# Attractor Landscapes¶

The most common use of dynamical network models is the analysis of their attractor landscape. In many cases, the attractors are associated with some form of functionally important network state, e.g. a cell type in a gene regulatory network. Neet provides standard landscape analysis methods via the LandscapeMixin from which Network derives.

## State Transitions¶

The starting point for all of these analyses are the state transitions: where does each state of the network go upon update?

>>> s_pombe.transitions
array([  2,   2, 130, 130,   4,   0, 128, 128,   8,   0, 128, 128,  12,
0, 128, 128, 256, 256, 384, 384, 260, 256, 384, 384, 264, 256,
...
208, 208, 336, 336, 464, 464, 340, 336, 464, 464, 344, 336, 464,
464, 348, 336, 464, 464])


Each element of the resulting array is the state to which the index transitions, e.g. $$0 \mapsto 2$$, $$2 \mapsto 130$$, etc. The indices and values are, of course, encoded states. You can always decode them:

>>> for x, y in enumerate(s_pombe.transitions):
...     print(s_pombe.decode(x), '→', s_pombe.decode(y))
[0, 0, 0, 0, 0, 0, 0, 0, 0] → [0, 1, 0, 0, 0, 0, 0, 0, 0]
[1, 0, 0, 0, 0, 0, 0, 0, 0] → [0, 1, 0, 0, 0, 0, 0, 0, 0]
[0, 1, 0, 0, 0, 0, 0, 0, 0] → [0, 1, 0, 0, 0, 0, 0, 1, 0]
...
[1, 1, 1, 1, 1, 1, 1, 1, 1] → [0, 0, 0, 0, 1, 0, 1, 1, 1]


Given state transitions, the next question you might ask is how to compute sequences of state transtions — a trajectory — by applying the network update scheme recursively?

>>> s_pombe.trajectory([0, 0, 0, 0, 0, 0, 0, 0, 0], timesteps=2)
[[0, 0, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 0, 0], [0, 1, 0, 0, 0, 0, 0, 1, 0]]
>>> s_pombe.trajectory([0, 0, 0, 0, 0, 0, 0, 0, 0], timesteps=2, encode=True)
[0, 2, 130]


Notice that if you request a trajectory with $$t$$ time steps, the resulting trajectory will have $$t+1$$ elements in it; the first element is the initial state. If you want the trajectory for every state of the network, you can use the timeseries method.

>>> series = s_pombe.timeseries(2)
>>> series
array([[[0, 0, 0],
[1, 0, 0],
[0, 0, 0],
...,
[1, 0, 0],
[0, 0, 0],
[1, 0, 0]],

...

[[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
...,
[1, 1, 1],
[1, 1, 1],
[1, 1, 1]]])
>>> series.shape
(9, 512, 3)
>>> series[:, 0, :].transpose()
array([[0, 0, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 0, 0, 0, 0, 1, 0]])


The resulting $$3$$-D array is indexed by the nodes, state and timestep; in that order. For a more wholistic description of the state transitions, you can construct a landscape graph.

>>> import networkx as nx
>>> g = s_pombe.landscape_graph()
>>> len(g)
512
>>> nx.shortest_path(g, 0, 130)
[0, 2, 130]


The landscape graph, much like the network topology, can be drawn if you’ve installed pygraphviz. See Getting Started.

## Attractors and Basins¶

With the state transitions under our belt, we can start computing landscape features such as the attractors.

>>> s_pombe.attractors
array([array([76]), array([4]), array([8]), array([12]),
array([144, 110, 384]), array([68]), array([72]), array([132]),
array([136]), array([140]), array([196]), array([200]),
array([204])], dtype=object)


Each element of the resulting array is an array of states in a fixed-point attractor or limit cycle. Beyond this, you can determine which of the attractor’s basin each state is in.

>>> s_pombe.basins
array([ 0,  0,  0,  0,  1,  0,  0,  0,  2,  0,  0,  0,  3,  0,  0,  0,  0,
0,  4,  4,  0,  0,  4,  4,  0,  0,  4,  4,  0,  0,  4,  4,  4,  4,
...
0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
0,  0])
>>> s_pombe.basins[18]
4


That is, state $$18$$ is in basin $$4$$, and so is fated to land in the cycle $$\{144, 110, 384\}$$.

The LandscapeMixin provides a whole host of other properties, so check out the API Reference for the full list.

## Landscape Data¶

A key feature of the LandscapeMixin is that it tries to compute as much as it can, as efficiently as it can. For example, when the attractors are computed, the basins of all of the states, the recurrence time, etc… can all be computed at the same time. These values are

1. Computed lazily, but preemptively when you first request any of the associated property.
2. Cached in a LandscapeData object stored in the LandscapeMixin.

This means, that the attractors are computed when you request them. A second request will simply use the cached values. Similarly, you get a cached value for the basins once you’ve accessed the attractors. The following only computes the attractors once, and the basins are computed at that call:

>>> s_pombe.attractors  # may take a moment
array([...], dtype=object)
>>> s_pombe.attractors  # almost instantaneous
array([...], dtype=object)
>>> s_pombe.basins  # almost instantaneous; computed on first call to attractors.
array([...])


The order you access the properties in does not matter, so don’t worry about that.

There may be cases when you want to

1. Compute some landscape features of a network
2. Modify the network in some way
3. Compute landscape features on the new network
4. Compare the results

Because you’ve modifed the network, you will need to reset the cached landscape data. Since you are going to be comparing features before and after, you need to extract the data before you do that. This is where LandscapeMixin.landscape(), LandscapeMixin.expound() and LandscapeMixin.landscape_data come into play.

import numpy
from neet.boolean.examples import s_pombe

# Compute all of the landscape properties
s_pombe.expound()
# Get the data out
before = s_pombe.landscape_data

# Modify the network
s_pombe.thresholds = numpy.zeros(s_pombe.size)
# Reset the landscape (notice the method chaining...)
s_pombe.landscape().expound()
# Get the new data
after = s_pombe.landscape_data

# Compare before and after as you so choose


The result of LandscapeMixin.landscape_data is a LandscapeData object which has all of the landscape features cached (provided they’ve been computed):

>>> s_pombe.attractors
array([...], dtype=object)
>>> s_pombe.landscape_data
<neet.landscape.LandscapeData object at 0x...>
>>> s_pombe.landscape_data.attractors
array([...], dtype=object)