matrix $J_\theta(x) = -J_\theta(x)^\top$ (reversible routing — signaling exchange and cell-cycle oscillation, which neither creates nor destroys energy), a positive semi-definite matrix $R_\theta(x) = L_\theta(x)L_\theta(x)^\top \succeq 0$ (dissipation — degradation, metabolic friction, irreversible commitment to arrest), and a port matrix $g(x)$ through which a drug dose schedule $u(t)$ enters:

$$\dot{x} = \big(J_\theta(x) - R_\theta(x)\big)\nabla_x H_\theta(x) + g_\phi(x)\,u(t).$$

The structural constraints carry the physics: energy $H$ can never increase from the internal $R$ term, and the drug acts only through the port, so it can push the system but cannot silently rewrite its internal wiring. This is exactly the decomposition proven to stabilize robot dynamics learning in Port-Hamiltonian Neural ODE Networks on Lie Groups For Robot Dynamics Learning and Control and to cut parameter needs roughly threefold in Symplectic ODE-Net: Learning Hamiltonian Dynamics with Control (0.13M versus 0.36M parameters against an unconstrained baseline). The claim to be tested is that this structure, transplanted onto GRN-constrained gene dynamics, predicts held-out perturbation trajectories and — critically — long-horizon and dose-extrapolation behavior that unconstrained gene ODEs get wrong, and that it does so with an interpretable decomposition: senescence-like arrest should appear as growth of $R$ along the trajectory and relaxation into a basin of $H$, whose depth the model estimates.

This is deliberately distinct from the closest existing work. Contact Wasserstein Geodesics for Non-Conservative Schrödinger Bridges shares the motivation that biological systems are non-conservative, but it operates on whole probability distributions, leaves its contact damping factor as a manual hyperparameter, and models no gene interactions or drug mechanism of action; we learn state-dependent dissipation on a gene network and represent drugs explicitly. A scalable gene network model of regulatory dynamics in single cells (FLeCS) shares the gene-space ODE and the GRN prior, and we adopt its experimental setting — but its dynamics are an unconstrained production-minus-decay equation with no potential, no PSD dissipation matrix, no passivity, and perturbations only as modified initial conditions rather than continuous control inputs. Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport (DeepRUOT) motivates us by showing that ignoring non-conservation of cell mass creates "false transitions," but its transport field is an unconstrained neural ODE in PCA space; we borrow its growth-rate idea as an optional extension (a scalar proliferation/death field alongside the transport above), not as the core.

Research Program

The build has three parts. First, the architecture: $J_\theta$ and $R_\theta$ are parameterized as low-rank corrections on the GRN adjacency, so that reversible routing and dissipation are sparse over known regulatory edges rather than dense $d \times d$ matrices; $R_\theta$ is produced by a Cholesky-factor network so PSD holds by construction, following Port-Hamiltonian Neural ODE Networks on Lie Groups. Second, the port: for each compound, $g_\phi(x)$ is a matrix whose columns are supported on the drug's documented target genes (DrugBank-style target lists, as used by Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways), with learned, state-dependent gains — so an unseen drug with known targets is a new $u$ and a new support pattern for $g$, not a new model. Third, the analysis layer: from the trained model we read out (a) the free-energy landscape $H$ restricted to the manifold of reachable states, (b) the magnitude and direction of dissipation $R\nabla H$ along trajectories, and (c) basin depth of the senescent attractor under sustained versus withdrawn drug input, which turns the "is arrest reversible?" question into a simulation: set $u(t) = 0$ after the basin is reached and measure whether the learned dynamics escape.

Training uses snapshot perturbation data — unpaired distributions at control and post-perturbation time points — which is the standard regime because single-cell measurement is destructive; we pair snapshots with the same minibatch-optimal-transport trick used by Departures: Distributional Transport for Single-Cell Perturbation Prediction with Neural Schrödinger Bridges, but the dynamics fitted to those pairs are port-Hamiltonian rather than an unconstrained bridge.

Proposed Experiments

Three experiments, in increasing order of biological specificity. All use the same gene set, GRN prior, compute budget, and tuning protocol (matched hyperparameter-search count) for every model compared.

The first experiment is trajectory prediction on the FLeCS benchmarks: the K562 CRISPRi knockdown panel (12 knockdowns) and the 72-hour small-molecule perturbation responses in A549 lung cancer cells, both from A scalable gene network model of regulatory dynamics in single cells, whose codebase is available at Bertinus/FLeCS for exact protocol reuse. Knockdowns enter as control inputs on the knocked-down gene. Metric: Wasserstein-1 distance between predicted and held-out post-perturbation distributions, plus directional accuracy on known marker changes. For reference, FLeCS reports alignment with known biology in 11 out of 12 knockdowns, but no reported result found on Lacuna for a scalar trajectory-error value for FLeCS, so we will re-run FLeCS ourselves under the matched protocol to get the number.

The second experiment is dose and drug generalization on sci-Plex3 (~650,000 cells, 188 drugs across three cancer lines), the shared benchmark of Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution and Departures. We hold out doses and held-out drugs with shared targets, and report median $r^2$ on differentially expressed genes at 10 µM, directly comparable to the reported pretrained chemCPA figure of $r^2 = 0.79$ on DEGs (0.82 on all genes), and to non-pretrained chemCPA at $r^2 = 0.64$ and a CPA-style baseline at $r^2 = 0.48$, all from Predicting Cellular Responses to Novel Drug Perturbations. The decisive sub-check is long-horizon stability: simulate 2x and 4x past the training horizon and measure whether trajectories diverge or relax — the regime where the physics literature documents failure of unconstrained models.

The third experiment is the biological claim: MDA-MB-231 triple-negative breast cancer cells under 1 µM palbociclib, using the day-1/day-7 mass-spectrometry panel from Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition. The model must reproduce the directional signature (downregulation of Cyclins E2, A2, B1, B2; loss of Ki67 and Lamin B1; RRM1/RRM2 repression; upregulation of MHC class I/II and Trop-2 — no reported scalar metric found on Lacuna for this panel, so scoring is by directional agreement counts). The novel output is the drug-withdrawal simulation: whether the learned basin at day 7 is deep enough that $u=0$ leaves the cell arrested, which is the model's answer to the open reversibility question and a testable prediction against the follow-up experiments that field is already running.

Baselines

The naive reference point is a mean-shift predictor: take the control population, add the average expression shift observed for the perturbation in training, and output that distribution. It is the floor every serious method must clear, and it is exactly the behavior that Learning single-cell perturbation responses using neural optimal transport shows collapses multimodal drug responses into a single average response.

The strongest existing competitor is FLeCS (A scalable gene network model of regulatory dynamics in single cells), because it shares everything with our proposal except the physics: the same gene space, the same GRN prior, the same ODE simulation of perturbations, and the same adjoint-gradient training. It is the cleanest possible test of whether the port-Hamiltonian structure itself earns anything, so we run it with its released codebase on identical data, gene sets, and tuning budget. Since no scalar trajectory-error value is reported for FLeCS on Lacuna, both numbers come from our own matched runs. Beating FLeCS means lower Wasserstein-1 on held-out perturbations at matched parameter count, and — the sharper criterion — materially better prediction at 2x–4x extrapolated horizons and unseen doses, where its unconstrained production-minus-decay field has nothing preventing drift.

The strongest static baseline is pretrained chemCPA (Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution), which reports median $r^2 = 0.82$ on all genes and $0.79$ on DEGs at 10 µM on sci-Plex3; we use its released model where possible and otherwise its protocol. It cannot produce trajectories at all, so the comparison isolates the value of dynamics: we compare on its home metric (endpoint $r^2$) to show we are not losing accuracy for the added structure, and on horizon/dose extrapolation where it has no answer. The key ablation replaces the structured $R_\theta$ with an unconstrained matrix and the port $g_\phi u$ with an additive shift in the drift — same capacity, same training — leaving a generic GRN neural ODE. If that ablation matches the full model, the physics is decoration and the proposal's core claim is wrong; the experiment is designed so that this outcome would be visible, most likely in the extrapolation regime where the physics literature predicts separation.

Why not just…?

Why not just train an unconstrained neural ODE or optimal-transport map? This is the most serious objection, and the honest answer is that on in-distribution endpoint prediction it might suffice — CellOT (Learning single-cell perturbation responses using neural optimal transport) already does respectable work there, while acknowledging that its straight-path assumption breaks under cell death and proliferation. The evidence that unconstrained models fail in exactly the regimes we care about is consistent across the physics literature: black-box neural ODEs violated both thermodynamic laws under out-of-distribution forecasting in Machine learning structure preserving brackets for forecasting irreversible processes, and diverged within 50-second simulations in Port-Hamiltonian Neural ODE Networks on Lie Groups where the structured model improved tracking by a factor of 2 to 4. Our extrapolation-horizon and dose-generalization experiments are designed precisely to settle whether this transfers to gene space; if the unconstrained ablation wins there too, that is a result worth reporting and would kill the proposal's motivation.

Why not just use the static autoencoder (chemCPA) on snapshots? Because the questions that distinguish receptor-negative-cancer therapies are temporal: the day-1-versus-day-7 distinction during CDK4/6 inhibition, the stability of senescent arrest after withdrawal, and cell-cycle oscillations that static models conflate with drift. The reported numbers also show the static approach is fragile without massive pretraining — $r^2$ falls from 0.79 to 0.64 on DEGs without L1000 pretraining, and a CPA-style baseline sits at 0.48, per Predicting Cellular Responses to Novel Drug Perturbations — and latent mean-shift models miss bimodal, subpopulation-specific responses per CellOT. Our design concedes chemCPA's advantage where it is real: we compare on its endpoint metric to check we are not regressing, and only claim the dynamic questions it cannot address.

Why not just add soft physics penalties to the loss instead of new architecture? Because this alternative has been directly evaluated and failed in the closest setting: in Machine learning structure preserving brackets for forecasting irreversible processes, the penalty version produced physically impossible negative entropy production and high variance out-of-distribution, while the structurally constrained GNODE maintained non-negative entropy production by construction. APIC: Orthogonalized Neuro-Symbolic Modeling for Nonlinear Dissipative Dynamics documents the underlying optimization pathology — the physics penalty and the data fit fight each other in backpropagation. We will nevertheless run a penalty-only ablation (unconstrained architecture plus an energy-dissipation loss term) so the comparison is empirical in our setting, not inherited.

Why not just screen combinations empirically with structured ML, as BFReg-NN already did? AI-powered omics-based drug pair discovery for pyroptosis therapy targeting triple-negative breast cancer achieved a C-index of 0.90 and a striking in vivo result, so this is a real alternative for which combination to try — but it predicts synergy scores from static omics, not what happens over time at which dose and schedule, and it cannot address arrest reversibility at all. We treat it as complementary: its output pairs (e.g., mitoxantrone plus gambogic acid) become candidate multi-port inputs our model schedules and simulates.

Decisive Checks

The proposal stands or falls on three falsifiable outcomes. First, on matched runs against FLeCS, the port-Hamiltonian model must win at extrapolated horizons and unseen doses, not merely tie in-distribution — a tie there means the structure adds cost without benefit. Second, the learned dissipation must be readable: $R$ must grow along palbociclib trajectories relative to control, and the day-7 state must sit in a basin whose depth is estimable with uncertainty; if $R$ and $H$ are uninterpretable despite fitting the data, the biological payoff evaporates and only the forecasting result remains. Third, the drug-withdrawal simulation must make a commitment — reversible or not — that a collaborator's washout experiment can confirm or refute. Success on all three gives oncology a model class where a therapy is a control signal with guaranteed-bounded side effects on the internal dynamics, and gives non-conservative force modeling its first genuinely biological application where the learned dissipation law, not a hyperparameter, does the work.