The Clinical Utility of Microarray-Based Gene Expression Profiling in the Diagnosis and Sub-Classification of Leukemia: Final Report on 3252 Cases from the International MILE Study Group

During the years of 2005 to 2008, the MILE (Microarray Innovations in LEukemia) study research program was performed in 11 laboratories across three continents: 7 from the European Leukemia Network (ELN, WP13), 3 from the US and 1 in Singapore. The first stage was designed as biomarker discovery phase to generate whole-genome gene expression profiles (GEP) from recognized categories of clinically relevant leukemias and myelodysplastic syndromes (MDS). These were C1: mature B-ALL with t(8;14), C2: pro-B-ALL with t(11q23)/MLL, C3: c-ALL/pre-B-ALL with t(9;22), C4: T-ALL, C5: ALL with t(12;21), C6: ALL with t(1;19), C7: ALL with hyperdiploid karyotype, C8: c-ALL/pre-B-ALL without specific genetic abnormalities, C9: AML with t(8;21), C10: AML with t(15;17), C11: AML with inv(16)/t(16;16), C12: AML with t(11q23)/MLL, C13: AML with normal karyotype or other abnormalities, C14: AML with complex aberrant karyotype, C15: CLL, C16: CML, C17: MDS, and C18: non-leukemic and healthy bone marrow samples as controls and were compared to conventional diagnostic assays (“Gold Standard”). Data from the completed MILE Stage I included 2143 retrospectively collected adult and pediatric samples tested with HG-U133 Plus 2.0 microarrays (Affymetrix). In total only 47 analyses (2.1%) failed technical quality criteria. Cross-validation accuracy (average of three 30-fold cross-validations) of the final 2096 MILE Stage I samples was 92.1% concordant with the center-specific “Gold Standard” diagnosis (average call rate 99.4%). In nine classes the sensitivity was ≥94.3%: C2, C3, C4, C5, C9, C10, C11, C15, and C16. Lower sensitivities were observed for C7, C8, C14, and C17; which can largely be explained by the biological heterogeneity and non-standardized “Gold Standard” definitions for these entities. Yet, it is notable that all these classes showed specificities above 98.1%. In order to assess the clinical utility of microarray-based diagnostics a prospective Stage II was subsequently performed using a customized microarray representing 1480 probe sets. Overall, 1156 high quality GEP have been generated in MILE Stage II and represent an independent and blinded test set for the algorithms developed. A focused classification scheme aimed at accurately addressing only acute leukemias resulted in a 95.5% median sensitivity and a 99.5% median specificity for the 14 classes included in the classifier (C1 – C14, n=696). Lower accuracies were observed for the interface of C7–C8 in ALL, as well as C12 and C14 in AML. Interestingly, during the process of discrepant results analyses, it was observed that for 7.5% (n=52) of acute leukemias microarray results were correctly diagnosing samples as compared to the initial “Gold Standard” diagnoses entered into the study database, either because of erroneous entries into case report forms (24%) or subsequent re-testing of left-over material following the suggested diagnosis from the microarray (76%). In addition, predicted accuracies for CLL, CML and MDS in Stage II were 99.2%, 95.2%, and 81.5%, respectively. In conclusion, the MILE research study confirms in a final cohort of 3252 patients that microarrays accurately classify acute and chronic leukemia samples into known diagnostic and prognostic sub-categories. This final report underlines that the standardized method of gene expression profiling with low technical failure rate and simplified standard operating procedures may improve current “Gold Standards” as an adjunct to conventional diagnostic algorithms and potentially offers a reliable diagnostic/prognostic tool for many patients who don’t have access to a state-of-the-art “Gold Standard” workup. Our gene expression database, intended to be submitted to the public domain, will further contribute to research that aims to elucidate the molecular understanding of leukemias.