Matching Models Across Abstraction Levels with Gaussian Processes

Giulio Caravagna, Luca Bortolussi, Guido Sanguinetti

TL;DR

This paper introduces a novel machine learning framework to reconcile predictions from computational models of the same system that operate at different levels of abstraction. By learning a "correction map" using Gaussian Processes, the method allows a simpler, reduced model to accurately mimic the behavior of a more complex, detailed model, enabling efficient analysis and simulation.

What Problem Does This Paper Address?

In fields like systems biology, researchers often create multiple models of the same phenomenon, each at a different level of detail or abstraction. While these models might qualitatively agree, it's often unclear if their quantitative predictions align, especially across different parameter settings. This discrepancy makes it difficult to trust the predictions of simplified models or to use them as reliable surrogates for computationally expensive detailed models. The paper tackles the challenge of automatically bridging this gap, ensuring that a reduced model's output can be quantitatively corrected to match that of a more complex model.

Methods

The core of the proposed methodology is a correction map, which is essentially a learned function that adjusts the output of a reduced model ($m$) to match the statistics of a more detailed model ($M$). This correction map is framed as a non-parametric regression problem solved using Gaussian Processes (GPs).

Here's a breakdown:

  • Data Generation: The process starts by sampling a set of parameters ($\theta_m$) for the reduced model and then simulating both the reduced model ($m$) and the detailed model ($M$) using these parameters. The outputs (statistics) from both models are collected.
  • Correction Map Definition: The correction map, $M_\eta$, is defined to reconcile a specific statistic $\eta$. It aims to satisfy $M_\eta(\theta_m) \approx E_{\theta_f}[M_\eta(\theta_m; \theta_f)] - m_\eta(\theta_m)$, where $E_{\theta_f}[\cdot]$ denotes expectation over the free parameters ($\theta_f$) of the detailed model $M$.
  • Gaussian Process Regression: A GP is used to learn this correction map. GPs are powerful tools for regression that provide not only predictions but also uncertainty estimates. The paper explores different ways to handle the variance in the data, leading to heteroscedastic regression techniques (Estimator 1, 2, and 3). This is crucial because the discrepancy between models can vary depending on the parameters.
  • Marginalization: A key challenge is dealing with the "free parameters" of the detailed model ($M$) that are not present in the reduced model ($m$). The paper discusses approximating the expectation over these free parameters, either by direct sampling or by assuming their distribution is concentrated around a mode.

What's novel is the application of GPs not just for standard emulation, but specifically for learning a correction to reconcile models at different abstraction levels, explicitly accounting for the varying discrepancy (heteroscedasticity).

Datasets

The paper does not introduce new benchmark datasets. Instead, it uses two biological examples for proof-of-principle demonstrations:

  1. Enzymatic Reaction (Henri-Michaelis-Menten kinetics): Comparing a full model of irreversible canonical enzyme reaction with its simplified Henri-Michaelis-Menten (QSSA) form.
  2. Protein Translation Network: Comparing a detailed gene switching model with a reduced model where transcription and mRNA decay are omitted, relying on time-scale separation.

For each example, data is generated through simulations of the respective models.

Key Insights and Ideas

  • Bridging the Abstraction Gap: The most significant insight is the development of a statistically principled method to quantify and correct the discrepancies between models operating at different levels of abstraction. This moves beyond qualitative agreement to quantitative reconciliation.
  • Correction as a Regression Task: Framing the problem as learning a correction map using Gaussian Processes is elegant. It leverages the power of GPs for flexible, uncertainty-aware function approximation.
  • Heteroscedasticity is Key: The realization that the discrepancy between models is often not uniform but parameter-dependent (heteroscedastic) is crucial. Accounting for this varying variance leads to more accurate and robust corrections.
  • Mechanistic Insight: The learned correction map can reveal where and why simplified models deviate from detailed ones, offering insights into the model structure and the impact of specific simplifications.
  • Efficient Surrogate Models: The corrected reduced model can serve as a computationally efficient surrogate for the complex model, enabling faster inference, parameter estimation, and simulation.

Key Results

The paper demonstrates the effectiveness of its approach on the two biological examples. While specific quantitative metrics like "accuracy" are not presented as single numbers, the figures show:

  • Successful Correction: The learned correction maps (e.g., Figure 4, 6, 7) effectively adjust the predictions of the reduced models to closely match the statistics of the detailed models.
  • Variance Capture: The heteroscedastic regression methods (point-wise $\sigma$-estimator) are shown to better capture the variance of the correction compared to fixed-variance approaches, especially in scenarios with significant parameter-dependent discrepancies (Figure 6, 7).
  • Model Equivalence: The framework provides a way to define and assess "n-bisimilarity" between models, offering a quantitative measure of their equivalence with respect to a specific statistic.

Limitations and Future Work

  • Computational Cost: The computational cost of generating data (simulations) and training the GP can still be significant, especially for high-dimensional parameter spaces of the detailed model. The paper notes that costs grow exponentially with the number of free parameters in $M$.
  • Assumptions: The method relies on the assumption that the detailed model $M$ can be simulated and that a set of common observables/state variables exist. It also assumes that the reduced model $m$ is a reasonable starting point.
  • Future Work:
    • Extending the framework to handle more complex statistics or vector-valued outputs.
  • Applying the method to more challenging biological systems, such as cardiac electrophysiology models.
  • Investigating the use of this approach for model reduction and synthesis in a more automated fashion.
  • Exploring its application in scenarios where the "complex model" is replaced by experimental data.

Summary updated

Abstract

Biological systems are often modelled at different levels of abstraction depending on the particular aims/resources of a study. Such different models often provide qualitatively concordant predictions over specific parametrisations, but it is generally unclear whether model predictions are quantitatively in agreement, and whether such agreement holds for different parametrisations. Here we present a generally applicable statistical machine learning methodology to automatically reconcile the predictions of different models across abstraction levels. Our approach is based on defining a correction map, a random function which modifies the output of a model in order to match the statistics of the output of a different model of the same system. We use two biological examples to give a proof-of-principle demonstration of the methodology, and discuss its advantages and potential further applications.
BibTeX
@inproceedings{caravagna2016matching,
  author = {Giulio Caravagna and Luca Bortolussi and Guido Sanguinetti},
  title = {{Matching Models Across Abstraction Levels with Gaussian Processes}},
  booktitle = {CMSB},
  year = {2016},
  doi = {10.1007/978-3-319-45177-0\_4},
}

Loading related content...

Figure