Clustering Real cDNA Data

A test of clustering real cDNA data was performed. The input contained 2329 cDNAs, originating from 18 genes. The true clustering, obtained by hybridization with long, unique sequences, is given in table 11.2.

 

Table 11.2: True clusters in a cDNA fingerprinting experiment. Results are shown in figure 11.12.

Cluster	Cluster size	Gene name
T18	709	Ef1 alpha
T17	285	clone 190B1
T16	284	Cytochr c oxi
T15	213	tubulin beta
T14	187	40SRibo protS6
T13	146	40SRibo protS3
T12	108	40SRibo protS4
T11	91	GAPDH
T10	86	60SRibo protL4
T9	67	Ef1 beta
T8	43	Human calmodulin
T7	39	heat shock cogKD71
T6	32	heat shock cogKD90
T5	14	Human TNF recep
T4	12	Human AEBP1
T3	10	clone 244D14
T2	2	clone 241F17
T1	1	Human anion ch

 

The high variability in abundance of genes can be easily seen. The results of the test are summarized in figure 11.12. In 14 out of the 17 clusters generated by the algorithm, over 92% of the entities belong to the same gene (true cluster). Those clusters are called almost pure. As the correct clusters are not known (in real experiments), and the main goal of cDNA clustering is to avoid repeated sequencing of cDNAs originating from the same gene, the following strategy can be used: From each cluster, up to 10 cDNAs are randomly chosen to be sequenced. If, for example, 9 out of 10 give the same sequence, the cluster is with high certainty almost pure, and no more sequencing of its members is needed. Otherwise, all the members of the cluster are sequenced. This strategy may save about 75% of the sequencing cost.

  

Figure: Clustering results on real cDNA data (true clusters are listed in table 11.2). A: The binarized similarity matrix. A block point appears at position (i,j) iff $S_{ij}\geq 110$. B: Reordering of A according to the true clustering. cDNAs from the same t rue cluster appear consecutively, and the black lines are the borders between the different clusters. C: Reordering of A according to the clustering produced by the HCS algorithm. Clusters appear in the order of detection. D: Comparison of the algorithmic solution and the true solution. Rows and columns are ordered as in B. Position (i,j) is black iff the algorithm put i and j in the same cluster.