3 Results
3.1 The effect of oil feeding rate on SLs production during fed-batch fermentation
During 5 L fed-batch fermentations, SLs were synthesized with different rapeseed oil feeding rates (high, medium, and low modes) for 168 h. There were no significant differences between the three modes with respect to the titer of SLs, total glucose consumption, and total oil consumption (Fig. 2A, Table 1), however, the phased synthesis rate of SLs differed substantially during the fermentation process. The higher the oil feeding rate, the faster the accumulation of SLs in the early stages of fermentation, although the accumulation rate decreased during the middle and late stages. We also found that the phased productivity of SLs differed significantly between the three modes (Fig 2B). The highest productivity of SLs was 3.53 g/L/h under the high mode, while the highest productivity in the medium and low modes was 2.63 and 2.09 g/L/h, thus representing reductions of 25.5% and 40.8%, respectively. Notably, after 48 h, the productivity of SLs declined rapidly under the high mode until 110 h and was then maintained at 0.98 g/L/h to the end of the fermentation process. In contrast, the productivity of SLs under the low mode was maintained at an average of 2.04 g/L/h until 120 h and then began to decrease. As shown in Fig 2C, the difference between the high and low modes reached a maximum of 139.7 g at around 87 h, while the maximum difference between the high and medium modes was 85.7 g at 73 h. Otherwise, there were significant differences in the OUR and CER profiles when compared between the three modes during the fermentation process. In high mode, the maximum OUR reached up to 64 mmol/L/h, gradually decreased after 48 h, and was then maintained at 47 mmol/L/h after 110 h. In contrast, the OUR in low mode was maintained at approximately 54 mmol/L/h for 120 h. The CER showed a similar trend to OUR (Fig 2D).
3.2 Cellular metabolic characteristics in different stages of SLs production
The fed-batch fermentation process underlying the production of SLs predominantly consisted of two phases: the primary metabolism of cell growth and the secondary metabolism of SLs synthesis. We found that the maximum productivity of SLs was related to the oil feeding rate and that the highest value was 3.53 g/L/h. However, a further increase in the oil feeding rate did not enhance SLs productivity, meaning that the maximum specific productivity for SLs reached 0.15 g/gDCW/h due to limitations imposed by the production capacity of the cells. Moreover, we observed that the production of SLs in the shake flask began to decrease significantly (by 15.6%) when the concentration of SLs exceeded 75 g/L (non-normalized) (Fig. 3A); at this point, the productivity of SLs in the fermenter under the three modes also gradually decreased, thus inferring that product inhibition may have occurred. Furthermore, the degree of inhibition increased with an increase in the concentration of SLs. Interestingly, although the OUR and CER decreased during this phase, the RQ remained unchanged, thus showing that product inhibition only limited the metabolic capacity of the cells, while the metabolic pathway remained unchanged. Notably, when the concentration of SLs in the fermentation broth exceeded 140 g/L (non-normalized), although the OUR and CER continued to decrease with the decline of SLs productivity, the RQ began to increase from approximately 0.70 to 0.76 (Fig. 3B). This result demonstrated that the metabolic pathways of cells underwent significant changes at this time. By further analyzing the viscosity of the fermentation broth, we found that the viscosity of the fermentation broth would rise sharply during the middle and late phases of fermentation, and when the concentration of SLs exceeded 140 g/L (non-normalized), the viscosity exceeded 20 cP (Fig. 3C).
The fermentation of SLs is a high oxygen consumption process. High viscosity can exert serious effects on oxygen supply, thus resulting in alterations in cell metabolism and the synthesis of SLs. Further analysis of the RQ profiles demonstrated that when the average RQ was approximately 0.70 (high mode: 24-96 h; medium mode: 24-110 h; low mode: 24-120 h), the average yields of SLs to glucose and rapeseed oil were 1.21 g/g and 1.39 g/g, respectively. However, when the concentration of SLs exceeded 140 g/L (non-normalized) (high mode: after 96 h; medium mode: after 110 h; low mode: after 120 h), the average yields of SLs to glucose and rapeseed oil changed to 0.82 g/g and 1.46 g/g, respectively, and the average RQ reached 0.76 (Table 2). This implied that once the oxygen supply became limited, the yield of SLs to glucose was reduced, while the yield to rapeseed oil was increased, thus leading to an increase in RQ value.
In summary, the synthesis of SLs during fed-batch fermentation could be divided into three phases (Fig. 4). During the initial phase, the synthesis of SLs was mainly limited by the production capacity of the cells as well as the oil supply. The maximum specific SLs productivity could reach up to 0.15 g/gDCW/h but was affected by the oil feeding rate. During the second phase, when the concentration of SLs in the fermentation broth exceeded 75 g/L (non-normalized), the production capacity of the cells began to be inhibited by high product concentration. The higher the concentration of SLs, the stronger the degree of inhibition. However, cellular metabolic pathways remained unchanged during this phase. During the last phase, the concentration of SLs reached 140 g/L (non-normalized) and the production capacity of the cells became significantly restricted by oxygen supply. On the basis of these analyses, the first and second phases were considered to represent the main stages of SLs accumulation during fed-batch fermentation. Increasing the productivity of SLs as much as possible during these phases via the fine control of substrate feeding strategy was of great significance to the improvement of production efficiency. In terms of the third phase, although it was difficult to improve production efficiency without increasing oxygen supply during the conventional fermentation mode, the semi-continuous fermentation mode, involving thein-situ separation of SLs, which was regulated by the concentration of oil in the fermentation broth (Chen et al., 2021b), could effectively alleviate product inhibition and significantly reduce the viscosity of the fermentation broth to eliminate oxygen limitation (Liu et al., 2019).
3.3 Construction of an intelligent feedback feeding model based on multi-parameter process data
Next, we constructed and optimized an efficient, stable, and rational fermentation process for SLs production by regulating feeding during the first and second phases. First, we established correlations between the process macroscopic parameters and the production capacity of cells. All the on-line parameters, including agitation, temperature, pH, DO, glucose concentration, oil concentration, SLs concentration, OUR, CER, and RQ were applied as independent variables. The oil consumption rate, glucose consumption rate, and SLs productivity were set as dependent variables. Then, six linear and non-linear mathematical equations were adopted to establish the correlation between independent and dependent variables, including unary linear regression equations (LR), multiple linear regression equations (MLR), support vector machines (SVMs), partial least squares (PLS), random forest algorithms (RFs) and gradient boosting regression algorithms (GBRs). Finally, the pros and cons of the mathematical equations were compared by the correlation coefficient (R2) between the predicted values and the real values (Fig. 5). Approximately 110 sets of data were used as the training set, and 20 sets of data were used as the testing set. Three dependent variables relating to SLs (productivity, oil consumption rate, and glucose consumption rate) were all strongly correlated with the actual values when applying the six equations (R2> 0.96), thus showing that these six algorithms accurately reflected the relationship between the consumption of substrates and the synthesis of product and fermentation process parameters (Table 3). Notably, the R2 for the RF and GBR algorithms decreased slightly after the testing set was added, while the R2 for the other four equations increased as the number of parameters increased. It is possible that with the continuous input of parameters, the database of equations also increased; consequently, the accuracy of the predicted value would also be improved. Non-linear equations might be associated with over-fitting problems; therefore, when there is less data, the predicted values of ULR, MLR, SVM, and PLS, were more reliable. In contrast, with a continuous increase in sample number, the prediction values for RF and GBR became more accurate.
When the model produced a predicted value (SLs productivity, glucose consumption rate, and oil consumption rate) with an optimal R2, then the model would add data to the database, thus increasing the data sample. This feedback feeding model was combined with the data provided by on-line sensors. This meant that the model could read the on-line data, perform equation fitting, and output the optimal predicted value by the computer. By integrating the working volume, the required concentrations of glucose and oil, as well as the amounts of consumed glucose and oil, the model could automatically guide the precise feeding rate via a computer-controlled feedback module. Simultaneously, the model performed data overlay and self-learning based on the continuous accumulation of data, thereby further improving the accuracy of the model.
The feedback feeding model was validated in a 5 L fermenter. As shown in Fig. 6A and B, the concentration of residual oil was well controlled at 2 g/L and 10 g/L respectively, and the concentration of residual glucose was maintained at 40 g/L from 24 h until the RQ was increased to end the fermentation process. In addition, by applying the feedback feeding model, the SLs titer with a residual oil concentration at 2 g/L was improved by 4.9% during the first and second phases of fed-batch fermentation, when compared to the high oil feeding mode. More importantly, the feedback feeding model shortened the feeding interval from 6 to 12 h by conventional manual intervention to 1 h by automatic feeding. The model also reduced the error related to feeding control from more than 15% to less than 5%, thus providing significant savings in terms of labor and time costs.
3.4 An intelligent feedback feeding model facilitated semi-continuous fermentation for the efficient production of SLs
To alleviate oxygen limitation in the late phase of fed-batch fermentation, we used gravity sedimentation for the in-situ separation of SLs, thus permitting semi-continuous fermentation. When the concentration of SLs reached 140 g/L (non-normalized), we began the in-situ separation process and allowed this to proceed until the concentration of SLs fell below 60 g/L, thus alleviating the influence of a high SLs concentration on the rheological properties, as well as relieving product inhibition. As illustrated in Fig. 7A, the feedback feeding model could stably control the concentrations of residual glucose and residual oil at appropriate levels during the entire fermentation process, so as to ensure the efficient in-situ separation of SLs (Chen et al., 2021b). After two rounds of in-situ separation of product, the total production of SLs reached 1348.3 g within 234 h. Moreover, the overall SLs productivity and yield were 2.30 g/L/h and 0.57 g/g, respectively, thus representing increases of 40.2% and 18.7% in comparison to the fed-batch fermentation with a high oil feeding mode (Fig. 7B). Notably, the productivity of SLs did not exhibit a downwards trend at the end of semi-continuous fermentation (Fig. 7C). Therefore, it was reasonable to speculate that the production efficiency could be further improved by performing longer operations with the assistance of an intelligent feedback feeding model.