Machine learning predicts the behavior of biological chains

Duke University biomedical engineers have developed a machine learning approach to model interactions between complex variables in engineered bacteria which should be too difficult to predict. Their algorithm can be generalized to many types of biological systems.

In the new study, researchers are training neural networks to predict circular patterns produced by biological chains embedded in bacterial cultures. The system runs 30,000 times faster than the existing calculation model.

To further improve accuracy, the team developed a method to train the machine learning model several times and compare their answers. They then use it to solve the second biological system that is computationally different, and to show that the algorithm is suitable for various challenges.

This work was inspired by Google and showed that neural networks can learn to beat someone in a board game, said Lingchong You, professor of biomedical engineering at Duke. Even though this game has simple rules, there are too many ways for computers to calculate the next variant that is best deterministically, they said. I wonder if such an approach could be useful for dealing with the biological complexity that we face.”

The challenge for you and your doctoral student, Shanging Wang, is to determine what parameters can be produced by certain models in bacterial culture by following the technical gene chain.

In previous research, you programed bacteria in a laboratory that produces proteins that interact with each other and form rings depending on the specific growth of the culture. By controlling for variables such as the size of the growth medium and the amount of nutrients delivered, the researchers found that they could control the thickness of the ring and how long other features appeared.

By changing a number of potential variables, the researchers found they could do more to make two or even three rings. However, because computer simulations take five minutes, it is not practical to search every large design space for specific results.

For their study, the system consisted of 13 bacterial variables such as growth rate, diffusion, protein degradation and cell movement.

Just calculating six parameter values will require more than 600 years on one computer. By running in a parallel cluster of computers with hundreds of nodes, execution time can be reduced to several months, with machine learning up to hours.

“The model we use is slow because it has to take intermediate steps in time with a speed slow enough to be accurate,” they said. “We don’t always care about intermediary steps, we only want the final results for certain applications, and we can understand the next steps if we find an interesting end result.”

To achieve the final result, Wang turned to a machine learning model called the Deep Neural Network, which made work estimates faster than the original model.

The network accepts model variables as input, assigns weights and arbitrary deviations first, and predicts which model will form a bacterial colony, completely eliminating intermediate steps leading to the final model.

Although the initial results are far from the right answers, the weights and deviations can be adjusted each time new training data is sent to the network. With a sizeable training set, the neural network finally learns to make accurate predictions almost every time.

Sources / Further Reading: Duke University

Leave a Reply

Your email address will not be published. Required fields are marked *