Session: 637. Myelodysplastic Syndromes—Clinical Studies: Personalized Clinical-Decision Tools and treatment of lower risk MDS
Methods: Peripheral blood and bone marrow (BM) biopsy specimens from cytopenic patients, along with clinical history, CBC, and other results including karyotyping, FISH and pathology reports from local pathologists were reviewed by central pathologists. The updated 2016 WHO classifications were used to diagnose MDS. Targeted exon sequencing of 96 genes was performed using BM specimens. A subset of 648 individuals that were classified as MDS (n=212) or Other (n=436, including 90 CCUS and 89 individuals with other cancers) by pathology assessments were selected. A mean coverage of 1,317X was achieved and variants had a minimum variant allele frequency (VAF) of 2% (except FLT3). Variants for 596 subjects were manually reviewed to retain likely disease-causing variants to build a binary classifier (MDS vs. Other) using the maximum VAF per gene as input (Figure 1). Subjects diagnosed with MDS or Other by both central and local pathology were used for training, validation, and testing, and were considered “gold standard” (GS) cases (n=546). These subjects were split into 4 random groups with equal proportions of MDS cases. 75% of the GS cases were used to train and validate lasso-regularized logistic regression models using 3-fold cross validation. ROC curve analysis was carried out using the remaining 25% of GS cases (Test Set 1) on the best model to identify an optimal probability cut off point for classifying subjects as MDS. Model performance was then tested on 50 subjects for which the central and local pathology diagnosis disagreed (Test Set 2), as well as on 52 additional subjects irrespective of agreement (Test Set 3).
Results: The best performing logistic regression model retained 7 genes as most informative in a discriminating diagnosis of MDS from Other based on their VAFs, in order of impact: TP53, SF3B1, U2AF1, ASXL1, TET2, STAG2, and SRSF2. We used this model to assign probabilities for each of the subjects in Test Set 1 and to estimate the performance using ROC analysis (Figure 1), resulting in a high area under the curve (AUC) of 0.89. We chose a probability cut-off of ≥0.17, being associated with a high percentage of correct classification of MDS with a sensitivity and specificity of 0.90 and 0.81, respectively. Among the cohort of 50 subjects with a discordant local and central pathology diagnosis (Test Set 2), the classifier accurately reassigned 37 subjects (accuracy = 74%) from the local to the central pathology. The blinded tertiary pathology reviewer agreed with central in all Test Set 2 cases. This included 24/34 MDS cases that had been labeled as Other by local pathology (positive predictive value [PPV]=0.89). 3/16 final pathology-classified Other cases were mis-classified as MDS by the local pathologist (negative predictive value [NPV] = 0.57). Next, we assessed the ability of the model to predict MDS vs. Other for 52 additional independent subjects using the third pathologist’s diagnosis to break any ties (Test Set 3). The classifier correctly predicted 15/21 MDS cases (PPV=0.83) and misclassified 6/31 Others as MDS (NPV=0.82). The overall accuracy was 83%.
Conclusions: We identified that VAFs for 7 genes can correctly re-classify subjects as either MDS or Other in 74% of cases that were misclassified between local and central pathology review. Further assessment on an independent cohort showed an accuracy of 83% of the model. Taken together, these data suggest that complementing pathology reviews with targeted sequencing of 7 genes could improve MDS diagnosis.
Disclosures: Lindsley: Bluebird Bio: Consultancy; MedImmune: Research Funding; Jazz Pharmaceuticals: Consultancy, Research Funding; Takeda Pharmaceuticals: Consultancy. Bejar: Aptose Biosciences: Current Employment; AbbVie/Genentech: Honoraria; Astex/Otsuka: Honoraria; Takeda: Honoraria, Research Funding; Celgene/BMS: Honoraria, Research Funding; Daiichi-Sankyo: Honoraria; Forty-Seven/Gilead: Honoraria; Genoptix/NeoGenomics: Honoraria. DeZern: MEI: Consultancy; Astex: Research Funding; Abbvie: Consultancy; Celgene: Consultancy, Honoraria. Foran: H3Biosciences: Research Funding; Aptose: Research Funding; Kura Oncology: Research Funding; Trillium: Research Funding; Takeda: Research Funding; Revolution Medicine: Consultancy; Xencor: Research Funding; Agios: Honoraria, Research Funding; Aprea: Research Funding; Actinium: Research Funding; Servier: Membership on an entity's Board of Directors or advisory committees; Novartis: Membership on an entity's Board of Directors or advisory committees; Abbvie: Research Funding; BMS: Membership on an entity's Board of Directors or advisory committees; Pfizer: Membership on an entity's Board of Directors or advisory committees; Boehringer Ingelheim: Research Funding. Gore: Abbvie: Consultancy, Honoraria, Research Funding. Komrokji: Acceleron: Honoraria; Incyte: Honoraria; Abbvie: Honoraria; Agios: Speakers Bureau; BMS: Honoraria, Speakers Bureau; Jazz: Honoraria, Speakers Bureau; Geron: Honoraria; Novartis: Honoraria. Maciejewski: Alexion, BMS: Speakers Bureau; Novartis, Roche: Consultancy, Honoraria. Padron: Novartis: Honoraria; BMS: Research Funding; Incyte: Research Funding; Kura: Research Funding. Starczynowski: Captor Therapeutics: Consultancy; Tolero Therapeutics: Research Funding; Kurome Therapeutics: Consultancy, Current equity holder in private company, Research Funding. Sekeres: BMS: Consultancy; Takeda/Millenium: Consultancy; Pfizer: Consultancy.
See more of: Oral and Poster Abstracts