融合计划采样的增强物理信息神经网络用于单克隆抗体细胞培养过程建模

An improved physics-informed neural network with scheduled sampling for modeling the cell culture process for production of monoclonal antibodies

摘要:

物理信息神经网络(physics-informed neural networks, PINN)将机理方程融入神经网络以提升模型预测精度，已初步应用于细胞培养过程建模。虽然PINN凭借物理规律对解空间进行约束减少了对数据量的依赖，但在小样本场景下，仍有可能收敛到次优解。本研究提出了一种改进的混合物理信息神经网络(improved hybrid-physics-informed neural network with scheduled sampling, SS-IHPINN)，通过物理引导初始化和计划采样模块为单克隆抗体细胞培养过程建立更稳健的模型。通过对比实验和消融实验对SS-IHPINN进行了评估和比较，结果表明这2个功能模块的协同作用使SS-IHPINN具有更高的预测性能。即使细胞培养过程参数发生变化，该框架也表现出良好的泛化和迁移学习能力。本研究提出的方法有效提升了小样本条件下的建模性能，为生物过程的监测提供了可靠的技术支撑。

Abstract:

Physics-informed neural networks (PINNs) integrate mechanistic equations into neural networks to enhance the predictive accuracy of models, and has found initial application in cell-culture modeling. While PINNs reduce data dependency by constraining the solution space with physical laws, it can still converge to suboptimal solutions in small-sample scenarios. This paper proposes an improved hybrid-physics-informed neural network with scheduled sampling (SS-IHPINN) to build a more robust model for the cell culture processes for producing monoclonal antibodies through physics-guided initialization and scheduled sampling. The proposed model was evaluated and compared through comparative and ablation experiments, which demonstrated that the two functional modules worked synergistically to endow SS-IHPINN with higher predictive performance. Even with variations in cell culture process parameters, the framework exhibits good generalization and transfer learning capabilities. The method proposed in this study effectively enhances modeling performance under small-sample conditions, providing reliable technical support for bioprocess monitoring.