Can AI Predict Behavior of Complex Biological Systems?
Duke University’s novel AI method for predicting biological systems behavior.
Posted Oct 09, 2019
Biological systems are inherently complex. Identifying patterns in biological systems is a daunting, time consuming endeavor. Biomedical engineers at Duke University have created a novel artificial intelligence (AI) machine learning methodology that can predict behaviors of biological circuits in orders of magnitude faster than standard computational methods, and published their findings in Nature Communications on September 25, 2019.
In scientific research in pharmaceuticals, disease treatments, and biomedicine, mathematical modeling is used to understand the processes for the particular biological system. Different systems require separate approaches. For example, mechanistic modeling may be used for research on metabolic networks and genetic circuits. An analytic approach suits biological systems that are well-understood, and logic can be used to calculate a precise outcome.
However, biological systems are often unpredictable and need to account for randomness. A numerical approach is a more suitable approach where there are no exact solutions, such as mechanism-based models that involve spatial or unpredictable characteristics. Numerical problem solving uses trial and error search to find suitable solutions. A major drawback to numerical simulation is its inefficiency in performing complex problem solving.
“For example, consider a model with 10 parameters,” wrote the Duke researchers in their study. “To examine six values per parameter, there will be 6^10 parameter combinations. If each simulation takes 5 min, which is typical for a partial differential equation (PDE) model, the screening would require 575 years to finish.”
The researchers’ strategy was to use a limited number of simulations generated by a mechanistic model to train a neural network — an amount large enough to enable accuracy, and could be completed in a relatively short time period.
“A major innovation of our approach is the combined use of the mechanistic model and the neural network,” wrote the Duke researchers. “The mechanistic model is used as a stepping stone for the latter by providing a sufficient data set for training and testing. This training set is extremely small compared with the possible parameter space.”
In this study, the team used 13 bacterial variables that include cellular movement, synthesis rate, inhibition factors on growth, diffusion, and protein degradation. First, they trained a deep neural network using a limited number of simulations generated by a mechanistic model to predict pattern formation and stochastic gene expression. The Duke researchers used a specific type of recurrent neural network (RNN), for the prediction of the normalized distribution.
RNNs are deep neural networks that will send predicted output back unto itself, as opposed to a feedforward neural network, in order to process sequential data. In essence, RNNs exhibit a form of “memory.” However, regular RNNs are not well-suited for long-sequence data because the “memory” of early inputs fade as data is transformed through the RNN. Thus, the Duke researchers opted to use a Long-Short-Term Memory (LSTM) network, an RNN specifically designed to memorize long-term dependencies.
“Our deep neural network consists of an input layer with inputs to be the parameters of mechanism-based model, connected to a fully connected layer, and the output layer consists of two types of outputs, one for predicting the logarithm of the peak value of the profile, directly connected to the fully connected layer, the other for predicting the normalized profile, connected to LSTM cell arrays, which was fed by the output from fully connected layer,” the researchers reported.
The Duke biomedical engineers then trained four separate neural networks for comparison. The result was that when the trained neural networks had similar predictions, those were nearer to the correct answer. In this manner, the team found a way to validate the predictions that is much faster than using a standard computational method.
“For a pattern formation circuit, our approach leads to ~30,000-fold acceleration in computation with high prediction accuracy,” wrote the Duke researchers. “Our results demonstrate the tremendous potential of deep learning in overcoming the computational bottleneck faced by many mechanistic-based models. Depending on the application context, this capability can facilitate the engineering of gene circuits or the optimization of experimental conditions to achieve specific target functions (e.g., generation of multiple rings from our circuit), or to elucidate how a biological system responds to environmental perturbations (e.g., drug treatments).”
By applying AI machine learning and mathematics to complex biological systems, the Duke researchers have created a method that can accelerate future research in biomedicine, neuroscience, pharmacokinetics, de novo drug discovery, and other industries.
Copyright © 2019 Cami Rosso All rights reserved.
Wang, Shangying, Fan, Kai, Luo, Nan, Cao, Yangxiaolu, Wu, Feilun, Zhang, Carolyn, Heller, Katherine A. Heller, You, Lingchong. “Massive computational acceleration by using neural networks to emulate mechanism-based biological models.” Nature Communications. September 25, 2019.