betaFeedback

WorksAgricultural and Biological Sciences

Co2^2ODE: Conditionally Coarse-Grained Model Discovery for Rewiring in Biochemical Networks

Abstract

Biochemical networks propagate the signals that regulate cellular behaviour, and which routes carry them depends on cellular context: a shift in context rewires the network's effective dynamics. Mechanistic simulators are misspecified here because the effective dynamics emerge from an interaction structure that is only partially known. Model discovery instead learns the dynamics from data, but sparse, correlated measurements poorly identify a full network model. What matters is narrower: the portion the context rewires. We introduce conditionally coarse-grained ODEs (Co2^2ODE), which model that portion as a sparse neural or symbolic ODE: a conditional mask selects the markers carrying the rewiring, and a variational autoencoder compresses the remaining markers into a learnt closure term. On single-cell phosphoproteomic data spanning 30 cancer-relevant overexpressions, Co2^2ODE compresses a 30-marker panel into a 10-dimensional state while matching the predictive accuracy of a neural ODE fit to the full panel. The closure absorbs what the mask drops, hence the accuracy is insensitive to which markers are kept. Adding a prior score to the loss shifts the selected markers at no cost in fit, indicating the data are not interventional enough to identify the underlying biology.

Introduction

Biochemical networks underlie most cellular regulation, and their dysregulation underpins many diseases, cancer in particular [1, 2]. Dysregulation arises when a network’s topology or its surrounding context is perturbed: protein abundances and feedback strengths shift, the effective interaction strengths change, and with them how signal propagates [3]. The path to a target node is then rerouted or interfered with by alternative pathways, what we call rewiring. It limits targeted therapy for cancer: inhibiting a protein is a known change, but the rewiring that follows is not, and signal can reach the target by a route the inhibitor does not block [4, 5], differently in each patient.

Modelling rewiring means recovering, for each context, which interactions carry the signal, and what shape their dynamics take. Hand-built mechanistic simulators encode prior knowledge and are fit to data [6], but incomplete knowledge and poor transferability between contexts limit their predictiveness, and their scale leaves them weakly identifiable from partial, noisy measurements [7]. Data-driven model discovery avoids relying on prior knowledge entirely [9, 10], but its difficulty grows with the dimensionality of intracellular networks. Existing methods for discovering rewiring learn interaction structure from perturbation response [11, 12], but they assume the intervention acts on a node the model contains and the system sits at steady state.

In this study, we learn the coarse-graining and the dynamics simultaneously, removing the need for a trusted simulator or a full-dimensional learnt model. This is tractable because we do not need a general coarse-graining, only the markers carrying signal to the target in a given context. Conditionally coarse-grained ODE (Co2ODE\mathrm{Co}^{2}\mathrm{ODE}) is an end-to-end differentiable architecture that learns such a reduced model directly from data.

Our contributions are threefold. First, we learn a context-conditional coarse-graining jointly with the dynamics, such that a reduced model can be found where no parent model exists to reduce. Second, we build two prior-knowledge scores that evaluate the recovered structure against a curated network. These scores can also be added to the model as inductive biases. Our third contribution is to apply the model to real signalling data where we find that the node score relocates that structure at no cost in fit, showing these data do not identify the true rewiring, and agreement with known biology cannot validate what such a model recovers.

Related Work

Coarse-graining biochemical networks. There are many different approaches to model reduction in the literature: time-scale separation [13, 14], lumping [15], pruning [8], profile-likelihood reduction [16], manifold boundary approximation [17], and projection methods [18, 19, 20]. Most often they reduce an already existing parent model, and preserve their behaviour under conditions prescribed in advance. Here there is no such model, and the behaviour of interest is what changes with context.

Model discovery. This approach infers dynamics from data with little or no prior structure, from symbolic regression to neural ODEs [9, 10, 21, 22, 23], with sparsity regularisation giving interpretable structure in the learnt network [24, 25]. However the symbolic search space grows combinatorially with input dimension and neural formulations remain weakly identified in the same regime as the simulators they replace.

Closure. Removing variables from a dynamical system destroys Markovian closure: the retained variables no longer evolve autonomously, which can dampen predictiveness. The Mori–Zwanzig formalism restores this information by constructing a memory integral over the history of the retained variables together with a fluctuating remainder [35, 36, 38], and recent work learns it from data instead of deriving it [39, 40]. In our case, the discarded markers enter the model through the present-time latent evolved alongside the reduced state; the closure is therefore Markovian.

Modelling rewiring from perturbation data. Modular-response analysis recovers connection strengths from steady-state responses to module-specific perturbations [12], applied per cell line to quantify EGFR signalling and predict resistance-breaking combinations [27]. cSTAR pairs that inference with a supervised axis between known cell states, reducing to a core circuit that differs between cell lines and between receptor variants of one line [34]. These approaches need a perturbation specific to each module, and the modules are prescribed in advance and shared across contexts. CellBox instead learns a full interaction matrix at panel dimension [11, 26]. Our contexts act from outside the measured state and are seen only through a reporter, the data are out-of-equilibrium time courses, and the reduction is learnt per context rather than fixed.

Methods

Co2ODE\mathrm{Co}^{2}\mathrm{ODE} takes trajectories from a biochemical panel measured across contexts and returns, per context, a low-dimensional ODE around a designated target node. The architecture has three components, trained jointly (Figure 1). A context-conditional selector chooses which measured markers enter the model. The markers left out are handled by a latent closure. The retained markers and the latent are then evolved by a sparse dynamics module, conditioned on context and realised either as a neural or a symbolic ODE.

Co²ODE architecture

Figure 1. Co²ODE. A context-conditional selector keeps the markers that carry the rewiring. The discarded markers are compressed into a latent closure, which joins the selected markers to form the coarse state. That state evolves under a context-conditioned sparse neural or symbolic ODE.

Problem setup

For each context cc we observe a panel of MM markers, xc,s(t)∈RMx^{c,s}(t) \in\mathbb{R}^{M}, over t∈[0,T]t \in[0,T]. The scalar s∈[0,1]s \in[0,1] is the strength at which the context is induced. Each context is measured at a finite number of strengths, and we assume xc,sx^{c,s} varies smoothly in ss, so intermediate strengths interpolate between the levels observed. One designated coordinate yc,s(t)=xic,s(t)y^{c,s}(t)=x^{c,s}_{i}(t) is the target. We want to learn a small closure around yc,sy^{c,s}: a reduced state uc,s∈Rdu^{c,s} \in\mathbb{R}^{d}, d≪Md \ll M, comprising the target and a few other coordinates, evolving as u˙c,s=fc(uc,s,s)\dot{u}^{c,s}=f_c(u^{c,s},s). Contexts differ in which coordinates are selected and in the field fcf_c. That difference is the rewiring.

Context-conditional state selection

We select the markers entering the reduced state in two parts. We always keep the target rr and a small anchor set A⊂[M]\mathcal{A}\subset[M], the canonical routes to the target. The rest are chosen by a straight-through top-kk (k=3k=3) over per-context logits ℓc,i\ell_{c,i}, i∈[M]∖({r}∪A)i\in[M]\setminus(\{r\}\cup\mathcal{A}). Forward, the hard mask mc∈{0,1}Mm_c\in\{0,1\}^{M} keeps the kk largest; backward, gradients flow through softmax⁡(ℓc/τ)\operatorname{softmax}(\ell_c/\tau), where τ\tau is the annealing temperature [28]. Masks are indexed by context alone so a single selection applies across the context strength gradient. Thus Sc={r}∪A∪{i:mc,i=1}S_c=\{r\}\cup\mathcal{A}\cup\{i:m_{c,i}=1\} has 1+∣A∣+k1+|\mathcal{A}|+k elements, only kk of them learned. Throughout, xIx_I denotes the subvector on an index set II, entries in index order, such that the selected markers are xScc,sx^{c,s}_{S_c} and the discarded markers are xSˉcc,sx^{c,s}_{\bar{S}_c}, with Sˉc=[M]∖Sc\bar{S}_c=[M]\setminus S_c.

Closure by a latent

An ODE over the selected markers alone is not necessarily closed: The discarded markers could still act on them, and the ODE has no term for that influence. We therefore encode xSˉcc,sx^{c,s}_{\bar{S}_c} into a latent zSˉcc,sz^{c,s}_{\bar{S}_c} with a conditional variational autoencoder (CVAE) [31, 32], and split it into z∥c,sz^{c,s}_{\parallel} and z⊥c,sz^{c,s}_{\perp} to discourage it from becoming a second route for the context signal. We achieve this by regularising the CVAE with two heads. One of them predicts ss from z∥c,sz^{c,s}_{\parallel}, while a gradient-reversal layer [33] trains the encoder so that ss is not predictable from z⊥c,sz^{c,s}_{\perp}. Only z⊥c,sz^{c,s}_{\perp} joins the selected markers, giving the coarse state uc,s(t)=[xScc,s(t),z⊥c,s(t)]u^{c,s}(t)=[x^{c,s}_{S_c}(t),z^{c,s}_{\perp}(t)]: a learned Markovian closure [40] whose latent supplies background dynamics.

Context-conditioned dynamics

The coarse state evolves under u˙c,s=fc(uc,s,s)\dot{u}^{c,s}=f_c(u^{c,s},s) and is integrated with a fixed-step Runge–Kutta scheme to predict the target trajectory. The vector field fcf_c is learnt in one of two ways: (i) as a sparse neural ODE carrying a group-sparsity (L21) penalty on its Jacobian (Appendix L), or (ii) a conditional SINDy module [41] that regresses over a non-linear function library sparsely with per-context coefficients. In both cases, the strength modulates the coefficients continuously. The rewiring structure is read off the Jacobian elements ∂fc,i/∂uc,s\partial f_{c,i}/\partial u_{c,s}. We report the neural implementation in the remainder of the paper unless stated otherwise; the symbolic version is discussed in Appendix § 16.

Training

All components are differentiable and are trained jointly against a weighted sum of the mean squared error of the target trajectory integrated from t=0t = 0, the autoencoder’s negative ELBO (Gaussian reconstruction and KL to a standard normal prior), the mean-squared supervised and adversarial strength regressions, and the dynamics sparsity penalty. The neural variant trains in one stage with Adam [42]; the SINDy variant adds a second stage that freezes the selector and autoencoder and sparsifies the dynamics. Schedules are annealed over training; hyperparameters are in Appendix § 10.

Scoring the model against prior knowledge

Our prior is a curated directed interaction network over the markers and overexpressed genes, built with OmniPath [43]. The node score charges each selected marker for the detour it forces: for a marker ii, we compare the shortest path from the perturbed node pp to the target rr that passes through ii against the shortest path overall. The edge score penalises each dependency the model learns—a nonzero Jacobian entry, or an active symbolic term—by the shortest-path distance between the two markers on the prior network. Direct edges are free; longer paths cost more, and pairs with no path at all cost the most. Both the prior and the distance are given in Appendix § 13.

Experiments

Data, metrics, and baselines

We use the single-cell mass-cytometry dataset from Lun et al. [3] [3], in which single-gene overexpression induces a set of contexts in a common cell background. A 30-marker phospho panel is read out in 6 snapshots over one hour, following stimulation upstream of the network by the growth factor EGF. Two conditions are dropped for mismatched time grids, leaving 38: the 30 overexpressions we report on, and 8 controls that carry no overexpression. Trajectories are Gaussian-process smoothed [44]. The context strength is a GFP-derived proxy for induction, the target is p-S6 (a growth-signalling readout), and the anchors are p-MEK and p-AKT (upstream pro-proliferation nodes); the 80/10/10 split gives 2919/365/365 (Appendix § 9).

We report range-normalised mean absolute error (rnMAE) on the target trajectory from t=0t = 0; the cross-seed overlap of the selected marker set; and the two prior scores of § 4 (Appendices § 11, § 15).

Two ablation studies each change one part of the model. The first varies how much of the panel enters the state: anchors and target alone (3), plus the markers the selector adds (6), plus the four-dimensional closure (10, Co2ODE\mathrm{Co}^{2}\mathrm{ODE}), and the panel unreduced (30). The 3- and 6-dimensional arms are closure-free, not latent-free: they train the same autoencoder, but its latent never enters the ODE. The second varies how the bias enters: not at all, node score, edge score, or both. Every arm is swept over the same grid and selected on validation (Appendices § 12, § 17).

Results

Coarse-graining costs almost nothing. We first asked what the reduction costs. The three reduced models from the first ablation study sit within half a point of one another, and the full panel is no better on either metric (17.1%/14.8%; Figure 2a). The best fit of the three is the 3-state control (−0.4 points against Co2ODE\mathrm{Co}^{2}\mathrm{ODE}, p = 0.004), which selects nothing and so recovers no context-specific structure. A ceiling in the data explains the flatness: a probe given the anchors, context and strength predicts the target at R2=0.917R^2 = 0.917, and the other 27 markers add nothing (0.913). The symbolic implementation, with a 1-dimensional latent instead of 4, compresses the 85-term library to 3–4 terms per equation and sits 7 points behind on the target (Appendix § 16).

Four-panel plot showing state ladder, marker swap, reselection, closure, and prior scores

Figure 2. rnMAE (Appendix C); lower is better. Boxes in (a, b) show the 38 per-context values, each a median over seeds (10 per arm; 5 for the 30-marker arm); whiskers 1.5×IQR. (a) Reducing the state from 30 to 10, 6 or 3 variables costs at most half a point. (b) Swapping in another context’s or random markers costs about ten points. (c) Seed-to-seed Jaccard overlap of selections, same vs. mismatched context; permutation test. (d) Extra error of runs departing from the modal selection, with and without closure; bootstrap 95% CIs. (e) Both prior scores at one bias strength, in hops on the prior graph; bars are means, lines min–max over 10 seeds.

Prediction runs through the selected markers. We next asked whether the model uses the markers it selects. We gave a trained model another context’s selected markers at inference, leaving its context embedding and closure untouched. The median target error rises from 15.9% to 25.0%, and to 26.2% with random markers (Figure 2b); the swap is worse than the model’s own markers in 36 of 38 contexts. The cost is the same in the closure-free models, so the selection and not the latent carries the prediction. Selections also repeat across seeds: the Jaccard overlap of the three chosen markers is 0.56, against 0.12 for mismatched contexts (permutation p=2×10−4p = 2 \times10^{-4}; Figure 2c). For 19 of the 30 perturbations a single marker is kept in at least 8 of 10 seeds while rarely appearing elsewhere; no control has one.

The closure buys robustness. The closure encodes exactly the markers the coarse state excludes, so the model sees the whole panel either way and only the split between state and latent changes. An unusual split should then cost little. A seed that departs from the most common selection predicts 4.1% worse without the closure ([1.0, 7.4]) and no worse with it (−2.3%, [−6.9, 2.7]), at equal mean accuracy (Figure 2d). The closure therefore makes error uniform across seeds rather than lower. The effect replicates in an independent out-of-distribution test (Appendix § 14).

A prior rearranges the structure. We last asked how tightly the data constrain the structure the model recovers, using each prior score as a probe. Fit barely moves across the four bias settings: p-S6 rnMAE spans 16.1–16.5% and the three-marker average 14.1–14.5%. The structure moves instead. The node bias cuts the node score from 1.06 to 0.10. It also moves the edge score, which it does not optimise, from 2.85 to 2.66 (p=0.002p = 0.002 over 10 seeds), and both biases together reach 0.10 and 2.63 (Figure 2e). The edge bias alone moves neither score. A prior therefore rearranges the recovered structure at no cost in fit.

Discussion

We introduce conditionally coarse-grained ODEs (Co2^{2}ODE) to decode rewiring in biochemical networks across different contexts. Our approach matches a 30-state model on the target, but so does a three-state control: on these data the reduction costs nothing and buys nothing. Its value is that the context-specific selection is explicit and reproducible across seeds where the full model leaves it implicit—and that the same machinery then shows how weakly these data constrain what it selects. Indeed, adding a prior score to the loss moves the recovered structure onto known pathways at no cost in accuracy. The fit is therefore near-indifferent between structures that disagree on the prior graph, so agreement with known biology is not evidence that the structure is right.

Deciding between the structures these data leave open needs interventions they do not contain. We are designing a mass cytometry experiment after Lun et al. (2019) [3], adding an additional drug panel targeting phospho-proteins in the measured panel. Cutting an edge inside the state breaks the correlations that make competing structures indistinguishable: the 3-state control should then stop being sufficient on its own, and the markers this reduction discards should start to matter.

Acknowledgments and Disclosure of Funding

F.F. and T.P. were supported by the Francis Crick Institute, which receives its core funding from Cancer Research UK (CC2242), the UK Medical Research Council (CC2242), and the Wellcome Trust (CC2242). This work was funded by the European Union (ERC, DeepMechanism, 101163005). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. F.F. reports consulting honoraria from Deep Origin, which had no influence on this study.

References

Data and preprocessing

Split structure. The split is random over trajectories, grouped by GFP bin and seeded identically for every arm and every model seed. All 38 contexts therefore appear in all three splits, and each trajectory appears in exactly one. Generalisation reported here is across trajectories within a context; the architecture is conditioned on a context embedding and does not transfer to a context it has never seen, so held-out-context evaluation is not applicable.

SourceLun et al. [3] [3], single-cell mass cytometry
PreprocessingGP-smoothed trajectories, GFP-binned, with a derived strength column
Panel30 phospho-markers
Conditions40 total; 38 retained: 30 single-gene overexpressions and 8 controls
Dropped2 contexts whose time grids do not match the rest (DUSP4, MAP4K2)
StimulusEGF, upstream of the network
Timesnapshots over 1 h post-stimulation
SmoothingGaussian-process fit per context [44]
Targetp-S6
Anchorsp-MEK1-2, p-AKT (fixed across contexts, always retained)
Strength ssGFP-derived induction proxy, scaled to [0,1][0, 1], predicted rather than observed
Split80/10/10 by trajectory: 2919 / 365 / 365
Split seedfixed at 42 across every arm and every model seed

Table 1. Dataset and preprocessing. Unit of splitting is the trajectory.

Architecture and hyperparameters

Base preset, with the per-arm overrides in Table 3.

ArmState dimLatent in ODEComposition of the state
3-state3notarget and 2 anchors only; no marker selection and no latent closure
6-state6notarget, anchors, and the 3 selector-chosen markers; no latent closure
Co2^{2}ODE10yes, dim 4the 6 markers above plus the 4-dimensional z⊥z_\perp evolved inside the ODE
Full panel30—all 30 markers as ODE states, context-conditioned, no selection or closure

Table 3. The four arms of the state-dimension ablation.

Table 2: Architecture and optimisation, identical across arms except where Table 3 states otherwise. HSIC is the Hilbert–Schmidt independence criterion [37].

ComponentSettingValue
Selectorconditioning dim16
hidden128
number of conditions38
relaxation temperatureannealed 2.5 →\to 0.6, rate 0.95, from epoch 10
Gumbel samplingoff (straight-through top-kk)
explicit slots6 = 3 fixed (target + 2 anchors) + k=3k=3 learned
Conditional VAEencoder hidden(128, 64), dropout 0.15
decoder hidden(64, 64), dropout 0.15
z⊥z_\perp (adversarial)4
z∥z_\parallel (strength-supervised)3
gradient-reversal scale3
DynamicstypeMLP, hidden 192
predictor hidden192
context embedding dim32
solverRK4, fixed step
substeps per observed interval2
OptimiserAdam, learning rate1.622×10−31.622 \times10^{-3}
weight decay2.542×10−72.542 \times10^{-7}
gradient clip0.5
batch size128
lr schedule10 warmup epochs, min scale 0.5
epochs / early stop400 / 50
Loss weightsλrec,λs\lambda_{\mathrm{rec}}, \lambda_{\mathrm{s}}1.0, 1.0
λadv\lambda_{\mathrm{adv}}0.1018 (trajectory 0.2607), warmup 10 epochs from 0.2
βKL\beta_{\mathrm{KL}}0.02
λsp\lambda_{\mathrm{sp}}swept; see Table 5
latent invariance, correlation / HSIC2.23×10−3/4.59×10−32.23 \times10^{-3} / 4.59 \times10^{-3}
latent trajectory, corr / HSIC4.14×10−2/8.28×10−24.14 \times10^{-2} / 8.28 \times10^{-2}
latent L2 / velocity4.64×10−5/3.42×10−54.64 \times10^{-5} / 3.42 \times10^{-5}
trajectory latent supervision0.5
fit schedulemix over 60 epochs, λ\lambda 0.5 →\to 0.2

Table 2. Architecture and optimisation, identical across arms except where Table 3 states otherwise.

Metric

Range-normalised MAE (rnMAE) is computed per trajectory and per marker:

rnMAE=100×1T∑t∣y^t−yt∣max⁡(max⁡tyt−min⁡tyt,δ),δ=0.5.(1)\mathrm{rnMAE} = 100 \times\frac{\frac{1}{T}\sum_t \lvert\hat{y}_t-y_t\rvert}{\max\left(\max_t y_t-\min_t y_t,\delta\right)}, \qquad\delta=0.5. \tag*{(1)}

The floor δ\delta prevents blow-up on degenerate flat trajectories. The floor does engage. The median test-trajectory range for p-S6 is 0.79, and 27% of test trajectories fall below δ\delta. On that subset the denominator is the floor, not the trajectory’s own range, which compresses the differences between all models. Trajectories are integrated from t=0t = 0; no teacher forcing is used at evaluation. Cells are aggregated by averaging per-trajectory rnMAE, never by pooling absolute errors and dividing by a pooled range. “mean3” is the unweighted mean of rnMAE over {p-S6, p-MEK1-2, p-AKT}.

Aggregation. Figure 2a,Figure 2b box the 38 per-context values, each a median over model seeds. Numbers quoted in § 5 use the same reduction: rnMAE is first reduced per context as the median over model seeds, and the reported figure is the median over the 38 contexts. The two statistics will not agree to the decimal; both are given so the figure and the text can be read against each other.

Trivial baselines. rnMAE is relative, so its scale needs anchoring. These rows give what it returns for predictors that model no dynamics at all, on the same test split, target and aggregation as every other number in the paper.

The per-context mean trajectory is the context lookup table: it is given the context identity and returns that context’s average behaviour, with no dependence on the initial state. It reaches 36.7%. Co2^{2}ODE reaches 16.0%, so the reduced model is not reproducing a per-context average, and the flatness of the state ladder in Figure 2a is not the flatness of a metric that cannot separate anything.

Training, sweeps, and model selection

Selection metric. Early stopping and best-checkpoint selection use a reporters-only validation metric, computed on the target and anchors alone rather than uniformly over all state slots. With a uniform-over-slots metric the reporter share of the selection signal shrinks as the state grows (100% at 3 states, ∼50% at 6, ∼10% at 30), so larger arms have their checkpoint chosen on markers the paper never forecasts. Every number reported here uses reporters-only selection. For the 3-state arm the two metrics coincide, since all three of its slots are reporters, so that arm is an unchanged control.

Sweep. Both nuisance hyperparameters are swept identically for every arm and selected per arm on validation loss; test is reported once.

Swept quantityGrid
Reporter weighting of the trajectory loss{0.5, 0.7, 0.9}
Jacobian group-sparsity strength λsp\lambda_{\mathrm{sp}}{0, 1×10−21 \times10^{-2}, 3×10−23 \times10^{-2}}
Model seeds0…90 \ldots9

Table 5. Sweep grid, applied identically to the three reduced arms (3×3×10=903 \times3 \times10 = 90 runs each); the full panel had its own width and learning-rate sweep.

Full-panel baseline tuning. The full panel received the same reporter-weighted objective and reporters-only selection, the same training budget (400 epochs / early stop 50), a width sweep (hidden ∈{192,384,768}\in\{192, 384, 768\}) and a learning-rate sweep ({8×10−4,1.6×10−3}\{8 \times10^{-4}, 1.6 \times10^{-3}\}). Hidden 192 with lr 8×10−48 \times10^{-4} won; more capacity made it worse (hidden 768 reached a validation training-objective loss of 0.222 against 0.161 at 192; this is the training loss, not rnMAE), so the baseline is not capacity-limited.

Prior graph and the two scores

ResourceOmniPath [43], directed literature-curated interactions
AccessOmniPath enzsub endpoint via the omnipath Python client, downloaded 28 April 2026
Layer and filtersthe enzyme–substrate (enzsub) layer, human, restricted to modifications containing “phosphorylation”, taken directed; unsigned (activation and inhibition are not distinguished); self-edges dropped. Retained edges carry curation effort 3–44 (median 4)
Graph size70 directed edges over 28 of the 30 panel markers; the remaining 2 have no retained edge and are unreachable by construction
Marker mappingcurated per-marker gene-symbol sets, so a phospho-marker maps to every gene of its family (p-RAF → ARAF, BRAF, RAF1; p-AKT → AKT1–3; p-ERK1-2 → MAPK1, MAPK3), and an edge is kept only when both endpoints land on panel markers
Cost basispairwise shortest-path distance on the directed graph
Latent coordinatestreated as unreachable_cost (no graph identity)
Self loopsself_loop_cost, set separately from the distance lookup

Table 6. Prior interaction network.

Node score. Mean prior cost of the three markers the selector chooses, anchors and target excluded, scored by position on the shortest path from the intervened gene to the target: on a short route scores low (good), unreachable scores at the ceiling. The 8 control contexts have no intervened gene, so the node score is defined only for the 30 overexpressions and the prior-score results are over those.

Edge score. Mean prior cost of the dependencies the dynamics express — a nonzero Jacobian entry ∂fc,i/∂uj\partial f_{c,i}/\partial u_j or an active symbolic term — scored by the shortest-path distance from jj to ii. A direct edge costs 0; each additional hop beyond free_hops costs more; an unconnectable pair costs unreachable_cost (shortest-path costs precomputed once per context).

Prior scores by bias setting. Node score, then edge score, means over 10 seeds, in hops on the prior graph: none 1.06 / 2.85; node bias 0.10 / 2.66; edge bias 1.09 / 2.87; both 0.10 / 2.63. The edge bias moves the edge score only at a thirty-fold stronger setting (2.70, p=0.004p = 0.004). That setting is a separate sweep at 4 substeps rather than 2, so it is compared only against its own unbiased control, against which it costs +0.8% of validation loss over 10 seeds (p=0.52p = 0.52); it is not comparable with the four matched-strength settings above. Fit across the same four settings spans 16.1–16.5% on p-S6 and 14.1–14.5% on mean3^{3}.

Where the scores are evaluated. When a bias is added to the loss it uses the same cost matrix, so reporting the biased model’s own penalty would be circular. Figure 2e therefore scores every arm, biased and unbiased, with the identical unbiased graph cost.

Probe and ablation protocols

Table 7 gives, for each claim in § 5 that does not come from the fit table, what was measured and the value obtained. All probes are fitted on the training split and evaluated on the held-out test split.

ProtocolWhat is doneReported
Conditioned completenessPredict p-S6 at t=1…5t=1\ldots5 from t=0t=0. The probe is a random-forest regressor (300 trees, minimum leaf size 2) predicting all five future time-points jointly, scored by 5-fold cross-validated R2R^2 averaged over outputs; context enters one-hot and strength as a scalar. A linear probe understates the anchors and so makes the panel look informative; the result below requires a nonlinear probeR2=0.796R^2=0.796 anchors only; 0.917 adding context and strength; 0.913 full panel with conditioning
Random-triple controlSelector’s markers against random marker triples, under the completeness probe0.883 against 0.889: the selected markers are no more informative than random ones. The marker-swap test asks the different question of whether the trained model uses them
Marker swapAt inference, give a trained model another context’s markers, leaving its context embedding and latent closure untouchedp-S6 rnMAE rises from 15.9 to 25.0 (median over contexts), and the swap is worse than the model’s own markers in 36 of 38 contexts; random markers are worse in 38 of 38
Off-consensus costSplit runs by whether their selection matches the most common one for that context, and compare error within each group+4.1% [1.0, 7.4] closure-free against −2.3% [−6.9, 2.7] with the closure; mean error 14.39 against 14.44
Selector-variability costRank-correlate a run’s selection agreement with the most common selection against its error, per (seed, context); 300 points per arm, bootstrap CIs clustered over seeds; on the p-S6/MEK/AKT average. Percentages are relative error changes; rnMAE differences elsewhere are in pointsclosure-free ρ=−0.134\rho=-0.134 [−0.186, −0.084]; closure −0.000 [−0.131, +0.105]. Replicated in an independent strength-OOD grid (closure-free +11.8% [+5.1, +18.3]; closure −3.8% [−10.7, +3.2]). Pinning the closure arm’s selections into the closure-free model costs +0.51 points [−0.54, +1.54] against a peer seed’s, so the picks are different, not worse
Strength and context probePredict ss and context identity from z⊥z_{\perp}, against the same from the selected markers. Continuous targets: standardised ridge regression, penalty chosen by generalised cross-validation over 13 log-spaced values, scored by 5-fold R2R^2. Categorical targets: standardised multinomial logistic regression, scored by 5-fold cross-validated accuracyR2=0.71R^2=0.71 for strength; 0.20 against 0.83 for context identity on the encoder’s sample, 0.80–0.98 once evolved in the ODE (chance 0.03)
Seed reproducibilityOverlap of selected free-marker sets for the same context across seeds, against a mismatched-context null0.56 against 0.12 (4.7×); permutation p=2×10−4p=2\times10^{-4}; 36–37 distinct supports of 38
Random-support retrainRetrain with a random support of equal sizep=0.026p=0.026
Full-panel context specificityCross-context cosine of non-anchor dependency patterns0.997 over 5 seeds: one pattern for all 38 contexts

Table 7. Protocols behind each structural claim.

Statistical treatment

Unit of analysis. Each arm is trained with 10 seeds (full-panel control: 5). rnMAE is first reduced per context, as the mean over the test trajectories of that context, giving 38 values per seed. Comparisons between arms are therefore paired: the same 38 contexts and the same test split are seen by every arm.

Test. Arms are compared by a two-sided Wilcoxon signed-rank test over the 38 paired per-context values. We report the mean paired difference, the pp-value, and the number of contexts favouring each arm, since with 38 paired samples and a consistent sign a difference of a few tenths of a point is detectable while remaining small: significance and magnitude are reported separately and should be read separately.

Table 9: Paired comparisons of p-S6 rnMAE over the 38 contexts, all arms at the matched protocol. On the three-marker average the same comparisons give −0.23 (pp = 0.005), −0.11 (pp = 0.194), −0.61 (pp = 0.046) and −0.85 (pp = 0.002): Co²ODE separates from the full panel on mean3 but not on the target alone.

Comparison, p-S6Δ meanWilcoxon ppContexts favouring smaller
3-state against Co²ODE−0.440.00429/38
6-state against Co²ODE−0.230.24322/38
Co²ODE against full panel−0.330.24322/38
3-state against full panel−0.770.00529/38

Table 9.

Inductive-bias arms. The four bias settings are compared to the unbiased arm by the same paired test over the 38 contexts, at the matched protocol and with 10 seeds each. On p-S6 the node bias gives −0.32 (pp = 0.005), the edge bias −0.19 (pp = 0.040) and both together −0.22 (pp = 0.118); on the three-marker average none of the three separates (pp = 0.171, 0.966, 0.099). The biases therefore do not cost fit — if anything they improve it slightly — but the effect is ∼2% of the error, against a 90% reduction in the node score, so the objective is close to indifferent between structures that differ wholesale on the prior graph.

Whole-state loss. The masked trajectory loss differs by ∼36% between the unbiased and node-biased arms (median 0.136 against 0.087), but this is not a like-for-like comparison: the bias changes which markers occupy the state, so the two arms are scored on different slots. Only the reporter metrics, which are fixed by construction, are comparable across arms, and those are the ones reported.

Structural claims. The seed-reproducibility result uses a permutation test against a mismatched-context null (p=2×10−4p = 2 \times10^{-4}, 10410^{4} permutations); the random-support retrain comparison uses a paired test over contexts (p=0.026p = 0.026); the out-of-distribution comparison reports per-protocol paired tests (p=0.093p = 0.093 on p-S6, 0.013 on mean3 for the hardest protocol). Probe R2R^{2} and accuracy values are single point estimates on the held-out split and are reported without intervals.

Multiplicity. All tests are two-sided and treat the context as the unit of analysis. Reported pp-values are unadjusted: the comparisons are pre-specified but exploratory. Several claims we report sit between 0.01 and 0.05 and should be read as such.

The symbolic variant

The dynamics module has two interchangeable implementations (§ 3.4). Every result in the main text uses the sparse neural one. This section reports the conditional SINDy variant, which returns the recovered structure as equations rather than as Jacobian entries.

The symbolic runs use a different Co2^2ODE configuration. The 6-marker state is the same: the target, the two anchors, and the 3 selected markers. The latent is 1-dimensional rather than the 4 of Table 3, so the ODE has 7 states rather than 10. Each latent state adds an equation and its couplings to the library, which makes the coefficients harder to read. The gap to the neural numbers is therefore partly configuration and not only the dynamics module. The latent enters the target’s equation in 0–6 of the 38 contexts.

All 260 symbolic runs use the same data, split and metric as the main text. The table below rests on the 205 that are compatible with the current implementation and do not diverge, spanning 72 configurations at 1 to 4 seeds each. Fifteen runs diverge and are excluded; a further 40 belong to an earlier version of the dynamics module and are not comparable with it. That second exclusion is not random: those runs are the only ones trained without a learned selector, so the symbolic arm has no selector ablation, and its results should be read as conditional on the selector rather than as a test of it.

Table 10: The symbolic variant, 205 non-diverged runs grouped by active terms per equation, on the same metric and split as the neural arms at the foot of the table; Table 4 gives the trivial baselines for the same metric. Both error columns are medians over the 38 contexts and then over runs, with the range across runs in brackets; mean3 averages the target and the two anchors. Compressing an 85-term library to 3–4 active terms costs 2.2 points of target error against the densest runs and 1.3 points of mean3 against the best band, so the reduction is inexpensive but not free. The neural rows are the state-ladder sweep of Figure 2a.

PredictorrnMAE p-S6 (%)
Global mean trajectory86.0
Linear extrapolation from t0t_0, t1t_137.1
Hold the initial value36.8
Per-context mean trajectory36.7
Best flat line per trajectory, fitted in hindsight28.4
Symbolic variant, 3–4 terms per equation (Appendix § 16)23.4
Co2^{2}ODE16.0

Table 4. What rnMAE gives for predictors that learn nothing, on p-S6. “Best flat line” is the constant minimising absolute error for each trajectory, chosen in hindsight, so it is a lower bound on any constant prediction and not available to a causal predictor. The learnt arms are included for scale.

ModelRunsTerms/eq.p-S6 (%)mean3 (%)Target eq. non-empty
Symbolic, 3–4 terms623.523.4 [21.9, 27.4]23.5 [21.7, 26.4]38/38
Symbolic, 4–8 terms454.523.2 [20.8, 30.5]23.3 [20.9, 28.3]37/38
Symbolic, 8–20 terms3611.825.5 [21.3, 32.4]23.9 [21.3, 30.0]36/38
Symbolic, 20–60 terms4523.423.3 [19.9, 25.9]22.2 [18.1, 24.6]38/38
Symbolic, 60–85 terms1785.021.2 [19.7, 21.5]23.4 [19.9, 23.6]38/38
Neural, Co2^2ODE16.014.3
Neural, full panel17.114.8

Table 10.

Compression is cheap here too, though its cost depends on the metric. On the target, going from 20–60 active terms to 3–4 costs 0.1 points (Mann–Whitney p=0.03p = 0.03 on a 0.1-point difference) and only runs keeping essentially the whole library are meaningfully better, by 2.2; the intermediate bands differ by 0.1–0.2. On the three-marker average the ordering differs: the 20–60 band is best at 22.2%, compressing to 3–4 terms costs 1.3 points (p<0.001p < 0.001), and the densest runs are no better than the sparsest. Neither metric shows error growing steadily with sparsity: among models under 20 terms, the correlation between terms and target error is +0.24 (n=143n = 143, p=0.005p = 0.005). The symbolic variant is behind the neural one throughout, 23.4% against 16.0% on the target, which is why the main results use the neural implementation.

The configuration we would report is the cap-scheduled, graph-biased one. Three of its four seeds reach 5.6–5.8 terms per equation at 22.6–23.1% rnMAE, with the target’s own equation non-empty in all 38 contexts and carrying a numerically non-negligible coupling in 37 of them. For the ALPK2 context, seed 0 returns

uS6=(0.040−0.009s)−(0.017+0.012s)uS6+(−0.008+0.012s)uAKT−(0.010+0.024s)uS62+(0.005−0.006s)uAKTuJNK,u_{\mathrm{S6}} = (0.040 - 0.009s) - (0.017 + 0.012s)u_{\mathrm{S6}} + (-0.008 + 0.012s)u_{\mathrm{AKT}} \\ \qquad- (0.010 + 0.024s)u_{\mathrm{S6}}^{2} + (0.005 - 0.006s)u_{\mathrm{AKT}}u_{\mathrm{JNK}},

with the induction strength ss modulating every coefficient. Across the 38 contexts the target’s equation couples most often to the two anchors, p-MEK1-2 in 15 contexts and p-AKT in 11. Both anchors are in the state by construction in every context, so this is a statement about which couplings the sparsifier keeps, not about which markers were available.

Sparsity comes from the prior graph and the cap schedule, not from the sparsity penalty: the unstructured proximal sparsifiers (L1, elastic net, group L1) sit at 62–85 terms per equation whatever their penalty strength, which is no reduction against the library.

Three failure modes qualify this. Fifteen of the 220 evaluated runs diverge outright (test trajectory MSE above 10, against roughly 0.12 for healthy runs) and are excluded above; in the reported configuration that is one seed in four. Sparsity can also come from emptying the target’s equation rather than compressing it: the most aggressive second-pass sparsifier reaches 3.5 terms per equation but reduces the target to a strength-modulated constant in 11 of 38 contexts. And the accuracy-capped configurations delete the target’s equation outright in 4 of 38 contexts. A term count alone therefore cannot distinguish compression from deletion, and we report the count of contexts with a surviving, non-negligible target equation alongside it.

Compute

HardwareCPU only; no GPU used at any stage
Per run8 CPU cores, 24 GB RAM
Wall-clock, Co2ODE\mathrm{Co}^{2}\mathrm{ODE} and reduced controls≈50 min per run, ≈6.5 CPU-hours
Wall-clock, full-panel control≈30 min per run, ≈4 CPU-hours
Reported state-dimension experiment270 sweep runs + 29 of 30 full-panel runs (one failed to finish), ≈1900 CPU-hours
All experiments including exploratory≈2700 CPU-hours
SchedulerSlurm, shared institutional CPU partition

Table 11. Compute. The dominant cost is the identical hyperparameter sweep applied to every arm rather than any single model: one Co2ODE\mathrm{Co}^{2}\mathrm{ODE} model trains in under an hour on 8 cores.

Code, data, and licences

Code. Code is not released for review, and will be released on publication. Section K gives the exact commands and Section B the full hyperparameter set, so the results are reproducible from the description alone.

Data. Single-cell mass-cytometry measurements from Lun et al. (2019) [3], obtained from the resource published with that paper (Mendeley Data, 10.17632/3kh7ypz232.1). We use it unmodified apart from the preprocessing in Section A. The dataset contains no human-subject or personally identifying information; it comprises measurements on a cultured cell line.

Prior network. OmniPath [43], enzsub endpoint, accessed 28 April 2026. OmniPath aggregates literature-curated resources that retain their original licences; we use it for scoring only and redistribute no part of it.

Software. PyTorch 2.8.0, NumPy 1.26.4, SciPy 1.16.1, pandas 2.3.1, scikit-learn 1.7.1, matplotlib 3.10.5, NetworkX 3.5 and the omnipath client — all under BSD-, MIT- or Apache-family licences and used as published. The conditional SINDy module is our own implementation of the method of Brunton et al. [41]; no symbolic-regression package is used in any reported run.

Exact run configuration

Sections B–D state the configuration in full; the invocation below is included only so that it is unambiguous which settings were non-default. One Co2^{2}ODE seed:

python -m train --model_name PRESET \
  --data_path data/experimental/lun_2019_100gfp_with_gfpbin_strength_gpfit.csv \
  --perturbation_col scenario --strength_col strength \
  --val_size 0.1 --test_size 0.1 --split_random_state 42 --seed 0 \
  --drop_mismatched_scenarios --no-explicit_inputs_as_exogenous --conditional_vae \
  --explicit_input_features "p-AKT,p-MEK1-2" --explicit_target_features "p-S6" \
  --strength_source pred --max_explicit 6 --evolve_adv_in_ode --latent_adv 4 \
  --lambda_neural_jacobian_l21 3e-2 --traj_reporter_frac 0.9 --select_on_reporters \
  --beta_kl 0.02 --epochs 400 --early_stop 50 --device cpu --output_dir OUT

The reduced controls differ only in the state composition given in Table 3, with the latent closure disabled. The full-panel control is a separate 30-state model, context-conditioned, at hidden width 192 and learning rate 8×10−48 \times10^{-4} (the winner of its own width and learning-rate sweep), with the same group-sparsity penalty and the same reporters-only selection.

Limitations

Out-of-distribution behaviour. Across four strength-extrapolation protocols Co2^{2}ODE is worse than the 3-state control on mean3 in all four; p-S6 is mixed and never significant. On the hardest protocol, with dose bins 85–99 held out and validation on 70–84: 3-state 30.64 / 34.91 against Co2^{2}ODE 32.17 / 36.97 (p=0.093p = 0.093 / 0.013).

Coupling-level reproducibility. The overlap of 0.56 in § 5 is over selected marker sets. The recovered couplings themselves do not reproduce across seeds: pairwise agreement of the dependency support is 0.032 closure-free and 0.036 with the closure, both at chance. Both arms behave the same way, so this is a property of the Jacobian readout. Which markers a context uses repeats. Which couplings the model writes down among them does not.

Latent leakage. The adversarial head acts on the encoder’s sample; no penalty in the reported runs acts on the evolved latent. A sweep that does, λinv,ode∈{0.03,0.1,0.3,1.0}\lambda_{\mathrm{inv,ode}} \in\{0.03, 0.1, 0.3, 1.0\} at 10 seeds each, reduces the strength correlation of the evolved latent from 0.61 to 0.47 at λ=0.3\lambda= 0.3, with validation fit slightly better (0.0291 against 0.0311). The leak is reduced, not removed.

Sparsity. No sparsity advantage is claimed. We measure concentration of the target’s Jacobian row by k90k_{90}, the number of entries, ranked by absolute value, needed to reach 90% of the row’s absolute mass. Within an arm the L21 penalty changes accuracy (ΔrMMAE\Delta r\mathrm{MMAE} at λ=3×10−2\lambda= 3 \times10^{-2}: 3-state +0.72, 6-state +0.35, Co2^{2}ODE −0.51) but does not reduce k90k_{90}, so it regularises rather than prunes. We make no cross-arm comparison, since k90k_{90} is a raw count and a wider state can spread the same mass over more entries. The full-panel control carries the same penalty, so a sparsity-penalised 30-state model of the kind proposed in [24] is already in the comparison and fits no better (Table 8); we do not report its k90k_{90}, so whether that penalty leaves a 30-state structure readable is open.

Armp-S6 (%)mean3 (%)
3-state control15.613.8
6-state (closure-free)16.114.0
Co²ODE (10-state)16.014.3
Full panel (30-state)17.114.8

Table 8. The state ladder, medians over the 38 contexts. These are the levels behind the paired differences in the next table, and behind “within half a point” in § 5.

Discretisation. Results are reported at the training discretisation (RK4, 2 substeps per observed interval), and every arm was trained at that setting with no teacher forcing. Evaluating a trained model under a finer integrator degrades fit and shifts the recovered dependencies, but the integrator is part of the fitted model, so this measures sensitivity to swapping it and not convergence of a vector field. We do not run the convergence test, which would retrain across substep counts and compare the recovered structure. Comparisons here are therefore valid between runs at matched substeps. Runs at 4 substeps exist for the reduced arms and are excluded, since no full-panel counterpart was trained that way.

Seed counts. Co2^{2}ODE and the reduced controls use 10 seeds each; the full-panel baseline uses 5. The baseline’s spread is small relative to the gap: test-set R2R^2 ranges over 0.936–0.941 across its seeds.

References

  1. [1]B. Alberts, A. Johnson, J. Lewis, D. Morgan, M. Raff, K. Roberts, and P. Walter. Molecular Biology of the Cell. Garland Science, 6th edition, 2015.
  2. [2]D. Hanahan and R. A. Weinberg. Hallmarks of cancer: the next generation. Cell, 144(5):646–674, 2011.DOI
  3. [3]X.-K. Lun, D. Szklarczyk, A. Gábor, N. Dobberstein, V. R. T. Zanotelli, J. Saez-Rodriguez, C. von Mering, and B. Bodenmiller. Analysis of the human kinome and phosphatome by mass cytometry reveals overexpression-induced effects on cancer-related signaling. Molecular Cell, 74(5):1086–1102.e5, 2019.
  4. [4]A. Prahallad, C. Sun, S. Huang, F. Di Nicolantonio, R. Salazar, D. Zecchin, R. L. Beijersbergen, A. Bardelli, and R. Bernards. Unresponsiveness of colon cancer to BRAF(V600E) inhibition through feedback activation of EGFR. Nature, 483(7387):100–103, 2012.DOI
  5. [5]S. Chandarlapaty. Negative feedback and adaptive resistance to the targeted therapy of cancer. Cancer Discovery, 2(4):311–319, 2012.DOI
  6. [6]F. Fröhlich, T. Kessler, D. Weindl, A. Shadrin, L. Schmiester, H. Hache, A. Muradyan, M. Schütte, J.-H. Lim, M. Heinig, F. J. Theis, H. Lehrach, C. Wierling, B. Lange, and J. Hasenauer. Efficient parameter estimation enables the prediction of drug response using a mechanistic pan-cancer pathway model. Cell Systems, 7(6):567–579.e6, 2018.
  7. [7]A. Raue, C. Kreutz, T. Maiwald, J. Bachmann, M. Schilling, U. Klingmüller, and J. Timmer. Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics, 25(15):1923–1929, 2009.DOI
  8. [8]T. J. Snowden, P. H. van der Graaf, and M. J. Tindall. Methods of model reduction for large-scale biological systems: a survey of current methods and trends. Bulletin of Mathematical Biology, 79(7):1449–1486, 2017.DOI
  9. [9]M. Schmidt and H. Lipson. Distilling free-form natural laws from experimental data. Science, 324(5923):81–85, 2009.DOI
  10. [10]R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud. Neural ordinary differential equations. In Advances in Neural Information Processing Systems 31 (NeurIPS), 2018.arxiv.org/abs/1806.07366
  11. [11]B. Yuan, C. Shen, A. Luna, A. Korkut, D. S. Marks, J. Ingraham, and C. Sander. CellBox: interpretable machine learning for perturbation biology with application to the design of cancer combination therapy. Cell Systems, 12(2):128–140.e4, 2021.DOI
  12. [12]B. N. Kholodenko, A. Kiyatkin, F. J. Bruggeman, E. Sontag, H. V. Westerhoff, and J. B. Hoek. Untangling the wires: a strategy to trace functional interactions in signaling and gene networks. Proceedings of the National Academy of Sciences, 99(20):12841–12846, 2002.pubmed.ncbi.nlm.nih.gov/12242336
  13. [13]L. A. Segel and M. Slemrod. The quasi-steady-state assumption: a case study in perturbation. SIAM Review, 31(3):446–477, 1989.DOI
  14. [14]A. Goeke, S. Walcher, and E. Zerz. Classical quasi-steady state reduction—a mathematical characterization. Physica D, 345:11–26, 2017.DOI
  15. [15]M. S. Okino and M. L. Mavrovouniotis. Simplification of mathematical models of chemical reaction systems. Chemical Reviews, 98(2):391–408, 1998.pubmed.ncbi.nlm.nih.gov/11848905
  16. [16]T. Maiwald, H. Hass, B. Steiert, J. Vanlier, R. Engesser, A. Raue, F. Kipkeew, H. H. Bock, D. Kaschek, C. Kreutz, and J. Timmer. Driving the model to its limit: profile likelihood based model reduction. PLoS ONE, 11(9):e0162366, 2016.DOI
  17. [17]M. K. Transtrum and P. Qiu. Model reduction by manifold boundaries. Physical Review Letters, 113(9):098701, 2014.pubmed.ncbi.nlm.nih.gov/25216014
  18. [18]B. C. Moore. Principal component analysis in linear systems: controllability, observability, and model reduction. IEEE Transactions on Automatic Control, 26(1):17–32, 1981.DOI
  19. [19]W. Liebermeister, U. Baur, and E. Klipp. Biochemical network models simplified by balanced truncation. FEBS Journal, 272(16):4034–4043, 2005.pubmed.ncbi.nlm.nih.gov/16098187
  20. [20]M. Budišić, R. Mohr, and I. Mezić. Applied Koopmanism. Chaos, 22(4):047510, 2012.DOI
  21. [21]C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ramadhan, and A. Edelman. Universal differential equations for scientific machine learning. arXiv preprint arXiv:2001.04385, 2020.arxiv.org/abs/2001.04385
  22. [22]M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.DOI
  23. [23]A. Yazdani, L. Lu, M. Raissi, and G. E. Karniadakis. Systems biology informed deep learning for inferring parameters and hidden dynamics. PLoS Computational Biology, 16(11):e1007575, 2020.DOI
  24. [24]H. Aliee, T. Richter, M. Solonin, I. Ibarra, F. Theis, and N. Kilbertus. Sparsity in continuous-depth neural networks. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022.arxiv.org/abs/2210.14672
  25. [25]H. Aliee, F. J. Theis, and N. Kilbertus. Beyond predictions in neural ODEs: identification and interventions. arXiv preprint arXiv:2106.12430, 2021.arxiv.org/abs/2106.12430
  26. [26]W. Ji, B. Yuan, C. Shen, A. Regev, C. Sander, and S. Deng. Inference of cell dynamics on perturbation data using adjoint sensitivity. In ICLR SimDL Workshop (Deep Learning for Simulation), 2021.arxiv.org/abs/2104.06467
  27. [27]B. Klinger, A. Sieber, R. Fritsche-Guenther, F. Witzel, L. Berry, D. Schumacher, Y. Yan, P. Durek, M. Merchant, R. Schäfer, C. Sers, and N. Blüthgen. Network quantification of EGFR signaling unveils potential for targeted combination therapy. Molecular Systems Biology, 9:673, 2013.DOI
  28. [28]Y. Bengio, N. Léonard, and A. Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.arxiv.org/abs/1308.3432
  29. [31]D. P. Kingma and M. Welling. Auto-encoding variational Bayes. In International Conference on Learning Representations (ICLR), 2014.arxiv.org/abs/1312.6114
  30. [32]K. Sohn, H. Lee, and X. Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems 28 (NIPS), pages 3483–3491, 2015.
  31. [33]Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning (ICML), PMLR 37:1180–1189, 2015.arxiv.org/abs/1409.7495
  32. [34]O. S. Rukhlenko, M. Halasz, N. Rauch, V. Zhernokov, T. Prince, K. Wynne, S. Maher, E. Kashdan, K. MacLeod, N. O. Carragher, W. Kolch, and B. N. Kholodenko. Control of cell state transitions. Nature, 609(7929):975–985, 2022.DOI
  33. [35]H. Mori. Transport, collective motion, and Brownian motion. Progress of Theoretical Physics, 33(3):423–455, 1965.DOI
  34. [36]R. Zwanzig. Memory effects in irreversible thermodynamics. Physical Review, 124(4):983–992, 1961.
  35. [37]A. Gretton, O. Bousquet, A. Smola, and B. Schölkopf. Measuring statistical dependence with Hilbert–Schmidt norms. In Algorithmic Learning Theory (ALT), pages 63–77, 2005.
  36. [38]A. J. Chorin, O. H. Hald, and R. Kupferman. Optimal prediction with memory. Physica D, 166(3–4):239–257, 2002.DOI
  37. [39]C. Ma, J. Wang, and W. E. Model reduction with memory and the machine learning of dynamical systems. Communications in Computational Physics, 25(4):947–962, 2019.arxiv.org/abs/1808.04258
  38. [40]A. Gupta and P. F. J. Lermusiaux. Neural closure models for dynamical systems. Proceedings of the Royal Society A, 477(2252):20201004, 2021.DOI
  39. [41]S. L. Brunton, J. L. Proctor, and J. N. Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016.DOI
  40. [42]D. P. Kingma and J. Ba. Adam: a method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015.arxiv.org/abs/1412.6980
  41. [43]D. Türei, T. Korcsmáros, and J. Saez-Rodriguez. OmniPath: guidelines and gateway for literature-curated signaling pathway resources. Nature Methods, 13(12):966–967, 2016.DOI
  42. [44]C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. MIT Press, 2006.DOI

Paper details

vlnc:W_986vcb95
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

Not recorded for this paper

View statement
Looking for related papers…
Contents