Session: 20-17-01: Rising Stars of Mechanical Engineering

Paper Number: 173230

Sequential Bayesian Learning on Multi-Scale Probabilistic Knowledge Graph for Biomanufacturing Mechanisms Federated Learning 

Unlike traditional pharmaceuticals, biopharmaceuticals use living organisms, e.g., cells, as factories to provide essential life-saving treatments for severe and chronic diseases (including cancers, metabolic diseases, and infectious diseases such as COVID-19) often with advantages such as increased efficacy and reduced side effects. However, current manufacturing systems have high variability and lack the flexibility to quickly produce existing and new biopharmaceuticals on demand. These issues arise because bioprocessing in biomanufacturing is enormously complex and there is a lack of a deep, systemic understanding of underlying mechanisms. With cells (or other living organisms) as factories, biomanufacturing involves Biological Systems-of-Systems (Bio-SoS) with hundreds of biological, physical, chemical factors dynamically interacting with each other at molecular, cellular, and macroscopic scales and impacting production outcomes. Further, bioprocessing mechanisms are not systematically understood, and data are often very limited, sparse, and heterogeneous. To address these challenges, the proposed sequential Bayesian learning on a multi-scale probabilistic knowledge graph (KG), as a foundation model, for Bio-SoS, enables us to quickly fuse sparse and heterogeneous data collected from different production processes and advance scientific understanding of bioprocessing mechanisms across scales. This KG with a modular design enables us to organize multi-scale Bio-SoS into manageable and reusable building blocks and modules, which facilitates assembling digital twins for various production processes to enhance biomanufacturing systems integration, scalability, and flexibility. In specific, we introduce a novel sequential Bayesian inference framework, called Langevin diffusion-based linear noise approximation (LD-LNA), on a general multi-scale probabilistic KG as mechanistic or hybrid (mechanistic + statistical) foundation model in stochastic differential equations (SDEs) form with a modular design capable of flexibly representing spatial-temporal causal interdependencies from molecular to cellular to macroscopic scales. Built on domain knowledge and structure information on bioprocessing nonlinear dynamics accounting for molecule-to-molecule interactions, the proposed Bayesian learning framework integrates physics-informed mechanisms, i.e., Langevin diffusion (LD) and linear noise approximation (LNA), to efficiently fuse heterogeneous data and generate posterior samples quantifying mechanistic parameter estimation uncertainty and predicting latent state variables. LD takes advantage of the gradient information from the derived posterior of mechanistic parameters to speed up the learning compared with classic Bayesian inference and sampling approaches, while LNA helps bypass the difficulty in selecting the step size for solving SDEs and inferring heterogeneous latent states. The proposed LD-LNA provides fast provable convergence accelerating digital twin development and mechanisms learning for biomanufacturing systems. Numerical experiments inspired by real-world biomanufacturing problems demonstrate the effectiveness of the proposed framework, showing its potential to guide the most informative data collection to reduce model uncertainty and support optimal robust interpretable decision making.

Presenting Author: Wei Xie Northeastern University

Presenting Author Biography: Prof. Xie is an associate professor in Mechanical and Industrial Engineering at Northeastern University. She received PhD degree from Northwestern University 2014. Prof. Xie’s research interests focus on computer simulation, reinforcement learning, multi-scale bioprocess modeling and optimal design/control, federated learning, and design of experiments. She has published more than 70 peer-reviewed papers. Prof. Xie received 2025 NSF CAREER award, the 2015 Outstanding Publication Award from the INFORMS Simulation Society, and the 2023 Outstanding Translational Research Faculty Award from Northeastern University College of Engineering. Her research on process analytical technologies for biomanufacturing has been reported by a well-known News Magazine, Genetic Engineering & Biotechnology News, in healthcare and biopharmaceutical industry. Dr. Xie is an associate editor for INFORMS Journal on Computing and ACM Transactions on Modeling and Computer Simulation. She also serves as Northeastern University representative Technical Activity Committee for National Institute for Innovation in Manufacturing Biopharmaceuticals (NIIMBL).

Authors:

Wandi Xu Northeastern University
Wei Xie Northeastern University