Presentation Information
[P04-465]Development of a data-driven framework integrating deep learning and multi-objective optimization for cost-effective medium design in recombinant Escherichia coli
○Kazuki Watanabe1, Tomoko Kagenishi1, Konishi Masaaki1 (1. Kitami Institute of Technology, Division of Applied Chemistry (Japan))
Keywords:
Medium optimization,Machine learning,Escherichia coli,Multi-objective optimization
[Purpose]
In recent years, efficient methods for optimizing microbial culture media using machine learning with single-objective optimization (SOO) have been well studied. However, SOO approaches focused only on improving target product productivity with ignoring trade-off factors, i. e. medium cost. To break through the limitation of SOO based optimization, a multi-objective optimization (MOO), which simultaneously optimizes multiple objectives, was attempted to applied for microbial medium composition. In this study, In this study, MOO applied to increase expression of heterologous protein and to reduce the medium cost, in case of an engineered Escherichia coli secreted green fluorescent protein (GFP), as a model bioprocess.
[Methods]
Based on thirteen medium components with the ranges of the concentration, the 64 experimental conditions were defined by an orthogonal array (L64). To obtain training data, the recombinant cells were cultivated in the experimental conditions with deep-well plates at 37°C with shaking at 1,400 rpm, and GFP fluorescence was detected by a multiplate reader. Subsequently, deep learning (DL) models were constructed between the medium compositions and the intensities of GFP fluorescence. In case of MOO, NSGA-II (Non-dominated Sorting Genetic Algorithms II) with the trained DL models was applied to search the maximize GFP intensity and minimalize amount of medium as cost index. For comparison, optimization was also performed using a conventional SOO approach based on improving the GFP fluorescence intensity. After the optimization, the algorithm-proposed media were experimentally evaluated.
[Results and Discussion]
The total concentration of medium components ranged from 6.1 to 11.0 g/L in the SOO-optimized media and from 1.78 to 8.22 g/L in the MOO media. Therefore, MOO identified lower-cost media than SOO. Zinc chloride concentrations were mostly zero in the MOO-optimized media, suggesting that incorporating medium cost into MOO reduced less important components, consistent with the DL model sensitivity analysis predicting a minimal contribution of zinc chloride to GFP expression. In addition, even for highly important components such as tryptone and glucose, lower concentrations were observed compared with the control medium. These results indicate that the algorithm effectively identified compositions that balance productivity and material input. Experimental cultivation using the optimized media showed that, whereas the highest GFP fluorescence intensity in the training dataset was 11.3 × 10³ a.u., the SOO-designed medium achieved 13.7 × 10³ a.u. and the MOO-designed medium achieved 14.1 × 10³ a.u. These results demonstrate that GFP expression was improved to a similar extent in both optimized media. Furthermore, the total concentration of medium components in the MOO media was reduced to 70% of that in the control medium. Based on reagent prices, the cost-effectiveness relative to the control increased to 112% of that of the control. Therefore, only MOO was able to achieve both increased GFP expression and reduced medium cost.
[Conclusion]
The present study demonstrated that medium optimization with MOO enables the simultaneous improvement of recombinant protein production and reduction of medium cost, which could not be achieved by SOO. Thus, MOO represents a powerful strategy for cost-effective medium design and offers a promising framework for accelerating bioprocess development.
In recent years, efficient methods for optimizing microbial culture media using machine learning with single-objective optimization (SOO) have been well studied. However, SOO approaches focused only on improving target product productivity with ignoring trade-off factors, i. e. medium cost. To break through the limitation of SOO based optimization, a multi-objective optimization (MOO), which simultaneously optimizes multiple objectives, was attempted to applied for microbial medium composition. In this study, In this study, MOO applied to increase expression of heterologous protein and to reduce the medium cost, in case of an engineered Escherichia coli secreted green fluorescent protein (GFP), as a model bioprocess.
[Methods]
Based on thirteen medium components with the ranges of the concentration, the 64 experimental conditions were defined by an orthogonal array (L64). To obtain training data, the recombinant cells were cultivated in the experimental conditions with deep-well plates at 37°C with shaking at 1,400 rpm, and GFP fluorescence was detected by a multiplate reader. Subsequently, deep learning (DL) models were constructed between the medium compositions and the intensities of GFP fluorescence. In case of MOO, NSGA-II (Non-dominated Sorting Genetic Algorithms II) with the trained DL models was applied to search the maximize GFP intensity and minimalize amount of medium as cost index. For comparison, optimization was also performed using a conventional SOO approach based on improving the GFP fluorescence intensity. After the optimization, the algorithm-proposed media were experimentally evaluated.
[Results and Discussion]
The total concentration of medium components ranged from 6.1 to 11.0 g/L in the SOO-optimized media and from 1.78 to 8.22 g/L in the MOO media. Therefore, MOO identified lower-cost media than SOO. Zinc chloride concentrations were mostly zero in the MOO-optimized media, suggesting that incorporating medium cost into MOO reduced less important components, consistent with the DL model sensitivity analysis predicting a minimal contribution of zinc chloride to GFP expression. In addition, even for highly important components such as tryptone and glucose, lower concentrations were observed compared with the control medium. These results indicate that the algorithm effectively identified compositions that balance productivity and material input. Experimental cultivation using the optimized media showed that, whereas the highest GFP fluorescence intensity in the training dataset was 11.3 × 10³ a.u., the SOO-designed medium achieved 13.7 × 10³ a.u. and the MOO-designed medium achieved 14.1 × 10³ a.u. These results demonstrate that GFP expression was improved to a similar extent in both optimized media. Furthermore, the total concentration of medium components in the MOO media was reduced to 70% of that in the control medium. Based on reagent prices, the cost-effectiveness relative to the control increased to 112% of that of the control. Therefore, only MOO was able to achieve both increased GFP expression and reduced medium cost.
[Conclusion]
The present study demonstrated that medium optimization with MOO enables the simultaneous improvement of recombinant protein production and reduction of medium cost, which could not be achieved by SOO. Thus, MOO represents a powerful strategy for cost-effective medium design and offers a promising framework for accelerating bioprocess development.
Comment
To browse or post comments, you must log in.Log in