陈煜成

Hybrid models integrate mechanistic and data-driven components, effectively addressing the challenges of limited process understanding and data availability typical of biopharmaceutical processes.

In this study, we applied a hybrid modeling framework named differentiable physics solver-in-the-loop (DP-SOL) to describe the reversed-phase chromatographic purification of an oligonucleotide, overcoming the mentioned limitations of purely mechanistic and data-driven models. The framework establishes a connection between neural networks (NNs) and mechanistic models through differentiable physical operators and their gradients. We first collected a data set comprising six linear gradient elution experiments at different resin loadings and gradient slopes, split in three experiments each for training and testing, for few-shot learning. The hyperparameters were determined through a grid search, resulting in a NN with two hidden layers and 14 nodes. Compared to a calibrated mechanistic model used for initialization of NN, the DP-SOL hybrid model showed significant performance improvement on both training and testing sets, with R2>0.97 for the former. The good predictivity of DP-SOL is attributed to the combination of mechanistic models and NNs at the solver level.

As a novel and versatile hybrid modeling paradigm, DP-SOL has the potential to significantly impact modeling approaches in the downstream processing field and the broader biopharmaceutical sector.

A hybrid (gray-box) modeling framework, pysics-informed neural networks (PINNs), has garnered significant attention. However, a huge challenge in applying PINNs to bioprocesses is developing a loss function that synergizes different bioprocess dynamics.

To mitigate this challenge, a novel physics-based deep learning method was developed by integrating order-of-magnitude analysis to the loss function of PINNs (oPINNs) using biological first principles. Compared to standard PINNs and numerical methods for solving the forward problem of linear chromatographic models, oPINNs demonstrated notable improvements: an order-of-magnitude enhancement in accuracy with an equivalent sample size, or a 32-fold reduction in sample size for equivalent accuracy, along with a 1000-fold acceleration in computational speed for millisecond-scale simulation. Moreover, oPINNs showed exceptional robustness in weight determination and hyperparameter selection amidst variations in chromatographic model parameters.

In summary, oPINNs represent a significant advancement in integrating physics-based deep learning into hybrid modeling of bioprocesses, particularly for developing real-time digital twins.

Development of a next-generation chromatographic model, capable of simultaneously meeting academic demands for thermodynamic consistency and industrial requirements in everyday project work, has become a focal point of research.

In this study, anti-Langmuirian to Langmuirian (AL-L) elution behavior was observed in cation-exchange chromatographic separation of charge variants of industrial Fc-fusion proteins. To characterize this behavior, the multi-protein Mollerup activity model was integrated into the steric mass action (SMA) model, resulting in a new model named the generalized ion-exchange (nGIEX) isotherm for multi-protein systems. An R2 exceeding 0.95 calibrated by three elution experiments indicates an effective description of the AL-L behavior (dynamic adsorption). Using isotherm sampling, the nGIEX model exhibited sigmoidal AL-L isotherms (static adsorption). Finally, the model’s extrapolation capability was externally validated through process optimization, resulting in an optimal two-step elution condition and a yield improvement of the main variant from 25.9% to 89.1% within purity specifications (>70%).

Regulatory authorities recommend using residence time distribution (RTD) to address material traceability in continuous manufacturing. Continuous virus filtration is an essential but poorly understood step in biologics manufacturing in respect to fluid dynamics and scale-up.

Here we describe a model that considers nonideal mixing and film resistance for RTD prediction in continuous virus filtration, and its experimental validation using the inert tracer NaNO(3). The model was successfully calibrated through pulse injection experiments, yielding good agreement between model prediction and experiment ( R2 > 0.90). The model enabled the prediction of RTD with variations-for example, in injection volumes, flow rates, tracer concentrations, and filter surface areas-and was validated using stepwise experiments and combined stepwise and pulse injection experiments. All validation experiments achieved R2 > 0.97. Notably, if the process includes a porous material-such as a porous chromatography material, ultrafilter, or virus filter-it must be considered whether the molecule size affects the RTD, as tracers with different sizes may penetrate the pore space differently. Calibration of the model with NaNO(3) enabled extrapolation to RTD of recombinant antibodies, which will promote significant savings in antibody consumption.

This RTD model is ready for further application in end-to-end integrated continuous downstream processes, such as addressing material traceability during continuous virus filtration processes.

Ion exchange chromatography (IEC) is one of the most widely-used techniques for protein separation and has been characterized by mechanistic models. However, the time-consuming and cumbersome model calibration hinders the application of mechanistic models for process development.

A new methodology called "parameter-by-parameter method (PbP)" was proposed with mechanistic derivations of the steric mass action (SMA) model of IEC. The protocol includes four steps: (1) first linear regression (LR1) for characteristic charge; (2) second linear regression (LR2) for equilibrium coefficient; (3) linear approximation (LA) for shielding factor; (4) inverse method (IM) for kinetic coefficient. Four SMA parameters could be one-by-one determined in sequence, reducing the number of unknown parameters per species from four to one, and predicting almost consistent retention. Numerical single-component experiments were investigated firstly, and the PbP method showed excellent agreement between experiments and simulations. The effects of loadings on the PbP and Yamamoto methods were compared. It was found that the PbP method had higher accuracy and robustness than the Yamamoto method. Moreover, a five-experiment strategy was suggested to implement the PbP method, which is straightforward to reduce the cost of calibration experiments. Finally, a real-world multi-component separation was challenged and further confirmed the feasibility of the PbP method.

In general, the proposed method can not only reliably estimate the SMA parameters with comprehensive physical understanding but also accurately predict retention over a wide range of loading conditions.