CoODE: Conditionally Coarse-Grained Model Discovery for Rewiring in Biochemical Networks
Organizations: Dynamics of Living Systems Laboratory, The Francis Crick Institute, London, UK · Department of Mathematics, Imperial College London, London, UK
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 (CoODE), 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, CoODE 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.
Figures & tables
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Source | Lun et al. (2019) [ 3 ] , single-cell mass cytometry |
|---|---|
| Preprocessing | GP-smoothed trajectories, GFP-binned, with a derived strength column |
| Panel | 30 phospho-markers |
| Conditions | 40 total; 38 retained: 30 single-gene overexpressions and 8 controls |
| Dropped | 2 contexts whose time grids do not match the rest (DUSP4, MAP4K2) |
| Stimulus | EGF, upstream of the network |
| Time | snapshots over 1 h post-stimulation |
| Component | Setting | Value |
|---|---|---|
| Selector | conditioning dim | 16 |
| hidden | 128 | |
| number of conditions | 38 | |
| relaxation temperature | annealed , rate , from epoch 10 | |
| Gumbel sampling | off (straight-through top- ) | |
| explicit slots | fixed (target 2 anchors) learned |
| Arm | State dim | Latent in ODE | Composition of the state |
|---|---|---|---|
| 3-state | 3 | no | target and 2 anchors only; no marker selection and no latent closure |
| 6-state | 6 | no | target, anchors, and the 3 selector-chosen markers; no latent closure |
| Co 2 ODE | 10 | yes, | the 6 markers above plus the 4-dimensional evolved inside the ODE |
| Full panel | 30 | — | all 30 markers as ODE states, context-conditioned, no selection or closure |
| Predictor | rnMAE p-S6 (%) |
|---|---|
| Global mean trajectory | 86.0 |
| Linear extrapolation from | 37.1 |
| Hold the initial value | 36.8 |
| Per-context mean trajectory | 36.7 |
| Best flat line per trajectory, fitted in hindsight | 28.4 |
| Symbolic variant, 3–4 terms per equation (Appendix H ) | 23.4 |
| Swept quantity | Grid |
|---|---|
| Reporter weighting of the trajectory loss | |
| Jacobian group-sparsity strength | |
| Model seeds |
| Resource | OmniPath [ 43 ] , directed literature-curated interactions |
|---|---|
| Access | OmniPath enzsub endpoint via the omnipath Python client, downloaded 28 April 2026 |
| Layer and filters | the 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 size | 70 directed edges over 28 of the 30 panel markers; the remaining 2 have no retained edge and are unreachable by construction |
| Marker mapping | curated 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 basis | pairwise shortest-path distance on the directed graph |
| Latent coordinates | treated as unreachable_cost (no graph identity) |
| Protocol | What is done | Reported |
|---|---|---|
| Conditioned completeness | Predict p-S6 at from . The probe is a random-forest regressor (300 trees, minimum leaf size 2) predicting all five future timepoints jointly, scored by 5-fold cross-validated 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 probe | anchors only; adding context and strength; full panel with conditioning |
| Random-triple control | Selector’s markers against random marker triples, under the completeness probe | against : 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 swap | At inference, give a trained model another context’s markers, leaving its context embedding and latent closure untouched | p-S6 rnMAE rises from to (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 cost | Split runs by whether their selection matches the most common one for that context, and compare error within each group | closure-free against with the closure; mean error against |
| Selector-variability cost | Rank-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 points | closure-free ; closure . Replicated in an independent strength-OOD grid (closure-free ; closure ). Pinning the closure arm’s selections into the closure-free model costs points against a peer seed’s, so the picks are different, not worse |
| Strength and context probe | Predict and context identity from , 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 . Categorical targets: standardised multinomial logistic regression, scored by 5-fold cross-validated accuracy | for strength; against for context identity on the encoder’s sample, – once evolved in the ODE (chance ) |
| Arm | p-S6 (%) | mean3 (%) |
|---|---|---|
| 3-state control | 15.6 | 13.8 |
| 6-state (closure-free) | 16.1 | 14.0 |
| Co 2 ODE (10-state) | 16.0 | 14.3 |
| Full panel (30-state) | 17.1 | 14.8 |
| Comparison, p-S6 | mean | Wilcoxon | Contexts favouring smaller |
|---|---|---|---|
| 3-state against Co 2 ODE | 29/38 | ||
| 6-state against Co 2 ODE | 22/38 | ||
| Co 2 ODE against full panel | 22/38 | ||
| 3-state against full panel | 29/38 |
| Model | Runs | Terms/eq. | p-S6 (%) | mean3 (%) | Target eq. non-empty |
|---|---|---|---|---|---|
| Symbolic, 3–4 terms | 62 | 3.5 | 23.4 [21.9, 27.4] | 23.5 [21.7, 26.4] | 38/38 |
| Symbolic, 4–8 terms | 45 | 4.5 | 23.2 [20.8, 30.5] | 23.3 [20.9, 28.3] | 37/38 |
| Symbolic, 8–20 terms | 36 | 11.8 | 25.5 [21.3, 32.4] | 23.9 [21.3, 30.0] | 36/38 |
| Symbolic, 20–60 terms | 45 | 23.4 | 23.3 [19.9, 25.9] | 22.2 [18.1, 24.6] | 38/38 |
| Symbolic, 60–85 terms | 17 | 85.0 | 21.2 [19.7, 21.5] | 23.4 [19.9, 23.6] | 38/38 |
| Neural, Co 2 ODE | 16.0 | 14.3 |
| Hardware | CPU only; no GPU used at any stage |
| Per run | 8 CPU cores, 24 GB RAM |
| Wall-clock, Co 2 ODE and reduced controls | min per run, CPU-hours |
| Wall-clock, full-panel control | min per run, CPU-hours |
| Reported state-dimension experiment | 270 sweep runs 29 of 30 full-panel runs (one failed to finish), CPU-hours |
| All experiments including exploratory | CPU-hours |
| Scheduler | Slurm, shared institutional CPU partition |
Dynamical Systems
Dynamical systems studies how states evolve over time, including stability, feedback, chaos, and transitions in physical, biological, and engineered systems.
See all 512 papers