http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
- 中文 을 입력하시려면 zhongwen을 입력하시고 space를누르시면됩니다.
- 北京 을 입력하시려면 beijing을 입력하시고 space를 누르시면 됩니다.
RISS 인기검색어
- 검색키워드
- 저자 : Massaki KONISHI (검색결과 1 건)
- 무료
- 기관 내 무료
- 유료
-
Culture Medium Profiling and Design Assisted by Machine Learning
Massaki KONISHI 한국생물공학회 2021 한국생물공학회 학술대회 Vol.2021 No.10
Production media for microbial cultivation is a significant factor to perform efficient cell growth and production. In industrial fermentation processes, raw materials including yeast, malt, and meat extracts, peptone, molasses, and the other agricultural and industrial waste are often used. The compositions can be varied in seasonal and geological varieties, brands, and production-lots, and influence microbial growth and material production. To understand the varieties of medium components, we have suggested that an analytical method assisted by metabolomics-like finger printing using gas-chromatogram mass spectrometry (Tachibana et al. 2019), and by deep neural networks (DNN) architecture (Tachibana et al. 2021) can apply to culture medium profiling. In the study, it was analyzed as a typical microbial cultivation that various brands of yeast extract were influenced to Escherichia coli growth and green fluorescent protein (GFP), a model of foreign protein production. According to our procedure, the bacterial growth and protein production was accurately estimated from the initial medium components profiles measured by GC-MS. Furthermore, significant components were estimated by a permutation algorithm using DNN model. The results indicated that the initial medium components can sufficiently explain the cultivation results including growth and protein production. As well, bioethanol production can be explained by the composition of toxic materials in lignocellulosic hydrolysates (Watanabe et al. 2019; Konishi 2020). To design optimal culture medium for engineered E. coli producing GFP, L81 Latin square design with 3 levels was applied to minimal medium M9 with supplemental components including amino acids and vitamins. To compare suitable machine learning algorithms for estimating GFP production, 12 algorithms, linear regression (LR), Ridge regression (Ridge), Lasso regression (Lasso), support vector machine (SVM), partial least square regression (PLS), decision tree regression (dtree), random forest regression (RFR), neural networks (NN), deep neural networks (DNN), Gradient tree boosting regression (gbr), K neighbor regression (kbr), and voting regression (vtr) were applied. According to evaluate the algorithms by cross validation (supervised: 85% and validation: 15%), although mean square errors between measured and estimated values of test data (MSEtest) were approximately 1.0-1.5 in case of LR, Ridge, Lasso, SVM, and PLS, those of dtree, RFR, gbr, NN, and DNN were in range below 0.06. On the other hand, considering interaction terms of independent variables, the data accurately fit to the all tested algorithms. The MSEtest were in range between 0.03 and 0.07. The results meant that interaction among medium components were strongly influenced to GFP production. Gaussian process optimization using trained DNN model as objective function were applied to exploring the optimal medium composition for GFP production. Based on the experimental confirmation, the improved composition increased GFP fluorescence to 117% against the best in the original experimental dataset in fact. The machine-learning-associated optimization of culture medium can provide high-throughput explore of the optimal medium compositions for not only microbial culture but also mammalian culture in theoretically. Furthermore, the idea will contribute to promote digital transformation for wide range of bioproductions. This research was partly supported by New Energy and Industrial Technology Development Organization (NEDO) project of Ministry of Economy, Trade and Industry (METI), Japan.
연관 검색어 추천
이 검색어로 많이 본 자료
활용도 높은 자료
