Enhanced Machine Learning Models Streamline Predictions for Monoclonal Antibody Designs
In a significant breakthrough, a team of researchers from the Max Planck Institute for Biochemistry, Technical University of Munich, Imperial College London, and GSK have developed a hybrid approach to improve design space identification for monoclonal antibodies. The researchers focused on single-column antibody Protein A affinity chromatography.
Dr. Michael Schmidt, currently at the Max Planck Institute, and Dr. Laura Becker from the Technical University of Munich, were two key contributors to this hybrid approach.
The hybrid model, pretrained on synthetic data generated by a high-fidelity process model and fine-tuned using wet-lab data, offers improved predictive accuracy for low-density data sets. This is particularly beneficial in situations where experimental data is scarce. For such cases, the approach augments the data with a machine learning-enhanced in silico model.
The hybrid model also improves the accuracy and reliability of design space predictions, even for low-density datasets. Its second-order F1 scores can be particularly useful where true experimental labels are scarce, incomplete, or unavailable.
The enhanced hybrid model significantly improves the predictive capability of data-driven modeling and reduces the need for process knowledge. It generates a larger design space than a purely data-driven approach, and its performance was evaluated using first-order and second-order F1 scores.
The hybrid model's improved predictive accuracy for medium-density datasets ranged from a decline of 9% to an improvement of 5%. However, for medium-density regions, one run showed less accuracy with the hybrid model. There was no significant improvement in predictive accuracy for high-density data with the hybrid model.
The approach informs in silico models with results from data-driven experiments. It can help manufacturers expedite process development by enabling quantitative knowledge transfer and reducing experimentation. The hybrid model is useful in early-stage biopharmaceutical process development, where true experimental labels are often limited.
Moreover, the hybrid model delivers reliable predictions for the target system. Compared to the data-driven model, the hybrid model improved predictive accuracy for high-, medium-, and low-density data by 6-27%, 7-15%, and 6-41% respectively, for medium-density regions and low-density datasets.
Second-order F1 scores quantify agreement between a target model's predictions and those of a reference model. Wet-lab data can be limited in quantity and quality, but knowledge can be leveraged from other systems without being parameterized to the system of interest.
In summary, the hybrid model offers a promising solution for improving design space identification for monoclonal antibodies, particularly in situations where experimental data is scarce. Its ability to generate a larger design space, deliver reliable predictions, and reduce the need for process knowledge makes it a valuable tool in early-stage biopharmaceutical process development.
