Inflammatory Transcriptional Signature Can Discriminate De Novo and Secondary AML

Acute Myeloid Leukemia (AML) is the most common leukemia in adults and is defined by > 20% myeloblasts in the peripheral blood or bone marrow. AML can arise de novo or can arise after a diagnosis of Myelodysplastic Syndrome (MDS). MDS is a myeloid neoplasm characterized by bone marrow dysplasia and cytopenia with a 30% AML transformation rate. Previous studies have demonstrated that de novo and secondary AML could be determined without clinical history if the presence of ‘secondary-like’ mutations were identified. However, whether transcriptional changes can refine the discrimination of these entities is unknown. Here, we hypothesize that inflammation-driven MDS will be transcriptionally reprogrammed toward a pro-inflammatory myeloid lineage that will be imprinted in secondary AML transcriptomes. To evaluate this point, we performed a gene expression analysis of the patient’s bone marrow samples with either MDS, secondary AML, and de novo AML from BeatAML2 (N = 671) and GSE15061 (N = 366) datasets.

We conducted a gene set enrichment analysis of inflammation pathways from the Molecular Signatures Database (MSigDB). Consistently, MDS and AML patients with prior MDS had a higher enrichment for hallmark inflammation p = 1.1e-04 and p = 8.4e-05, respectively. Several cytokine receptors were consistently enriched in MDS and prior MDS AML, including IL18RAP, IL18R1, IL1R1, IL2RB, and IL4R. Next, we performed a systematic association of inflammation with somatic events. Based on a Fisher’s Exact test, ASXL1 (p = 0.02937) and RUNX1 (p = 8.3e-4) were enriched for inflammation-high samples irrespective of diagnosis consistent with prior reports. We additionally compared transcriptomes in each disease category in the context of signatures that are associated with specific myeloid lineage enrichment, including HSC, CMP, GMP, and monocytes. We found MDS and secondary AML to be strongly enriched for the monocytic lineage and strongly correlated with inflammation (p = 6.1e-15). Collectively, these data suggest that MDS-associated inflammatory signatures are imprinted in secondary AML patients.

To determine the additive value of transcriptional signatures on genomic data on discriminating between de novo and secondary AML with prior MDS, we first trained an XGBoost model, a gradient boosting machine algorithm, to predict prior MDS based on genomic lesions by splitting the BeatAML cohort into training (n = 469) and testing (n = 201). Feature selection was performed to select the top 13 genomic features. This initial training model demonstrated an AUC ROC of 0.7012 with marginal performance in correcting predicting 36.8% (7 / 19) of the AML patients with prior MDS. XGBoost recapitulated the enriched features identified in prior MDS in BeatAML2, such as somatic mutations of ASXL1 and RUNX1. Notably, FLT3-ITD and CBFB-MYH11 fusions were not associated with prior MDS and are known to be associated with de novo AML. To determine whether transcriptional signatures can refine mutationally directed prior MDS assignment, our second model was a combination of these mutations and transcriptional signatures. Prediction performance improved in this joint model with an AUC ROC of 0.8122. Notably, 78.9% (15 / 19) AML patients were accurately predicted as prior MDS. Thus, transcriptional signatures improve the performance of the model and is needed to accurately discriminate between de novo and secondary AML with prior MDS. Taken together, we developed a preliminary machine learning algorithm to integrate genomics and transcriptomic data to predict AML patients with prior MDS. We plan to evaluate other external datasets from the NCBI Genome Expression Omnibus and NCI’s Genomic Data Commons database.