Abstract: I will introduce Bayesian methods tailored for information borrowing from external source and sample size estimation/re-estimation that empowers adaptive decision making during the course a clinical trial. For information sharing, the idea is to apply Bayesian nonparametric and parametric models to match the distributions of covariates and responses, achieving subgroup matching and accounting for unmeasured confounders. The particular models are related to the Bayesian addictive regression trees (BART) and meta-analytic predictive (MAP) priors. For sample size estimation/re-estimation, the adaptation is centered at balancing the observed and predicted population summary, confidence in decision making, and (re-)estimated sample size. I will provide some numerical examples and case studies illustrating these powerful approaches.
Abstract: On March 13, 2025 issue of FDA’s Prevision Policy, FDA highlighted the Sanofi’s acid sphingomyelinase deficiency (ASMD) treatment Xenpozyme (olipudase alfa) as a model that demonstrates the ability to conduct robust dose optimization work in the context of a rare disease development program. ASMD is a rare autosomal recessive lysosomal disorder that results in liver and respiratory failure, which are the primary causes of death in patients with ASMD. In this presentation, the study design and dose optimization strategy from Animal model with ASM Knockout mouse in non-clinical study, first-in-human single ascending dose (SAD) and multiple ascending dose (MAD) phase 1 studies, as well as pivotal phase 2/3 adult and phase 1/2 pediatric study will be discussed. Relevant regulatory discussions, final trial results and conclusions will also be shared. In addition, statistical considerations beyond dose optimization strategy will also be highlighted, particularly regarding the selection of primary endpoints, as well as utilization of modeling and external natural history data to support pediatric indication.
Abstract: Clinical trials in rare diseases often face significant challenges due to low prevalence, making it difficult to recruit sufficient number of patients for adequate statistical power. This constraint calls for innovative design strategies that can maximize the efficiency under limited total sample sizes. Platform trial designs offer a promising solution by evaluating multiple investigational drugs simultaneously against a shared reference arm, thereby reducing the overall sample size required. In this presentation, we investigate a platform trial framework under a fixed total sample size, reflecting practical constraints in rare disease research. A key practical consideration is the staggered availability of investigational drugs — not all investigational drugs may be ready to be evaluated concurrently, and some may be dropped early from the platform due to futility or external factors. To address these dynamics, we develop and evaluate design strategies that determine when to add or drop arms, and more critically, how to optimize patient allocation across treatment arms over time to maximize statistical power for treatment comparisons. Our findings provide practical guidance for designing more informative and resource-efficient platform trials under real-world operational limitations.
Abstract: Prognostic scores (PS) can improve statistical efficiency in clinical trials by reducing the variance of treatment effect estimates via covariate adjustment, leading to trial derisking and cost savings. This potential illustrated through an example in Alzheimer’s disease where PS adjustment reduced treatment effect variance by 20%, yielding a power increase from 80% to 87%, followed by a discussion of opportunities and challenges for developing and implementing prognostic scores in the rare disease space.
Abstract: TBD
Abstract: TBD
Abstract: TBD
Abstract:Model Informed Drug Development (MIDD) is the practice of developing quantitative models from preclinical and clinical data to inform decision making throughout the drug development process. It is a general framework being implemented across pharmaceutical sponsors that has been a joint effort from industry and academia with support from regulators. MIDD methods can improve clinical trial efficiency, increase the probability of trial and regulatory success, and optimize dosing. MIDD for advanced therapeutics such as cell and gene therapies offers an opportunity where a comprehensive approach for integrating relevant data into quantitative models could greatly aid in the evaluation of safety and efficacy of these drugs. Recent papers in this area highlight the needs and approaches of MIDD in this space. There is a shared outlook from sponsors and regulators that MIDD approaches can provide insight and evidence for product safety and efficacy, and this talk will focus on the workflow and opportunities for MIDD practice in the evolving space of cell and gene therapies and other novel modalities.
Abstract: In this talk, we will provide an overview of the current state of deep learning models for medical image analysis focusing on foundation models. These models are trained on a huge amount of unlabeled data allowing the model to extract general visual features. We will begin by describing the core components of foundation models such as choosing a model architecture, performing self-supervised learning, and applying adaptation approaches which allows the model to adapt to downstream imaging tasks. Furthermore, we will discuss some current benefits and challenges when using foundation models for medical image analysis. We will then conclude by illustrating a recent application of foundation models to pathology specifically in the area of rare cancer detection.
Abstract: Rare diseases affect over 300 million people worldwide, yet 95% lack approved treatments due to small patient populations, limited research funding, and prolonged diagnostic journeys. This presentation looks at how artificial intelligence and machine learning stand to transform rare disease research across three critical domains: revolutionizing drug discovery, accelerating accurate diagnosis, and optimizing treatment strategies. Using several examples, we examine how AI-powered diagnostic tools are reducing the average 5–7-year diagnostic odyssey. Additionally, AI/ML continues to improve treatment protocols and, we will look at how computational approaches will speed the identification of novel therapeutic targets and repurposing of existing drugs. In addition to current successes, the presentation will highlight implementation challenges including data scarcity and regulatory hurdles faced in the way of the transformative potential of AI to change the lives of underserved patient communities.
Abstract: TBD
Abstract: To address the challenge of information overload in scientific research, we present how one can use AI agents to automate literature curation and augment the discovery process. The system's core is a dual-mode AI agent architecture. For deep exploration, a single agent performs sequential, iterative analysis. For comprehensive synthesis, a multi-agent framework deploys specialized agents such as for analysis, verification, and discovering alternative perspectives to deconstruct research questions and work in parallel. This architecture supports both local and cloud-based Large Language Models, ensuring flexibility and robust performance. The agent framework is complemented by analytical tools designed for strategic insight. Researchers can uncover latent bibliographic and semantic connections through interactive network visualizations, and identify emerging research topics using an automated trend analysis dashboard with temporal clustering and forecasting. By integrating a sophisticated multi-agent system with these powerful tools, this work demonstrates a shift from passive information retrieval to an active, AI-augmented discovery process, enabling researchers to efficiently synthesize knowledge and strategically navigate the evolving scientific landscape.
Abstract: Population-adjusted indirect comparisons (PAICs) have emerged as essential techniques for assessing the comparative effectiveness and safety of interventions, especially in cases of rare diseases where direct head-to-head trials are either unavailable or impractical. The most widely used PAIC methods include matching-adjusted indirect comparison (MAIC) and simulated treatment comparison (STC), both of which have gained traction in health technology assessment (HTA) submissions. However, the application of PAIC in rare disease contexts presents unique challenges, including reliance on unanchored comparisons from single-arm trials, small sample sizes, and heterogeneity in patient characteristics. Unanchored indirect treatment comparisons face significant hurdles due to the lack of a shared comparator arm and the strict assumptions necessary for valid comparisons. The limited patient populations exacerbate typical small-sample issues, such as poor overlap between trial populations, difficulties in fitting robust regression models for STC, and reduced capability to account for all prognostic factors and effect modifiers. This talk will highlight these challenges and explore practical considerations for conducting PAICs. Furthermore, it will discuss innovative methods like two-stage MAIC, weight truncation, and STC with G-Computation to enhance the robustness of PAIC findings.
Abstract: The robust integration of natural history (NH) data is essential for designing efficient clinical trials for one-time gene therapies in rare neurodegenerative disorders. We present a pharmacometrics-informed clinical scenario evaluation framework (CSE-PMx) that formally incorporates quantitative NH models directly into the prospective trial planning process. This presentation showcases an application of the framework to optimize a hypothetical gene therapy trial in Autosomal-Recessive Spastic Ataxia of Charlevoix-Saguenay (ARSACS). Through in-silico evaluation, we demonstrate how insights from the ARSACS disease progression model can be used to compare competing design and analysis strategies, ultimately identifying a nonlinear mixed-effects model (NLMEM) as a more powerful approach. This methodology provides a quantitative roadmap for translating NH insights into more efficient clinical development programs, helping to transform and accelerate the path forward for rare disease therapies.
Abstract: Little knowledge of natural history of disease, heterogeneity in disease presentation and response to treatment, difficulty in enrolling and retaining patients in randomized controlled trials pose challenges to evaluating treatment effects in patients with a rare disease. Targeted learning provides a roadmap for addressing these challenges, and sophisticated methodologies for estimating causal effects of treatment even when there is informative loss to follow-up. Simulation studies based on actual patient data will compare performance of the MMRM method often used with longitudinal data and targeted machine learning using targeted maximum likelihood estimation and super learning (TMLE+SL).
Abstract: Randomized controlled trials (RCTs) are considered the gold standard to evaluate the efficacy and safety of a new therapy for regulatory approval, though in rare diseases there are special situations where maintaining a concurrent control arm is unethical, impractical, or infeasible and can lead to increased patient burden and threaten the completion of the RCTs. Externally controlled designs are gaining a greater attention for their potential to overcome these challenges by supplementing or replacing the concurrent control arm in the RCTs with external data. A major challenge in the externally controlled designs is its inability to eliminate systematic imbalances in measured confounding factors between the current study patients and the external controls, which may cause a serious bias in treatment effect estimation. In these contexts, propensity score matching is known to help reduce systematic imbalances in measured confounders; however, it causes a challenge of unequal matching where fewer external controls are to be matched to more current study patients, leading to the case with a matching ratio of less than 1. This talk presents several approaches for the unequal matching, including an assignment algorithm, a genetic algorithm, and methods with searching all possible combinations. For example, the assignment algorithm is a method finding an optimal set of external controls while minimizing the overall absolute difference of the propensity score. Through a simulation study, we compare the methods in terms of their performance in the propensity score balancing and the treatment effect estimation (e.g., bias, type I error rate, and power).