Invited Session 8: Frontiers in Statistical and AI Methods for Cancer Research

1:40-3:00 PM, November 23, 2026; Room 7C-D

Moderator
Speakers

Byung Park

Moderator: Oregon Health & Science University
Biography

Byung Park

Byung Park is a professor of Biostatistics in the OHSU-PSU School of Public Health, Cancer Data Science Shared Resource (CDSSR) with the Knight Cancer Institute, and he is an associate director of Biostatistics and Bioinformatics Core (BBC) at the Oregon National Primate Research Center. He has a long-standing collaboration with investigators from Knight Cancer Institute (KCI) and Oregon National Primate Research Center (ONPRC), since he joined the Oregon Health & Science University as a collaborative biostatistician (2002-present). He has a broad range of expertise including clinical trials, high-dimensional genomic data modeling, general linear models, mixed effect models and statistical genetics. His current research interest focuses on designs of earlier phases of oncology clinical trials for targeted agents, statistical analysis of clinical trials, and statistical methods for high-dimensional genomic data. As the CDSSR at the KCI, he collaborates with investigators by designing and developing statistical analysis plans for clinical trial protocols as well as he serves as a reviewer for the Clinical Research Review Committee (CRRC). He served as a faculty for the AACR/ASCO Methods in Clinical Cancer Research Workshop from 2018 to 2021. He had been a member of ASCO Conquer Cancer Grant Selection Committee and Annual Meeting Scientific Program Committee.


Conditional Approval of Cancer Drugs in China: Evidence and Implementation Challenges

Speaker: Ji-bin Li, Department of Clinical Research, Sun Yat-sen University Cancer Center, Guangzhou, China

Authors: Ying-ying Zhu; Ji-bin Li

Biography

Ji-bin Li

Ji-Bin Li is a Professor of Biostatistics and Clinical Research Methodology in the Department of Clinical Research at the Sun Yat-sen University Cancer Center, Guangzhou, China. He has served as a statistician and methodologist on a series of cancer trials. His research focuses on biostatistical methods for clinical trials and longitudinal studies, with particular expertise in trial design, longitudinal data modeling, survival analysis, and regulatory science. He has published more than 40 peer-reviewed articles as first or corresponding author, in journals including Nature. His work has been cited more than 8,000 times, with an h-index of 49 (Google Scholar). He is the chief editor of the textbook Clinical Research Methodology (1st and 2nd editions) and has contributed to more than ten other textbooks and monographs. He was named to the Stanford University/Elsevier list of the World’s Top 2% Scientists in 2025 and 2026.

Abstract

Background: To address unmet clinical needs, regulatory agencies have advanced policies enabling earlier access to innovative medicines, especially for cancer drugs. The U.S. FDA pioneered accelerated approval mechanisms for novel therapeutics in 1992. Until 2019, the China National Medical Products Administration (NMPA) officially launched the conditional approval pathway for new drugs. Similar to the U.S. FDA accelerated approval pathway, mandatory confirmatory trials are then required after conditional approval to confirm patient benefit, and indications are either converted to regular approval or withdrawn. However, the evidence and challenges of pivotal and confirmatory trials supporting conditional approval of cancer indications by NMPA have not been systematically characterized.

Methods: The data on cancer drug-innovation pairs granted conditional approval between Dec. 1, 2019 and Dec. 31, 2024 in mainland China were identified based on the review documents from the NMPA website and PubMed. The characteristics of pivotal and confirmatory trials, and NMPA’s decisions of regular approval for indications were extracted and analyzed. The clinical benefit of indications converted to regular approval was evaluated using the ESMO-Magnitude of Clinical Benefit Scale version 2.0 (ESMO-MCBS v2.0).

Results: In total, 90 cancer drugs for 115 indications were granted conditional approval by NMPA, supported by 146 pivotal trials and 111 confirmatory trials. Among pivotal trials, 63.1% were with sample size<100, 83.6% single-arm design and 86.4% with response-related rate (such as objective response rate, complete response rate, etc.) as the primary outcome. Conversely, more than 70% of confirmatory trials were randomized control design, and 54.1% using progression-related survival (such as progression-free survival, disease-free survival, etc.) as the primary endpoint. Notably, 20.9% of trials adopted response-related rate as the primary endpoint in both the pivotal and confirmatory stages, implying that these drugs without using such as overall survival as primary endpoints in post-approval confirmatory trials. At the cut off date (Mar. 1 2026), 12.2% (14/115) of confirmatory stage were not finished within the the validity period specified by NMPA, with an extra median years of 0.88 (0.08-2.94). Of 115 conditionally approved indications, four indications were finally withdrawn (Serplulimab, Mobocertinib, Pralsetinib, and Copanlisib) that failed to improve the primary endpoint in confirmatory trials, while one indication had no confirmatory trial initiated, and 74 were still ongoing. Noticeably, 36 (31.3%) fulfilled their post-marketing requirements and were converted to regular approval, among which 43.7% (n=17) had broadened indications. Surprisingly, only 7 (19.4%) indications used overall survival as the primary endpoint. 8 of 36 (22.2%) showed clinical benefit assessed by ESMO-MCBS.

Conclusion: More emphasis should be paid to the existing problems in the evidence of confirmatory trials (e.g., primary endpoints, study overdue), the uncertainty of clinical benefit, and the change of indications from conditional to regular approval. It is critical for indications with conditional approval to complete the confirmatory stage with primary endpoints definitely demonstrating clinical benefit in time. Regulatory agencies should further clarify the evaluation criteria for clinical benefit, set standardized requirements for indications expansion, and review rigorously for the confirmatory stage, thereby protecting patients from exposure to drugs with unconfirmed clinical benefit.


Event-Driven Trial Design for Recurrent Event Outcomes: A Blinded Monitoring Approach

Speaker: Jingwen Zhang, Department of Biomedical Statistics, The University of Osaka

Authors: Jingwen Zhang and Satoshi Hattori

Biography

Jingwen Zhang

Jingwen Zhang is a PhD student in the Department of Biomedical Statistics at the University of Osaka. Her research focuses on statistical methodology for clinical trials, particularly the design and analysis of recurrent event outcomes. She has an undergraduate background in medicine and holds a Master of Public Health (MPH) degree, with a focus on cardiovascular epidemiology.

Abstract

Background: Event-driven designs are widely used in randomized clinical trials with time-to-first-event outcomes to ensure adequate statistical power for a specified Type I error rate to detect the target treatment effect. In clinical trials evaluating chronic conditions, including oncology settings, recurrent event outcomes can provide additional information about treatment effects beyond the first event. However, extending conventional event-driven designs to recurrent events presents challenges because within-subject correlation structure among event times affects statistical power and is often difficult to specify at the design stage. Misspecification of this correlation may risk underpowered trials.

Methods: We propose an event-driven type design for recurrent event outcomes based on the marginal rates/means model. Rather than relying on a prespecified number of events, the proposed method derived a robust variance estimator free of treatment allocation under a local alternative. At the design stage, given a prespecified treatment effect, we set the target for variance. After the trial is initiated, we monitor the proposed robust variance estimator in a blinded manner using pooled data. The timing of the final analysis is conducted when the value of the proposed robust variance estimator reaches the target value calculated at the design stage. This procedure aims to achieve the target statistical power without requiring accurate specification of nuisance parameters, including within-subject correlation. We evaluated the operating characteristics of the proposed design through simulation studies under various scenarios and illustrated its application using data from a randomized clinical trial, of which the primary outcome is the number of severe infections of patients with a chronic disease.

Results: Simulation studies demonstrated that the proposed method maintained statistical power close to the target level across a range of scenarios by adjusting the durations of follow-up, including those with misspecified planning assumptions, without substantial inflation of the Type I error rate. The clinical trial application illustrated how blinded variance monitoring can guide the timing of the final analysis based on accumulating statistical information.

Conclusions: The proposed design offers a flexible, information-based alternative to conventional event-driven designs for clinical trials with recurrent event outcomes. By accommodating uncertainty in within-subject correlation through blinded monitoring, the method can help maintain the desired statistical power while providing a practical approach to determining analysis timing. This framework has potential applications in oncology clinical trials involving recurrent outcomes, supporting more reliable trial design and evidence generation.


TWICEBEE: A Two-stage Intra-patient Curve-free Bayesian Decision-Theoretic Dose Escalation Design

Speaker: Dehua Bi, Stanford University
Biography

Dehua Bi

Dehua Bi is a biostatistician specializing in Bayesian parametric and nonparametric methods for clinical trial design, survival analysis, and machine-learning applications. He is currently a postdoctoral fellow at the Stanford Cancer Institute, jointly mentored by Dr. Ying Lu and Dr. Crystal Mackall. Dr. Bi earned his PhD from The University of Chicago under the mentorship of Dr. Yuan Ji. His research focused on developing novel Bayesian methods for information borrowing, Bayesian transfer learning, digital twins, designing early-phase clinical trials, and creating flexible sample-size estimation and adaptation approaches under the Bayesian framework.

Abstract

We propose TWICEBEE, a two-stage Phase I intra-patient dose-escalation design for multi-cycle treatment settings. Motivated by repeated CAR T-cell administration, the design assumes a clinically justified nonincreasing working toxicity ordering across cycles at a fixed dose. Stage 1 uses accelerated titration, whereas Stage 2 adapts the calibrated curve-free Bayesian decision-theoretic design to a partial order on dose-cycle combinations, with patient-specific safety restrictions and dose-cycle-specific overdose protection. At trial completion, cycle-specific toxicity-based recommendations are made when direct information is available and admissible doses remain. We evaluated TWICEBEE in 12 history-independent scenarios and 17 history-dependent settings with up to eight treatment cycles, comparing safety, treatment exposure, and terminal recommendations according to each method’s dose-finding objective. TWICEBEE generally enrolled fewer patients, with trade-offs in treatment exposure, toxicity burden, and terminal dose selection across settings. The resulting recommendations identify candidate dose schedules for further clinical evaluation.


Stein-Encoder: A White-Box Supervised Encoder via Stein Identities in Multi-Modal Studies

Speaker: Xinzhou Guo, Hong Kong University of Science and Technology
Abstract

In multi-modal biomedical research, integrating high-dimensional genomic data with clinical baselines is essential for precision medicine. However, standard deep neural network approaches often entangle these modalities, obscuring the specific predictive impact of genetic features and leading to possibly suboptimal predictive performance. Motivated by the landmark METABRIC cohort primary breast tumors study, we propose the Stein-Encoder, a white-box supervised framework designed to isolate the genetic signal driving clinical outcomes conditional on nuisance covariates. By leveraging Stein’s method and residualization techniques, our approach constructs an interpretable single index that summarizes relevant biological heterogeneity while flexibly incorporating clinical factors and can be used to improve downstream prediction. We establish theoretical guarantees for identification, consistency and efficiency improvement. Applied to the METABRIC cohort, the Stein-Encoder outperforms unsupervised benchmarks in predictive accuracy. Crucially, it achieves structural disentanglement by revealing response-specific biological mechanisms: we find that tumor size is driven primarily by mitotic networks, whereas prognostic indices rely on a distinct proliferation-versus-immune axis. This work contributes a unified, computationally efficient framework that bridges statistical rigor with the representational power of neural networks, enabling interpretable, task-specific and efficient compression of multi-modal health data for a wide range of precision medicine applications, beyond biomarker discovery.

Biography

Xinzhou Guo

Xinzhou Guo is an Assistant Professor in the Department of Mathematics at the Hong Kong University of Science and Technology. He received his B.S. in Applied Mathematics from Peking University and Ph.D. in Statistics from the University of Michigan. Prior to joining HKUST in 2021, he did a postdoc at Harvard University. His main research interests are subgroup analysis, resampling methods, precision medicine and regulatory decision-making.


Multimodal AI Analysis Reveals Distinct Cancer Molecular Subtypes with Genomic and Epigenetic Interpretation

Speaker: Zheng Chen, Keiji AI Japan; The University of Osaka
Biography

Zheng Chen

Dr. Zheng Chen is the lead of Keiji AI Japan and an Associate Professor at the University of Osaka, specializing in artificial intelligence for biomedical research and drug development. His research focuses on multimodal representation learning, cancer multi-omics integration, and AI-driven analysis of complex biological and clinical data. He has published research in leading venues, including Nature Machine Intelligence, Nature Biomedical Engineering, NeurIPS, and ICLR. His work bridges cutting-edge AI research and real-world applications in precision oncology and clinical studies.

Abstract

Serial circulating tumor DNA measurements can track tumor burden during treatment, but whether joint genetic and epigenetic changes in cell-free DNA (cfDNA) are associated with treatment duration remains unclear. We developed a multimodal representation-learning framework integrating seven molecular and clinical inputs using a resource of 129,631 tests across four cancer types, without treatment-outcome supervision. We quantified baseline-to-early-treatment changes in the learned representations in independent patients who were excluded from encoder training and tuning. Greater representation change was associated with longer subsequent treatment duration in non-small-cell lung cancer patients receiving immune checkpoint inhibitors and in colorectal cancer patients receiving anti-EGFR therapy, even after accounting for tumor-fraction changes and sampling time. Regional methylation changes showed complementary associations with treatment duration, while representation-level associations remained detectable even in encoders pretrained without methylation.


Bayesian Framework for Clinical Validation of Multiple Companion Diagnostics with Limited Samples

Speaker: Songbai Wang, Johnson & Johnson Innovative Medicine
Biography

Songbai Wang

Dr. Wang is Executive Director and Head of Precision Informatics at Johnson & Johnson Innovative Medicine. Over a 40-year career spanning industry and healthcare research, he has built deep expertise in biostatistics, study design, and regulatory submissions for both drugs and diagnostic assays. He has provided analytical expertise and leadership for more than 35 regulatory submissions across the United States, Japan, China, and Europe. Combining medical training with expertise in biostatistics and informatics, he brings an integrated perspective to both therapeutic and diagnostic product development. Songbai holds a Master of Science in Public Health (MSPH) with a focus in Biostatistics from the University of South Florida and a medical degree from Tongji Medical University in Wuhan, Hubei, China.

Abstract

Clinical validation of multiple companion diagnostic tests (CDx) within a global oncology trial is challenging when assay-specific cohorts and residual specimens are limited. We propose a fit-for-purpose Bayesian framework in which observed outcomes for each candidate CDx constitute the primary evidence, while relevant reference-CDx data from the same randomized trial are borrowed when supported by analytical validation, diagnostic concordance, and clinical commensurability. Essential safeguards include prespecified estimands, priors, borrowing rules, and decision criteria, together with simulation-based evaluation of operating characteristics, transparent assessment of prior influence, and comprehensive sensitivity analyses. Analyses of clinical trial data illustrate implementation of the framework. Without replacing analytical validation or lowering evidentiary standards, this approach can improve the precision and efficiency of assay-specific clinical validation, expand the availability of validated CDx options, and ultimately broaden patient access to targeted therapies.


Optimized Adaptive Oncology Clinical Trials with a Surrogate Endpoint-Based Drop-The-Losers Design

Speaker: Chenguang Wang, Regeneron Pharmaceuticals
Biography

Chenguang Wang

Dr. Chenguang Wang is the Head of Quantitative Innovation and Statistical Strategy (QISS) at Regeneron. Previously, he served as an Associate Professor at Johns Hopkins University and worked as a Mathematical Statistician at the Center for Devices and Radiological Health (CDRH), FDA. Dr. Wang has extensive experience in clinical trial design and analysis, particularly in regulatory settings. He is also an elected Fellow of the American Statistical Association.

Abstract

Typical phase 1 oncology studies identify the maximum tolerated dose as the “optimal” dose for subsequent phases. With the advancement of molecular targeted agents and immunotherapies, however, evaluating two or more doses has become increasingly critical for dose selection. Such evaluation is often done in phase 2 studies in a randomized manner. In this paper, we evaluate the strategy of applying an adaptive phase 2/3 seamless design for dose selection in oncology studies. Specifically, we consider the “drop-the-losers” design, where multiple treatment arms and a control arm are administered during the initial stage, and a more effective arm is identified for later stages by a surrogate endpoint such as overall response. We derive the theoretical type I error inflation scale and conduct simulation studies to illustrate the impact of various factors on the type I error inflation in such designs. Furthermore, we demonstrate the findings through the design of a lung cancer trial and introduce a software that implements the proposed design.


Literature-based meta-analysis of adverse events accounting for heterogeneous follow-up duration in oncology clinical trials

Speaker: Satoshi Hattori, Department of Biomedical Statistics, Graduate School of Medicine, The University of Osaka, Japan
Biography

Satoshi Hattori

Satoshi Hattori is Professor of Biostatistics at The University of Osaka, Japan. His research focuses on statistical methodology for clinical trials and evidence synthesis, with particular interests in survival analysis, meta-analysis, causal inference, and methods for addressing publication bias. He has contributed to the development of statistical methods for the design and analysis of clinical studies, particularly in oncology. He is currently President of the Japanese Region of the International Biometric Society.

Abstract

Assessing the safety profile of a drug based on a single clinical trial is challenging, and meta-analysis can provide valuable evidence by combining information from multiple studies. In oncology clinical trials, the follow-up duration for each patient is often determined by the occurrence of the primary time-to-event endpoint. Consequently, simply pooling incidence proportions of adverse events across studies may be misleading, particularly when follow-up durations differ between treatment groups. In this talk, we discuss methods for addressing this issue. Specifically, we propose novel approaches to meta-analysis of adverse events that account for heterogeneous follow-up durations by utilizing Kaplan–Meier estimates of the primary efficacy endpoint to obtain information on follow-up duration.


Bayesian Framework for Clinical Validation of Multiple Companion Diagnostics with Limited Samples.

Speaker: Jingjing Ye, DataDoctor LLC
Biography

Jingjing Ye

Dr. Jingjing Ye is a biopharmaceutical executive and former senior FDA regulatory leader with more than 20 years of experience, specializing in digital transformation, advanced analytics, and data-driven drug development. During her 10-year tenure at FDA, she has been the statistical lead in setting up multiple regulatory precedents and innovations, including companion diagnostics, AI-assisted imaging devices, benefit-risk assessment in primary screening test in CDRH, Bayesian analysis in pediatric extrapolation, external control in rare diseases and concurrent global review and decision (Project Orbis, Real-time Oncology Review (RTOR) and Assessment Aid) in CDER. Dr. Ye is an elected Fellow of the American Statistical Association (ASA). As a founder and principal consultant of DataDoctor LLC, Dr. Ye is driven by a commitment to improving patient outcomes and is passionate about multidisciplinary collaboration to advance innovations in drug development.

Abstract

TBD