[P004] Poster

1:00 PM - 1:50 PM JST(4:00 AM - 4:50 AM UTC)

[P02-178] Feasibility Study of Machine Learning Data Reconstruction with Compartmental kinetic Modeling for Optimal Biotransformation of Microbial Fermentation of Perilla frutescens

○BorYann Chen¹ , ShiZhen Hong¹, ShinRu Leng¹, KengWei Liu¹, ChengYang Hsieh¹, PoWei Tsai², Adityas Agung Ramandani³, KuanShiong Khoo³, ChungChuan Hsueh¹, (1. Department of Chemical and Materials Engineering, National I-Lan University, I-Lan 260, Taiwan (Taiwan), 2. Department of Food Science, National Taiwan Ocean University, Keelung 202, Taiwan (Taiwan), 3. Department of Chemical Engineering and Materials Science, Yuan Ze University, Taoyuan320, Taiwan (Taiwan))

Keywords:

Microbial fermentation,Machine learning,Compartmental kinetic modeling

This study proposed a hybrid AI framework that integrated machine learning (ML) of data reconstruction and compartmental kinetic modeling to “blind review” predict transient dynamics of microbial fermentation. The study investigated the serially-acclimated fermentation of Perilla frutescens extract by Lactobacillus curvatus through ML-based time-series analyses upon transient evolution of rosmarinic acid (RA) and its metabolites (e.g., caffeic acid (CA)) for system optimization.Due to non-growth-associated characteristics of RA degradation, compartmental kinetic modeling was first employed to indicate transient dynamics of RA via serially-acclimated fermentation, effectively revealing behaviors of derived intermediates and products. Next, Support Vector Regression (SVR) was competitively introduced as the ML-based method of data reconstruction to capture non-linear characteristics of practical data with effective prediction of missing data even lack of complete information of cellular metabolism. The findings indicated that at 500 mg L^-1 RA, the fourth cycle of acclimation could minimize area under the curve (AUC) for maximal performance of degradation of RA. Although AUC values obtained from ML-reconstructed data and incomplete experimental measures were 10.370 and 8.982, respectively, the predicting power of system dynamics still reveal consistent rakings and optimization for serial metabolic acclimation. Compared to others ML methods, SVR model seemed to be more promising for continuous time-series data of P. frutescens fermentation and overall optimization possibly due to SVR more appropriate for this dynamics data. This suggested that the feasibility of ML-assisted microbial fermentation would be technically promising. Apparently, the leading performance indices strongly depended upon nature of raw data set for ML. In addition, microbial fuel cell (MFC) was utilized as a platform of bioenergy expression to indicate overall performance of operation. The results showed a decrease in MFC power after fermentation (ca. 16% reduction of power generation), clearly suggesting the degradation of electron-shuttle-related compositions. HPLC-MS analysis also provided further validation for biotransformation of RA into various derivatives. Effective increases of acetylcholinesterase (AChE) inhibitory activities of fermented samples evidently suggested considerable stimulation of neuroprotective properties through fermentation. In conclusion, through ML-oriented data reconstruction to couple with compartmental kinetic modeling, this first-attempt study indicate significant power for system optimization (e.g., optimal operation and rankings of serial acclimation). This approach could provide potent analytical tool to decipher not well-characterized systems for complex biotransformation via data reconstruction scenarios, and still effectively exhibit appropriate potential for microbial fermentation optimization.