Gene-based outcome prediction in multiple cohorts of pediatric T-cell acute lymphoblastic leukemia: a Children's Oncology Group study

T-cell acute lymphoblastic leukemia (T-ALL) affects approximately 15% of newly diagnosed pediatric ALL patients. Continuous complete clinical remission (CCR) in T-ALL patients is now approaching 80% due to the implementation of aggressive chemotherapy protocols [1–6]. However, patients that relapse (R) have poor prognosis and aggressive therapy can lead to long-term side effects in those that achieve CCR [7]. In the clinical setting, age and white blood cell count (WBC) at diagnosis are used to stratify B-lineage ALL patients as either standard or high risk, significantly impacting on the type and intensity of post-induction therapy used. However these NCI-defined criteria have been shown to have little prognostic value in T-ALL disease [1–3]. Improved markers are needed for outcome prediction to improve T-ALL patient stratification.

Common karyotypic abnormalities have been identified in some forms of leukemia and have proven useful for outcome prediction [8–12]. In precursor B-lineage ALL (pre-B ALL), the presence of hyperdiploidy or translocations such as E2A-PBX1, BCR-ABL, or ETV6-RUNX1 contribute to the severity of disease and response to chemotherapy [8, 9]. In T-ALL, increased expression of TLX1/HOX11 has been associated with favorable outcome [10, 11, 13, 14], whilst aberrant expression of TAL1, LYL1 and TLX3 and deletions at 6q15-16.1 have been linked to poor prognosis [11, 15, 16]. Recent work by Coustan-Smith and colleagues [17] has led to the identification of a new very high risk subset of T-ALL (early T-cell precursor leukemia) that has a distinct expression profile and immunophenotype. However, due to the lack of consensus between studies and the small proportion of T-ALL patients that carry these genetic or molecular aberrations, the identification of a universal molecular signature has become a priority. Several studies have attempted to identify gene signatures that predict induction failure and/or relapse in T-ALL [8, 18, 19], but have had limited success verifying their findings in other patient cohorts. The current study aimed to identify robust gene signatures that could be used for the accurate prediction of relapse at the time of diagnosis, in independent patient cohorts, and across different experimental platforms.

The fate of children who relapse with T-ALL remains dismal. This has fuelled considerable research into the discovery of 'risk factors' that are indicative of a patient's likelihood of relapse before post-induction therapy is prescribed. Our study statistically modeled a 5-GC that successfully predicted T-ALL patient outcome in four independent studies across different platforms. In a complementary approach we used biological signatures to successfully model patient outcome in these studies. To our knowledge this is the first time gene classifiers have been developed that accurately model ALL relapse in more than two independent cohorts. It is important to note that the models described here could not be used to down-grade a patient's risk classification since predictions of outcome under these models is in the context of the therapy each patient actually received. Thus CCR patients in these cohorts may potentially have relapsed if treated with lesser therapy. However, it would be possible to use such models to augment therapy. In the case of patients already stratified as high-risk (the majority of T-ALLs), this could include bone marrow transplant in first remission [6] or the use of experimental therapeutics. The best models generated in this study (the 5-GC and Cell Adhesion Pathway) were associated with good average specificity across the four cohorts (88-93%), but achieved lower average sensitivity (47-70%). In clinical terms this means that application of these models as a diagnostic test could have successfully identified up to two-thirds of the patients in these cohorts destined to relapse, whilst potentially over-treating only a small percentage of patients (7-12%) that would have achieved CCR under current protocols. Although higher sensitivity would be desirable, the correct identification of even a few patients destined to relapse could further improve cure rates.

In recent years much criticism has been directed towards microarray studies aiming to identify gene markers from small cohorts. Owing to the dimensionality of the data it is often possible to select genes at random that can discriminate between two phenotypes or patient subgroups with surprising accuracy. Furthermore, many statistical tools over-fit data such that the ability of classifiers to discriminate between phenotypes only extends to the cohort in which they were developed. However the probability of selecting gene classifiers at random that can discriminate between phenotypes in more than one cohort is vanishingly small. Validation of classifiers across multiple cohorts as described here (especially those identified using a permutative resampling algorithm such as the Random Forest) is empirical evidence of their robustness.

In this study low expression of the IL-7R was recorded in diagnostic T-ALL specimens from patients who later relapsed, linking low IL-7R expression to eventual therapy failure. The IL-7 cytokine is normally essential for T-cell development, survival and proliferation [27], and can inhibit both spontaneous apoptosis [28] and the apoptotic responses to chemotherapeutic agents in T-ALL [30], with the level of expression of the IL-7R correlating with these responses [28]. As such, IL-7 has been proposed as an important factor supporting leukemogenesis [31], but a proportion of T-ALL patients have blasts that do not respond to IL-7 [28]. This latter observation has been correlated with tumor maturation stage but it is also possible that it represents the acquisition of growth-factor independence. Growth-factor independence is a classical hallmark of a successful cancer cell [32] and indicates the development of potent pro-survival mechanisms. Importantly, T-ALL patients with an IL-7 non-responsive phenotype demonstrate poorer clinical responses to glucocorticoid therapy and thus have an adverse prognosis [28], consistent with our own findings in the present study.

Despite the obvious relevance of the IL-7R for T-ALL, the genes of the 5-GC were not selected on the basis of biological function. As such the 5-GC is considered as a tool for prognosis rather than for the interpretation of mechanisms of relapse, although the individual genes themselves do have links with cancer. The ABTB2 gene is involved with protein-protein interactions through its ankyrin repeats and BTB (POZ) domains. Although its specific function is unknown, one study has reported the up-regulation of ABTB2 in gastric tumor metastasis, highlighting the possible role of this gene in aggressive malignant phenotypes [33]. The FAM13A1 gene has an unknown function but is induced in various cell types exposed to hypoxic conditions [34]. It has been reported to be down-regulated in malignant thyroid tissue [35] but up-regulated in ovarian and breast cancer with links to poor prognosis [34]. PLAC8 is conserved in all vertebrates and is expressed at high levels in immune cell types [36]. The function of this gene is also unclear but has been linked to proliferation and apoptosis. PLAC8 is over-expressed in hepatocellular carcinoma tumours [37] and reduced in Paclitaxel-resistant prostate cancer [38].

Our alternative approach focused on identifying biological pathways that are involved in the progression to therapy failure in T-ALL. The NFκB and Wnt signaling pathways both had significant predictive power in this regard. The NF-κB pathway is highly active in T-ALL and is one of the major mediators of NOTCH1-induced transformation, establishing NFκB as a potentially promising target for T-ALL therapy [39]. The Wnt pathway is also important for T cell development and proliferation and is deregulated in several types of leukemia [40]. Although few studies have directly reported a role for Wnt signaling in the pathogenesis of T-ALL, antagonism of Wnt signaling has been shown to lead to chemotherapy resistance in a model of acute myeloid leukemia, via the downstream action of NFκB [41]. The pathway model that predicted relapse with the highest accuracy across all four cohorts in the present study was the Cell Adhesion Receptor geneset, with 12 out of the 14 genes representing integrins (Table S3, Additional file 1). Interestingly, LGALS8, the final member of the 5-GC, codes for a secreted mammalian beta-galactosidase binding protein (galectin-8) that binds with high affinity to a variety of cell surface integrins, thereby modulating cell adhesion and cell survival [42, 43]. Adhesion between host and tumor cells, and extrinsic signals within the tumor microenvironment can promote an optimal niche for tumor cell survival and is an essential component of tumor invasion and metastasis [44, 45]. New strategies for therapy have consequently been designed to disrupt these tumor-stromal cell interactions. For example, the inhibition of CXCR4 (a key receptor for tumor cell migration and adhesion) has been shown to overcome stromal-cell mediated drug resistance in acute myeloid leukemia and chronic lymphocytic leukaemia [46]. Clinical trials using specific integrin inhibitors have also shown promise in different types of solid tumours [47, 48]. Clearly cell adhesion interactions have an important role to play in tumor progression; the observations from the present study indicate that they may also contribute to the mechanisms that lead to disease recurrence in ALL.

We have used two different approaches to identify gene signatures that can successfully discriminate relapse and CCR patients at the time of diagnosis across multiple patient cohorts and platforms. Defined gene classifiers (such as the 5-GC) containing a smaller number of genes may be useful to augment existing risk stratification regimens for patients diagnosed with ALL as they can easily be adapted to qRT-PCR technology [18, 22]. The complementary method we present here uses larger, biologically defined genesets that provide important clues to the underlying mechanisms of relapse. Such insights may provide for the development of improved therapies for T-ALL.

The authors declare that they have no competing interests.