Acquired regions of homozygosity in adult B-cell acute lymphoblastic leukemia: Insights from a machine learning–powered analysis of genomic data from a multicenter cohort Free

INTRODUCTION

Acquired regions of homozygosity (aROHs), due to mitotic recombination between homologous chromosomes, may contribute to the expansion of a neoplastic clone if they involve mutated tumor suppressor genes (TSGs) or oncogenes (OGs), with subsequent loss of the wild-type allele and duplication of the aberrant one. aROHs can be readily detected across the genome of patients (pts) with B-cell acute lymphoblastic leukemia (B-ALL) by single nucleotide polymorphism (SNP)-array analysis. However, their potential contribution to leukemogenesis remains poorly investigated. Our previous work (Blood (2024) 144 (Supplement 1): 2844) demonstrated the recurrence of aROHs in a large cohort of adult Philadelphia chromosome-negative (Ph-) B-ALL pts enrolled in the PETHEMA LAL-19 trial (NCT 04179929). Most of these alterations did not involve genes related to B-ALL leukemogenesis and, according to conventional statistical analyses, did not impact prognosis. On the contrary, due to their association with a low rate of mutation and with preserved diploidy, we hypothesized aROHs without driver genes for B-ALL could play a beneficial role for the genome integrity, although this was not reflected in pts' survival. Based on these contradictory findings, we aimed to evaluate if these aberrations could actually have a prognostic impact through a machine learning-mediated approach.

METHODS

aROHs were identified by SNP-array analysis on DNA extracted from infiltrated bone marrow, using the CytoScan™ 750K platform, according to the manufacturer's instructions (ThermoFisher Scientific, MA, USA). Briefly, ROHs ≥3 Mb in size, not associated with copy number variations, located in regions covered by ≥20 consecutive probes and not overlapping germline polymorphisms in ≥50% of their length were selected to identify only putatively acquired events. Following conventional statistical analysis, we designed an analytical strategy using the LightGBM machine learning model provided by Microsoft (WA, USA). Death and relapse were predefined as target variables (TVs), and the ability of aROHs to predict these two outcomes was assessed. Based on the median follow-up (FU) of the cohort, a survival time cutoff was established at 60, 100, and 180 days (d), and six predictive models for the TVs were generated (one for each cut-off). Each model included an average of 37 informative clinical and molecular variables (range: 29–43), including the total size of aROHs per pt (ΣaROHs/pt), the involvement of pan cancer TSGs and OGs within aROHs, and their possible extension to telomeres. Multiple rounds of hyperparameter tuning were performed using grid search combined with an 80/20 cross-validation split to identify the best-performing predictive model in each case.

RESULTS

We previously described a total of 332 aROHs in 156/277 pts (56%), included in our study from December 2019 to March 2024, with a median FU of 17 months. Median age was 40 years (range 18-60). The male/female ratio was 1:1. 94 % of aROHs were segmental, distributed across the whole genome. Involvement of pan cancer TSGs, OGs and telomeres was reported in 44%, 57% and 19% of cases, respectively. Based on the ROC curves derived from the six models using the survival time cut-offs previously described, d+100 was identified as the most effective time point for accurate prediction of the TVs. The study assessing the prognostic weight at d+100 of informative variables, based on absolute values, demonstrated that ΣaROHs/pt was the most important factor (importances 22 and 7 for death and relapse, respectively), after minimal residual disease (MRD) levels at the end of induction (end-Ind) (importances 45 and 27 for death and relapse, respectively). Spearman's correlation analysis revealed that ΣaROHs/pt was strongly negatively associated with both the risk of relapse and the risk of death (Spearman's ρ: death –0.1 for ΣaROHs/pt vs. +0.3 for MRD positivity at end-Ind; relapse –0.6 for ΣaROHs/pt vs. +0.6 for MRD positivity at end-Ind). In contrast, the presence of pan-cancer TSGs, OGs, and telomeric regions within aROHs was only marginally associated with an increased risk of both events.

CONCLUSIONS

Our machine-learning mediated approach revealed that ΣaROHs/pt may represent an important variable for early risk stratification of Ph- B-ALL pts, with favorable impact on their survival.