The document discusses using Restricted Boltzmann Machines (RBMs) to analyze regulation of metabolism through biological signals, transcription factors, and enzymes. An RBM model is presented with signals and enzymes as visible units and transcription factors as hidden units. The document provides an example application with 8 visible and 2 hidden units to learn weights between units from sample data. The results show a strong dependency between one signal and enzyme, while others have little effect, demonstrating RBMs can learn regulatory patterns from data. Further applications of deep learning strategies using RBMs are mentioned.
3. Page 31/31/2012 |
Author
Department Biological Problem
Analysing the regulation of metabolism
Signal
Regulation
Metabolism
A linear metabolic pathway of enzymes (E) …
4. Page 41/31/2012 |
Author
Department Biological Problem
Analysing the regulation of metabolism
Signal
Regulation
Metabolism
… is regulated by transcription factors (TF) …
5. Page 51/31/2012 |
Author
Department Biological Problem
Analysing the regulation of metabolism
Signal
Regulation
Metabolism
… which respond to signals (S)
13. Page 131/31/2012 |
Author
Department
Restricted Boltzmann Machines
In the most common model all units have binary values …
𝑉 ≔ set of visible units
𝐻 ≔ set of hidden units
𝑠 𝑣: = value of 𝑣, ∀𝑣 ∈ 𝑉
𝑠ℎ ≔ value of ℎ, ∀ℎ ∈ 𝐻
𝑠 𝑣 ∈ 0, 1 , ∀𝑣 ∈ 𝑉
𝑠ℎ ∈ 0, 1 , ∀ℎ ∈ 𝐻
14. Page 141/31/2012 |
Author
Department
Restricted Boltzmann Machines
… and arbitrary tresholds which allow us …
𝑉 ≔ set of visible units
𝐻 ≔ set of hidden units
𝑠 𝑣: = value of 𝑣, ∀𝑣 ∈ 𝑉
𝑠ℎ ≔ value of ℎ, ∀ℎ ∈ 𝐻
𝑠 𝑣 ∈ 0, 1 , ∀𝑣 ∈ 𝑉
𝑠ℎ ∈ 0, 1 , ∀ℎ ∈ 𝐻
𝜃 𝑣 ≔ threshold of 𝑣, ∀𝑣 ∈ 𝑉
𝜃ℎ ≔ threshold of ℎ, ∀ℎ ∈ 𝐻
15. Page 151/31/2012 |
Author
Department
Restricted Boltzmann Machines
… to define energy functions: Local energies Ev and Eh …
𝐸 𝑣 ≔ − 𝑤 𝑣, ℎ 𝑠 𝑣 𝑠ℎ
ℎ
+ 𝜃 𝑣 𝑠 𝑣
𝐸ℎ ≔ − 𝑤 𝑣, ℎ 𝑠 𝑣 𝑠ℎ
𝑣
+ 𝜃ℎ 𝑠ℎ
𝑤 𝑣, ℎ ≔ weight of 𝑒𝑑𝑔𝑒(𝑣, ℎ)
Local Energy
16. Page 161/31/2012 |
Author
Department
Restricted Boltzmann Machines
… and the global Energy E. We want to minimize E
𝐸 ≔ 𝐸 𝑣
𝑣
+ 𝐸ℎ
ℎ
𝑤 𝑣, ℎ ≔ weight of 𝑒𝑑𝑔𝑒(𝑣, ℎ)
Local Energy
Global Energy
𝐸 𝑣 ≔ − 𝑤 𝑣, ℎ 𝑠 𝑣 𝑠ℎ
ℎ
+ 𝜃 𝑣 𝑠 𝑣
𝐸ℎ ≔ − 𝑤 𝑣, ℎ 𝑠 𝑣 𝑠ℎ
𝑣
+ 𝜃ℎ 𝑠ℎ
1
18. Page 181/31/2012 |
Author
Department
Restricted Boltzmann Machines
… we can use the Boltzmann Factor …
∆𝐸𝑣 = 𝐸 𝑣, 𝑜𝑓𝑓 − 𝐸 𝑣, 𝑜𝑛
= −𝑘 𝐵 𝑇 𝑙𝑛 P 𝑣,off
−(−𝑘 𝐵 𝑇 𝑙𝑛 P 𝑣,on )
Energy Delta for visible units
19. Page 191/31/2012 |
Author
Department
Restricted Boltzmann Machines
… to get a term for the probability [v, on]
P 𝑣,off = 1−P 𝑣,on P 𝑣,on =
1
1 + 𝑒−
∆𝐸𝑣
𝑘 𝐵
𝑇
∆𝐸𝑣 = 𝐸 𝑣, 𝑜𝑓𝑓 − 𝐸 𝑣, 𝑜𝑛
= −𝑘 𝐵 𝑇 𝑙𝑛 P 𝑣,off
−(−𝑘 𝐵 𝑇 𝑙𝑛 P 𝑣,on )
Energy Delta for visible units
21. Page 211/31/2012 |
Author
Department
Restricted Boltzmann Machines
With (1) an (2) we can perform Simulated Annealing
set T= TMax
while (T> TMin)
forall v
if (P[v,on] > rand(0,1)) set sv = 1
forall h
if (P[h,on] > rand(0,1)) set sh = 1
set Tsmaller
Simulated Annealing
𝐸 → min
29. Page 291/31/2012 |
Author
Department
Example
Let‘s feed the machine with learning samples …
S E
1,0,0,1 1,0,0,0
1,0,0,1 1,1,0,0
1,0,0,1 1,0,1,0
1,0,0,1 1,0,0,1
1,0,1,1 0,0,0,0
1,0,1,1 0,1,0,0
1,0,1,1 0,0,1,0
1,0,1,1 0,0,0,1
Learning samples
30. Page 301/31/2012 |
Author
Department
Example
.. to get the calculated weight matrix
TF1 TF2
S1 0,3 0,8
S2 0,5 0,6
S3 0,9 0,1
S4 0,3 0,8
E1 1,0 0,0
E2 0,1 0,0
E3 0,1 0,0
E4 0,2 0,0
Weight matrix
31. Page 311/31/2012 |
Author
Department
Example
The weights are visualized by the intensity of the edges
S
E
TF
TF1 TF2
S1 0,3 0,8
S2 0,5 0,6
S3 0,9 0,1
S4 0,3 0,8
E1 1,0 0,0
E2 0,1 0,0
E3 0,1 0,0
E4 0,2 0,0
Weight matrix
32. Page 321/31/2012 |
Author
Department
Example
Now we can compare the results with the samples
S E
1,0,0,1 1,0,0,0
1,0,0,1 1,1,0,0
1,0,0,1 1,0,1,0
1,0,0,1 1,0,0,1
1,0,1,1 0,0,0,0
1,0,1,1 0,1,0,0
1,0,1,1 0,0,1,0
1,0,1,1 0,0,0,1
Learning samples
S
E
TF
33. Page 331/31/2012 |
Author
Department
Example
There‘s a strong dependency between S3 an E1
S E
1,0,0,1 1,0,0,0
1,0,0,1 1,1,0,0
1,0,0,1 1,0,1,0
1,0,0,1 1,0,0,1
1,0,1,1 0,0,0,0
1,0,1,1 0,1,0,0
1,0,1,1 0,0,1,0
1,0,1,1 0,0,0,1
Learning samples
S
E
TF
34. Page 341/31/2012 |
Author
Department
Example
S1, S2 and S4 do almost not affect the metabolism
S
E
TF
S E
1,0,0,1 1,0,0,0
1,0,0,1 1,1,0,0
1,0,0,1 1,0,1,0
1,0,0,1 1,0,0,1
1,0,1,1 0,0,0,0
1,0,1,1 0,1,0,0
1,0,1,1 0,0,1,0
1,0,1,1 0,0,0,1
Learning samples
35. Page 351/31/2012 |
Author
Department
Further Objectives
Since 2006 RBMs have successfully be used to train (pre-
train) Multi-Layer ANNs (Hinton, Osindero, 2006)
This new branch in machine learning („deep learning“) already
has a wide area of applications, incuding:
• Face recognition / Voice recognition
• Unsupervised detection of features
• Imagetransformation
It has to be asumed that deep learning strategies also provide
further capabilities in regulatory analysis