ESPE Abstracts (2016) 86 FC8.1

ESPE2016 Free Communications Growth: Clinical (6 abstracts)

Transcriptomics and Machine Learning Methods Accurately Predict Diagnosis and Severity of Childhood Growth Hormone Deficiency

Philip Murray a , Adam Stevens a , Ekaterina Koledova b , Pierre Chatelain c & Peter Clayton a

aInstitute of Human Development, Faculty of Medical and Human Sciences, University of Manchester and Manchester Academic Health Science Centre, Royal Manchester Children’s Hospital, Central Manchester, Manchester, UK; bGlobal Medical, Safety & CMO, Merck, Darmstadt, Germany; cDepartment Pediatre, Hôpital Mère-Enfant, Université Claude Bernard, Lyon, France

Background: The diagnosis of Growth Hormone Deficiency (GHD) involves the use of GH stimulation tests that require day case admission, multiple blood sampling and are associated with significant adverse effects.

Aim: To assess the utility of gene expression (GE) profiling and candidate SNP analysis for the diagnosis of and classification of GHD.

Method: Pre-pubertal treatment-naïve children with GHD (n=98) were enrolled from the PREDICT study and controls (n=26) acquired from online datasets. Whole blood gene expression (GE), determined with Affymetrix HU133v2.0 microarrays, was correlated with peak GH using rank regression and a Random Forest algorithm tested for prediction of the presence of GHD and in classification into severe (peak GH<4 μg/L) and non-severe (peak≥4 μg/L). For GHD severity classification data on age, gender, baseline IGF-I and IGFBP-3 levels was added to the Random Forest model along with SNP genotype for 97 growth-related candidate genes. Performance was assessed using Area under the Receiver Operating Characteristic Curve (AUC-ROC). A biological network of the GE related to peak GH levels was generated and cluster hierarchy assessed.

Results: Rank regression identified 347 probesets representing 271 genes where expression correlated with peak GH concentrations: (R+0.28, P<0.01). These 347 probesets gave an AUC of 0.98 (sensitivity 100%, specificity 96%) for predicting GHD status versus controls, while using only the top 10 probesets ranked by network centrality gave an AUC of 0.94. Random Forest analysis was also able to accurately predict GHD severity with an AUC of 0.93 using transcriptomic data, not improved with addition of demographic, biochemical or SNP genotype data.

Conclusion: GE profiling differentiates normal subjects from those with GHD and accurately predicts GHD severity. It may therefore be a useful tool to aid in the diagnosis of GHD, potentially replacing two GH stimulation tests with a single blood sample.

Article tools

My recent searches

No recent searches.