SARNAclust

This repository contains code for both RNApeakFold and SARNAclust. RNApeakFold is an RNA secondary structure prediction tool tailored to CLIP peaks. SARNAclust is a clustering algorithm for CLIP peaks that typically uses the secondary structure prediction from RNAfold.

SARNAclust

SARNAclust requires the following python libraries:

• networkx
• EDeN (graph kernel)
• sklearn
• matplotlib
• forgi (bulge graph)
• Biotools

It also requires Locarna 1.8.12 to be installed. Moreover, it imports dpcluster, an in-house implementation [1] of Density Clustering also available here. There are also a few fixed parameters that the user can change in the code:

• MINL = 10 - min length of a peak for it to be considered
• MAXL = 320 - max length of a peak for it to be considered
• MAXP = 1400 - maximun number of peaks to be clustered in the same iteration

SARNAclust inputs an extended fasta file (-file FileName, default GAGAandRandom.efa)) (i.e., the output from RNAfold) where there has to be a set of sequence/structures each composed of three lines: fasta comment, sequence and secondary structure. It also requires a set of parameters:

• -r - radius (see EDeN library, default 2)
• -d - distance (see EDeN library, default 3)
• -cluster_option - depends on the method selected, but the format is "(method,param1,param2, ... ,paramN)" (default DBSCAN '(1,0.6,10)'):
  • 0 - MeanShift (sklearn)
  • 1 - DBSCAN (sklearn)
    • similarity - maximum distance between 2 sequences in the same cluster (0:1)
    • minsamples - minimum number of samples for a cluster to be considered
  • 2 - Affinity Propagation (sklearn)
  • 3 - Prints similarity matrix
  • 4 - KMeans (sklearn)
  • 5 - Spectral Clustering (sklearn)
  • 6 - Density Clustering (dclust in-house implementation of [1], read paper to understand parameters)
    • Flag for adding halo as part of the cluster (0,1)
    • lowerbound percentile
    • dc
• -graph_option (default 9):
  • 0 - Complete graph with no structural information
  • 1 - graphProt minus directional
  • 2 - Bulge graph plus Complete Graph
  • 3 - Bulge graph
  • 4 - Bulge graph plus sequence in HP loops
  • 5 - Bulge graph plus sequence in IL loops
  • 6 - Bulge graph plus sequence in external loops
  • 7 - Bulge graph plus sequence in double stranded
  • 8 - Bulge graph plus sequence in HP and IL
  • 9 - Bulge graph plus sequence in single stranded
  • 10 - Bulge graph plus sequence everywhere
  • 11 - Same as 2 but no sequence in double stranded
• -thClus - distance threshold for merging clusters (0:1, default 0.6)
• -iterations - number of iterations (default 5)
• -verbose (default 0)
• -alpha - alphabet (0- A,C,G,U; 1- G,R, default 0)

Since python can only handle a certain number of graphs loaded in memory (around 800 depending on the machine running the code), SARNAclust performs a number of iterations selecting each time MAXP random sequences and then merging clusters among iterations if the average distance between sequences in 2 clusters is lower than thClus. Each sequence/structure is decomposed into substructures that are automatically detected and the clustering is performed on those, unless graph option 0 is selected. To disallow structure decomposition just change readFile(fn,gopt) in the main function for readFile(fn,0). This should allow the user to up MAXP to around 2000.

At the end, SARNAclust returns a set of clusters along with their consensus sequence and structure. You can try with our example file GAGAandRandom.efa and DBSCAN (default options), for instance, with the following command line:

python SARNAclust.py

References

• [1] - Rodriguez A, Laio A. 2014. Machine learning. Clustering by fast search and find of density peaks. Science 344: 1492–1496.