Knowledge transfer across cell lines using hybrid Gaussian process models with entity embedding vectors.

Area of Science:

• Biochemical process development
• Machine learning applications in chemistry
• Data-driven scientific discovery

Background:

• Biochemical process development relies heavily on experimental data, often used only once.
• Existing data from developed processes are underutilized for predicting novel process outcomes.
• Effective cross-process learning requires accurate representation of product identity.

Purpose of the Study:

• To develop a method for predicting novel biochemical process outcomes using data from previously developed processes.
• To introduce a new way to represent product identity for improved machine learning model performance.
• To reduce the number of experiments required in biochemical process development.

Main Methods:

• Product identity represented by learned embedding vectors.
• Gaussian process regression model incorporating product embeddings.
• Comparison with traditional one-hot encoding using simulated cross-product learning tasks.

Main Results:

• Learned embedding vectors capture interpretable product similarity.
• The proposed method significantly outperforms one-hot encoding in cross-product learning.
• Demonstrated feasibility of learning product embeddings from process data.

Conclusions:

• Product embedding vectors offer a powerful way to represent categorical features in process development.
• This data-driven approach can substantially decrease the need for wet-lab experiments.
• The method facilitates more efficient and predictive biochemical process optimization.