Mutational dynamics of early and late relapsed childhood ALL: rapid clonal expansion and long-term dormancy

Childhood acute lymphoblastic leukemia (cALL) is the most frequent pediatric cancer, accounting for ∼25% of all cases.¹  Despite improved treatment strategies, ∼20% of patients are either refractory or ultimately relapse, making cALL the second cause of cancer-related mortality among children and adolescents.^2-5 

Tumor plasticity, based on the random accumulation of somatic mutations combined with external pressures, such as chemotherapy, is a breeding ground for clonal competition and evolutionary adaptation leading to treatment resistance.^6,7  Over the past few years, next-generation sequencing (NGS) has provided meaningful information toward the characterization of these evolutionary processes and the identification of somatic events providing selective advantages. In the context of hematological malignancies, mutations in the RAS pathway genes have been associated with early relapse and chemoresistance in cALL.⁸  Hyperdiploid leukemia cells mutated in the histone acetyltransferase (HAT) domain of CREB-binding protein (CREBBP) were identified in glucocorticoid-resistant patients,^9,10  whereas the mutated cytosolic 5′-nucleotidase II gene (NT5C2) was shown to be involved in resistance to treatment using nucleoside analog therapies.^11,12  Such studies indicated that treatment could either lead to the eradication of nonresistant clones and the emergence of fitter and more aggressive subclones or, in rarer cases, to the persistence of an already existing resistant and dominant clone.^7,13-17 

Although important, these studies generally offered only snapshots of tumors and therefore failed to capture the complexity of their temporal evolution.⁶  The latter is essential to decipher the full spectrum of relapse determinants. Furthermore, early cALL relapses, occurring during treatment, and late relapses, occurring at least 36 months after the initial diagnosis, are likely to present different evolutionary mutational processes. However, large scale genomic studies have thus far focused on early events only¹³  and have not addressed the yet elusive origin of long-term cALL relapses.

To decipher the full spectrum of somatic events that lead to treatment failure in cALL and shed light on the spatiotemporal clonal evolution of cALL relapse, we performed both whole-genome sequencing (WGS) and whole-exome sequencing (WES) on a series of primary tumor (PT) samples and multiple relapses (≤4 relapses [R1-R4]) from 19 pre–B-cell cALL patients. Based on variant allele frequency (VAF) analysis, we inferred the global clonal population frequencies as well as evolutionary or ancestral relationships. By dividing our cohort into early (<36 months postdiagnosis, n = 8) or late bone-marrow relapse events (n = 11), we captured significant different clonal dynamics and showed a variable clonal origin of late relapse events. This study is the first to use serial sampling to show the difference of clonal evolution in late vs early pre–B-cell cALL relapses.

All study subjects were French Canadians of European descent. Incident cases were diagnosed in the Division of Hematology-Oncology at the Sainte-Justine Hospital (Montreal, Canada) as part of the Quebec cALL cohort (Table 1).¹⁸  Bone marrow samples were collected at diagnosis and/or relapse. Paired normal DNA specimens were collected from peripheral blood samples without blast cells. The CHU Sainte-Justine Research Ethics Board approved the protocol. Informed consent was obtained from the parents of the patients to participate in this study and for publication of this report and any accompanying images.

Exomes (Nextera Rapid Capture Exome Enrichment) and genomes were sequenced on Illumina HiSequation 2000/2500 systems to reach an average coverage of 200× and 80×, respectively. Resulting reads were aligned to the hg19 (Human Genome version 19) reference genome using Bowtie2 (version 2.2.3)¹⁹  and Casava software, respectively. Single nucleotide variants (SNVs) and small indels were called using Strelka²⁰  (supplemental Figure 1). To uncover any missed subclonal mutations, the variations identified in each tumor event were screened directly in pileup files of previous and/or next events using an adapted version of SNooPer.²¹  Copy number variants (CNVs) were called using the Varscan2 copyCaller option²²  (supplemental Figure 1). Additional structural variations (SVs) were called from WGS using the Casava software (supplemental Figure 1). Variants were prioritized as previously described.²³  See supplemental Data (“Variant annotation and prioritization of cancer driver gene mutations”) for details.

To investigate the temporal evolution of the tumors, we performed serial sampling of 19 pre–B-cell cALL cases who suffered 1 or multiple relapses (R1-R4) and obtained a total of 63 samples (Table 1). The cytogenetic information regarding genome ploidy and large chromosome abnormalities is summarized in Table 2. Using deep WES and/or WGS associated with a strict data reduction strategy (supplemental Methods; supplemental Figure 1), we generated a comprehensive repertoire of somatic mutations, including SNVs and insertions/deletions (indels) (Figure 1; supplemental Table 1), as well as structural variations (SVs), such as CNVs and cryptic inversions or translocations (Figure 1; supplemental Table 2). VAFs of both SNVs and indels were calculated from the distribution of reads either supporting the mutated or the reference allele. The comparison of VAFs obtained for samples for which both WES and WGS were performed indicated that sequence capture had limited influence on read distributions (supplemental Figure 2).

A mean of 145 (range: 5-1852, median = 21) and 891 (range 8-9,985, median = 31) coding somatic mutations were identified per PT and first relapse event, respectively (supplemental Table 3). Sixty-eight percent (in PT) and 52% (in relapse) were considered subclonal mutations (VAFs adjusted for tumor purity <0.30; supplemental Table 3). Of note, local or complete deletions of the short arm of chromosome 9 was one of the most recurrent SVs observed with 16% and 21% of cases presenting a variation at diagnosis and relapse, respectively (Table 2; supplemental Table 2). Nonsynonymous SNVs, including missense and nonsense variants, were the most prevalent class of mutation in all events and for all cases (supplemental Figure 3). On average, 26% and 46% were persistent between PT/R1 and R1/R2, respectively (supplemental Table 3). All patients presented shared mutations between the PT and subsequent events, suggesting a common ancestral origin of lymphoblasts, as previously reported.^13,24  Notably, all deletions of chomosome 9p persisted to relapse (Table 2; supplemental Table 2). Five relapse events (case 325 R3, case 579 R2, case 684 R2, case 772 R2, and case 808 R1) carried >10 times the number of coding somatic mutations identified in the precedent event (supplemental Table 3). These patients were considered hypermutable and were either refractory or suffered multiple relapse events. Importantly, 4 of these (579 R2, 684 R2, 772 R2, and 808 R1) were consecutive to a BMT (Table 1), suggesting an association between their aggressive chemotherapy regimens and the phenotype (P = .004, 2-tailed Fisher's exact test). Of note, we identified no association between this phenotype and irradiation (P = .262). The mutation spectra of the 5 hypermutable cases revealed a substantial increase of transition mutations [A(T)>G(C) and C(G)>T(A)] from the PT (mean percentage of total spectrum = 29%) to the hypermutated event (72%) (supplemental Figure 4), previously identified as a chemotherapy-related mutational signature in other cancer types and reinforced by defective DNA repair mechanisms.^25-27  Except for these cases, no significant shift in mutation number was observed (supplemental Table 3). The hypermutated cases 684, 772, and 808 were the only patients from this cohort to harbor a p.G39E germ line polymorphism (rs1042821) in the DNA repair gene MSH6, a known predisposing gene in Lynch syndrome^28-30  that is also associated with an increased risk of several cancer types.³¹  As for patient 325, who did not undergo transplant, the acquisition at R3 of a somatic homozygous loss of function mutation in the mismatch repair endonuclease PMS2 (p.R134*, VAF = 0.77), which presented a modest positive drift at R4 (VAF = 0.85; Figure 1; supplemental Table 1), could have contributed to the hypermutator phenotype. Case 808 also carried the p.R20Q polymorphism in PMS2 that has been associated with reduced apoptotic function.³² 

Putative driver mutations were grouped into relevant signaling (“Methods”; Figure 1; supplemental Tables 1 and 2) and therapy-related pathways (“Methods”; supplemental Figure 5; supplemental Tables 1 and 2). Six signaling pathways were identified as recurrent targets (Figure 1; supplemental Table 1). Of these, epigenetic regulation and MAP kinase–RAS signaling showed the highest mutation rates with, overall, 84.2% and 78.9% of patients presenting ≥1 altered gene, respectively. Interestingly, these pathways showed numerous putative driver events, yet counter-selected at relapse (13 of 47 mutations vs 3 of 36 mutations in other pathways; Figure 1; supplemental Table 1). Although each pathway showed a significant proportion of relapse-specific mutations (range: 22.2%-75.0%), most alterations identified in MAP kinase–RAS signaling were already present at diagnosis (8.7% relapse-specific mutations; Figure 1; supplemental Table 1). Multisubclonal hits in 5 genes (KDM6A, SETD2, NRAS, NR3C1, and TP53) were identified in 4 early relapse cases (cases 325, 382, 447 and 772) and 1 late relapse patient (case 659) (Figure 1; supplemental Table 1). Of note, mutations in NR3C1 (p.D724E and p.Y641H) and KDM6A (p.R1351*) were selected at the expense of co-occurring mutations (p.P626S in NR3C1 and p.R519* in KDM6A).

Overall, a slightly higher proportion of early relapse cases compared with late ones were identified as mutated for 4 of the 6 pathways: 100% vs 72.7% of mutated cases for the epigenetic regulation pathway, 75.0% vs 54.5% for the hemopoiesis–immune system development pathway, 87.5% vs 72.7% for the MAP kinase–RAS signaling pathway, and 62.5% vs 36.4% for the cell cycle pathway. A strong yet not significant effect was observed for the DNA repair signaling pathway, with 62.5% of early relapse cases mutated compared with only 18.2% of patients with late relapses (P = .07, 2-tailed Fisher's exact test). Of note, 4 of the 5 early relapse cases harbored these DNA repair pathway mutations within the dominant clonal population. Finally, 80% of patients who presented a hypermutator phenotype harbored ≥1 mutation in the DNA repair pathway vs 21% for other cases (P = .04). No association was identified between relapse types and cytogenetic subgroups, such as hyperdiploid (P = 1.0) and normal karyotypes (P = .3).

To evaluate the influence of tumor plasticity on outcome, we divided our patients into groups according to their progression-free survival (early vs late event), their response to treatment (nonrefractory vs refractory), and their survival status. A “dynamic mutation rate” (number of dynamic mutations per Mb per day), considering only mutations showing significant VAF shifts, was calculated for each relapse (supplemental Data [“Dynamic and subclonal mutation rates”]; supplemental Figure 6). Each value was representative of the clonal dynamics of a tumor between 2 events (either PT vs R1 or R[n] vs R[n + 1]) (Figures 1 and 2).

This way, we observed marked reduction in clonal dynamics in late relapses (>36 months postdiagnosis) compared with early relapse events (WES data, P = .0094, Mann-Whitney U test) and for nonrefractory compared with refractory events (P = .0473) (Figures 2 and 3). The same pattern was observed using WGS data for which first relapses (R1) only were available (P = .0079; supplemental Figure 7). Given that R1 samples were obtained at the moment these relapses were diagnosed, before any relapse-associated treatment, this therapy cannot explain the observed difference in clonal dynamics between early and late subgroups. Of note, the absolute number of somatic mutations was not significantly different between early and late relapse groups (P = .9, Mann-Whitney U test), and no difference in the distribution of mutation classes (eg, synonymous, missense, nonsense) was observed (supplemental Figure 3). These results illustrate the plasticity of the earliest events, with rapid clonal switches over short periods of time (eg, case 808 with a mean dynamic mutation rate 1.4e-01 mutations/Mb per day and case 325 with 2.3e-02 mutations/Mb per day) compared with those associated with long-term and quasi-inert relapses (eg, case 34 with 1.3e-04 mutatons/Mb per day). Interestingly, a similar trend of reduced clonal dynamics was observed in WGS data for relapse events of patients who survived compared with those who did not (P = .0159, supplemental Figure 7). We also observed a progressive increase of dynamics at each new relapse event for the same patient (eg, case 325; Figure 1 and 2).

To assess the predictive value of the clonal dynamics in regards to relapse-free survival (first relapse event) while taking variables such as therapy differences into account, we conducted a multivariable analysis (Cox proportional hazard model) adjusted for sex, age at diagnosis, white blood cell (WBC) count, cytogenetic risk, as well as treatment protocol (Dana-Farber Cancer Institute 1991, 1995 2000, and 2005). We showed that a high dynamics (greater than the median dynamic mutation rate value) was independently associated with a significantly lower relapse-free survival (adjusted hazard ratio = 9.92 [95% confidence interval, 1.48-66.52], P = .01). Treatment protocols and clinical characteristics showed no significant influence (P = .4, .4, .7, .2, and .6 for sex, age at diagnosis, clinical risk group, WBC, and treatment protocol, respectively).

Assuming that subclonal diversity plays a central role in therapeutic failure and relapse,^3,6,7,33  we quantified the number of newly acquired subclonal mutations representative of the general expansion of subclones at each event. The estimated “subclonal mutation rate” revealed the exact same pattern as reported above with a significantly reduced subclonal expansion in late events compared with early events (P = .0053 and P = .0317 for WES and WGS data, respectively; Figure 1; supplemental Figures 8 and 9) and in nonrefractory compared with refractory events (P = .0022, WES data; supplemental Figure 8). Again, cases with higher subclonal mutation rates were also the ones who did not survive (P = .0159, WGS data; supplemental Figure 9).

Based on the analysis of somatic VAFs in loss of heterozygosity–free regions of the genome, we inferred the number of subclones and tracked their dynamics over time (supplemental Data [“Spatial and temporal analysis of clonal populations”]).³⁴  We also built clonal trees to determine the ancestral relationships between somatic mutations (supplemental Data [“Spatial and temporal analysis of clonal populations”]).³⁵  Contrasting clonal evolution phenotypes are illustrated by 3 cases, 34, 325, and 62 (Table1; Figure 4). Case 34 presented the longest period before relapse (close to 6 years), but also showed the slowest evolving tumor of our cohort (Figure 3). As for case 325, one of the patients with a hypermutator phenotype, he presented a first and early relapse event less than a year after PT (day 306) followed by multiple relapse events (R1-R4) (Table 1). With a relapse occurring 1025 days after PT, case 62 was the last patient of the early event group (Table 1). He showed the slowest evolving tumor of this group (supplemental Figure 7) and presented an overall number of mutations comparable to patient 34 (supplemental Table 3). Case 34 showed a near perfect clonal equilibrium with the same tumor architecture reoccurring 6 years postdiagnosis (Figure 4A-B). Six clusters of clones were detected. The evolution of the mutation p.W2006* in MLL2/KMT2D (VAF-PT = 0.31, VAF-R1 = 0.55) was the only difference observed between the 2 events. The re-emergence of subclonal mutations at relapse with similar frequencies (eg, KRAS p.K117N, VAF-PT = 0.03, VAF-R1 = 0.05) further highlighted the inertia of this tumor. Clonal evolution in this patient was reminiscent of an escape from long-term dormancy of premalignant neoplastic stem cells that also initiated the primary leukemia. Although this case was the most extreme example of long-term clonal equilibrium and the only one that presented the exact same clonal architecture at diagnosis and relapse, all late relapses presented common mutations with PTs (supplemental Table 3). On the other hand, with >25 SNVs/indels identified in tumor suppressors, oncogenes, and therapy-related genes, case 325 presented a complex clonal architecture composed of 9 clusters of clones spatially evolving over the 5 time points (Figures 1 and 4C), including the hypermutable event R3. Although case 325 did not present the highest evolutionary dynamics within this cohort (Figure 2), these multiple relapses in a short period of time illustrated the step-by-step replacement of the dominant clonal population to the benefit of a fitter clonal progeny under selective pressures, such as chemotherapy. Finally, case 62 showed an intermediate phenotype with ∼50% of the mutations identified at diagnosis that persisted at relapse (supplemental Table 3). Four clusters were identified (Figure 4E-F) and an accumulation of newly acquired mutations was observed at relapse. Although modest in comparison with other patients of the early relapse group, the clonal evolution pattern of patient 62 was still clearly distinguishable from the late relapse cases.

Although the observed effects require validations in larger cohorts to generalize our findings, to our knowledge, this is the first study capturing 2 clonal dynamics governing late and early relapses through the study of somatic allele frequencies. However, some limitations have to be noted: first, although the mean coverage on the targeted region of WES reached 200×, bulk cell sequencing is limiting and excludes an exhaustive characterization of the subclonal architecture. Secondly, RNA sequencing was not performed, which did not allow for the full characterization of cases and a comprehensive analysis of early and late relapse-specific expression profiles.

Five of the 6 pathways identified as frequently targeted in our cALL cohort (epigenetic regulation, hemopoiesis–immune system development, MAP kinase–RAS signaling, DNA repair, and cell cycle) confirmed their relapse-driving potential.¹³  The JAK-STAT/hemopoiesis SH2B3 gene was found mutated in 2 cases (Figure 1; supplemental Table 1). However, we did not observe the recently reported enrichment of mutations in the JAK-STAT pathway in relapsed cALL patients.¹³  Alterations of this pathway are common genomic features of Ph-like acute lymphoblastic leukemia (ALL),³⁶  and although we missed expression data allowing a classification of cases in this particular subgroup, we suspect the observed divergence to rely on a lack of representatives in our cohort. With the rise and fall of numerous mutated clones, epigenetic regulation and MAP kinase–RAS signaling were the 2 most represented (84.2% and 78.9% of mutated patients, respectively) and dynamic signaling pathways. They perfectly illustrated the clonal competition that occurs under external selective pressures, such as chemotherapy. Of particular interest, we also identified a mutational turnover of the glucocorticoid receptor NR3C1 with a succession and a co-occurrence of missense mutations (p.P626S, p.Y641H, and p.D724E) in the dominant clones of case 325. Despite its limitation to a single case in this cohort, the identified reduction of transactivation by glucocorticoids due to a mutation of the residue P626 in CV-1 cells³⁷  suggests a possible involvement in resistance to therapy and thus could explain the selection of mutations. Further suggesting a role of NR3C1 in glucocorticoids resistance and relapse, a very recent study characterized a novel transcriptional program in B-lymphoid cells controlling malignant transformation through the repression of glucose and energy supply and identified the product of NR3C1 as a central effector of this restriction.³⁸ 

We characterized 2 general evolution patterns: a highly dynamic/adaptive clonal pattern specific to early relapses, illustrating the quick emergence of fitter clones and the eradication of other subclones, and a slow-to-quasi-inert evolution pattern associated with late events. By quantifying both the general clonal dynamics and the newly acquired subclonal diversity, we captured the underlying differences of resistance mechanisms between early and late pre–B-cell ALL relapse. In the most plastic tumors, the accumulation of somatic mutations allows rapid subclonal diversification and increases the chance of acquisition of resistant mutations. It leads to a postchemotherapy turnover of predominant populations and to the selection of the fitter clone. Interestingly, we found an overrepresentation of cases from this early relapse group with mutation in DNA repair genes (62.5% vs 18.2% in the late relapse group). Furthermore, most early relapses presented dominant mutations in this pathway. Although this observation requires validations in larger cohorts, a defect in the DNA repair process might explain the fast evolutionary adaptability of early relapse cases. On the other hand, the limited occurrence of hypermutator phenotypes (26% of the cohort) suggests that a strong accumulation of somatic mutations is not a prerequisite mechanism for the acquisition of resistance, but obviously increases the odds of acquisition of advantageous mutations and leads to a higher tumor plasticity, likely giving rise to resistant clones. In regards to these hypermutable patients, 4 of them (2 early and 2 late relapsers) underwent a BMT and the associated aggressive chemotherapy regimens, including high doses of alkylating agents, such as busulfan or cyclophosphamide. In other cancer types, somatic mutations in the DNA repair pathway, such as the ones observed here for hypermutable patients, were suspected to confer resistance to alkylator therapy and to generate hypermutations because of strong exposure to mutagenic agents. ^25-27  Similar to our observations, these cases also presented an excess of transition mutations. Importantly, given that as many early as late relapsers were affected, this particularity is unlikely to have contributed to the captured difference of clonal evolution. As for case 325, who presented a hypermutator phenotype at his R3, which was not preceded by a transplant, the loss of function of the DNA repair gene PMS2 combined with repetitive treatments were likely responsible for this phenotype. This confirmed the recently identified hypermutator potential of the somatic loss of function of the mismatch repair endonuclease PMS213, that, until recently, was mostly associated with Lynch syndrome.³⁹  Although 3 of the 5 hypermutables cases were the only patients of the cohort identified as carriers of germ line mutations in MSH6 and PMS2, these polymorphisms alone can’t be considered responsible for the phenotype. This is further illustrated by the absence of a significant difference in the number of mutations harbored at the PT by these cases compared with the nonhypermutable ones. However, as pointed out previously, an association between the pretransplant intensive treatment regimen and a defect in DNA repair mechanisms, caused either by somatic mutations or, less likely, by germ line variants, could have led to this phenotype. Although further investigation is required to substantiate the role of these variants, in this context, a predisposition for a destabilization of the DNA repair functions should not be disregarded as a contributing factor.

As for the significantly reduced mutational dynamics of late events, it suggests a different mechanism of relapse than early ones, less likely relying on mutation-acquired chemoresistance. Until recently, data that attributed late recurrences to the emergence of clones derived from the original population at the time of diagnosis were generally limited to the use of rearrangements of immunoglobulin genes (IGH/IGK) as specific markers of clonal populations.^40-42  Although little is known regarding cell-of-origin and the process of long-term cALL relapses, recent studies failed to identify shared mutations between diagnosis and late relapse, except for the most ancestral events (eg, fusion BCR-ABL1).⁴³  This led to the suggestion that late relapse cells likely derive from stem cells of a minor clone at diagnosis.^43,44  Here, we showed that all late events harbored somatic alterations that are shared with the PT. Some of these cases presented no variation of allele frequency between primary and relapse leukemias for most of the identified mutations, illustrating the conservation of the same clonal architectures. This provides evidence of a possible re-emergence of the predominant population at the time of diagnosis. The absence of clonal evolution implies that the leukemic cells have not undergone any biological changes during their prolonged dormancy. If so, it is likely that the reappearance of leukemia has been caused by a relaxation of the host's immune surveillance or an environmental change allowing the reactivation/re-expansion of dormant cells spared from the initial cytoreductive therapy⁴¹  rather than a genetic evolution enabling an escape from immune surveillance, as previously suggested.⁴³  This is a clinical asset because the leukemia retains its chemosensitivity. The treatment quickly leads to a new remission, as observed in our cohort. Altogether, these characteristics suggested that late events could originate from a subpopulation of relapse-inducing cells exhibiting properties of long-term dormancy and stemness, as recently isolated from pediatric and adult patients with minimal residual disease (MRD).⁴⁵  These cells also displayed in vivo drug resistance due to limited proliferation when located in their niche and were sensitive when dissociated from this protective environment.⁴⁵ 

Overall, if confirmed in other cohorts, our results have direct clinical implications considering that rapid subclonal expansion and high tumor plasticity are key determinants of rapid cALL evolution and predictive factors of early therapeutic escapes. This is in line with previous findings associating shorter remissions with the early presence of subclonal driver mutations in leukemia samples^7,46  or with MRD levels on day 19 of remission induction.⁴⁷  In this context, appropriate clinical use of novel genomic technologies could be an asset and, as previously suggested,¹³  would allow a systematic postdiagnosis driver mutation–oriented survey of MRD to detect early re-emergence of cancer cells. As for the prevention of long-term relapses that are less likely relying on mutation-based chemoresistance, patients would benefit from therapeutic strategies that release MRD cells from the niche.

The authors thank the patients and their parents for participating in this study. Patient tissue samples were provided by the CHU Sainte-Justine Pediatric Cancer biobank. Whole-exome sequencing was performed at the Integrated Clinical Genomics Centre In Pediatrics, CHU Sainte-Justine. Whole-genome sequencing was performed by Illumina, Inc. (San Diego, CA). Whole-genome genotyping was performed at the McGill University and Genome Quebec Innovation Centre. Computations were made on the supercomputer Briarée from the University of Montreal, managed by Calcul Québec and Compute Canada.

This work was supported by research funds provided by the Canadian Institutes for Health Research. D.S. holds the François-Karl-Viau Research Chair in Pediatric Oncogenomics. The operation of the Calcul Québec supercomputer is funded by the Canada Foundation for Innovation, NanoQuébec, Réseau de médecine génétique appliquée, and the Fonds de Recherche du Québec–Nature et technologies.

Contribution: D.S. is the principal investigator and takes primary responsibility for the paper; J.-F.S. and D.S. contributed to the conception and design of the study; P.C., C.R., and M.O. were involved in sample and library preparation for the whole-exome sequencing; J.-F.S. performed bioinformatics analysis, data integration, and analysis; J.-F.S. interpreted the data and wrote the paper; J.H. and D.S. were involved in critical revision of the manuscript; and all authors approved the final version.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Daniel Sinnett, Division of Hematology-Oncology, CHU Sainte-Justine Research Center, 3175 Côte Sainte-Catherine, Montreal, QC H3T 1C5, Canada; e-mail: daniel.sinnett@umontreal.ca.