Fig. 3

Retinitis pigmentosa mechanistic model from the inside. A Overview of the Retinitis Pigmentosa machine learning-driven approach to drug repurposing. A Multi-output random Forest (MORF) regressor was trained, with gene expression from the GTEx dataset, and the circuit’s activity values, to assess the predictive power of 711 KDTs over the activity of the 226 circuits that configure the RP mechanistic map. Feature selection of highly predictive KDTs for a given circuit was calculated based on the R2 score and up to the top 5% KDTs ranked by the mean absolute Shapley value. B Machine learning model (MORF) stability vs R2 score circuit metrics. C Boxplots representing the distribution of the absolute model relevances (Y-axis) across the top 30 ranked KDT (X-axis). The rank is obtained by means of the L1 norm and the values are colored by their sign (positive = red; negative = blue)