Subgrouping by gene expression profiles to improve relapse risk prediction in paediatric B-precursor acute lymphoblastic leukaemia

Abstract

Relapsed acute lymphoblastic leukaemia (ALL) remains a prevalent paediatric cancer and one of the most common causes of mortality from malignancy in children. Tailoring the intensity of therapy according to early stratification is a promising strategy but remains a major challenge due to heterogeneity and subtyping difficulty. In this study, we subgroup B-precursor ALL patients by gene expression profiles, using non-negative matrix factorization and minimum description length which unsupervisedly determines the number of subgroups. Within each of the four subgroups, logistic and Cox regression with elastic net regularization are used to build models predicting minimal residual disease (MRD) and relapse-free survival (RFS) respectively. Measured by area under the receiver operating characteristic curve (AUC), subgrouping improves prediction of MRD in one subgroup which mostly overlaps with subtype TCF3-PBX1 (AUC = 0·986 in the training set and 1·0 in the test set), compared to a global model published previously. The models predicting RFS displayed acceptable concordance in training set and discriminate high-relapse-risk patients in three subgroups of the test set (Wilcoxon test p = 0·048, 0·036, and 0·016). Genes playing roles in the models are specific to different subgroups. The improvement of subgrouped MRD prediction and the differences of genes in prediction models of subgroups suggest that the heterogeneity of B-precursor ALL can be handled by subgrouping according to gene expression profiles to improve the prediction accuracy.

Keywords: B-precursor acute lymphoblastic leukaemia; gene expression profiles; minimal residual disease; non-negative matrix factorization; relapse.

FIGURE 1

Flowchart of the prediction model combining NMF for subgrouping and logistic and Cox regressions for predicting MRD and RFS. Beside 10‐fold cross‐validations, performance of the model is validated by predicting the early response and relapse in the test set

FIGURE 2

Comparison of NMF of different ranks shows that the training set constitutes of four subgroups. (A) The description length of the gene expression profiles of NMF of different ranks, calculated by methods based on empirical histogram and based on fitted distribution. (B) Consensus matrices for ranks 2 to 7, averaging over 30 runs. A pixel is coloured from blue, when the pair of samples are never in the same cluster, to red, when the pair of samples are always in the same cluster. In consensus matrices for ranks 5 to 7, circles indicate the blocks getting merged

FIGURE 3

Subgrouping of the training set by NMF and the association of subgroups with genetic subtypes. (A) A consensus matrix for rank 4 NMF, averaging the top 20 from 60 runs. A pixel is coloured from blue, when the pair of samples are never in the same cluster, to red, when the pair of samples are always in the same cluster. On the left of the consensus matrix, the hierarchical clustering of samples is shown. (B) Annotation tracks for consensus, subgroups, genetic subtypes, organism parts, white blood cell at diagnosis, age, race and gender. Legends and scales of the tracks are shown on the right, where “Hyperdiploidy” is short for “Hyperdiploidy without trisomy of both chromosomes 4 and 10”, and “Trisomy of 4/10” is short for “Trisomy of both chromosomes 4 and 10”

FIGURE 4

Performance of logistic regression models predicting MRD. (A, C, E and G) ROC curves (red solid lines) for subgroups I, II, III and IV in the training set. The value of λ, the number of coefficients (k) and the area under the ROC curve (AUC) are shown for each subgroup. The diagonal dashed line is the no discrimination line. (B, D, F and H) ROC curves (red solid lines) comparing predicted MRD with early response for subgroups I, II, III and IV in the test set. The AUC is shown for each subgroup

FIGURE 5

Performance of the Cox regression models predicting RFS. (A, C, E and G) Time‐dependent ROC curves for subgroups I, II, III and IV in the training set. Solid, dashed and dotted red lines indicate 1‐year, 2‐year and 5‐year RFS respectively. The value of λ, the number of coefficients (k) and concordance statistic (c) of the Cox regression model are shown for each subgroup. The diagonal dashed line is the no discrimination line. (B, D, F and H) Wilcoxon tests of the Cox regression models in predicting relapse within 3 years for subgroups I, II, III and IV in the test set. Points indicate the risk predicted by the Cox regression model. Horizontal bars indicate the averages among CCR patients and among relapsed patients. The p value of a Wilcoxon test is shown for every subgroup