Utilizing Large Language Models for Clinical Development: Insights on Clinical Text Mining

Kyubum Lee, Principal Data Scienctist, Amgen

Kyubum Lee, PhD, is a biomedical data scientist currently serving as a Principal Data Scientist at Amgen. He earned his bachelor's degree from Korea University in Seoul, South Korea, majoring in both Computer Science and Biology. He went on to receive his master's degree in Bioinformatics and a PhD in Computer Science from the same institution, with a focus on machine learning and text mining.

Dr. Lee completed his postdoctoral fellowship at the National Center for Biotechnology Information (NCBI) at the National Institutes of Health (NIH) and Moffitt Cancer Center. During his postdoctoral training, he led numerous cross-functional projects in collaboration with other biomedical researchers from various institutes, including the CDC, EBI, SIB, and NCI. His projects primarily utilized machine learning to extract valuable information from extensive text datasets, such as over 30 million biomedical articles in PubMed and PMC, as well as biomedical image and genomic data.

Currently, at Amgen, he is part of the Center for Design and Analysis, applying machine learning and artificial intelligence to enhance clinical development and the design of clinical trials.

Abstract:

Large Language Models (LLMs) are revolutionizing the fields of text mining and natural language processing. LLMs can be used to understand and analyze biomedical text from sources such as ClinicalTrials.gov, PubMed, and proprietary internal documents. LLMs have extensive background knowledge by training on a large volume of biomedical text data. For these reasons, LLMs do not need to be trained on a specific training set, making them more efficient and invaluable in performing biomedical tasks.

In this talk, I will discuss the practical applications of LLMs for text mining within the realm of clinical development, and highlight both their strengths and limitations. Additionally, I will share our experiences and insights we gained from using LLMs to identify biomarkers in clinical trial text data, and discuss their role in advancing clinical development.

Large Language Models and their Applications in Drug Development

Shams Zaman, Senior Director of Data Science & Statistical Methodology, BMS

Shams Zaman is Senior Director of Data Science & Statistical Methodology. He is leading data driven advance analytics in Cell Therapy and Hematology areas within GBDS supporting various stages of drug development and post approval activities. His team is also developing state-of-the-art clinical natural language processing (NLP) solutions to generate insights and competitive intelligence from clinical documents. Before joining BMS, he was a fellow at FDA and a research scientist in AI lab Philips Healthcare Research. He worked heavily on developing several NLP driven products and clinical decision support (CDS) algorithms from electronic health record (EHR) data in his previous positions.

Abstract:

Recent advancements in generative AI have pushed the boundaries of what is possible in terms of generating realistic and creative content. These advancements have been driven by the development of large-scale language models and improved training techniques. One significant advancement is the development of models with unprecedented size and complexity. Models like OpenAI's GPT-3, with billions of parameters, have demonstrated remarkable language generation capabilities. These models can understand and generate human-like text, making them useful in a wide range of applications, from chatbots to content generation. Another notable advancement is the improvement in fine-tuning techniques. Fine-tuning involves training a pre-trained language model on specific tasks or domains, allowing it to specialize in particular areas. This approach has led to better performance on various natural language processing tasks, such as text classification, sentiment analysis, and question answering. Techniques like prompt engineering and few-shot learning aim to improve the models' ability to generalize and adapt to new tasks with minimal training examples. This has made generative AI models more versatile and applicable to a wider range of real-world scenarios.

Generative AI, particularly in the form of large language models, has found several applications in drug development. These models have the potential to accelerate various stages of the drug discovery and development process. One key application is in analyzing and extracting insights from vast amounts of scientific literature. Generative AI models can process and summarize research papers, clinical trial data, and patents, helping researchers identify potential drug targets, understand disease mechanisms, and explore new therapeutic approaches. By automating the analysis of this information, these models save time and provide valuable insights. Another application is in the design of new molecules. Generative AI models can generate novel chemical structures and predict their properties, aiding in the discovery of potential drug candidates. These models can also suggest modifications to existing molecules to improve their efficacy, safety, or pharmacokinetic properties. By accelerating the molecule design process, generative AI can help researchers identify promising candidates more efficiently. Generative AI can also assist in predicting drug-drug interactions and adverse effects. By analyzing drug databases and medical literature, these models can provide insights into potential risks and help optimize treatment plans. This can enhance patient safety and improve the effectiveness of drug therapies. In summary, generative AI has emerged as a valuable tool in drug development. By assisting in literature analysis, molecule design, prediction of drug interactions, and optimization of drug formulations, these models have the potential to accelerate the discovery and development of new therapeutics, ultimately benefiting patients and improving healthcare outcomes.