Nucleic Acids Research

DISCUSSION

One of the main challenges in molecular biology is to understand what determines the selection of DNA target sites by a regulatory protein. Examination of the solved protein-DNA complexes shows a stereochemical complementarity between the elements involved in forming the complex. These data and results of binding experiments of sequence variants indicate that the compatibility between the interacting groups in the protein and DNA plays a major role in dictating specific recognition. In its simplest view, a protein binding site will favor DNA target sites in which there is a one-to-one compatibility between the amino acids used for interaction and the DNA bases that they contact. Understanding how this compatibility is determined is the key to discerning specific recognition, and the ability to quantify amino acid-base compatibility is the basis for any predictive algorithm that will identify favorable binding sites. In the present study we took advantage of the growing number of solved protein-DNA co-crystals and extracted a quantitative measure for amino acid-base compatibility based on the observed frequencies of pairs of amino acid-base that are in contact in the solved complexes.

Table 3. Calculated scores and ranking for selected fingers by DNA triplets [based on data from Jamieson et al. (13) (A) and Rebar and Pabo (12) (B)]

	DNA triplet			Amino acid triplet			Score	Percentile	Percentile
	1	2	3	-1	3	6		(rank/8000)	(rank/512000)
A	G	A	C	E	N	R	5.22	1	1
	G	C	A	Q	E	R	4.45	1	1
	G	C	G	R	E	R	6.03	1	1
				R	D	R	6.49	1	1
				G	E	R	-0.64	9	14
	G	G	G	R	H	R	7.04	1	1
	G	G	A	K	H	R	4.22	1	1
	G	T	G	E	A	R	-0.53	13	14
	G	T	T	T	A	R	3.34	2	1
B	G	A	C	D	N	R	5.68	1	1
	G	C	A	R	D	R	4.09	1	1
				Q	S	R	3.22	1	1

Amino acid triplets listed in the second column are those selected experimentally for given DNA triplets (listed in the first column). Scores for each combination of the listed amino acid-base triplets were calculated based on the matrix in Table 2. Ranking of the experimentally selected amino acid triplets by the scores out of all possible 8000 combinations of amino acid triplets is given in column 4, while ranking out of all possible 512 000 combinations of amino acid-base triplets is given in column 5.

Our approach was to count the number of all different pairs of amino acid-base in the solved complexes and to extract a measure by computing the log odds of the observed frequencies and those expected if these interactions were random. The log odds matrix for base-amino acid interactions quantifies the preferences of the pairwise interactions and can be used to evaluate compatibility between protein and DNA binding sites. Log odds matrices were generated in a variety of computational studies that attempted to derive knowledge-based parameters from a data set of sequences or structures, given a particular biological question. They have been employed in the derivation of scoring matrices for amino acid-amino acid substitutions, such as the Dayhoff matrix, used for protein sequence alignment (reviewed in 21). Information content calculations in aligned sequences and generation of a specificity matrix to define a particular functional site are based on the same concept (reviewed in 22). Scores that describe the compatibility between the different amino acids and defined structural environments were extracted similarly by Eisenberg and colleagues from a database of solved protein structures and have been used to evaluate sequence-structure fit (23,24). Also, as commented by Jones and Thornton (10), many of the derived amino acid-amino acid contact potential matrices can be considered as log odds matrices, as they encode observed distributions of residue pairs in real proteins and do not seek to measure energy. Likewise, in the present study there is no attempt to ascribe an energetic meaning to the extracted values, but merely to consider them as quantitative scores that reflect compatibility between amino acids and bases.

In all examples above, the likelihood ratios compare the probability of an event occurring under two alternative hypotheses. In the case of amino acid-base interaction we compare the frequencies of pairs that appear in solved structures with the expected frequencies if these interactions were random. The expected frequency of a pair is calculated under the assumption that there is no preference for any amino acid to interact with any base, by multiplying the expected frequencies of bases and amino acids. These latter frequencies may be defined in different ways. One approach is to use the total frequencies of amino acids and bases in the data. However, by this approach interactions that involve bases and amino acids that are frequent in protein-DNA complexes become artificially weaker. Also, by using the frequencies from the data set itself the fact that some bases and amino acids participate in protein-DNA complexes at frequencies well above average is masked. To overcome these drawbacks general frequencies of amino acids and bases were used. As shown in Table 2, the quantitative measures obtained succeed in reflecting reasonably well the pair preferences. For example, the highest measures were obtained for Arg-G, Lys-G and Asn-A. The preference for these pairs in protein-DNA recognition has been shown in many structures and their possible role in protein-DNA recognition has been suggested (7,8,20). In their scoring scheme Suzuki and Yagi (8) assigned to all of them an equal score (the highest score in their scheme). The current measures, however, indicate a hierarchy of these pairs: Arg-G > Lys-G > Asn-A. Another example is hydrophobic interactions. In the scoring scheme of Suzuki and Yagi (8) most of those interactions were scored equally. Here, Ala-T and Ile-T seem to be the most favorable among hydrophobic interactions, followed by Tyr-T and Met-T. However, because of the limited size of the data set used to derive these parameters, such conjectures should be drawn with caution. It is expected that with the accumulation of more solved complexes the accuracy of such quantitative parameters will increase. The advantage of the current computational approach is that the quantitative measures can be refined systematically when the pair frequency of amino acids and bases is updated. We have recently explored the possible role of CH^...O interactions in protein-DNA recognition and concluded that inclusion of these interactions in the amino acid-base frequency matrix results in a more consistent pattern of the preferred pairs, which can be explained on the basis of electrostatic considerations (25). Thus, inclusion of these interactions in future derived log odds matrices may also improve the parameters. Nevertheless, as demonstrated above and discussed below, despite the relatively small size of the structural data set used and with all reservations taken into account, using such parameters for prediction looks promising.

The usual application of log odds matrices in the context of sequence and structure analysis is to evaluate compatibility either between two sequences or between a sequence and a structure. The approach is to sum the appropriate log odds along the alignment to a score that can be used to compare different alignments. The inherent assumption in all these studies is that the contributions of the different positions along the alignment are independent. The search for compatible protein and DNA sequences that will form the most favorable protein-DNA complex can also be viewed as an alignment problem, where the log odds matrix is used to find the best match of protein and DNA sequences among several possibilities. Additivity in the contributions of the pairs involved in the complex is assumed here also, however, in this case it has some experimental support. In at least one experimental binding study additivity was inferred based on differences in the dissociation constants of all possible sequence variants of OR1 bound to the Cro protein of phage [lambda] (26). In other studies consistent interpretations of binding results of substituted sequences (amino acids and/or DNA bases) are possible when each pair interaction is considered as independent of the other interactions (see for example 27). Ideally, possible interdependence between binding residues should be taken into account and is expected to improve prediction. This may be feasible with a significantly larger data set of solved complexes, enabling quantitative evaluation of the mutual effects of different pairs of amino acid-base on binding.

In the current study the validity of the computed scores is demonstrated by their consistency with experimental binding data of variant sequences of zif268 zinc fingers and their DNA binding sites. In most cases the computed scores succeeded in ranking highly those combinations of amino acid and DNA base triplets that were selected experimentally, although not always in the highest rank (Fig. 4). In some instances the amino acids or bases that were selected experimentally were not those that are the most compatible according to the log odds matrix. This is probably due to the effect of other factors on binding. Note that these quantitative parameters were derived from the pair interactions in a variety of protein-DNA complexes and reflect the likelihood of interaction in general. Consequently, possible position-dependent effects specific to each binding motif are masked. For example, in the zif268-like zinc fingers steric constraints that are position dependent are probably imposed by the specific orientation of the protein binding element relative to the DNA (16). Conceivably, incorporation of position-dependent effects specific to each binding motif together with the quantitative parameters will yield better predictions (6,16). In addition, there are other factors that affect binding, such as the sequence context of the binding sites, coupled interactions, where one amino acid is assisted by another in contacting the DNA (see for example 15,28), and the structure of the DNA binding site (see for example 29-31). In the case of the zif268-like zinc fingers all these additional factors were not taken into account and yet the scores calculated by the quantitative measures gave quite satisfactory predictions. This is probably due to the simple binding framework of these proteins. In more complicated cases the other factors may carry more weight and should also be considered. This requires quantitation of the different parameters, like position-dependent effects and coupled interactions, as well as prediction of the DNA structure in the binding site. Still, for all families of transcription factors in which the framework of amino acid-base interactions is defined, quantitative parameters such as those extracted here are useful for first screening and narrowing down the number of candidate sequences.

ACKNOWLEDGEMENTS

We thank Robert Jernigan, Victor Zhurkin and Ora Schueler for helpful discussions. This study was supported by a grant to H.M. from the Israel Science Foundation, administered by the Israeli Academy of Sciences and Humanities.