Machine Learning predicts the behavior of biological cycles
4. Oktober 2019Machine Learning predicts the behavior of biological cycles
New York, October 4.10. 2019
Lingchong You, Professor of Biomedical Engineering at Duke University (North Carolina) and his postdoctoral fellow Shangying Wang have trained a neural network to predict the circular patterns produced by a biological cycle embedded in a bacterial culture. They developed a machine learning approach to model the interactions between complex variables in technical bacteria that would otherwise be difficult to predict. The results make people sit up and take notice: Their algorithms are transferable to many types of biological systems and their system works 30,000 times faster than the existing calculation model.
The researchers trained four different neural networks and compared their responses for each case. They found that when the trained neural networks made similar predictions, they were close to the correct answer. „We found that we didn’t have to validate every answer with the slower standard calculation model,“ said You. „We essentially used the wisdom of the set instead.“
With the trained and confirmed model of machine learning, the researchers set out to gain new insights into their biological cycle. In the first 100,000 data simulations used to train the neural network, only one produced a bacterial colony with three rings. But with the speed of the neuronal network, You and Wang could not only find many more triplets, but also determine which variables were crucial for production. „The neural network was able to find patterns and interactions between the variables that would otherwise not have been detectable,“ said Wang.
At the end of their study, the two researchers tested their approach on a biological system that works randomly. The solution of such systems requires a computer model that repeats the same parameters several times in order to achieve the most probable result. While this is a completely different reason for long computing times than their original model, the researchers still found their approach functional and showed that it can be generalised for many different complex biological systems.
The researchers are now trying to apply their new approach to more complex biological systems. In addition to running on computers with faster GPUs, they are trying to program the algorithm as efficiently as possible. „We trained the neural network with 100,000 data sets, but that could have been an overkill,“ said Wang. „We’re developing an algorithm where the neural network can interact with real-time simulations to speed things up.“
„Our first goal was a relatively simple system,“ said You. „Now we want to improve these neural network systems to open a window into the underlying dynamics of more complex biological cycles. The researchers trained four different neural networks and compared their responses for each case. They found that when the trained neural networks made similar predictions, these predictions were close to the correct answer.
„We’ve found that we don’t have to validate every answer with the slower standard calculation model,“ said You. „We essentially used the wisdom of the set instead.“
With the trained and confirmed model of machine learning, the researchers set out to gain new insights into their biological cycle. In the first 100,000 data simulations used to train the neural network, only one produced a bacterial colony with three rings. But with the speed of the neural network, you and Wang could not only find many more triplets, but also determine which variables were critical for production. „The neural network was able to find patterns and interactions between the variables that would otherwise not have been detectable,“ said Wang.
At the end of their study, You and Wang tested their approach on a biological system that works randomly. Solving such systems requires a computer model that repeats the same parameters several times to achieve the most likely result. While this is a completely different reason for long computing times than their original model, the researchers still found their approach functional and showed that it can be generalised for many different complex biological systems.
The researchers are now trying to apply their new approach to more complex biological systems. In addition to running on computers with faster GPUs, they are trying to program the algorithm as efficiently as possible.
„We trained the neural network with 100,000 data sets, but that could have been an overkill,“ said Wang. „We’re developing an algorithm where the neural network can interact with real-time simulations to speed things up.“
„Our first goal was a relatively simple system,“ she said. „Now we want to improve these neural network systems to open a window into the underlying dynamics of more complex biological cycles.