Team:UC Davis/Signal Test

From 2014.igem.org

(Difference between revisions)
(Created page with "{{Team:UC_Davis/Head}} {{Team:UC_Davis/projectMenu}} <html> <head> </head> <body> <!--Project Menu Start--> <div class="floatboxProjectMenu"> <a href="https://2014.igem.org/Team:U...")
 
(13 intermediate revisions not shown)
Line 32: Line 32:
  <div class="switchbox1">
  <div class="switchbox1">
-
   <h2>Machine Learning</h2>
+
   <h2>Olive Oil</h2>
-
   <a href="https://2014.igem.org/Team:UC_Davis/Signal_Machine">
+
   <a href="https://2014.igem.org/Team:UC_Davis/Signal_Oil">
-
     <span><h2>Machine Learning</h2></span>
+
     <span><h2>Olive Oil</h2></span>
   </a>
   </a>
  </div>
  </div>
</div>
</div>
-
<div class="mainContainer">
+
<div class="mainTitleHeader">
 +
<p>Combinations</p>
 +
</div>
 +
 
 +
<div class="mainContainer"><p>
 +
To test our model we built a combinatorial set of aldehydes. We chose three representative aldehydes from each bin, saturated medium, saturated long, and unsaturated. The three aldehydes were chosen such that the enzyme's response to each would best represent their respective groups. We created a total of 64 different combinations by mixing Pentanal, Decanal, and <i>E</i>-2-Decenal in four different values, 0 µM, 10µM, 100µM, & 1000µM. Three combinatorial well plates were made and mixed with each enzyme separately. The observed velocity is recorded in each well.</p><br>
 +
<div style="margin:auto;display:block;float:center">
 +
<p align="center"><img src="https://static.igem.org/mediawiki/2014/1/1d/64Combinations.png" width="800px" style="margin-left:auto;margin-right:auto;"/></p>
 +
<br></div>
 +
<p><br>
 +
The catalytic matrix is inverted and multiplied by the observed velocity in each well and out comes our predicted concentrations!<br>
 +
</p><br>
 +
<div style="margin:auto;display:block;float:center">
 +
<p align="center"><img src="https://static.igem.org/mediawiki/2014/1/1f/Inverse.png" width="400px" style="margin-left:auto;margin-right:auto;border:1.5px solid #212f20;"/></p><br.
 +
</div><br>
 +
<p>
 +
In high concentrations, we found that aldehydes would crash out of solution. We chose to focus on only the data set below 1000 µM. A more serious issue was brought to light however. The observed velocities from the combinatorial plates suggested competitive inhibition was occurring when <i>E</i>-2-Decenal was present in solution. This obfuscated our model considerably. Our primary assumption was that competitive inhibition would not come into play. We needed to think more abstractly. We asked a simple question: If measured catalytic values would not work in our suggested model, was there still a set of numbers that would work?<br><br></div>
 +
</div>
 +
<div class="mainTitleHeader">
 +
<p>Machine Learning</p>
 +
</div>
 +
<div class="mainContainer"><p>
 +
Instead of using the measured K<sub>cat</sub>/K<sub>m</sub>, we randomized the catalytic matrix and tested the variants for prediction accuracy. One million variants later and we were starting to produce consistently better predictions.</p><br>
 +
 
 +
<div style="margin:auto;display:block;float:center">
 +
<p align="center"><img src="https://static.igem.org/mediawiki/2014/c/cd/SomeStuff.png" width="400px" style="margin-left:auto;margin-right:auto;"/></p><br>
 +
 
 +
<p>We found that with decreasing aldehyde concentrations, our model accuracy would increase. This is most likely due to the inherent non-linearity of Michaelis Menten kinetics. We redesigned the experiment; this time we would vary aldehyde in the following concentrations: 0 µM, 12.5 µM, 25 µM, & 50 µM. We would also train on an incomplete set and test the best matrices on the full set. This mimics a concept from protein crystallography, R-Free, where several atoms are removed from a model and later used to test the model's accuracy. The matlab code we used to produce these results is provided <a href="https://static.igem.org/mediawiki/2014/9/98/MutantVariation.txt">here</a>. The results were astonishing. With our best mutant matrix, we were able to predict aldehyde concentration with an average error of only 6.25 µM!<br><br>
 +
</p>
 +
<div style="margin:auto;display:block;float:center">
 +
<p align="center"><img src="https://static.igem.org/mediawiki/2014/3/3e/AveragepredictionError.png" width="400px" style="margin-left:auto;margin-right:auto;border:1.5px solid #212f20;"/></p>
 +
</div>
 +
<br><p><b>To see how our model preformed in olive oil, click <a href="https://2014.igem.org/Team:UC_Davis/Signal_Oil" class="brightlink">here</a>.</b></p>
 +
</div>

Latest revision as of 03:10, 18 October 2014

UC Davis iGEM 2014

Mathematical Approach

Mathematical Approach

Testing Our Model

Testing Our Model

Olive Oil

Olive Oil

Combinations

To test our model we built a combinatorial set of aldehydes. We chose three representative aldehydes from each bin, saturated medium, saturated long, and unsaturated. The three aldehydes were chosen such that the enzyme's response to each would best represent their respective groups. We created a total of 64 different combinations by mixing Pentanal, Decanal, and E-2-Decenal in four different values, 0 µM, 10µM, 100µM, & 1000µM. Three combinatorial well plates were made and mixed with each enzyme separately. The observed velocity is recorded in each well.




The catalytic matrix is inverted and multiplied by the observed velocity in each well and out comes our predicted concentrations!



In high concentrations, we found that aldehydes would crash out of solution. We chose to focus on only the data set below 1000 µM. A more serious issue was brought to light however. The observed velocities from the combinatorial plates suggested competitive inhibition was occurring when E-2-Decenal was present in solution. This obfuscated our model considerably. Our primary assumption was that competitive inhibition would not come into play. We needed to think more abstractly. We asked a simple question: If measured catalytic values would not work in our suggested model, was there still a set of numbers that would work?

Machine Learning

Instead of using the measured Kcat/Km, we randomized the catalytic matrix and tested the variants for prediction accuracy. One million variants later and we were starting to produce consistently better predictions.



We found that with decreasing aldehyde concentrations, our model accuracy would increase. This is most likely due to the inherent non-linearity of Michaelis Menten kinetics. We redesigned the experiment; this time we would vary aldehyde in the following concentrations: 0 µM, 12.5 µM, 25 µM, & 50 µM. We would also train on an incomplete set and test the best matrices on the full set. This mimics a concept from protein crystallography, R-Free, where several atoms are removed from a model and later used to test the model's accuracy. The matlab code we used to produce these results is provided here. The results were astonishing. With our best mutant matrix, we were able to predict aldehyde concentration with an average error of only 6.25 µM!


To see how our model preformed in olive oil, click here.

Retrieved from "http://2014.igem.org/Team:UC_Davis/Signal_Test"