ML Predicts Charged Antibody Distribution Across a Filter Membrane

By Gail Dutton

Credit: ustas/Adobe Stock

Machine learning is being used more and more often to replace mechanistic models or traditional laboratory experiments in bioprocessing. Now, scientists in Japan are using explainable machine learning to describe the distribution of charged antibodies across a semi-permeable membrane.

Accurate insight into this distribution, known as the Gibbs–Donnan effect, is particularly important in monoclonal antibody (mAb) manufacturing. That’s because antibody distribution shifts significantly during ultrafiltration and diafiltration (UF/DF).

“The semi-permeable membrane used in UF/DF allows excipient molecules to pass through, while retaining the larger antibodies. This leads to shifts in excipient distribution that affect the final balance of excipients and mAb in the final product,” explains Chyi-shin Chen, PhD, purification process development scientist at Chugai Pharmaceutical.

In two pilot-scale manufacturing cases, Chen tells GEN, “We have observed estimated savings of up to 30 g of material and a reduction in UF/DF process development time of up to four weeks, compared to the conventional approach.” Those time savings occurred in sample preparation, UF/DF experiments, and analysis. “While the study of potential applications in process validation is ongoing, we expect similar benefits in process validation studies.”

Chugai’s model, described in a recent paper, alleviates the need to perform experiments to determine excipient displacement, or to develop mechanistic models to describe solute distribution based on protein charges. At a processing level, it eliminates the need for engineers to know the specific biology of the antibody being developed, or the specific equations used for modeling.

The model uses eXtreme Gradient Boosting with a low-dimensional data set of common amino acids of mAbs, their biophysicochemical descriptions, and process conditions. Shapley additive explanations of feature importance translated the model’s decisions into molecular-level protein and excipient interactions.

The team used several different molecular descriptors to capture a broad range of biophysicochemical effects during UF/DF that, they say, may otherwise have been overlooked. Ultimately, they chose the antibody’s isoelectric point, retention conductivity from hydrophobic interaction chromatography for hydrophobicity, and viscosity.

Those data points provided insights on the molecule’s structure and electrostatic properties, surface hydrophobicity distribution, the size and location of hydrophobic patches, and protein-protein interactions. To let the model compensate for nonlinear portions of the Donnan slopes, the team also included target concentration data.

“Developers can create their own models based on their mAb pipelines and process conditions to optimize buffer compositions…or use the insights from this research to inform their buffer selection and experimental design, even without building a separate model,” Chen says.

“Even with limited data points, using common descriptors obtained during downstream process development, the concentration shift arising from the Gibbs–Donnan and volume exclusion effects can be predicted with an error of less than 5%,” the scientists report.

For this study, all data has a pH of 6. Currently, Chen says the Chugai scientists are expanding the model to account for different pH conditions, making it more applicable across a wider range of bioprocessing conditions. Eventually, pH dependency, and thus additional molecular descriptors, also may be added to the model.