Extended Data Fig. 1: Overview of the DRACO algorithm.

By default, a window of a size equal to 90% of the median read is slid along the transcript, in 5% increments. For each window, a mutation map is generated using only the reads covering the entire window. Bases that are mutated with respect to the reference are assigned a value of 1, while not mutated bases are assigned a value of 0. By using this map, a graph is built, in which each vertex is a base of the transcript, and edges connecting two vertices are weighted proportionally to the number of reads in which the two connected bases have been observed to co-mutate. Starting from the adjacency matrix of the graph, the normalized Laplacian matrix is calculated and used for spectral deconvolution. A null-model is derived by repeating the same procedure after shuffling the mutations in the original mutation map. Analysis of the distance between consecutive eigenvalues (eigengaps) for the experimental data with respect to the null model allows identifying the number of informative eigengaps, corresponding to the number of coexisting RNA conformations (clusters). Once the number of clusters has been defined, fuzzy clustering is performed using a custom graph cut approach, that enables the weighting of vertices in accordance with their affinity to each cluster. This analysis is repeated across the whole transcript. Consecutive windows showing a compatible number of clusters are merged. Then, reads are re-assigned to the respective cluster, allowing the deconvolution of the cluster reactivity profiles and relative abundances.