Parameter Inference in Mechanistic Biological Models

38 concepts35 papers1 proposal

Mathematical models are highly effective at simulating biological processes when their parameters are known. Across systems biology, researchers routinely build ordinary differential equations (ODEs) and process-based models to describe blood cell production (Ivana Drobnjak), plant growth dynamics (Amélie Rostand-Mathieu et al.), gene regulatory networks (Pablo Meyer et al.), and complex ecosystem food webs (Neo D. Martinez et al.).

However, a fundamental bottleneck exists: while running these models forward is straightforward, running them backward—inferring the underlying physical, chemical, or biokinetic parameters from observed macroscopic data—is an incredibly challenging inverse problem. Parameters like cellular binding affinities, stem cell cycling rates, or soil microbial growth rates (Paul Collart et al.) are frequently impossible to measure directly in vivo or in the field. Instead, they must be inferred indirectly from noisy, high-level observations.

The Geometry of Ill-Posedness: Why Calibration Fails

To understand why parameter estimation is so difficult, we must look at the mathematical structure of these models. Biological systems are typically governed by non-linear differential equations that lack closed-form analytical solutions (Jérémie Guedj et al.).

More fundamentally, these models exhibit a universal property known as sloppiness. As demonstrated by James P. Sethna in Universal Sloppy Parameter Sensitivities in Systems Biology Models, parameter sensitivities in complex biological models are highly unevenly distributed. The system's behavior is tightly constrained along a few "stiff" parameter combinations, but highly insensitive along a vast majority of "sloppy" directions.

Consequently, traditional optimization algorithms that search for a single "best-fit" parameter set fail. Because many distinct parameter combinations yield identical macroscopic observations, the inverse problem possesses a neutral landscape where multiple model structures and parameter configurations can explain the same experimental time-series data (Marco Vilela et al.).

Embracing Ambiguity via Bayesian Inversion

If pinpointing a single true parameter set is impossible due to sloppiness, how can we make reliable predictions? The solution is to shift from frequentist point estimation to a Bayesian framework. Instead of finding a single vector of parameters, Bayesian inversion maps out a probability distribution over the entire parameter space, explicitly capturing the correlations and uncertainties inherent to the sloppy directions.

For example, Nikos E. Papanikolaou et al. demonstrate that combining a Bayesian framework with Optimal Experimental Design (OED) allows researchers to decouple complex biological behaviors (like predator handling time vs. digestion) and determine exactly what data volume is required for calibration. Similarly, Jean-Charles Croix et al. apply infinite-dimensional Bayesian inversion to successfully infer unobservable diffusion and self-regulation rates from noisy experimental measurements, providing a robust method for uncertainty quantification.

Bypassing the Computational Bottleneck

While Bayesian methods solve the problem of ill-posedness, they introduce a severe computational challenge. Sampling from a posterior distribution via Markov Chain Monte Carlo (MCMC) requires evaluating the underlying differential equations hundreds of thousands of times. When the model lacks an analytical solution, each evaluation requires an expensive numerical simulation.

To make this tractable, researchers use surrogate models and neural calibration:

  • Gaussian Process Emulators: L. Mihaela Păun et al. combine MCMC with Gaussian Process (GP) emulators. The GP acts as a fast mathematical surrogate, learning to predict the model's output for any given parameter set instantly, thereby accelerating uncertainty quantification by orders of magnitude.
  • Neural Parameter Calibration: Thomas Gaskin et al. integrate neural networks directly with differential equations. By training a neural network to map observed data distributions directly to parameter probability distributions, they bypass traditional iterative sampling entirely, enabling real-time calibration for large-scale multiagent models.

Tracking Hidden States and Invariants

Another powerful strategy for solving the inverse problem involves transforming raw, noisy data into invariant representations or tracking them dynamically through state-space formulations.

In neural systems governed by complex dynamics like limit cycles, standard gradient descent gets stuck or oscillates. To resolve this, Xuyang Lou et al. utilize multi-innovation algorithms that leverage both current and past data states simultaneously to smooth out the trajectory toward true parameter values. For high-dimensional biophysical neuron models, Dimitrios V. Vavoulis et al. frame parameter estimation within a self-organizing state-space framework, using adaptive sampling to infer parameters directly from noisy electrophysiological data.

Alternatively, researchers can transform data to filter out confounding variables, such as unknown initial conditions. Consider developmental biology, where mathematical models generate complex spatial patterns (like animal coat markings). Inferring the underlying chemical parameters from a final pattern is notoriously difficult because the exact outcome depends on random initial fluctuations.

To solve this, David Schnörr et al. represent Turing patterns as Resistance Distance Histograms (RDHs). Because the RDH captures the underlying topological signature of the pattern rather than its exact spatial coordinates, it remains invariant to initial conditions. This allows a machine learning model to accurately reverse-engineer the physical parameters from just a single observed image.

[Physical Parameters] ───(Forward Model + Random Noise)───► [Messy Turing Pattern]
                                                                   │
                                                       (Topological Transform)
                                                                   ▼
[Inferred Parameters] ◄───(Machine Learning Model)───────── [Invariant RDH]

In a similar vein, when dealing with noisy, high-dimensional cellular data, Anahita Mirtabatabaei et al. utilize state aggregation techniques to bridge the gap between stochastic models and flow cytometry histograms, allowing model validation even when individual kinetic parameters remain hidden. For systems with mixed continuous and discrete dynamics, Bing Liu et al. employ $\delta$-complete decision procedures to systematically synthesize valid parameter spaces for complex cardiac cell models.

Key Foundational Work

Maximum Likelihood Estimation in Dynamical Models of HIV (2007) — contributed 2 concept elements to this direction.
Universally Sloppy Parameter Sensitivities in Systems Biology Models (2007) — contributed 1 concept elements to this direction.
Modeling and estimation of dynamic EGFR pathway by data assimilation approach using time series proteomic data. (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_parameter_estimation_challenges_ad6d7730,
  title = {parameter estimation challenges},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parameter-estimation-challenges/rdr_ad6d77304f574418aca880df17dbfb0f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_parameter_estimation_challenges_ad6d7730_v1,
  title = {parameter estimation challenges},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parameter-estimation-challenges/rdr_ad6d77304f574418aca880df17dbfb0f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake