AI Implementation in Downstream Bioprocessing : Opportunities, Capability Gaps, and Reconfiguration Needs in a Life Science Firm

Sammanfattning

There are many opportunities for implementing artificial intelligence (AI) within downstream bioprocessing hardware and related software products. However, while the technical usefulness of implementing AI in many cases is proven, the actual implementation within the pharmaceutical bioprocessing industry is limited. By conducting semi-structured interviews within a large life science company that provides downstream hardware and applying the Dynamic Capabilities View, several AI opportunities emerge, but also barriers that hinder adoption and organizational reconfiguration needs required for organizations to implement the opportunities. Life science firms show strong sensing capabilities by identifying various AI opportunities, but weaker seizing and reconfiguring capabilities to implement them. This gap between opportunity identification and implementation reflects challenges in gathering necessary training data and organizational adaptation. Firms should strengthen data infrastructure, enhance internal and external collaboration and redesign products to enable greater data collection. Additionally, a screening tool to differentiate which protein drug candidates would be easy or difficult to manufacture, was investigated in terms of technical feasibility. The sequence-based AI-model ESM-2 was investigated in terms of predicting protein properties, to evaluate whether that could be a part of such a tool. A ridge regression model trained on ESM-2 embeddings from a dataset consisting of antibodies and labels for hydrophobic interaction chromatography retention time, was able to predict retention time for antibodies in the test set with an R² of 0.35, implying that ESM-2 could be used to predict developability-related properties of proteins to some extent.