Chair: Jian Zhu, PhD (Servier)
Co-Chair: Cindy Gao, PhD (FDA)
Abstract: The use of external and literature data has greatly contributed to clinical development, and has been expanding to cover areas other than trial analysis. In this session the speakers will particularly focus on these new areas, sharing the use of AI/ML in enhancing clinical trial operations such as supply chain management, individual patient level data synthesis based on literature and signal detection using large language models for drug safety surveillance, in addition to recent advancement in Bayesian methods and tools for using external data in hybrid control clinical trials. The scope of the session will provide a broader understanding of these new trends to the audience and highlight the implementation of AI/ML in these areas.
Speaker: Brian Hobbs, PhD (Telperian Inc.)
Title: PD-L1 Subpopulations of Metastatic Urothelial Carcinoma Demonstrate Heterogeneity to Chemotherapy: A Systematic Review and Meta-Analysis of Synthetic Trial Data
Abstract: Efforts to translate advances in immunology into anti-cancer immunotherapies have progressed rapidly in recent years. Six antibodies acting on programmed death ligand 1 or programmed death 1 pathways were approved in 75 cancer indications between 2015 and 2021. Several of these therapies were granted accelerated approval for specific cancer indications on the basis of evidence acquired in single-arm phase II clinical trials. In the absence of randomization, however, patient prognosis for progression-free and overall survival may not have been studied under standard of care chemotherapies for PD-1 and PD-L1 biomarker subpopulations. In 2021, two immunotherapies were withdrawn from accelerated approval applications for treatment of metastatic urothelial carcinoma after randomized phase III trials failed to demonstrate evidence for survival advantage when compared to standard of care. The findings of the IMvigor210 (NCT02108652) and IMvigor211 (NCT02302807) trials of atezolizumab are reviewed for the purpose of elucidating the statistical implications of PD-L1 subpopulation heterogeneity. To place the findings into the context of external evidence, a systematic review was conducted to identify trials that assigned metastatic urothelial carcinoma patients to the same chemotherapy agents administered as a part of the control arm in IMvigor211. The systematic review searched PubMed for phase II/III clinical trials on the second-line treatment of urothelial carcinoma with single-agent chemotherapy drugs including paclitaxel, docetaxel and vinflunine, published in English between 1993 and 2016. Eleven trials including 742 patients met the inclusion and exclusion criteria. Of these, 11 were used to evaluate overall response rate (ORR) and 10 were used to evaluate overall survival (OS). This article defines the extent to which PD-L1 IC2/3 subpopulations outperformed their expectations in the IMvigor211study based on external trial evidence. The results found that the IC2/3 subgroup in IMvigor211 demonstrated statistically significantly higher ORR and prolonged OS as compared to prior mUC studies. Given the extent of PD-L1 heterogeneity identified by results, trial simulation was applied to define the probability that IMvigor211 would have resulted in a positive trial based on its actual design as well as alternative designs that enrolled a higher proportion of IC2/3 patients or had longer durations. The results indicated that as designed the IMvigor 211 trial had a <24% probability of success and would have required at least 559 patients in the PD-L1 high subgroup, as compared to the actual 234, or an additional 20 months of follow-up.
Speaker: Jinfeng Zhang, PhD (Insilicom LLC)
Title: Extracting Drug Adverse Events from Literature by Finetuning Large Language Models
Abstract: This In the past few decades, the biomedical research community has acquired a wealth of knowledge, much of which is stored in scientific literature as unstructured text. Converting this text into structured form is crucial for developing new methodologies and applications that can fully utilize this knowledge. To achieve this goal, two basic problems must be addressed: named entity recognition (NER) and relation extraction (RE). NER involves identifying the concepts or entities in texts, such as diseases, genes/proteins, and chemical compounds. RE, on the other hand, aims to extract the relationships between these entities. The information extracted from NER and RE can be used to create a knowledge graph, where nodes represent entities in the text and edges represent their relationships. This presentation will discuss our team’s work on the LitCoin NLP Challenge organized by NIH, for which we were awarded first place. Using pipelines developed for the challenge, we processed all PubMed articles and created a large-scale biomedical knowledge graph that integrates drug adverse event information. Recently, we enhanced this framework with a specialized method for extracting drug adverse events using state-of-the-art large language models (LLMs). This approach achieved a remarkable 97.8% recall and 87% precision, surpassing the performance of human experts. To facilitate user access, we developed a web portal and API for querying drug adverse events from newly published literature.
Speaker: Zoe Hua, PhD (Servier)
Title: Application of Artificial Intelligence in Clinical Trial Supply Chain Management
Abstract: This Emerging pivotal challenges from the intricate landscape of drug supply chain management can potentially offset the benefit in applying innovative adaptive designs in clinical trials. The challenges include the uncertainty of maximum drug supply needed, shifting of supply requirement, cost control, and high-dimension factors impacting the decision of drug resupply. To address these issues, we designed an optimization digital tool tailored for the efficient management of drug supply in clinical trials. The tool optimizes drug supply strategies in a sequential manner throughout the trial, with leveraging real-time data and statistical simulations to make informed decisions. Real world scenarios are integrated into the framework as pragmatic assumptions and setup. Statistical simulations are applied to optimize drug supply strategy in pre-study planning stage as well as during study monitoring stage. Drug supply is optimized via minimizing the total cost using real-time data from the study over time. An artificial intelligence model particle swarm optimization algorithm is applied to perform optimization, where feature extraction is implemented to reduce dimensionality and computational cost.
Speaker: Min Lin (University of Connecticut)
Title: Tool and Methods for Bayesian Divide-and-Conquer External Data Borrowing in Clinical Trials
Abstract: This talk introduces a Bayesian divide-and-conquer methodology for leveraging real-world data (RWD) in single-arm trials and randomized controlled trials (RCTs). Utilizing propensity-score stratification with tailored borrowing-by-parts power priors, this approach enables robust borrowing of external data while addressing challenges such as population heterogeneity and temporal effects. The first part provides an overview of the methodology’s application in both single-arm and RCT settings, highlighting key advances and simulation results. The second part demonstrates an R Shiny application that operationalizes these methods, offering an intuitive tool for trial design and analysis. Attendees will also preview features planned for an accompanying R package.
