Mathematical Modeling in Systems Biology: ODEs, PDEs, and Data

ODE models · PDE models · Parameter estimation · Bifurcation analysis · Hybrid mechanistic–data-driven modeling

Published 2026 · Last updated 2026-09-27 · Hailu Zhou

Systems biology began with a declaration that biology should be studied as a system of interacting components rather than a list of parts (Kitano 2002). Twenty years on, the mathematics that delivers on that promise is still largely differential equations — and that is a feature, not a failure. Mechanistic models encode what we know about a system's causal wiring, and that is precisely what purely data-driven models cannot provide. The interesting frontier is not replacement but coupling: mechanistic structure plus learned components.

Why mechanistic models still matter

A neural network trained on gene expression can predict an outcome but cannot tell you what happens when you knock out a kinase it has never seen inhibited. An ODE model with that kinase in it can, provided its parameters are constrained. Mechanistic models are extrapolative and interpretable in a way that black boxes are not, and in biology — where interventions are the whole point — extrapolation is the requirement. The cost is that mechanistic models have parameters, often many, and the parameters are rarely all identifiable from the data.

ODEs: the workhorse

An ODE model writes the rate of change of each species as a function of the current state: dx/dt = f(x, p). Mass-action kinetics give polynomial right-hand sides; Michaelis–Menten and Hill functions introduce saturation and cooperativity. Classic successes include the Hodgkin–Huxley model of the action potential, the cell-cycle models of Tyson and Novak, and the circadian clock models of Goldbeter. Software ecosystems — COPASI, tellurium, BioModels — have made it routine to assemble and share such models. The key modeling decision is not the solver but the level of abstraction: a coarse-grained model with ten species and well-constrained parameters beats a hundred-species model you cannot fit.

PDEs: when space matters

Spatial organization — morphogen gradients, tumor growth, biofilm formation, intracellular transport — requires reaction–diffusion equations, where concentration depends on both time and position. Turing's 1952 paper showed that diffusion coupled with reaction can produce spontaneous spatial patterns, and this remains the canonical explanation for many developmental prepatterns. PDE models add boundary conditions and diffusion coefficients as parameters, and their numerical solution (finite differences, finite elements, method of lines) is more demanding than ODE integration. Multiscale models that couple PDE fields to ODE or agent-based descriptions of cells are the current state of the art.

Parameter estimation: the actual hard part

Model structure is the fun part; parameter estimation is where projects die. The challenges are well catalogued (Gutenkunst et al. 2007): sloppy models, where a few parameter combinations are tightly constrained and most directions are nearly flat; structural and practical non-identifiability, where different parameter sets produce identical outputs; and noisy, sparse, absolute-unit-ambiguous measurements. Practical approaches:

The honest output of a fitting exercise is a confidence region, not a parameter table.

Bifurcation analysis: understanding qualitative behavior

Some of the most valuable results from systems biology are not numbers but structure: does the system have a stable steady state, bistability, or limit-cycle oscillations, and how does that change as a parameter varies? Bifurcation analysis answers exactly this by tracking equilibria and their stability as a bifurcation parameter crosses critical values (saddle-node, Hopf, pitchfork). Cell-cycle models famously show that increasing a parameter can drive transitions between quiescence and proliferation; bistable switches explain differentiation decisions. Tools like AUTO and PyDSTool automate continuation and stability computation. This is analysis that a regression cannot do, because it interrogates the model's logic rather than its fit.

Combining mechanistic and data-driven approaches

The sensible modern architecture is hybrid, and there are three common patterns. Sequential: use a data-driven model to estimate inputs or missing species, then feed into a mechanistic model. Embedded: replace an unknown rate law with a neural network while keeping the known mass-balance structure — neural ODEs and universal differential equations are the canonical tools, and they preserve conservation laws that a pure network would violate. Residual: fit a mechanistic model and learn a correction to its residual with a flexible model. The guiding principle is to encode everything you know mechanistically and let the network learn only the part you do not — this both reduces data requirements and improves extrapolation.

Practical guidance

Start with the smallest model that can produce the behavior you care about. Write down every assumption and every parameter's unit. Fit, then run parameter identifiability analysis before reporting. Validate against data not used in fitting, ideally an independent perturbation. Report the model file (SBML or equivalent) so others can reproduce. The culture of systems biology is refreshingly open about this, and it is the reason models from a decade ago are still being extended rather than rebuilt. Mathematics is not a competing explanation to machine learning here — it is the scaffold that keeps machine learning anchored to biology.

References

  1. Kitano H. Systems biology: a brief overview. Science. 2002;295(5560):1662–1664.
  2. Turing AM. The chemical basis of morphogenesis. Philos Trans R Soc Lond B Biol Sci. 1952;237(641):37–72.
  3. Hodgkin AL, Huxley AF. A quantitative description of membrane current and its application to conduction and excitation in nerve. J Physiol. 1952;117(4):500–544.
  4. Goldbeter A. A model for circadian oscillations in the Drosophila period protein (PER). Proc R Soc Lond B Biol Sci. 1995;261(1362):319–324.
  5. Tyson JJ, Chen KC, Novak B. Sniffers, buzzers, toggles and blinkers: dynamics of regulatory and signaling pathways in the cell. Curr Opin Cell Biol. 2003;15(2):221–231.
  6. Gutenkunst RN, Waterfall JJ, Casey FP, Brown KS, Myers CR, Sethna JP. Universally sloppy parameter sensitivities in systems biology models. PLoS Comput Biol. 2007;3(10):e189.
  7. Raue A, Kreutz C, Maiwald T, et al. Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics. 2009;25(15):1923–1929.
  8. Chen RTQ, Rubanova Y, Bettencourt J, Duvenaud D. Neural ordinary differential equations. Adv Neural Inf Process Syst. 2018;31:6571–6583.
  9. Bhattacharya D, Homsi M, et al. Computational modeling of biological systems. Nat Rev Mol Cell Biol. 2020;21(6):325–338.
  10. Ermentrout GB, Terman DH. Mathematical Foundations of Neuroscience. Interdisciplinary Applied Mathematics, Springer. 2010;35.