Description

Inside the nucleus, DNA is wrapped into a complex molecular structure called chromatin, whose fundamental unit is approximately 150 bp of DNA organized around the eight-histone protein complex known as the nucleosome. These tracks contains predicted nucleosome occupancy scores produced by three different computational models. Each model is a support vector machine classifier trained using microarray data from an MNase cleavage assay. Each SVM is trained to discriminate between 50 bp DNA sequences that show the strongest and weakest signals in the MNase assay. Although each model can predict regions of high and low nucleosome occupancy, one model (MEC) excels at recognizing regions of low nucleosome occupancy, whereas the other two (A375 and Dennis) are better at recognizing regions of high nucleosome occupancy.

The three models are as follows:

• A375 - This model was trained using data from the A375 cell line from Ozsolak et al. (2007). This cell line was prepared with weak MNase digestion. The A375 model excels at recognizing regions of strong protection from MNase cleavage; i.e., positions that are frequently occupied by a nucleosome.
• Dennis - This model was trained using data from MDA-kb2 cell line data from Dennis et al. (2007). This cell line was prepared with weak MNase digestion. Hence, like the A375 model, the Dennis model excels at recognizing regions that are frequently occupied by a nucleosome.
• MEC - This model was trained using data from the MEC cell line from Ozsolak et al. (2007). This cell line was prepared with strong MNase digestion. This model excels at recognizing regions of high accessibility to MNase cleavage; i.e., positions that are frequently nucleosome-free.

Display Conventions and Configuration

The output of the SVM is a unitless discriminant score. In the browser, the score of a 50-mer is assigned to its 26th base. Canonically, a score of 0 indicates an uncertain assignment; a score of 1.0 corresponds to a confident prediction for being in the positive class (i.e., a position of frequent nucleosome occupancy), and a score of -1.0 corresponds to a confident prediction for being in the negative class.

Methods

For a given microarray experiment, we identify the 1000 50 bp probes with the highest log intensity ratios. These comprise our positive training samples. In a similar fashion, we generate negative training samples with the lowest log intensity ratios. Each 50-mer in the training set is converted into a 2772-element vector of k-mer frequencies for k=1 up to 6 (collapsing reverse complements). A linear SVM is then trained to discriminate between the two classes. The SVM regularization parameter is selected by evaluating the entire regularization path on a held-out portion of the training data set. After training, each 50-mer in the human genome is converted to the 2772-element representation and scored using the trained SVM.

Detailed methods are given in Gupta et al. (2008), and supplementary data is available here.

Credits

This track was produced at the University of Washington by Shobhit Gupta and William Stafford Noble ([email protected]).

References

Ozsolak F, Song JS, Liu XS, Fisher DE. High-throughput mapping of the chromatin structure of human promoters. Nat Biotechnol. 2007 Feb;25(2):244-8.

Dennis JH, Fan HY, Reynolds SM, Yuan G, Meldrim JC, Richter DJ, Peterson DG, Rando OJ, Noble WS, Kingston RE. Independent and complementary methods for large-scale structural analysis of mammalian chromatin. Genome Res. 2007 Jun;17(6):928-39.

Gupta S, Dennis J, Thurman RE, Kingston R, Stamatoyannopoulos JA, Noble WS. Predicting human nucleosome occupancy from primary sequence. PLoS Comput Biol. 2008 Aug 22;4(8):e1000134.