Pfizer and AWS's PACT Prototyping Programme Brings Generative AI Search and Anomaly Detection to Pharmaceutical Sciences

In 2021 Pfizer's Pharmaceutical Sciences Small Molecule (PSSM) group and AWS launched the Pfizer-Amazon Collaboration Team (PACT), applying AWS analytics, machine learning, compute, storage and data warehousing to Pfizer's laboratory, clinical manufacturing and clinical supply chain work, according to an AWS case study. Pfizer teams propose ideas and AWS provides solutions architects and its Prototyping and Cloud Engineering team. Vijay Bulusu, Pfizer's head of data and digital innovation for PSSM, said building a prototype in-house would take at least three months, while a PACT prototype has typically taken no more than six weeks. The first use case aimed to cut the time 1,500 PSSM scientists spend on data discovery by up to 80 percent. Developing one drug can produce about 20,000 documents, and scientists often had to search for historical data manually with several tools; early experimentation led to the development of Amazon Kendra. PACT teams now use Anthropic's Claude 2.1 through Amazon Bedrock in an internal platform called Vox, which lets scientists search documents by voice command and chatbot in natural language. Pfizer estimates that scientists could save up to 16,000 hours of searching and extracting data each year, and it estimates a 55 percent reduction in infrastructure costs. Another PACT prototype detects anomalies in Pfizer's Portable Continuous Miniature and Modular (PCMM) manufacturing, a continuous process for producing solid oral-dose medicines. Built with Amazon SageMaker, Amazon Lookout for Equipment and Amazon Lookout for Metrics, it lets scientists detect anomalies as they occur, predict maintenance needs and reduce equipment downtime. The work led to a manufacturing analytics tool for time series data that Pfizer's large molecule group now also uses. From PACT's initial list of 14 projects, five ideas have moved into production. Not every prototype succeeded: a test of virtual and augmented reality for training showed users were not yet ready for the technology. The case study is undated; its earliest archived copy is from August 2024.