Machine learning-driven Interferon Signaling Gene Expression Score predicts Aicardi Goutières Syndrome
Adeline Vanderver1, Russell D'Aiello1, Asako Takanohashi1, Francesco Gavazzi1, Justine Shults1, Sarah Woidill1, Laura Adang1
1Children'S Hospital of Philadelphia
Objective:
To define the biomarkers in Aicardi Goutieres Syndrome (AGS).
Background:
Interferon Signaling Genes (ISG) scores have evolved as an important surrogate metric of interferon signaling, used in clinical trials and to measure disease activity in genetic interferonopathies. Recently, multiplexed approaches for gene expression have increased the need for standardization of ISG calculation from interferon response gene expression panels.
Design/Methods:
We measure interferon response gene expression using NanoString nCounter® Analysis System in 997 individual samples, of which 334 were from AGS affected subjects, and randomized samples into training and validation data sets. The training data set was used to develop several machine learning models. These models were assessed on the validation dataset along with the original 6 gene AGS score.
Results:
The random forest was the best performing model, achieving accuracy of 91.1% and area under the curve (AUC) of .962. This performance was better than the existing standard, 6 gene ISG AGS 6 gene score, which has an accuracy of 73.3% and AUC of .896. All classification methods evaluated perform worse in the AGS samples from patients with an RNASEH2 complex mutation. In addition, the longitudinal stability was lower for RNASEH2 complex mutation samples than for non- RNASEH2 related samples.
Conclusions:
This study validates a machine learning-driven ISG score as candidate diagnostic biomarker of AGS. Further study will be necessary to assess whether this score is useful in other genetic interferonpathies and whether there is any correlation with disease severity or therapeutic response.