A machine learning algorithm has been developed that can model relationships between complex variables of gene-edited bacteria that are difficult to predict.
Artificial neural networks reduced the modeling Times of circuits to illuminate the complex structures of biological circuits. A group of biomedical engineers from Duke University have developed machine learning that can model relationships between complex variables of gene-edited bacteria that are normally difficult to predict. It is noted that this machine learning algorithm, which they have developed, is generalizable to many biological systems.
In this new study, researchers aimed to be able to predict the circular pattern of the biological circuit in bacterial culture by training neural networks. This system worked 30,000 times faster than the current computational model.
The researchers developed a learning model that would compare learning outcomes to test the accuracy of the data. They observed the results using the algorithm in a second biological system that required different computational techniques. In this way, they proved that the algorithm works successfully in different computing systems, and the results of the study were published on the website of the journal Nature Communications on September 25.
Professor of Biomedical Engineering at Duke University Lingchong You; ” a study similar to this showed Google that neural networks can beat people in AlphaGo, a board game, after training, and this inspired Google for future research. Although the game had simple rules, the computer’s choice of the best move depended on many factors. We also wondered if these developments could be useful in illuminating biological events that we do not know about.” said. Prof.Lingchong You and his colleague Shangying Wang have the most difficulty; it was the determination of which parameters a gene-edited circuit could produce a specific model in bacterial culture.
In a previous study, You and his team had programmed bacteria to produce ring-forming and interacting proteins based on the growth characteristics of the culture. The researchers found that variables such as the size of the growth medium and the amount of nutrients provided affect the thickness, length and other properties of the ring. They found that if a large number of potential variables were regulated, they could form two-or three-ring structures. But because a single computer simulation took five minutes, observing the results was hardly practical.
A colony of bacteria genetically arranged to contain the gene circuit forms a purple ring as it grows.
Researchers are using machine learning to explore the thickness of these rings, how quickly they form, and the interaction between dozens of variables that affect properties such as the number of rings that form.
The system for bacterial study consisted of 13 variables, such as growth, diffusion, protein degradation, and cellular movement rates. Calculating six values per parameter took more than 600 years for a single computer. A parallel computer cluster could only perform this calculation in a long time, such as a few months. With machine learning, this time was reduced to a few hours.
“Dec deceptions are slow when we want to observe the system, so we usually want to get to the final results by bypassing the deceptions. If we find the results interesting, we repeat the steps to decipher the intermediate steps.”- Prof.Lingchong You
Wang developed a machine learning model called the deep neural network, which can predict much faster than the original model. This deep neural network takes model variables as input, initially giving random trends and revealing an estimate of what pattern the bacterial colony will form. In this process, it completely bypasses the intermediate steps and reaches the result.dec. If the results are too far from the correct answer, the neural network can be retrained by changing trends each time. It is possible for a neural network to always make accurate predictions after adequate training.
Sen and Wang came up with a way to quickly check the system to examine several instances where machine learning gave false results. For each neural network, the learning process contains a random element, and thus the neural network does not learn an input twice in the same way, even if it is trained on the same set of answers. The researchers trained four different neural networks and compared results in different conditions. When they trained them to make similar predictions, they found they were getting results very close to the correct answer.
“We discovered that we don’t have to validate every answer with a slow standard calculation model. We used the’ wisdom of the crowd ‘ instead.”- Researchers
The researchers decided to use the machine learning model while working on biological systems. Of the 100,000 data simulations used to train the neural network, only one produced a colony of three-ringed bacteria. Thanks to the speed of the neural network, Sen and Wang not only found more three-ring structures, but also determined which variables were more important in creating the ring structure. “The neural network was able to find interactions between variables that were not possible to uncover,” Wang said.
In the final step of their study, Sen and Wang experimented with their approach on a biological system that worked randomly. It took a computer model to repeat the same parameters many times to achieve the most accurate result. The researchers said their approach covered many complex biological systems. They want to use their new approach on more complex biological systems. The team’s next goal is to increase the efficiency of the algorithm and find ways to run it on computers with faster GPUs.
“We trained the neural network with 100,000 data sets, but this could also cause an overload. We are developing an algorithm where the neural network can accelerate the system by interacting simultaneously with simulations. Our first goal was a relatively simple system.”says You. “Now we want to illuminate the fundamental dynamics of more complex biological circuits with these neural network systems.”
The Office of Naval Research (N00014-12-1-0631), the National Institutes of health (1R01-GM098642) and supported by a David and Lucile Packard Fellowship.