基于机器学习的固态发酵非粮生物质产饲料蛋白条件优化

Machine learning-optimized solid-state fermentation of non-grain biomass for enhanced feed protein production

摘要:

我国饲用蛋白资源高度依赖进口，充分挖掘非常规蛋白资源有望从根源上解决我国饲用蛋白短缺的问题。然而，目前非粮生物质发酵产饲料蛋白过程多参数复杂耦合，传统发酵条件优化方法难以捕捉各变量间非线性关系，效率较低。本研究旨在通过机器学习构建精准高效的预测模型优化发酵条件，基于前期非粮生物质固态发酵产饲料蛋白的数据集，以发酵时间、温度、pH值、氮源添加量等13种发酵条件为变量，发酵后真蛋白含量为响应变量，分别采用留出法和留一交叉验证法，对比了3种线性模型和3种非线性模型对真蛋白产量的预测性能。结果表明，在留一交叉验证法下，非线性模型对非粮生物质固态发酵产饲料蛋白过程预测性能优于线性模型，其中类别型特征提升(categorical boosting, CatBoost)模型的决定系数R²高达0.84，均方根误差(root mean square error, RMSE)为0.98，具有最佳预测性能。利用训练的CatBoost模型，对固态发酵产饲料蛋白过程中8种变量设置步距，比较了8 225 290组发酵条件的真蛋白产量，发现3种底物和菌群组合的预测真蛋白含量高于数据集中的最高值。采用对氨预处理小麦秸秆和驯化菌群组合预测的最佳条件进行固态发酵实验，真蛋白含量为16.51%，比优化前提升了18.43%，同时发酵时间缩短了108 h，氮源添加量减少了40%，能大幅降低生产成本。本研究通过机器学习有效提高了固态发酵产饲料蛋白过程优化效率，将为开发非粮生物质资源、缓解我国饲料蛋白短缺奠定理论和应用基础。

Abstract:

The heavy reliance on imports for China’s feed protein resources necessitates the exploitation of unconventional protein sources to fundamentally address the domestic supply shortage. However, the bioconversion of non-grain biomass into microbial feed protein is characterized by complex multi-parameter coupling. Traditional optimization methods often struggle to decipher the intricate non-linear relationships among variables, resulting in limited efficiency. To overcome these bottlenecks, this study aims to develop high-fidelity predictive models based on machine learning for the precise and efficient optimization of fermentation conditions. To address these shortcomings, we employed machine learning for process optimization. On the basis of an existing dataset from solid-state fermentation of non-grain biomass for feed protein production, we used 13 fermentation parameters, including time, temperature, pH, and nitrogen source input, as input variables, with the true protein yield as the response variable. We used both hold-out and leave-one-out cross-validation methods to compare the predictive performance of three linear and three nonlinear models. Our results demonstrated that under leave-one-out cross-validation, nonlinear models outperformed linear models in predicting the true protein yield. The CatBoost model achieved the best performance, with a coefficient of determination (R²) of 0.84 and a root mean square error (RMSE) of 0.98. Using the trained CatBoost model, we evaluated 8 225 290 fermentation condition combinations by setting stepwise variations for eight key process variables. The model predicted maximum true protein yields for three specific substrate-microbial consortium combinations that exceeded the highest experimental values recorded in our original dataset. We then conducted experimental validation under the model-optimized conditions for the combination of ammonia-pretreated wheat straw and an acclimated microbial consortium. This experiment achieved a true protein yield of 16.51%, which represented a 18.43% improvement over that of the pre-optimization level. Furthermore, the optimized process reduced fermentation time by 108 h and nitrogen source input by 40%, indicating significant potential for cost reduction. This study demonstrates that machine learning can effectively improve the optimization efficiency of solid-state fermentation for feed protein production, thereby laying solid theoretical and practical foundations for developing non-grain biomass resources and alleviating the feed protein shortage in China.