Combining Gene Mutation with Transcriptomic Data Improves Outcome Prediction in Myelodysplastic Syndromes

Background and Aim. Myelodysplastic syndromes (MDS) are myeloid neoplasms characterized by peripheral blood cytopenias and risk of progression to acute myeloid leukemia (AML). Disease management is challenged by heterogeneity in clinical courses and survival probability. Recently, the genomic screening integration (by Molecular International Prognostic Scoring System, IPSS-M) into patient’s assessment has resulted into a significant improvement in predicting clinical outcomes compared to the conventional prognostic score (Revised IPSS, IPSS-R). Many of the consequences of genetic and cytogenetic alterations will affect gene expression by means of transcriptional and epigenetic instability and altered microenviromental signaling. The aim of this project conducted by GenoMed4All and Synthema EU consortia is to link genomic information with transcriptomic data for possibly improving the prediction of clinical outcomes in MDS patients.

Patients and Methods. Clinical, cytogenetic, genomic (somatic mutations screening of 31 target genes) and transcriptomic (bulk RNA-seq of CD34⁺ bone marrow cells) data were collected at diagnosis in 389 MDS patients. Transcriptomic and genomic profiles were processed and the former were normalized before Principal Component Analysis (PCA) dimensionality reduction to mine the interdependency of expression-wide perturbation and recurrent genomic alterations. The prognostic impacts of genetic, cytogenetic, transcriptomic, clinical and demographic features were assessed with a penalized Cox’s proportional hazards model [Gerstung M et al, Nat Commun. 2015. 6, 5901] considering the Overall Survival (OS) as primary end point. A 5‐fold cross‐validating (CV) scheme was exploited to control bias in risk estimation. Model accuracy was assessed using Harrell’s concordance index (C-index). An independent validation of the results on 202 patients was planned.

Results. We first processed each data layer assessing data robustness, removed not informative variables and scaled quantitative ones. We considered recurrent genomic and cytogenetic lesions (present in ≥5 patients), platelets, hemoglobin and bone marrow blasts (%), age and sex as covariates. To explore the main patterns of expression changes, PCA was performed to reduce multidimensional correlated expression features (20 PCs was selected, explaining 42% of the total transcriptomic variability). To evaluate the prognostic power of each data layer we grouped all available features into five groups: gene mutations (n=15), cytogenetic alterations (n=7), expression data (n=20), blood counts (n=3) and demographic variables (n=2). Within a 5-fold CV we combined these variables in our integrative model to calculate MDS patients risk. The obtained predictive accuracy (C-index) for OS was 0.83, underlying that transcriptomic data significantly improved the current standard prognostic scoring systems. Accordingly, in our patient population, the C-index of the conventional IPSS-R score and the new IPSS-M were 0.68 and 0.76, respectively. A similar improvement by adding transcriptomic data was observed in prediction of the risk of AML evolution. Moreover, by analyzing the contribution of each feature category to the OS probability (Figure 1), in term of explained variance, the relative impact of transcriptomic is 40%, with the remaining prognostic information distributed among genomic features (somatic gene mutations and cytogenetics lesions, 24%), demographics (20%) and clinical features (15%). An independent validation of these results on 202 patients is currently ongoing.

Figure 2 shows an example of personalized survival prediction using patients from the study population. In two subjects with same clinical phenotype and mutations leading to a similar IPSS-M prognosis, the integrative model captures additional prognostic information and efficiently predicts clinical outcome. Given the complexity of our model, specific technological support is needed to combine data at individual patient level and to translate it into a personalized outcome prediction. To this aim, we created a prototype web portal based on our dataset for user-defined genomic/transcriptomic and clinical features.

Conclusion. In predicting survival of MDS patients, genomic, transcriptomic and diagnostic clinical variables all have utility, with a significant contribution from the transcriptome.