Reliable prediction of transcription factor binding sites by phylogenetic verification

Supporting Information

Files in this Data Supplement:

Fig. 4. Ribosomal protein genes are often observed to show very similar expression patterns in various microarray studies. (A) Heat maps of clustered ribosomal protein genes are shown for an in vitro astrocyte and neural cells differentiation study (1). (B) Multiple liver and heart samples in studying circadian genes (2). (C) In vivo study of mouse preimplantation embryos (3).

1. Bachoo, R. M., Kim, R. S., Ligon, K. L., Maher, E. A., Brennan, C., Billings, N., Chan, S., Li, C., Rowitch, D. H., Wong, W. H. & DePinho, R. A. (2004) Proc. Natl. Acad. Sci. USA101, 8384–8389.

2. Storch, K. F., Lipan, O., Leykin, I., Viswanathan, N., Davis, F. C., Wong, W. H. & Weitz, C. J. (2002) Nature417, 78–83.

3. Wang, Q. T., Piotrowska, K., Ciemerych, M. A., Milenkovic, L., Scott, M. P., Davis, R. W. & Zernicka-Goetz, M. (2004) Dev. Cell6, 133–144.

Fig. 5. Evolution tree and the top motif found in each species. The species names are on top of each motif logo. After the species names, there are two numbers. The first number is the number of upstream sequences that contain the instances of this motif. The second number is the number of totally available upstream sequences in our study.

Table 5. Comparison of yeast ChIP–chip data

All transcription factors with known sites in ref. 1 and with experimentally verified sites in TRANSFAC (we can find the sites in TRANSFAC or the paper cited by TRANSFAC) will be used. But if the binding sites provided by TRANSFAC are too long or contradict with each other, we will not use them. By this criterion, we obtained above 53 transcription factors. Note that the numbers in the second column are the target gene numbers of that transcription factor in S. cerevisiae. Some genes may have no sequence in other species.

The criterion for matching with TRANSFAC motifs is that there should be at most one mismatch in the orange colored regions. Those orange regions must be continuous except the ambiguous positions, such as SIP4 (colored in orange according to ref. 1).

We know SWI4 works with SWI6, and CBF1 works with MET31 and MET32. So if we find any sites matching either of them for the corresponding transcription factors, we will claim matching.

We output 10 motifs from COMPAREPROSPECTOR, and if one of the top two outputs matches the known motifs, we claim that it made the correct predictions. For PHYLOCON, there is no order among motifs; thus, we manually check all the output motifs (on average, there are 60 output motifs for every transcription factor. Although there are some motifs appearing a few times in the output, we still give PHYLOCON a great advantage here, which may result in its higher specificity than COMPAREPROSPECTOR) ,and if one of its output matches the known motifs, we claim that it made the correct predictions. For our method, we output all motifs with P < 1 ´ 10^–19 and use the motif with the smallest P value. (We will say CSC made correct prediction for SUT1 in the above table because the known SUT1 motif is ranked as the second one and the first motif is unknown. We will not say CSC made correct predictions for MET4 and GCR1 either since the top one motifs have P > 1 ´ 10^–19 in both cases.)

The NULL in the fourth and fifth columns means the outputs only contain AAAAA or TTTTT or TATATA or CACACA-like segments. In AAAA and TTTTT case, the outputs must have >90% of A or T. In CACACA and TATATA case, all of them must be CA or TA except the two boundary nucleotides. For outputs as NULL, we treat them as no prediction. NULL in the last column means there is no motif with P < 1 ´ 10^–19.

Ambiguity codes: D = A or T; B= C, G, or T; S = C or G; W = A or T; R = A or G; Y = C or T; K = G or T; M = A or C; V = A, C, or G; N = A, C, G, or T.

In the last row of the table, 24 + 1 means there are 24 motifs output satisfying the criteria of matching defined above and there is another motif (SUM1) output that does not satisfy the criteria of matching defined above. In the last row, 29 + 1 has a similar meaning.

The true positive rate of CSC, COMPAREPROSPECTOR, PHYLOCON is 30/53 = 56.6%, 25/53 = 47.2%, and 25/53 = 47.2%, respectively. The false positive rate of CSC, COMPAREPROSPECTOR, and PHYLOCON is 5/35 = 14.3%, 21/46 = 45.7%, and 10/35 = 28.6%, respectively. We believe PHYLOCON has similar specificity as COMPAREPROSPECTOR, and the false positive rate of PHYLOCON shown here are greatly biased to match the known motifs with all output for one transcription factor.

In ref. 1, the authors used six different methods and did not find the correct predictions for 15 transcription factors in the above table from the output of any of the six methods they used. The 15 transcription factors are GCR1, HAC1, HAP5, MAC1, MET31, MET32, MOT3, MSN4, PDR3, RLM1, RPH1, ROX1, SKO1, SMP1, and SWI5. Without human intervention, CSC can identify the known motifs for MAC1, MET31, MET32, and SWI5 on the same data sets.

MEME did not output the known motifs for the following transcription factors: ACE2, CIN5, GAL4, GLN3, HAC1, HAP5, HSF1, MOT3, RCS1, PDR3, RLM1, ROX1, RPH1, RPN4, SIP4, SKO1, and SMP1. The known motifs of GCR1and MET4 are included in MEME output, and CSC ranks them as the first motif in the respective putative motif sets. Their P values are 6.893677 ´ 10^–10 and 4.718924 ´ 10^–12, respectively.

*Ref. 1 points out that the consensus site of CIN5 is TTACRTAA. We use TTACRTAA here instead of TRANSFAC one TTACTAA.

1. Harbison, C. T., Gordon, D. B., Lee, T. I., Rinaldi, N. J., Macisaac, K. D., Danford, T. W., Hannett, N. M., Tagne, J. B., Reynolds, D. B., Yoo, J., et al. (2004) Nature431, 99–104.

Table 6. The most significant motif in every species reported by CSC