Machine Learning to Identify Novel Serum Biomarkers in Lymphangioleiomyomatosis

Lymphangioleiomyomatosis (LAM) is a rare progressive cystic lung disease. While VEGF-D is a useful prognostic biomarker, up to 50% of patients have normal levels, indicating a need for other biomarkers. The aim of this study was to use machine learning to identify novel composite biomarkers to predict progression. An unbiased aptamer-based screen of >1000 proteins over a 7-log range was performed on serum samples from MILES trial participants. Placebo-recipients were classified as rapid or slow decliners based on above or below-mean FEV1 decline at 6-months. Baseline aptamer data was explored to identify potential prognostic biomarkers of decline. Initial group comparisons were made using Wilcoxon signed-rank test with significant features (p<0.05) undergoing further evaluation in a 5-fold cross-validation scheme using a) logistic regression, and b) random forests for feature ranking. Finally, sequential feature selection of the highest performing features in addition to VEGF-D were tested using cross-validation and logistic regression. Data for 32 patients in the placebo arm of the trial was available. Analysis identified and ranked the most significantly increased proteins. VEGF-D levels alone had a low precision (0.25) and sensitivity (0.4) in FEV1 decline at 6 months. When the most significantly identified protein (EFNA4 -ephrin A4) was included with VEGF-D, the precision increased to 0.8 and sensitivity to 0.73. By combining multiple features (VEGF-D+EFNA4 +IGHD+GDNF+TKT), the precision reached 0.95 and sensitivity 0.93. We identified previously unexplored protein biomarkers which may have relevance as composite biomarkers in LAM. Further mechanistic and longitudinal validation is required.