Tutorial on using spatial SSAs in DiffEqJump

This blog will show how to use the functionality I added to DiffEqJump over the summer. See the documentation for a tutorial on getting started with DiffEqJump.

Installing DiffEqJump

To use DiffEqJump you will need to have julia installed. Once in REPL, do ] add DiffEqJump. After the installation finishes, you will be able to use all the functionality described below.

Reversible binding model on a grid

A 5 by 5 Cartesian grid:


Suppose we have a reversible binding system described by A+Bundefinedk2k1CA+B \xleftrightarrow[k_2]{k_1} C, where k1k_1 is the forward rate and k2k_2 is the backward rate. Further suppose that all AA molecules start in the lower left corner, while all BB molecules start in the upper right corner of a 5 by 5 grid. There are no CC molecules at the start.

We first create the grid:

using DiffEqJump
dims = (5,5)
num_nodes = prod(dims) # number of sites
grid = CartesianGrid(dims) # or use LightGraphs.grid(dims)

Now we set the initial state of the simulation. It has to be a matrix with entry (s,i)(s,i) being the number of species ss at site ii (with the standard column-major ordering of the grid).

num_species = 3
starting_state = zeros(Int, num_species, num_nodes)
starting_state[1,1] = 25
starting_state[2,end] = 25

We now set the time-span of the simulation and the reaction rates. These can be chosen arbitrarily.

tspan = (0.0, 2.0)
rates = [3.0, 0.05] # k_1 = rates[1], k_2 = rates[2]

Now we can create the DiscreteProblem:

prob = DiscreteProblem(starting_state, tspan, rates)

Since both reactions are massaction reactions, we put them together in a MassActionJump. In order to do that we create two stoichiometry vectors. The net stoichiometry vector describes which molecules change in number and how much after each reaction; for example, [1 => -1] is the first molecule disappearing. The reaction stoichiometry vector describes what the reactants of each reaction are; for example, [1 => 1, 2 => 1] would mean that the reactants are one molecule of type 1 and one molecule of type 2.

netstoch = [[1 => -1, 2 => -1, 3 => 1],[1 => 1, 2 => 1, 3 => -1]]
reactstoch = [[1 => 1, 2 => 1],[3 => 1]]
majumps = MassActionJump(rates, reactstoch, netstoch)

The last thing to set up is the hopping constants – the probability per time of an individual molecule of each species hopping from one site to another site. In practice this parameter, as well as reaction rates, are obtained empirically. Suppose that molecule CC cannot diffuse, while molecules AA and BB diffuse at probability per time 1 (i.e. the time of the diffusive hop is exponentially distributed with mean 1). Entry (s,i)(s,i) of hopping_constants is the hopping rate of species ss at site ii to any of its neighboring sites (diagonal hops are not allowed).

hopping_constants = ones(num_species, num_nodes)
hopping_constants[3, :] .= 0.0

We are now ready to set up the JumpProblem with the next subvolume method.

alg = NSM() # Next Subvolume Method. Can also use DirectCRonDirect
jump_prob = JumpProblem(prob, alg, majumps, hopping_constants=hopping_constants, spatial_system = grid, save_positions=(true, false))

The save_positions keyword tells the solver to save the positions just before the jumps. To solve the jump problem do

solution = solve(jump_prob, SSAStepper())

Visualizing solutions of spatial jump problems is best done with animations.

This animation was produced by this script.

Making changes to the model

Now suppose we want to make some changes to the reversible binding model above. There are three "dimensions" that can be changed: the topology of the system, the structure of hopping rates and the solver. The supported topologies are CartesianGrid – used above, and any AbstractGraph from LightGraphs. The supported forms of hopping rates are Ds,i,Ds,i,j,DsLi,jD_{s,i}, D_{s,i,j}, D_s * L_{i,j}, and Ds,iLi,jD_{s,i} * L_{i,j}, where ss denotes the species, ii – the source site, and jj – the destination. The supported solvers are NSM, DirectCRDirect and any of the standard non-spatial solvers.


If our mesh is a grid (1D, 2D and 3D are supported), we can create the mesh as follows.

dims = (2,3,4) # can pass in a 1-Tuple, a 2-Tuple or a 3-Tuple
grid = CartesianGrid(dims)

The interface is the same as for LightGraphs.grid. If we want to use an unstructured mesh, we can simply use any AbstractGraph from LightGraphs as follows:

using LightGraphs
graph = cycle_digraph(5) # directed cyclic graph on 5 nodes

Now either graph or grid can be used as spatial_system in creation of the JumpProblem.

Hopping rates

The most general form of hopping rates that is supported is Ds,i,jD_{s,i,j} – each (species, source, destination) triple gets its own independent hopping rate. To use this, hopping_constants must be of type Matrix{Vector{F}} where F <: Number (usually F is Float64) with hopping_constants[s,i][j] being the hopping rate of species ss at site ii to neighbor at index jj. Note that neighbors are in ascending order, like in LightGraphs. Here is an example where only hopping up and left is allowed.

hopping_constants = Matrix{Vector{Float64}}(undef, num_species, num_nodes)
for ci in CartesianIndices(hopping_constants)
    (species, site) = Tuple(ci)
    hopping_constants[species, site] = zeros(outdegree(grid, site))
    for (n, nb) in enumerate(neighbors(grid, site))
        if nb < site
            hopping_constants[species, site][n] = 1.0

To pass in hopping_constants of form DsLi,jD_s * L_{i,j} we need two vectors – one for DsD_s and one for Li,jL_{i,j}. Here is an example.

species_hop_constants = ones(num_species)
site_hop_constants = Vector{Vector{Float64}}(undef, num_nodes)
for site in 1:num_nodes
    site_hop_constants[site] = ones(outdegree(grid, site))
hopping_constants=Pair(species_hop_constants, site_hop_constants)

We must combine both vectors into a pair as in the last line above.

Finally, to use in hopping_constants of form Ds,iLi,jD_{s,i} * L_{i,j} we construct a matrix instead of a vector for Ds,jD_{s,j}.

species_hop_constants = ones(num_species, num_nodes)
site_hop_constants = Vector{Vector{Float64}}(undef, num_nodes)
for site in 1:num_nodes
    site_hop_constants[site] = ones(outdegree(grid, site))
hopping_constants=Pair(species_hop_constants, site_hop_constants)

We can use either of the four versions of hopping_constants to construct a JumpProblem with the same syntax as in the original example. The different forms of hopping rates are supported not only for convenience but also for better memory usage and performance. So it is recommended that the most specialized form of hopping rates is used.


There are currently two specialized "spatial" solvers: NSM and DirectCRDirect. The former stands for Next Subvolume Method and was previously described here. The latter employs Composition-Rejection to sample the next site to fire, similar to the ordinary DirectCR method. For larger networks DirectCRDirect is expected to be faster. Both methods can be used interchangeably.

Additionally, all standard solvers are supported as well, although they are expected to use more memory and be slower. They "flatten" the problem, i.e. turn all hops into reactions, resulting in a much larger system. For example, to use the Next Reaction Method (NRM), simply pass in NRM() instead of NSM() in the construction of the JumpProblem. Importantly, you must pass in hopping_constants in the D_{s,i,j} or D_{s,i} form to use any of the non-specialized solvers.

Animation script

using Plots
is_static(spec) = (spec == 3) # true if spec does not hop
"get frame k"
function get_frame(k, sol, linear_size, labels, title)
    num_species = length(labels)
    h = 1/linear_size
    t = sol.t[k]
    state = sol.u[k]
    xlim=(0,1+3h/2); ylim=(0,1+3h/2);
    plt = plot(xlim=xlim, ylim=ylim, title = "$title, $(round(t, sigdigits=3)) seconds")

    species_seriess_x = [[] for i in 1:num_species]
    species_seriess_y = [[] for i in 1:num_species]
    CI = CartesianIndices((linear_size, linear_size))
    for ci in CartesianIndices(state)
        species, site = Tuple(ci)
        x,y = Tuple(CI[site])
        num_molecules = state[ci]
        sizehint!(species_seriess_x[species], num_molecules)
        sizehint!(species_seriess_y[species], num_molecules)
        if !is_static(species)
            randsx = rand(num_molecules)
            randsy = rand(num_molecules)
            randsx = zeros(num_molecules)
            randsy = zeros(num_molecules)
        for k in 1:num_molecules
            push!(species_seriess_x[species], x*h - h/2 + h*randsx[k])
            push!(species_seriess_y[species], y*h - h/2 + h*randsy[k])
    for species in 1:num_species
        scatter!(plt, species_seriess_x[species], species_seriess_y[species], label = labels[species], marker = 6)
    xticks!(plt, range(xlim...,length = linear_size+1))
    yticks!(plt, range(ylim...,length = linear_size+1))
    xgrid!(plt, 1, 0.7)
    ygrid!(plt, 1, 0.7)
    return plt

"make an animation of solution sol in 2 dimensions"
function animate_2d(sol, linear_size; species_labels, title, verbose = true)
    num_frames = length(sol.t)
    anim = @animate for k=1:num_frames
        verbose && println("Making frame $k")
        get_frame(k, sol, linear_size, species_labels, title)
# animate
anim=animate_2d(solution, 5, species_labels = ["A", "B", "C"], title = "A + B <--> C", verbose = false)
fps = 5
name = "ABC_anim_$(length(solution.u))frames_$(fps)fps.gif"
path = joinpath(name)
gif(anim, path, fps = fps)