C088 | OVERCOMING DATA LIMITATIONS BY BUILDING AND EXPANDING A UNIFIED CLINICAL-MOLECULAR RNA-SEQ DATASET IN ACUTE MYELOID LEUKEMIA

Acute Myeloid Leukemia (AML) is an aggressive, molecularly heterogeneous malignancy with limited survival despite recent advances in its characterization and treatment. Genomic data, including global expression profiles, are key for precision medicine, aiding patient stratification and target discovery. Access to datasets by artificial intelligence and machine learning algorithms promises to unveil otherwise hidden clinical and biological insights. However, this requires large, well-annotated datasets, which are lacking, and technical variability prevents RNA-seq dataset integration across studies. Here, we aim to address this issue i) by establishing a unified real-world RNA-seq AML cohort and ii) by generating, building on this, synthetic data to expand sample availability. Specifically, we analyzed 1 unpublished and 3 public RNA-seq datasets, forming a cohort of 918 patients that we reclassified by ICC/ELN 2022. As expected, standard normalization methods failed to integrate datasets, clustering samples by dataset of origin and losing any biological coherence (fig. 1a). Therefore, for data harmonization, we used a novel algorithm (named "Gu") combining 8 normalization methods with ComBat batch correction, which led to 27 integrated datasets (fig. 1b). In order to select the best, we tested various strategies, finally resulting in a unified real-world cohort of n=894 clinically and molecularly annotated cases. Importantly, this preserves biological integrity, consistently reflecting ICC 2022 classifications (fig. 1c). Moreover, pathway analysis revealed broader detection of dysregulated networks, reflecting improved stability from the larger cohort. Using this harmonized dataset, synthetic data were generated via the "synthpop" R package, incorporating clinical variables and 66 selected gene expression features, and performing a 300× augmentation that yielded n=2682 synthetic samples. Importantly, gene expression networks and ICC-based clustering were preserved, with improved subgroup separation (fig. 1d). Both real and synthetic datasets accurately replicated ELN risk stratifications (fig. 1e). At the best of our knowledge, our integrated fully-annotated real-world dataset represents the largest (n=894 AML cases) available in the setting of AML. From this, synthetic data generation allowed to reach a huge integrated dataset of n=2682 synthetic samples, supporting machine learning applications, particularly for rare but clinically relevant subgroups.