Towards a fully automated algorithm driven platform for biosystems design

Abstract

Large-scale data acquisition and analysis are often required in the successful implementation of the design, build, test, and learn (DBTL) cycle in biosystems design. However, it has long been hindered by experimental cost, variability, biases, and missed insights from traditional analysis methods. Here, we report the application of an integrated robotic system coupled with machine learning algorithms to fully automate the DBTL process for biosystems design. As proof of concept, we have demonstrated its capacity by optimizing the lycopene biosynthetic pathway. This fully-automated robotic platform, BioAutomata, evaluates less than 1% of possible variants while outperforming random screening by 77%. A paired predictive model and Bayesian algorithm select experiments which are performed by Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB). BioAutomata excels with black-box optimization problems, where experiments are expensive and noisy and the success of the experiment is not dependent on extensive prior knowledge of biological mechanisms.

Fig. 1

The overall workflow of BioAutomata. After setting the initial parameters, designing the sequence space of variable regions (such as promoter variants in a combinatorial pathway assembly), and defining the objective function, BioAutomata selects which experiments are expected to result in the highest improvement of yield, performs those experiments, generates data and learns from it, updating its predictive model given the newly presented evidence. It will then decide on the next experiments to perform to reach the goal set by the user while trying to minimize the number of experiments and the cost of the project

Fig. 2

Testing Bayesian optimization by finding the maximum of a two-dimensional function. a The acquisition function decides the next input to test and the output is used to refine the predictive model. Iterations 5, 10, 15 and 20 of this process are shown. b With increasing rounds of iteration, the predictive model grows more confident of the location of the global maximum and the distance between tested inputs decreases with each iteration. c The algorithm evaluated 9 points before finding the location of the maximum. Subsequent iterations tuned this approximation toward the true optimum. The algorithm evaluated 12 points before finding the maximum. The order in which each point is evaluated is shown on the graph

Fig. 3

Change in sampling behavior of Baysesian optimization of the lycopene production pathway. In the first round (a), all points were chosen to uniformally explore the landscape since it is completely unexplored and unknown (n = 46). In the second round (b), some information is acquired and the points picked by the algorithm are clearly skewed from the unifrom distribution (n = 45). However, since there is some uncertainty, it is still exploring the landscape. Finally, in the third round (c), a clear pattern is observed where the algorithm has determined the points in a particular area are more likely to be closer to the global optima and is actively exploring that area but still doing some minimal exploration (n = 45). Source data are provided as a Source Data file

Fig. 4

Lycopene production in different rounds of pathway optimization and random screening. The average and maximum points have increased after each round of pathway optimization. Moreover, although the average and maximum of evaluating 46 and 136 random points are a little more than the uniform distribution in round 1, they are significantly lower than the points picked by the algorithm in the subsequent rounds. The boxes of the plots contain data within the interquartile range (IQR), while whiskers spread from the boxes to 1.5 times IQR. The center line in the boxes is the median of the data and points above the whiskers are values which are higher than 1.5 times IQR above the third quartile. n = 46, 45, 45, 46 and 136 for each plot, respectively. Source data are provided as a Source Data file

Fig. 5

The overall fully automated pathway optimization workflow. The machine learning algorithm picks the plasmids to be assembled and returns the list to iBioFAB to perform the assembly. The assembled products are then transformed, and four single colonies are isolated for lycopene quantification and OD measurement. The resulting data are then given to the machine learning algorithm to pick the next set of points to be evaluated