Statistical Model

This document describes Maud from a statistical point of view.

Maud’s statistical model separates the information that might be available about a metabolic network into three different kinds:

  • structural information implicit in a kinetic model

  • information contained in directly modelled experiments

  • information from other sources

The role of the statistical model is to synthesise these different sources of information, making it possible to say exactly what is known about a metabolic network after a series of experiments.

More specifically, the statistical model is a joint probability distribution \(\pi: \Theta \times Y\rightarrow [0,1]\) that defines the probability density \(\pi(\theta, y)\) of any possible combination of unknown parameters \(\theta\) and observations \(y\). This model is written explicitly as a Stan program.

Given a kinetic model and a set of observations \(y\), Maud uses Stan’s adaptive Hamiltonian Monte Carlo algorithm to draw samples from the posterior distribution \(p(\theta\mid y)\). Each draw contains a configuration of unknown parameter values. Quantitative questions about the metabolic network can be answered by interrogating the ensemble of posterior draws.

Structural Information

Maud assumes that the following structural information is known in advance:

  • The volume of each compartment and the metabolites it houses

  • The network stoichiometry, i.e. the proportions in which each reaction in the network creates and destroys metabolites and which enzymes catalyse which reactions. This information is encapsulated in a stoichiometric matrix \(S\) with a row for each metabolite-in-compartment and a column for each enzyme.

  • Which metabolites-in-compartment modify which enzymes and how

  • Which metabolites-in-compartments are ‘balanced’, i.e. their concentrations do not change when the system is in a steady state.

We refer to this structural information collectively as a kinetic model. See [LINK] for a detailed description of these from a scientific point of view.

The kinetic model defines a system of ODEs - one differential equation for each metabolite-in-compartment - with the rate of change of each metabolite-in-compartment \(m_i\) described by the following equation:

\[\frac{dm_{i}}{dt} = \sum_{j} S_{ij} v_{j}(\theta, \mathbf{m})\]

When the system is at a steady state, all the balanced metabolites-in-compartments have unchanging concentrations, so their entries in the equation above are zero.

For a given set of parameters, enzyme concentrations and initial metabolite conditions, there should be a single steady state balanced metabolite concentration and set of fluxes.

The kinetic model’s role in Maud’s statistical model is to connect latent parameters - i.e. \(\theta\) above - with measureable quantities, i.e. \(\mathbf{m}\) and \(\mathbf{v}\).

Probabilistic Model

Maud aims to implement a Bayesian probabilistic model where the joint distribution \(\pi(\theta, y)\) of unknowns and observations is factored into a measurement model or likelihood \(\pi(y \mid \theta)\) and a prior model \(\pi(\theta)\). This section explains how each of these components is constructed.


Maud represents information from experiments that measure enzyme concentrations and metabolite concentrations using the following regression model, where \(y\) is the observation and \(\hat{y}\) is the unobserved true value of the experimentally measured quantity:

\[y \sim LogNormal(\log(\hat{y}), \sigma)\]

Measurements of reaction fluxes use the following similar regression model:

\[y \sim Normal(\hat{y}, \sigma)\]

In all three cases it is assumed that for each kind of measurement the error standard deviation \(\sigma\) is known (though this number is in general different for each measured quantity in each experiment). It is the user’s responsibility to choose values that reflect the accuracy of the measurement apparatus.

The log-normal distribution was chosen to represent experimental metabolite and enzyme concentration measurements because the apparatuses used to measure these quantities typically have multiplicative errors. In other words, if the measured value is large, then the associated error is also proportionally large.

Summary statistics

It is common for experimental results to be reported in the form of a sample mean and standard deviation. It is important to note that for non-negative quantities like metabolite and enzyme concentrations these summary statistics will generally not be good values use as \(y\) and \(\sigma\) above. If possible, non-summarised measurement results should be used instead.

Relative measurements

In many realistic cases a measurement apparatus will give comparatively accurate information about the relative concentrations of some metabolites or enzymes while being comparatively uninformative as to their absolute values. While Maud currently does not support this kind of measurement, support is planned and will take the following form.


Information that does not naturally take the form of an experimental measurement can be expressed in Maud’s prior model. Maud allows users to specify independent log-normal priors for the following quantities:

  • enzyme \(kcat\) parameters

  • enzyme/metabolite-in-compartment \(km\) parameters

  • enzyme transfer constants

  • enzyme/metabolite-in-compartment inhibition and dissociation constants

  • enzyme concentrations

  • unbalanced metabolite concentrations

The distinction between balanced and unbalanced metabolites is also found in the statistical model. Information about the unbalanced metabolites can be parsed in the form of a prior, however, due to the difficulty of non-linear transformations, balanced metabolites are always evaluated as part of the model likelihood. The distinction between unbalanced and balanced becomes aparent when considering what the unbalanced metabolites represent, which is a boundary condition. These define the outcome of systems of differential equations, in the case of Maud this happens to be balanced metabolite concentrations and fluxes. And, our knowledge about the state of each condition is only conveyed through priors on the boundary conditions: * unbalanced metabolite concentrations, * enzyme concentrations, * kinetic parameters, and, * drains.

For metabolite formation energies, which can be both negative and positive numbers, Maud allows users to specific independent normal priors.

Users are encouraged to choose prior locations and scales using the method by calculating quantiles. Prior information is often easiest to ellicit in the form of qualitative statements like “it is very unlikely that \(kcat_e\) is higher than 6.8 or lower than 0.4”. Information in this form naturally translates into restrictions on the quantiles of the corresponding marginal prior distribution - for example that the prior mass for the events \(kcat_e > 6.8\) and \(kcat_e < 0.4\) should each be about 1%. The prior values can then be calculated as roughly \(\mu_{kcat_e} = 0.5003\) and \(\sigma_{kcat_e} = 0.6089\).

Maud includes convenience functions for working out priors in this way, which can be used in a python environment as follows:

In [1]: from maud.utils import get_lognormal_parameters_from_quantiles

In [2]: get_lognormal_parameters_from_quantiles(0.4, 0.01, 6.8, 0.99)
Out[2]: (0.5003159401539531, 0.608940170915830)

Information about fluxes and balanced metabolite concentrations

It is currently not possible to include non-experimental information about fluxes and steady-state concentrations of balanced metabolites.

This is due to a technical limitatation. Since fluxes and steady state metabolite concentrations are calculated from the values of other parameters by finding the solution to the ODE system, directly setting priors would introduce a bias without a compensating Jacobian adjustment. We have not found a way to introduce this Jacobian adjustment, so Maud unfortunately cannot currently represent this information.

Multivariate priors

Sometimes the non-experimental information about two parameters is not independent. For example, some linear combinations of formation energies are known within a relatively small range even though the marginal value of each component of the linear combination is not well known.

In such cases a multivariate distribution is required in order to express the available information. This functionality is not yet supported, but will be soon.