Genotype-Phenotype Correlations in Patients with Myeloid Malignancies Using Explainable Artificial Intelligence

Background:

Myelodysplastic syndromes (MDS) and related myeloid malignancies are highly variable in both their clinical manifestations and underlying genetic abnormalities. While few mutations in myeloid malignancies are considered disease-defining, significant and complex associations between these genes exist and can influence the clinical characteristics and disease phenotype.

Here, we took advantage of a large, international cohort of patients with myeloid malignancies to define genotype-phenotype relationships using state of the art machine learning models.

Methods

Data were collected for patients (pts) from the Cleveland Clinic (CC; 652 pts), Munich Leukemia Laboratory (MLL; 1509 pts), and the University of Pavia in Italy (UP; 536 patients). Clinical data including CBC at time of diagnosis and a genomic panel of 20 commonly mutated genes in myeloid malignancies were analyzed.

Gene-gene correlations within disease subtypes, individual genes’ co-occurrence or exclusivity within disease subtypes, and the co-occurrence or exclusivity of individual genes with clinically meaningful features including karyotypic abnormalities and severe cytopenias (defined as hemoglobin < 8 g/dL, platelets < 50k/dL, and ANC < 1k/dL) were evaluated using multiple machine learning/correlation/feature extraction algorithms.

Results

2697 pts were included, 1630 (60%) with MDS, 399 (15%) with chronic myelomonocytic leukemia (CMML), 142 (5%) with idiopathic cytopenia of undetermined significance (ICUS), 129 (5%) with MDS-MPN overlap syndromes (MDS-MPN), 95 (4%) with primary myelofibrosis (PMF), 93 (3%) with clonal cytopenia of undetermined significance (CCUS), 52 (2%) with essential thrombocythemia (ET), 41 (2%) with polycythemia vera (PV), and 26 (1%) with other myeloproliferative neoplasms (MPNs). The median age at diagnosis for the entire cohort was 70 years [36 - 86]. Of patients with karyotype data available, 1091 pts (50%) had a normal karyotype, 17 (1%) had chromosome 17 abnormalities, 96 (4%) had chromosome 7 abnormalities, 145 (7%) had chromosome 5 abnormalities, and 123 (6%) had a complex karyotype. The most commonly mutated genes were: TET2 (28%), ASXL1 (22%), SF3B1 (22%), SRSF2 (19), JAK2 (11%), DNMT3A (9%), RUNX1 (9%), and U2AF1 (6%)

SF3B1 mutations were associated with normal karyotype (NK), age <65 years, ANC >1 k/dL, platelets (plts) >50 k/dL, marrow blasts (MB) <10% and hemoglobin (hb) <8 g/dL. TP53 mutations were associated with complex karyotype, chro 5, 7, or 17 abnormalities.

Clinical characteristics were also associated with specific genomic alterations (Figure 1). For example, NK correlated with the presence of SF3B1, ZRSR2, DNMT3A, a higher number of mutations, and absence of TP53, ASXL1, or KRAS; chromosome 5, 7, and 17 abnormalities were associated with a lower mutation number and the presence of TP53 mutations; complex karyotype correlated with the absence of TET2 and SF3B1 and the presence of TP53; age < 65 was associated with the presence of NRAS and JAK2 mutations and the absence of TET2, SF3B1, and SRSF2 mutations; hemoglobin < 8 g/dL positively correlated with mutation number and SF3B1 mutations and negatively correlated with TET2 mutations; ANC < 1 negatively correlated with JAK2, SF3B1, and DNMT3A mutations; platelets < 50k/dL negatively correlated with SF3B1 and JAK2 mutations, and positively correlated with the number of mutations; and MB <10% positively correlated with SF3B1 mutations and negatively correlated with number of mutations and ASXL1, RUNX1, TP53, and STAG2 mutations (Figure 1).

Conclusions

We applied machine learning techniques to reveal the complex relationships between mutational data and the clinical characteristics of several myeloid malignancies using a large, international patient cohort. In addition to correctly identifying previously described genotype-phenotype relationships, we identified several other intriguing relationships such as the relationship of particular mutations to the development of different cytopenias, demonstrating the potential utility for machine learning approaches in interrogating genomic data.