Introduction to episode: sparse estimation in ordinary differential equation systems

Numerical solver types

When using numsolve or the exact estimation procedures rodeo (described later), a numerical solver is employed. By default, when creating an ode object through its subclasses, the ODE system is given the Runge-Kutta-Fehlberg scheme of order 4/5. Other solver types are available, all of which are embedded pair solvers, a class of very accurate explicit ODE solver. You create them via solver and specify them for your ODE system when creating them:

solver("rk23")

## Solver type: Runge-Kutta of order 2/3 
##  Control parameters
##  ---------------------------------- 
##  Maximum no. steps           100
##  Tolerance level             1e-06
##  Initial step discretisation 1e-04

p <- plk(A, s = solver("rk23")) 

Additional arguments passed to solver include control parameters for the embedded pair solver.

Data format

Data is always a n-x-(d+1) matrix y, where d is the number of coordinates in the ODE system and n is the number of observations. The first column must be the time points at which the observations are collected. The remaining columns represent the observed coordinates. Missing values are marked with NA. Whole coordinates are allowed to be unobserved, i.e. latent, but extra care must be taken. See section on “Latent coordinates” for details.

Regularisation

The loss function consists of two parameter arguments (three for ratmak). The first is the initial state x0 and the remaining are the actual parameters. All parameter arguments can regularised, meaning a penalty function is added to the loss function. To specify a regulariation use reg. This function/object works just as solver in that you specify them when you create the ode object:

reg("elnet")

## Penalty: Elastic Net 
##  Control parameters
##  ---------------------------------
##  Maximal no. of steps      200
##  Steps per cycle           20
##  Maximal backtracking      50
##  Tolerance level           1e-05
##  Initial step length       0.1
##  Minimal step length       1e-10
##  Step length factor        0.5
## 

m <- mak(A, B, r = reg("elnet")) 

You can specify the reg object for each parameter argument seperately, including the initial state.

The following penalty functions are implemented:

• Lasso/l1 penalty, "l1"
• Ridge/l2 penalty, "l2"
• Elastic net penalty, "elnet"
• SCAD penalty, "scad"
• MCP penalty, "mcp"
• No penalty, "none"

Besides penalty type you can also specify (among others):

• Box constraints on the parameter via lower and upper
• Penalty weights via penalty_factor
• Different scales for the parameter coordinates via scales
• Various control parameters for optimisation

Including multiple experiments

The data may arise from multiple experiments performed on the same system. These experiments may also be refered to as “contexts” or “environments” depending on the scientific field. It is important to distinguish different experiments, as the system may have different initialisations or have been modified or intervened upon. To distinguish the experiments we use the time column, i.e., the first column of y. The convention is that a new experiment starts whenever time decreases. This is a natural way to define it, as time 0 often marks the beginning of an experiment. Thus, if you have s experiments, make sure that they all start at 0 and the observations are ordered by ascending time. Then y is supplied as the experiments stacked on top of each other. The time column of y now has a total of s-1 decreases.

Below we generate data from another experiment, where the reverse enzyme binding is inhibited:

# Generate intervened data with different initial state
y_int <- numsolve(m, time = Time, x0 = x0 + 1, param = k * c(1, 0, 1))  
y_int[, -1] <- y_int[, -1] + matrix(rnorm(length(y_int[,-1]), sd = .1), nrow = nrow(y_int))
y2 <- rbind(y, y_int)

# Create optimisation object with data from original system and intervened system
op2 <- opt(y2, nlambda = 10)

When optimising the loss function, each experiment is given its own initial state. However, if you know the mode-of-action of interventions in each experiment, you need to include it through contexts, which is an argument in the reg function.

contexts is a p-x-s matrix, where each row represents a parameter coordinate and each column represents the context. In the loss function the effective parameter used in context l is the coordinate wise product of the lth column in contexts and the estimatable baseline parameter. For instance in the data set y2, one would supply the following contexts:

# First column scales the parameter in the original system, the second in the intervened system
m2 <- mak(A, B, r = reg(contexts = cbind(1, c(1, 0, 1))))

Exact estimation

For doing exact estimation you use rodeo on your ode object:

rod <- rodeo(m2, op2, x0 = NULL, params = NULL)
rod$params$rate

## 3 x 10 sparse Matrix of class "dgCMatrix"
##                                                                   
## [1,] . 0.05040893 0.1195512 0.1953322 0.5381662 0.8262499 1.063134
## [2,] . .          .         .         .         .         .       
## [3,] . .          .         .         0.9337910 1.3535604 1.506730
##                                
## [1,] 1.254414 1.498628 1.748778
## [2,] .        0.621268 1.373936
## [3,] 1.554898 1.550173 1.539205

Note that parameter initialisation are set to 0 if explicitly set to NULL. Similarly, the initial state values are set to the first observations from each context, if explictly set to NULL.

Integral matching estimation

Alternatively you can use approximate integral matching estimation, which is faster and less likely to get stuck in local minima. You use it via aim:

a <- aim(m2, op2)
a$params$rate

## 3 x 10 sparse Matrix of class "dgCMatrix"

##    [[ suppressing 10 column names 's0', 's1', 's2' ... ]]

##                                                                        
## V1 . 0.69360617 1.0880222 1.2770790 1.345030 1.369423 1.378199 1.381347
## V2 . .          0.3081512 0.9898034 1.234672 1.322743 1.354325 1.365628
## V3 . 0.09867549 1.0314000 1.2771862 1.365549 1.397263 1.408685 1.412795
##                     
## V1 1.382475 1.382839
## V2 1.369805 1.371141
## V3 1.414243 1.414764

Note that integral matching relies on a non-parametric estimate of the trajectory, which you supply through the argument x. If not explicitly supplied (as above), the system uses linear interpolation of data. If you want to fiddle around with integral matching on your own, then study the two functions: imd and numint.

If you are not satisfied with the approximate estimates from aim, you can pass them to rodeo as initialisations for, say a non-regularised optimisation:

# Change regularisation type to "none"
a$o$rs$rate$reg_type <- "none"
rod <- rodeo(a)
rod$params$rate

## 3 x 10 sparse Matrix of class "dgCMatrix"
##                                                                    
## [1,] 2.028109 2.028109 2.028109 2.028109 2.028109 2.028109 2.028109
## [2,] 2.156663 2.156663 2.156663 2.156663 2.156663 2.156663 2.156663
## [3,] 1.528525 1.528525 1.528525 1.528525 1.528525 1.528525 1.528525
##                                
## [1,] 2.028109 2.028109 2.028109
## [2,] 2.156663 2.156663 2.156663
## [3,] 1.528525 1.528525 1.528525

All of the above estimates should be held against the true rate parameters:

matrix(k, ncol = 1)

##      [,1]
## [1,] 2.10
## [2,] 2.25
## [3,] 1.50

Latent coordinates

If some coordinates are completely unobserved, then rodeo still works. But if you wish to use aim you must supply a non-parametric estimate of the trajectory through the x argument to aim, including the latent coordinates.