AUTHORS
Karun Neupane, MD, Daniel Deavila, MS2, Ryan F. Gebert3, Adrian Baz4, Praneeth Reddy Sudalagunta, PhD5, Bailey J. Spence6, Hana Qasim, Fellow7, Rafael Renatino Canevarolo, PhD8, Mark B. Meads, PhD9, Matthew Froid10, Kinga M. Cid11, Pompom Ghosh, PhD12, Rachel Howard, PhD13, Phillip Reisman14, Brandon Blue, MD15, Ariel F. Grajales-Cruz, MD16, Melissa Alsina, MD17, Rachid Baz, MD18, Ariosto Siqueira Silva, PhD19, Kenneth H. Shain, MD, PhD20
AFFILIATIONS
1 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
2 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
3 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
4 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
5 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
6 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
7 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
8 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
9 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
10 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
11 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
12 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
13 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
14 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
15 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
16 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
17 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
18 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
19 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
20 H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States
ABSTRACT
Background: Precision medicine approaches in multiple myeloma (MM) require large, accurately annotated cohorts linked to longitudinal clinical and molecular data. However, assembling such datasets from electronic medical records remains labor-intensive. We applied our in-house artificial intelligence (AI) platforms, CUPI and ECHO, for automated cohort construction and clinical annotation to support development of a large language model (LLM)-based framework for predicting measurable residual disease (MRD) negativity following frontline therapy in newly diagnosed MM (NDMM) in the quadruplet induction and MRD era (2018 onwards).
Methods: CUPI and ECHO were deployed within the Moffitt Cancer Center Pentecost Myeloma Research Center (PMRC) database. CUPI automated abstraction of longitudinal clinical journeys for >18,000 MM-spectrum patients from the Moffitt Cancer Analytics Platform (MCAP), a PHI-compliant data warehouse integrating electronic health records and multi-omic data, achieving 93.6% concordance with human abstractors in a 1,145-patient validation cohort. ECHO decomposes complex clinical queries into auditable subtasks executed through structured pipelines integrating curated clinical data, external knowledge resources, and computational analyses. Leveraging the CUPI database, ECHO identified NDMM patients who underwent clonoSEQ-based MRD assessment during frontline therapy and integrated demographic, treatment, laboratory, cytogenetic, transcriptomic, molecular, MRD, and outcome data. ECHO performance was evaluated against manual curation.
Results: ECHO identified 9,247 patients with complete frontline treatment data among which, 1,145 patients had clonoSEQ-based MRD assessment during frontline therapy. Concordance with manual cohort identification was 91%. Median age was 66 years; 79.2% were White, 14.9% African American, and 10.9% Hispanic. Daratumumab-containing regimens were administered in 65.8% of patients, including daratumumab-based quadruplets in 44.3%, while 23.4% received non-daratumumab triplet therapy; 65.6% received high-dose melphalan.
MRD negativity was achieved in 588 patients (51.4%), while 557 (48.6%) remained MRD positive. Consistent with the prognostic significance of MRD status, median progression-free survival (PFS) was not reached in the MRD-negative cohort vs 73.7 months (95% CI, 59.2-86.1) in MRD-positive patients (HR 2.97, 95% CI 2.26-3.91; p<0.01). Median follow-up for MRD positive and negative cohort was 25.9 months and 30.9 months respectively. Next, we investigated clinical and molecular data from samples collected within 90 days before and 30 days after treatment initiation for association with outcome: longitudinal laboratory data were available for 966 patients, diagnostic FISH for 319, and RNA-seq for 62. Cytogenetic analyses showed associations between MRD status and specific abnormalities, with t(11;14) and del(17p) associated with lower rates of MRD negativity and t(4;14) associated with higher rates. Examination of bulk RNA-seq-derived single-sample gene set enrichment analysis (ssGSEA) demonstrated enrichment of MYC targets V2 (p=0.02), E2F targets (p=0.03), oxidative phosphorylation (p=0.05), and fatty acid metabolism (p=0.05) in MRD-positive patients. MRD-positive disease also demonstrated enrichment of the MS/MMSET molecular subtype signature (p=0.04) and a trend toward the PR/proliferation signature (p=0.07), consistent with proliferative and metabolically active residual disease. Consistent with the t(11;14) subtype, there was also a trend for CD1/CD2 transcriptional pattern in MRD-positive patients.
Cohort identification and multimodal data extraction were completed within minutes through prompting alone, without programming or database expertise. ECHO generated an auditable research notebook providing transparent, reproducible access to the underlying data and facilitating cross-disciplinary collaboration.
Conclusions: AI-enabled cohort construction and phenotyping facilitate rapid, scalable identification of clinically relevant NDMM populations and extraction of longitudinal multi-omic data with strong concordance to manual abstraction. Using ECHO, we generated a research-ready MRD cohort and identified clinical, cytogenetic, and transcriptomic correlates of MRD outcomes. Ongoing work focuses on expanding molecular profiling and integrating multimodal data into predictive models of MRD negativity following frontline therapy in NDMM.
DISCLOSURES
Karun Neupane, MD: No relevant conflicts of interest to declare.