1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
|
<html>
<head>
<title>Solvers</title>
</head>
<body>
<h1>Solvers</h1>
<p>
In this section we enumerate and briefly describe each of Cain's solvers.
See the <a href="Simulation.htm">Simulation Methods</a> section for more details
and derivations. In the <a href="UserMethod.htm">method editor</a> the
solvers are grouped according to four heirarchical categories: time
dependence, simulation output, algorithm, and option. One choose a
solver by selecting items from each of these in turn.
</p>
<a name="TimeDependence"><h2>Time Dependence</h2></a>
<p>
The first category differentiates between time homogeneous and time
inhomogeneous systems. In a time homogenous problem the reaction
propensities do not have explicit time dependence. Time inhomogenous
problems typically arise when the volume or temperature have a
prescribed time dependence. This implies that the reaction rates will
explicitly depend on time. Cain has many solvers for time homogeneous
problems, but few for time inhomogenous ones. While conceptually very
similar, one uses different algorithms for the two cases. Thus in
choosing a solver one first selects the time dependence category.
</p>
<a name="SimulationOutput"><h2>Simulation Output</h2></a>
<p>
Next one selects the desired simulation output. The various kinds of output
are described in the <a href="UserMethod.htm">Method Editor</a> section.
To recap, one can either collect time series or histogram data.
For time series data one can either record the state at uniformly spaced
time frames or one can record every reaction event. Although Cain is
primarily for stochastic simulations, it does offer a few simple deterministic
solvers. These generate time series data and are grouped in a separate
category. When the species populations are recorded in histograms,
the mean and standard deviation are also recorded. This is useful because
unless the bin width is unity one can only approximately calculate these
statistics from the histogram data.
For histogram output one can either study transient or steady
state behavior. For the former one records the state at uniformly spaced
frames in time. For the latter one averages the populations over
a specified time interval. Thus, in summary there are five simulation output
categories:
<ul>
<li>
<tt>Time Series, Uniform</tt>.
Here one takes snapshots of the state of uniformly spaced frames in time.
One can use either exact methods like the
<a href="SimulationDirect.htm">direct method</a>,
<a href="SimulationFirst.htm">first reaction method</a>, and
<a href="SimulationNext.htm">next reaction method</a>,
or approximate methods such as
<a href="SimulationTau.htm">tau-leaping</a> and
<a href="SimulationHybrid.htm">hybrid direct/tau-leaping</a>.
Recently there has been much research in optimized exact methods.
(See reference [<a href="Bibliography.htm#mauch2011">Mauch 2011</a>]
for an analysis.) Thus we
offer many different formulations of the direct and next reaction methods.
You may experiment to see how the various formulations perform. However,
the default formulation has good performance for most problems.
<li>
<tt>Time Series, All Reactions</tt>.
Recording every reaction event is useful when one wants a visual
representation of the amount of noise in a system. By contrast, when one
records the state at frames one does not get a picture of the magnitude of
the random fluctuations. Note that recording every reaction is expensive,
and plotting the results is even more so. For this output option we use
the 2-D search formulation of the direct method.
<li>
<tt>Time Series, Deterministic</tt>.
Deterministic systems are modeled with a set of ordinary differential
equations. Cain uses <a href="SimulationOde.htm">numerical integration</a>
to approximately solve the system. The state is recorded at uniformly
spaced time frames.
<li>
<tt>Histograms, Transient Behavior</tt>.
If one records the state at time frames one can of course plot the
trajectories. One can also analyze the species populations at these
frames. One can compute the mean and standard deviation of the population
or one can plot the distribution of populations in a histogram. Obtaining
accurate statistics or distribution data usually requires generating many
trajectories. In this case, the state information may require a lot of
storage capacity. It is more economical to dynamically update population
statistics and a histograms of the species populations as the trajectories
are being generated. For this output category populations statistics and
histogram data is stored at uniformly spaced frames in time. To generate
the trajectories we use the 2-D search formulation of the direct method.
<li>
<tt>Histograms, Steady State</tt>.
If one records every reaction event, one can analyze the steady state
behavior of a system. One could take the time average of the species
populations to determine the mean and standard deviation of the steady
state populations. One could also record the time weighted populations
in a histogram to determine the distribution of steady state populations.
However, storing every reaction event typically requires a lot of storage.
It is more efficient to dynamically update statistics and histogram data
as the trajectories are generated. For this output option the time averaged
statistics and histogram data is collected. Again we use the 2-D search
formulation of the direct method to generate the trajectories.
</ul>
</p>
<!--CONTINUE Show little plots for each output option.-->
<a name="Algorithms"><h2>Algorithms</h2></a>
<p>
The third category in selecting a solvers is the simulation algorithm.
There are algorithms for exact and approximate stochastic simulation.
Below is a list with some implementation details.
</p>
<ul>
<li>
<tt>Direct</tt>.
Gillespie's
<a href="SimulationDirect.htm">direct method</a>
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
A reaction dependency graph
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>]
is used to minimize the cost of recomputing propensities.
Unless otherwise indicated, the sum of the propensities and associated
quantities are dynamically updated.
At each step the time increment is determined by drawing an exponential
deviate. The mean of the distribution is the sum of the propensities.
The reaction is chosen by
generating a discrete deviate whose weighted probability mass function
is the set of reaction propensities. There are many options for
generating the discrete deviate. The 2-D search formulation is the default
method.
<li>
<tt>Next Reaction</tt>.
Gibson and Bruck's
<a href="SimulationNext.htm">next reaction method</a>
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>] stores reaction
times in a priority queue. There are several options for implementing
this data structure.
<li>
<tt>First Reaction</tt>.
Gillespie's
<a href="SimulationFirst.htm">first reaction method</a>
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<li>
<tt>Tau-Leaping</tt>.
We implement explicit <a href="SimulationTau.htm">tau-leaping</a>
with the step size selection presented
in reference [<a href="Bibliography.htm#cao2006a">Cao 2006a</a>].
<li>
<tt>Hybrid Direct/Tau-Leaping</tt>.
This hybrid algorithm uses the direct method for the volatile (low
population) species and tau-leaping for the rest.
<li>
<tt>ODE, Integrate Reactions</tt>.
For <a href="SimulationOde.htm">ODE integration</a>
we use
<a href="http://en.wikipedia.org/wiki/Runge-Kutta">Runge-Kutta</a>
schemes.
</ul>
<a name="Options"><h2>Options</h2></a>
<p>
Below is a list of the available solvers grouped according to the
heirarchical categories.
Each of the solvers use sparse arrays for the state change vectors
[<a href="Bibliography.htm#li2006">Li 2006</a>].
The methods that require exponential deviates use the ziggurat method
[<a href="Bibliography.htm#marsaglia2000">Marsaglia 2000</a>].
<ul>
<li>
<h3><tt>Time Homogeneous</tt></h3>
<ul>
<li>
<h4><tt>Time Series, Uniform</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
The computational complexities for generating discrete deviates
and for modifying propensities are indicated for each method in
terms of the number of reactions <em>M</em>.
<ul>
<li>
<tt>2-D search</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are stored in a 2-D table that stores the sum of each
row. The number of rows (and hence the number of columns) is
O(<em>M<sup>1/2</sup></em>). The discrete deviate is generated by
first determining the appropriate row with a linear search and then
performing a linear search within that row
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
<li>
<tt>2-D search, sorted</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are periodically ordered so the larger values have
smaller
<a href="http://en.wikipedia.org/wiki/Manhattan_distance">Manhattan
distance</a> from the lower corner of the table than smaller
values
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
For some problems this may reduce the costs of
searching.
<li>
<tt>2-D search, bubble sort</tt>
- Generate O(<em>M<sup>1/2</sup></em>). Modify O(1).
The propensities are dynamically ordered each time they are modified
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
<li>
<tt>Composition rejection</tt>
- Generate O(1). Modify O(1).
Propensities are placed in groups. In each
group the propensities differ by at most a factor of two. A deviate
is generated by a linear search on the groups followed by selecting
a reaction within a group with the method of rejection.
This solver implements a slightly modified version of the composition
rejection method presented in [<a href="Bibliography.htm#slepoy2008">Slepoy 2008</a>].
<li>
<tt>Binary search, full CMF</tt>
- Generate O(log <em>M</em>). Modify O(<em>M</em>).
The cumulative mass function (CMF) is stored in an array. This allows
one to generate a discrete deviate with a binary search on that array.
At each time step the CMF is regenerated. This is an implementation
of the logarithmic direct method [<a href="Bibliography.htm#li2006">Li 2006</a>].
<li>
<tt>Binary search, sorted CMF</tt>
- Generate O(log <em>M</em>). Modify O(<em>M</em>).
In this solver the reactions are arranged in descending order
according to the sum of the propensities of the influencing
reactions
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
This minimizes the portion of the CMF that one has
to recompute after firing a reaction and updating the
influenced propensities.
<li>
<tt>Binary search, recursive CMF</tt>
- Generate O(log <em>M</em>). Modify O(log <em>M</em>).
Instead of storing the full CMF, a partial, recursive CMF is used.
This approach is equivalent to the method presented in
[<a href="Bibliography.htm#gibson2000">Gibson 2000</a>]. A deviate can be generated
with a binary search. Modifying a propensity affects at most
log<sub>2</sub> <em>M</em> elements of the partial, recursive CMF.
<li>
<tt>Linear search</tt>
- Generate O(<em>M</em>). Modify O(1).
We use a linear search on the PMF to generate a deviate. The sum of
the propensities is dynamically updated.
<li>
<tt>Linear search, delayed update</tt>
- Generate O(<em>M</em>). Modify O(1).
This solver recomputes the sum of the propensities at each time step
as done in the original formulation of the direct method
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<li>
<tt>Linear search, sorted</tt>
- Generate O(<em>M</em>). Modify O(1).
The propensities are periodically sorted in descending order
[<a href="Bibliography.htm#mccollum2005">McCollum 2005</a>].
<li>
<tt>Linear search, bubble sort</tt>
- Generate O(<em>M</em>). Modify O(1).
The propensities are dynamically ordered by using swaps each time
a propensity is modified
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
</ul>
<li>
<h5><tt>Next Reaction</tt></h5>
Each formulation below uses a different data structure to
implement the priority queue. The computational complexities for
removing an element and modifying an element are given in terms
of the number of reactions <em>M</em>.
<ul>
<li>
<tt>Hashing</tt>
- Pop O(1). Modify O(1).
The indexed priority queue is implemented with a hash table
[<a href="Bibliography.htm#cormen2001">Cormen 2001</a>,
<a href="Bibliography.htm#mauch2011">Mauch 2011</a>]. We use hashing
with chaining,
which means that each bin in the hash table contains a list of
elements. The hash function is a linear function of the putative
reaction times, truncated to an integer to obtain a bin index.
The hash function is dynamically adapted to obtain the target load.
(The <em>load</em> is the average number of elements in each bin.)
<li>
<tt>Binary heap, pointer</tt>
- Pop O(log <em>M</em>). Modify O(log <em>M</em>).
This solver uses a binary heap that is implemented with an array
of pointers to the putative reaction times. This is the data structure
used in [<a href="Bibliography.htm#gibson2000">Gibson 2000</a>].
<li>
<tt>Binary heap, pair</tt>
- Pop O(log <em>M</em>). Modify O(log <em>M</em>).
We use a binary heap implemented with an array of index/time pairs
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
The pair version of the binary heap typically has better performance
than the pointer version on 64-bit architectures. Vice versa for
32-bit architectures.
<li>
<tt>Partition</tt>
- Pop O(<em>M<sup>1/2</sup></em>). Modify O(<em>M<sup>1/2</sup></em>).
A splitting value is used to divide the reactions two categories
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
The splitting value is chosen so that initially there are
O(<em>M<sup>1/2</sup></em>) reactions in the lower partition.
The remainder are in the upper partition. In determining the minimum
putative reaction time, one only has to examine reactions in the
lower partition. When the lower partition is empty, a new splitting
value is chosen.
<li>
<tt>Linear search</tt>
- Pop O(<em>M</em>). Modify O(1).
The putative reaction times are stored in an array. We use a linear
search to find the minimum
[<a href="Bibliography.htm#mauch2011">Mauch 2011</a>].
</ul>
<li>
<h5><tt>First Reaction</tt></h5>
Gillespie's first reaction method
[<a href="Bibliography.htm#gillespie1976">Gillespie 1976</a>]
[<a href="Bibliography.htm#gillespie1977">Gillespie 1977</a>].
<ul>
<li>
<tt>Simple</tt>
- A simple implementation of the method.
<li>
<tt>Reaction influence</tt>
- We use the reaction influence data structure to minimize recomputing
the propensities.
<li>
<tt>Absolute time</tt>
- We use absolute times instead of waiting times for each reaction.
This allows us to reduce the number of exponential deviates used.
After firing a reaction, we only generate new reaction times for the
influenced reactions.
</ul>
<li>
<h5><tt>Tau-Leaping</tt></h5>
All of the tau-leaping solvers have first order accuracy. They differ in
how they calculate the expected propensities, whether they correct
negative populations, and how they calculate the time step. For calculating
the expected propensities one may use a first, second, or fourth order
Runge-Kutta scheme. (The first and second order schemes are commonly
called forward and midpoint, respectively.)
While all yield a first order stochastic method, using
a higher order scheme is typically more efficient. Use the adaptive step
size solvers unless you are studying the effect of step size. With a fixed
step size it is difficult to predict the accuracy of the simulation.
With the adaptive step size solvers you may choose whether or not to
correct negative populations. Without correction, some portion of the
simulations may fail.
<ul>
<li>
<tt>Runge-Kutta, 4th order</tt>
<li>
<tt>Midpoint</tt>
<li>
<tt>Forward</tt>
<li>
<tt>Runge-Kutta, 4th order, no correction</tt>
<li>
<tt>Midpoint, no correction</tt>
<li>
<tt>Forward, no correction</tt>
<li>
<tt>Runge-Kutta, 4th order, fixed step size</tt>
<li>
<tt>Midpoint, fixed step size</tt>
<li>
<tt>Forward, fixed step size</tt>
</ul>
<li>
<h5><tt>Hybrid Direct/Tau-Leaping</tt></h5>
With these solvers you can choose the scheme for calculating the expected
propensities.
<ul>
<li>
<tt>Runge-Kutta, 4th order</tt>
<li>
<tt>Midpoint</tt>
<li>
<tt>Forward</tt>
</ul>
</ul>
<li>
<h4><tt>Time Series, All Reactions</tt></h4>
The direct method with a 2-D search is used when recording each reaction
event.
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<tt>2-D search</tt>
</ul>
</ul>
<li>
<h4><tt>Time Series, Deterministic</tt></h4>
When generating a single deterministic trajectory one can either use a
built-in solver or export the problem to a Mathematica notebook.
The solvers use
<a href="http://en.wikipedia.org/wiki/Runge-Kutta">Runge-Kutta</a>
schemes. For ordinary work, use the
adaptive step size
<a href="http://en.wikipedia.org/wiki/Cash-Karp">Cash-Karp</a> variant,
which is a fifth order method. The rest of the solvers are only useful
in studying how the order and step size affect the solution.
<ul>
<li>
<h5><tt>ODE, Integrate Reactions</tt></h5>
<ul>
<li>
<tt>Runge-Kutta, Cash-Karp</tt>
<li>
<tt>Runge-Kutta, Cash-Karp, fixed step size</tt>
<li>
<tt>Runge-Kutta, 4th order, fixed step size</tt>
<li>
<tt>Midpoint, fixed step size</tt>
<li>
<tt>Forward, fixed step size</tt>
</ul>
<li>
<h5><tt>Mathematica</tt></h5>
<ul>
<li>
<tt>NDSolve</tt>
</ul>
</ul>
<li>
<h4><tt>Histograms, Transient Behavior</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<tt>2-D search</tt>
</ul>
</ul>
<li>
<h4><tt>Histograms, Steady State</tt></h4>
By recording the average value of species populations you can determine
the steady state solution (if it exists).
<ul>
<li>
<h5><tt>Direct</tt></h5>
Each of the following use the direct method with a 2-D search.
<ul>
<li>
<tt>Elapsed time</tt> -
This is the most efficient method for most problems. The algorithm
keeps track of when species change values in order to minimize the cost
of recording the populations.
<li>
<tt>Time steps</tt> -
This is the traditional method. The species populations are recorded
at every time step.
<li>
<tt>All possible states</tt> -
This is an implementation of the all possible states method
[<a href="Bibliography.htm#lipshtat2007">Lipshtat 2007</a>].
It is usually less efficient than the traditional method.
</ul>
</ul>
</ul>
<li>
<h3><tt>Time Inhomogeneous</tt></h3>
<ul>
<li>
<h4><tt>Time Series, Uniform</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Time Series, All Reactions</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Time Series, Deterministic</tt></h4>
<ul>
<li>
<h5><tt>ODE, Integrate Reactions</tt></h5>
<ul>
<li>
<h6><tt>Runge-Kutta, Cash-Karp</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Histograms, Transient Behavior</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
<li>
<h4><tt>Histograms, Stead State</tt></h4>
<ul>
<li>
<h5><tt>Direct</tt></h5>
<ul>
<li>
<h6><tt>Constant propensities</tt></h6>
</ul>
</ul>
</ul>
</ul>
</p>
</body>
</html>
|
