Transfer learning across technologies for ATMPs

Important results from the project

The objectives have been achieved, and additional important validations of the results have been conducted to ensure that cells labeled “not approved” indeed have poor quality. This has increased the value of the project outcomes for users. Large amounts of transcriptomic data from cells of varying quality have been generated, and the hypothesis that an AI model can be trained using transcriptomic data to classify cells as “approved” or “not approved” based on gene expression works very well.

Expected long term effects

There is a significant need for robust, standardized quality assurance methods in the manufacturing of cells for therapies. The results demonstrate the usefulness of AI models for quality testing of starting materials The AI model developed within the project shows reproducible results, higher sensitivity than current methods, and is scalable for industrial cell production. Increased knowledge of processes that preserve stem cell pluripotency contributes to safer cell therapies.

Approach and implementation

The project was carried out according to plan despite one partner leaving halfway through. The remaining partners redistributed the work based on their respective expertise. Minor delays occurred toward the end of the project due to a relocation of laboratory facilities at RISE, but the plan was maintained: TakaraBio Europe supplied cell material, the University of Skövde analyzed the data, RISE developed qPCR tests, and MultiD Analysis implemented the AI model in its software.