Team:SJTang/Model
Overview
Our topic and experiment are about finding the optimal environment or setup for genetically engineered bacteria, including R. palustris and E. coli, to manufacture the greatest amount of hydrogen for the use of energy. In the early stages of the project, we discovered that the hydrogen yield process occurs primarily under an anaerobic environment, which is not ideal for direct quantitative experiments, and it hindered our data-collecting efficiency. In the early stages of the project, we discovered that the hydrogen yield process occurs primarily under an anaerobic environment, which is not ideal for direct quantitative experiments, and it hindered our data-collecting efficiency.
Considering the slow growth rate of cells under anaerobic conditions, we hope that by optimizing the culture conditions for their aerobic growth, a large number of cells can be obtained under aerobic conditions and finally used in the anaerobic hydrogen production fermentation process.For Escherichia coli, its aerobic culture conditions have been studied relatively clearly, but for Rhodopseudomonas palustris, its aerobic culture conditions have not yet been fully and quantitatively optimized. Therefore, we hope to explore the optimal culture conditions of Rhodopseudomonas palustris under aerobic conditions by constructing a mathematical model.
1. Parameter exploration before building the model
In experiments, we found that Rhodopseudomonas palustris can grow in YPMOPS medium under aerobic conditions. In the process of communicating with the Team SCAU-China, they believe that this phenomenon is worthy of in-depth study. Therefore, in order to further confirm the growth of Rhodopseudomonas palustris under aerobic conditions, we measured its growth curve in a variety of media. The test results are as follows:
Figure. 1Aerobic growth of R. palustris.
The results show that under aerobic conditions, the use of different media formulations will have a very large impact on the growth rate of Rhodopseudomonas palustris. So we decided to conduct further experiments.
2. Use TensorFlow to analyze the growth curve
The original design of the experiment is to try out all the combinations of different factors, plot all data points in a complete graph, and predict the best condition through machine learning algorithms. However, due to the limited amount of time and experimental apparatus such as anaerobic incubator and gas content analyzer, we are unable to try out all combinations and plot them into a complete graph. But still, we decided to utilize a new strategy: using few datapoints to simulate a curve on a graph by TensorFlow, which would be adjusted and verified continuously as more and more data points are added to the database.
Combined with the previous experimental results, we decided to test the carbon source, nitrogen source, buffer, metal ions and culture temperature during the growth of Rhodopseudomonas palustris. The specific test parameters are as follows:
Figure. 2Environmental factors and gradient settings.
As mentioned before, if all possible values for all these variables are considered in combination, there would be more than 6000 combinations, which is impossible to accomplish them all due to limited apparatus in the lab. Instead of testing them all, we randomly selected about 94 combinations cultivated in 30 degrees Celsius and another 94 combinations cultivated in 37 degrees Celsius. They are separated into 4 plates; each has 6 rows and 8 columns of spaces to place our medium. The arrangement is displayed below. For details, go to file:
Model Design
Figure. 3Experiment design.
Using this equation, we fit the growth curve obtained in the experiment, and get the A, B, C parameter values of each group of experiments. Then we use MLP, or multi-layer perceptron, to find out whether these parameters of A, B, and C, actually contributed to the growth of bacterium. Our model consists of 3 layers, with 7 environmental factors being the input layer, 12 perceptrons as the hidden layer, and 3 kinetic parameters as the output layer.
Figure.4Neural network for this model
As you can see from the picture above, every perceptron (circle) in this structure is linked to all perceptrons in the next layer. The reason we use MLP as our main analysis tool is to find out each parameter's importance to the curve From our results, we found that the parameter A contributes most to the growth, and parameters B and C’s contribution to the growth are fluctuating and unstable. The inaccuracy of training for parameters B and C might be caused by some measurements that are not accurate enough. <\P>
Figure.5Loss function of parameter. From Left fo Right are A, B, C's loss function
The red curves stand for the loss function of parameter A, B, and C, correspondingly. We observe a steady decrease of the loss value for A and a continuous oscillation for B and C. This suggests that A contributes much more than B and C.
We used logistic fitting for every combination of factors experimented in lab and training them to find the magnitude of contribution to the factors of logistic regression. By so, we gain a better insight on what factors and setting makes the R. palustris bacteria to grow best in aerobic environment. It is important since usually R. palustris is believed to be anaerobium. In future, we hope that applying bacterium in hydrogen manufacture would greatly solve the energy issue worldwide in present, and the optimal setup for them to grow would be an important research topic in future as higher efficiency are required.