Accelerating Clone Selection with Transfer Learning
Clone selection in biologics development is a time-intensive process for drug development programs that often operate under time and resource constraints. Process scientists face pressure to quickly identify a high-performing clone that meets critical quality attributes (CQAs) for productivity, product quality, and stability, all while ensuring scalability for manufacturing.
Balancing the need for speed with the rigor required to avoid costly setbacks later in development is a constant challenge, demanding both efficiency and precision under strict deadlines. To meet this challenge DataHows’ transfer learning approach, which leverages hybrid models, offers scientists the possibility to reliably accelerate clone selection by leveraging existing historical data rather than extensive, new experimental data alone.

Learning from historical clones to assess new process clones
To accelerate clone selection, DataHow utilizes its transfer learning capabilities to extract insights from historical clone data, minimizing the need for extensive experimentation with new process clones. This approach not only expedites clone selection but also transforms historical development data into a continuously growing operational asset that enhances decision-making over time.
At the core of this innovation are DataHow’s hybrid models, which seamlessly integrate mechanistic models—grounded in first-principles knowledge—with data-driven models that extract patterns directly from experimental data. By combining these complementary methodologies, DataHow leverages existing scientific understanding while harnessing machine learning to uncover insights beyond the reach of existing knowledge.
What is Transfer learning
Transfer learning is a machine learning methodology that utilizes data from historical processes to transfer knowledge horizontally to new developments. This approach is particularly effective when there is a high degree of similarity between processes or clones, enabling a more efficient application of prior insights to novel scenarios. Learn more about DataHow technologies.
Transfer learning can be applied accross molecules, by using machine learning techniques to transfer insights from historical molecules to novel developments, as well as within a project when transferring insights across scales during scale-up.
For further reading on transfer learning refer to the following publication.

Assessing the Impact of Transfer Learning via DataHowLab’s Hybrid Models
To illustrate the application and impact of transfer learning across molecules during clone selection, a highlights from industrial cases are presented:
Case 1: Assess the possibility of transferring process design knowledge across clones as part of an accelerated cell line development approach. Determine the top-performing clone (of 3 top clones) and optimal conditions by combining data from the 3 top clones at standard conditions only, and historical clone data where multiple conditions were tested.
Case 2: Assess which available information from four historical clones improves characteristic process understanding of a new clone. Model error (RMSE) was assessed when data from each of the historical clones was introduced individually, and when combined.
Case 3: Exploring early forecasting capabilities to accelerate clone selection. A predictive model was tasked with forecasting final clone performance at Day 14, leveraging historical data from two other molecules (each with 14-day profiles) and only 7 days of data from the new molecule.
Key Results
The clone selection process can be substantially optimized using a transfer learning approach. A summary of the insights from each case is highlighted. For the full results with supporting plots, download the full case study below:
Case 1: By combining new clone data at standard conditions, alongside historical clone data tested across multiple conditions, top clones and their optimal conditions could be identified simultaneously—significantly increasing development efficiency. Through transfer learning, historical clones serve as a valuable development asset, enabling both the prediction of high-performing clones and the optimization of experimental design.
Case 2: Incorporating historical clone data enhances the understanding of novel clones and improves model performance. However, the greatest impact comes from historical clones that are similar to the novel clones. Conducting a similarity assessment is crucial when applying transfer learning, as more representative historical data significantly improves prediction accuracy.
Case 3: Top-performing clones for a new molecule were correctly identified with only 7 days of data, and historical data from 2 other molecules. Scientists are able to accelerate clone selection by many days by leveraging historical clone data via transfer learning as part of their development approach.
Should you wish to further explore the case with the DataHow team, please contact us directly.
