Interpretable models for high-risk neuroblastoma stratification with multi-cohort copy number profiles

Although high-risk neuroblastoma (HR-NB) is clinically heterogeneous, it is treated in a similar fashion without additional risk stratification. Based on the 4 copy-number profiles with 556 HR-NB subjects and 14 potential risk factors of neuroblastoma, we develop an interpretable machine learning model with L ₀ penalized global AUC summary ( L ₀GAUCS) maximization, and identify 6 and 4 molecular factors associated with overall and event-free survivals (OS and EFS) of HR-NB, respectively. We further construct a six-factor model for OS and a four-factor model for EFS, and categorize HR-NB patients into 4 subtypes (Excellent, Good, Fair, and Poor) for both OS ( P = 1. 59 e − 11 ) and EFS ( P = 1. 73 e − 06 ). Particularly, 14.05% and 6.75% HR-NB patients are in the Excellent (I) and Poor subtype (IV) with median OS times of 137. and 14.5 months, respectively. Patients from such distinct subtypes may be assigned to different experimental therapies in future trials. Furthermore, although it is well known that infants (children less than 1 year) has significantly better prognosis in neuroblastoma, we discover that infants with MYCN amplification (MNA+) has unfavorable OS and EFS in HR-NB. Infants with MNA+ have the hazard ratio of 2.9 (95% CI: 1. 33 − 6. 34 ) and P value of 3. 67 e − 06 for OS, and hazard ratio of 2.61 (95% CI: 1. 05 − 6. 48 ) and P value of 0.0007 for EFS. The unexpected but important finding that the survival of infants with MNA+ has significantly worse prognosis in HR-NB may have clinical implications.