【作者】 丁健；

【导师】 史仲平；

【作者基本信息】 江南大学 ， 发酵工程， 2014， 博士

【摘要】 绝大多数的大宗发酵产品，如氨基酸、生物酶、生物药物蛋白等均依靠好氧发酵生产。除了选育优良菌种外，发酵过程控制技术也是提高发酵性能的重要手段。好氧发酵过程存在许多共性问题，例如：传统离线控制难以取得满意的效果，而许多重要的发酵参数又无法在线测量，实施发酵过程在线控制存在困难；发酵故障时有发生，严重影响发酵过程的稳定性和经济性；供氧压力巨大、溶解氧浓度（DO）难以控制、通纯氧发酵大大增加了发酵操作的成本和不安全因素。针对上述问题，本论文以利用两种典型的、好氧型大宗产品的发酵菌种（重组毕赤酵母和谷氨酸棒杆菌）生产药物蛋白和谷氨酸的过程为研究对象，深入研究了基于人工智能和代谢调控的发酵过程预测、在线控制和故障诊断的关键技术，旨在为好氧大宗产品发酵搭建共性平台，提高目标产物的浓度、发酵稳定性和发酵过程的整体性能。论文的主要研究内容总结如下：（1）利用毕赤酵母生产猪α干扰素（pIFN-α）的发酵过程中，发酵性能指标、pIFN-α抗病毒活性（pIFN-α-AVA）难以测量，发酵性能无法预测。为此，以易测量的发酵参数（诱导时间/温度、DO、O2消耗速率OUR、CO2释放速率CER、甲醇消耗速率、总蛋白浓度）为输入，以pIFN-α-AVA为输出，研究比较了基于多项式回归和人工神经网络（ANN）模型（传统BP-ANN和基于遗传算法的改良型ANN）的pIFN-α-AVA的预测性能。基于遗传算法的改良型ANN模型具有最准确的预测和泛化能力。分析pIFN-α-AVA对各发酵参数的感度发现，CER、OUR和甲醇消耗速率对pIFN-α-AVA影响最大。（2）利用MutS型毕赤酵母生产pIFN-α时，发酵性能不稳定。在甘油流加培养期，使用传统DO-Stat法流加甘油，存在乙醇生成积累的现象。如果细胞长时间地（>4h）处在高乙醇浓度（>6g L-1）的环境下，甲醇诱导无法启动。分析测定甲醇代谢途径关键酶的基因转录水平和酶活，结果发现并证实：培养期内、乙醇长时间处在高浓度下，醇氧化酶（AOX）的活性和基因转录水平不可逆转地受到抑制，这是pIFN-α发酵不稳定的最主要原因。利用商业化甲醇电极可以在线测量乙醇、并据此提出了基于乙醇在线测量的改良型自适应的DO-Stat甘油流加控制策略。该策略可以同时利用甘油和积累的乙醇、将乙醇浓度稳定在线地控制于低水平（约2g L-1），且不影响细胞生长速度。无乙醇浓度控制批次的最高pIFN-α浓度仅达到2.08g L-1，且不稳定。而只要乙醇浓度得到控制，pIFN-α浓度均可稳定在2.70~3.65g L-1的高水平，且受后续诱导条件的影响较小。pIFN-α浓度得到提高、发酵性能不稳定问题得到根本解决。（3）使用具有强AOX启动子的Mut+表型重组毕赤酵母发酵生产猪圆环病毒Cap蛋白时，甲醇利用效率低、耗氧严重、DO难以控制、Cap蛋白表达水平低。利用DO对添加等量甲醇和山梨醇的时间响应存在差异的特征现象，提出并构建了一种基于DO在线测量和模糊推理技术的甲醇/山梨醇共混流加自动控制系统。使用该控制系统、并将DO控制于10%时，甲醇代谢途径中的几个关键酶的活性均明显提高、甲醇流加速度得到适度限制、供氧压力得到缓解、DO可以稳定控制在期望水平。此时，甲醇利用效率显著改善、胞内毒副中间产物的积累缓解；Cap蛋白浓度达到198mg·L-1，比使用甲醇诱导策略最优批次的相应值121mg·L-1明显提高；诱导可在通空气条件下进行，整体发酵性能大幅改善。（4）利用生物素缺陷型谷氨酸棒杆菌发酵生产谷氨酸时，培养基中的生物素浓度波动严重影响发酵性能的稳定。为此，提出了基于智能型模式识别的发酵故障诊断系统。在某一时间窗口内、利用支持向量机（SVM）分类器对可在线测量的过程参数（发酵时间、搅拌转速、耗氨速率、OUR、CER）进行状态分类，并结合模糊推理技术将发酵状态划分成生物素“不足”、“适中”、“过量”的三种模式，构建了智能型的故障诊断系统。通过刻意制造“不足”、“适中”和“过量”的初始生物素浓度环境，获取大量、对应于不同发酵状态模式的数据，建立以生物素初始浓度和发酵时间为输入、其他可在线测量参数为输出的人工神经网络模型，自动生成大量数据对（datapairs），对上述故障诊断系统的有效性进行了仿真模拟研究。模拟结果表明：该故障诊断系统可以将不同初始生物素浓度下的发酵批次进行合理聚类，对发酵状态具有良好的识别判断能力。（5）将上述故障诊断系统实用于谷氨酸发酵过程。在初始生物素含量“适中”的条件下，该在线故障诊断系统自始至终没有发出预警信号；而在初始生物素含量不当的情况下，该系统均可以在发酵初期（6~8h）准确地识别判别出故障类型，并通过补加纯生物素或吐温40、对“错误”发酵批次进行补救。通过有效的故障识别和补救，该系统对“错误”发酵批次的识别结果逐步返回到生物素含量“适中”的范围内，所有发酵批次的谷氨酸最终浓度均达到75~80g·L-1的正常水平，发酵稳定性显著改善。更多还原

【Abstract】 Most of the bulk fermentation products, such as amino acids, enzymes, drug proteins, etc.are produced through aerobic fermentations. Besides strains breeding or screening, processcontrol technique is an alternative way for fermentation performance improvement. Aerobicfermentations suffer with many common problems, for example, off-line control has limitedpower in performance improvement, and implementing of on-line process control is difficultas many crucial fermentation parameters can not be on-line measured; the occurrence offermentation faults severely deteriorate fermentation stability and economics; extremely highoxygen supply load causes difficulty in dissolved oxygen concentration (DO) control and pureoxygen based aeration which largely increases operation cost and safety risk. Focusing onthose problems, in this thesis, drug proteins and glutamate fermentations by two typicalaerobic strains, recombinant Pichia patoris and Corynebacterium glutamicum, were used asthe prototypes, and the key techniques of intelligent and metabolic engineering based on-linestate prediction, adaptive control and fault diagnosis were investigated. The purpose ofdeveloping such key techniques is to supply a technological platform to the typical aerobicfermentation processes, to effectively improve targeted product concentrations, fermentationstabilities and performance. The major results of the dissertation were summarized as follows:(1) In porcine interferon-α (pIFN-α) production by Pichia pastoris, fermentationperformance is hardly to predict, as pIFN-α antiviral activity (pIFN-α-AVA) was difficult tomeasure. Prediction performance of various models, including polynomial regression basedmodel and ANN models (BP-ANN and “improved” ANN optimized by genetic algorithm),were compared, with the easily measurable parameters (induction time/temperature, DO, O2uptake rate OUR, CO2evolution rate CER, methanol consumption rate, and total proteinconcentration) as the inputs and pIFN-α-AVA as the output. Among the models, the improvedANN model indicated the most accurate and universal abilities. Sensitivity analysis suggestedthat CER, OUR and methanol consumption rate were closely correlated with pIFN-α-AVA.(2) pIFN-α production by MutSP. pastoris is unstable. Ethanol accumulation occurred iftraditional DO-Stat method was adopted for glycerol feeding during cultivation phase.Methanol induction could not be initiated if the cells were subject to high ethanolconcentration environment (>6g·L-1) for longer time (>4h). Analyzing the transcriptionallevels and activities of the key enzymes in methanol metabolism routes revealed that, thelong-term and high level accumulation of ethanol during cultivation phase irreversiblyrepressed the activity and transcriptional level of alcohol oxidase (AOX), which deterioratedpIFN-α fermentation stability in turn. Using a commercialized methanol electrode to on-linedetect ethanol, an on-line ethanol measurement based adaptive DO-Stat glycerol feedingstrategy was thus proposed. The strategy could alternatively utilize glycerol and theaccumulated ethanol to control ethanol concentrations at low level (2g L-1) without cellgrowth rate deterioration. The maximal pIFN-α concentration in runs without ethanolconcentration control was2.08g L-1, but unstable. pIFN-α concentrations in runs with low ethanol accumulation could be stabilized at higher levels of2.70~3.65g·L-1.Fermentationperformance could be successfully stabilized and enhanced.(3) Porcine circovirus Cap protein production by Mut+P. pastoris with strong AOXpromoter suffered with the problems: inefficient methanol utilization, extensively highoxygen supply requirement, difficulty in stably controlling DO, very low Cap protein titer.Based on DO response patterns against the addition of methanol or sorbitol with equal amount,a DO on-line measurement and fuzzy logical inference based automatic methanol/sorbitolco-feeding control system was proposed. With the aid of this control system by setting DOcontrol level at10%, key enzymes in methanol metabolism route were largely activated;methanol feeding rate was restricted at a moderate level; oxygen supply load was relieved;and DO was stably controlled at the desired level. Under this condition, methanol utilizationefficiency was greatly improved and accumulation of toxic intermediate metabolites wasrelieved. Cap protein concentration reached a level of198mg·L-1, which was much higherthan the maximal value of121mg·L-1in the runs by methanol induction. Induction could beconducted by air-aeration and overall fermentation performance enhanced significantly.(4) In glutamate fermentation by biotin-auxotroph C. glutamicum, the biotin contentvariation in culture medium severely affects fermentation stability. An intelligent patternrecognition based fault diagnosis system was thus proposed to solve the problem. A supportvector machine classifier (SVM) was firstly used to categorize the on-line measurableparameters (fermentation time, agitation rate, NH3consumption rate, OUR and CER) within amoving-window, and then the SVM was combined with fuzzy reasoning technique toconstruct an unique intelligent fault diagnosis system, which could classify the fermentationphysiological states into3catagories of biotin “in shortage”,“medium”, and “in excess”. Byexperimentally varying the initial biotin content to intently creat the three differentfermentation states and obtaining a large amount corresponding data, an artificial neuralnetwork (ANN) model was developed with initial biotin content and fermentation time as theinputs, the other on-line measurable parameters as the outputs. This ANN model couldautomatically generate a large amount of data pairs to simulate the effectiveness of the faultdiagnosis system. Simulation results indicated that the fault diagnosis system could clusterfermentation runs corresponding to different initial biotin content well and had goodfermentation states recognition ability.(5) The fault diagnosis system was applied to glutamate fermentation process for on-linefault diagnosis. In the cases of initial biotin content “medium”, the system did not send anywarning signals throughout the fermentations; while in the cases of improper initial biotincontents, the system could accurately identify the faults and their type in the earlyfermentation stage (6~8h), and rescue those failure-likelihood fermentations by adding purebiotin or tween40. After the rescue measures were taken, outputs of the system weregradually recovered to the normal range, final glutamate concentrations in all testing runsreached to normal levels of75~80g·L-1, and fermentation stability significantly enhanced.更多还原