Integrating hybrid modeling and high throughput screening: A modular process development platform for flowthrough...

As the biopharmaceutical industry continues to advance towards intensified manufacturing and increasingly complex therapeutic pipelines, there is a growing demand for more efficient processes and process development strategies. Flowthrough and frontal loading operations are particularly attractive due to their high loading capacities, low buffer consumption, and operational simplicity. High throughput screening combined with mechanistic modeling provides a powerful framework for rapid development of these processes, but model calibration needed across the multitude of resins and conditions during screening remains resource intensive, and accurately capturing transport and kinetic effects is challenging. In this work, neural network and closed-form symbolic regression models were developed to predict protein loading and product yield, which can exclusively use high throughput plate-based data as inputs. To support training these models, a large database of general rate model (GRM) simulations spanning both high molecular weight (HMW) and low molecular weight (LMW) separations was created. This was accompanied by development of novel analytical expressions for loading and yield derived from fundamental principles under ideal chromatography conditions, enabling transport limitations to be treated as perturbations from the ideal case. The resulting machine learning models trained with these data rely solely on easy-to-measure, experimentally accessible inputs from standard plate-based high throughput experiments and therefore provide accurate predictions without the full input parameter set required by the GRM. Neural networks achieved the highest accuracy, while symbolic regression models performed nearly as well with the added advantages of simplicity, interpretability, and ease of implementation. Further, Monte Carlo analysis demonstrated that predictive errors of the machine learning models were comparable to or smaller than those of mechanistic simulations when typical experimental uncertainty in difficult-to-measure input parameters was considered. Building on this, closed-form expressions for productivity were developed that use the machine learning models as inputs to identify resins and operating conditions that maximize performance as a function of residence time. Finally, a platform workflow is proposed that applies these models to translate high throughput data into practical guidance for resin and condition selection, with predicted purities, yields, and productivities aligning well with experimentally-measured column performances. Overall, this proposed platform provides a foundation for seamlessly integrating high throughput screening and hybrid modeling into process development.