Tutorial on using spatial SSAs in DiffEqJump

This tutorial shows how to use spatial solvers added to DiffEqJump in summer 2021. See the documentation for a tutorial on getting started with DiffEqJump.

Installing DiffEqJump

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:

<!– –><!– –><!– –><!– –><!– –>
....B
.....
.....
.....
A....

Suppose we have a reversible binding system described by $A+B \to C$ at rate $k_1$ and $C \to A+B$ at rate $k_2$. Further suppose that all $A$ molecules start in the lower left corner, while all $B$ molecules start in the upper right corner of a 5 by 5 grid. There are no $C$ 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)
A Cartesian grid with dimensions (5, 5)

Now we set the initial state of the simulation. It has to be a matrix with entry $(s,i)$ being the number of species $s$ at site $i$ (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
starting_state
3×25 Matrix{Int64}:
 25  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
0
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  2
5
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
0

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

tspan = (0.0, 3.0)
rates = [6.0, 0.05] # k_1 = rates[1], k_2 = rates[2]
2-element Vector{Float64}:
 6.0
 0.05

Now we can create the DiscreteProblem:

prob = DiscreteProblem(starting_state, tspan, rates)
DiscreteProblem with uType Matrix{Int64} and tType Float64. In-place: true
timespan: (0.0, 3.0)
u0: 3×25 Matrix{Int64}:
 25  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
0
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  2
5
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
0

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)
DiffEqJump.MassActionJump{Vector{Float64}, Vector{Vector{Pair{Int64, Int64}
}}, Vector{Vector{Pair{Int64, Int64}}}, Nothing}([6.0, 0.05], [[1 => 1, 2 =
> 1], [3 => 1]], [[1 => -1, 2 => -1, 3 => 1], [1 => 1, 2 => 1, 3 => -1]], n
othing)

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 $C$ cannot diffuse, while molecules $A$ and $B$ diffuse at probability per time 1 (i.e. the time of the diffusive hop is exponentially distributed with mean 1). Entry $(s,i)$ of hopping_constants is the hopping rate of species $s$ at site $i$ to any of its neighboring sites (diagonal hops are not allowed).

hopping_constants = ones(num_species, num_nodes)
hopping_constants[3, :] .= 0.0
hopping_constants
3×25 Matrix{Float64}:
 1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0  …  1.0  1.0  1.0  1.0  1.0  1.0  1
.0
 1.0  1.0  1.0  1.0  1.0  1.0  1.0  1.0     1.0  1.0  1.0  1.0  1.0  1.0  1
.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0     0.0  0.0  0.0  0.0  0.0  0.0  0
.0

We are now ready to set up the JumpProblem with the Next Subvolume Method.

alg = NSM()
jump_prob = JumpProblem(prob, alg, majumps, hopping_constants=hopping_constants, spatial_system = grid, save_positions=(true, false))
Number of constant rate jumps: 0
Number of variable rate jumps: 0
Have a mass action jump

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())
retcode: Default
Interpolation: Piecewise constant interpolation
t: 401-element Vector{Float64}:
 0.0
 0.01805156477589096
 0.021927208873289517
 0.025769109693334135
 0.03515984244149274
 0.041941982193058974
 0.052156972679011
 0.07268000778197724
 0.08516521051878839
 0.09139750379283877
 ⋮
 2.959136931928465
 2.9646237117842693
 2.9651086896415113
 2.9728489281258383
 2.9739124562005443
 2.9757635813669086
 2.9813576277964904
 2.994984074931063
 3.0
u: 401-element Vector{Matrix{Int64}}:
 [25 0 … 0 0; 0 0 … 0 25; 0 0 … 0 0]
 [24 1 … 0 0; 0 0 … 0 25; 0 0 … 0 0]
 [24 0 … 0 0; 0 0 … 0 25; 0 0 … 0 0]
 [24 0 … 0 0; 0 0 … 0 24; 0 0 … 0 0]
 [24 0 … 0 0; 0 0 … 0 23; 0 0 … 0 0]
 [24 0 … 0 0; 0 0 … 0 22; 0 0 … 0 0]
 [23 1 … 0 0; 0 0 … 0 22; 0 0 … 0 0]
 [22 2 … 0 0; 0 0 … 0 22; 0 0 … 0 0]
 [22 2 … 0 0; 0 0 … 0 23; 0 0 … 0 0]
 [21 2 … 0 0; 0 0 … 0 23; 0 0 … 0 0]
 ⋮
 [1 4 … 0 0; 0 0 … 1 0; 0 0 … 0 0]
 [1 4 … 0 0; 0 0 … 1 0; 0 0 … 0 0]
 [2 3 … 0 0; 0 0 … 1 0; 0 0 … 0 0]
 [2 3 … 0 0; 0 0 … 2 0; 0 0 … 0 0]
 [2 3 … 0 0; 0 0 … 2 0; 0 0 … 0 0]
 [2 3 … 0 0; 0 0 … 2 0; 0 0 … 0 0]
 [2 3 … 0 0; 0 0 … 2 1; 0 0 … 0 0]
 [2 2 … 0 0; 0 0 … 2 1; 0 0 … 0 0]
 [2 2 … 0 0; 0 0 … 2 1; 0 0 … 0 0]

Animation

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

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)
        else
            randsx = zeros(num_molecules)
            randsy = zeros(num_molecules)
        end
        for k in 1:num_molecules
            push!(species_seriess_x[species], x*h - h/4 + 0.5h*randsx[k])
            push!(species_seriess_y[species], y*h - h/4 + 0.5h*randsy[k])
        end
    end
    for species in 1:num_species
        scatter!(plt, species_seriess_x[species], species_seriess_y[species], label = labels[species], marker = 6)
    end
    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
end

"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)
    end
    anim
end
# animate
anim=animate_2d(solution, 5, species_labels = ["A", "B", "C"], title = "A + B <--> C", verbose = false)
fps = 5
gif(anim, fps = fps)
Plots.AnimatedGif("/var/lib/buildkite-agent/builds/7-amdci4-julia-csail-mit
-edu/julialang/scimltutorials-dot-jl/markdown/jumps/tmp.gif")

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 $D_{s,i}, D_{s,i,j}, D_s * L_{i,j}$, and $D_{s,i} * L_{i,j}$, where $s$ denotes the species, $i$ – the source site, and $j$ – the destination. The supported solvers are NSM, DirectCRDirect and any of the standard non-spatial solvers.

Topology

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
num_nodes = prod(dims)
grid = CartesianGrid(dims)
A Cartesian grid with dimensions (2, 3, 4)

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
{5, 5} directed simple Int64 graph

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 $D_{s,i,j}$ – 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 $s$ at site $i$ to neighbor at index $j$. 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
        end
    end
end

To pass in hopping_constants of form $D_s * L_{i,j}$ we need two vectors – one for $D_s$ and one for $L_{i,j}$. 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))
end
hopping_constants=Pair(species_hop_constants, site_hop_constants)
[1.0, 1.0, 1.0] => [[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.
0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0, 1
.0]  …  [1.0, 1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.
0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0,
 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]

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

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

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))
end
hopping_constants=Pair(species_hop_constants, site_hop_constants)
[1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0; 1.0 1.0 … 1.0 1.0] => [[1.0, 1.0, 1.
0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0,
 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1
.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0, 1.0]  …  [1.0, 1.0, 1.0, 1.0, 1.0]
, [1.0, 1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1
.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0],
 [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]

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.

Solvers

There are currently two specialized "spatial" solvers: NSM and DirectCRDirect. The former stands for Next Subvolume Method [1]. 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.

References

  • 1Elf, Johan and Ehrenberg, Mäns. “Spontaneous separation of bi-stable biochemical systems into spatial domains of opposite phases”. In: Systems biology 1.2 (2004), pp. 230–236.