structural similarity of DNA-binding domains of Current Biology 1993, 3:141-148 Background ~hough the number of protein structures deposited in he 8rookhaven protein database (PDB) has grown rapidly in recent years [l], the subset of new pro- tein folds hzs grown at a significantly slower rate [ 21. This rate Merence still persists after allowing for the many structural determinations of homologous, mutant and drugcomplexed versions in the same basic pro- tein f-. Therefore, assuming there is no systematic bias in the selection criteria in deciding which particu- lar protein strucmre is to be determined, it has been suggested that we are ‘closing-in’ on the complete repertoire of folds that are allomable from the multi- rude that constitute dl possible protein structures [3]. The limited number of these folds may be due to evolu- tion. once there are enough foIds to create all possible protein functions there is then no pressure to evolve new folds. On the other hmd, the limit to the num- ber of folds may be due to the existence of basic stmJd limitarions that dictate, and thus relate, the three-dimensional srructures of proteins. Finding and understanding such principles of protein construction will help in the design of new and variant proteins. Assuming that the reservoir of unobserved folds is de pleting rapidly, any structural constraints should be de retable in the srructural database presently available to us. Suitable and exhaustive comparisons of these structures against each other could reveal unexpected similarities that could help catalogue and, perhaps, de- fine structunl principles. In this context, it is worth noting that anmaiogous studies of the one-dimensional DNA and protein sequence databases, made possible by the development of elegant computer algorithms, Correspondence to: M. Levitt i have borne much fruit ‘in identifying and catalogu- ing many novel sequence motifs of functional interest [4,5]. With regard to the problem of comparing two different three-dimensional protein ,$mctures consid- ered here, despite early (and more kcently plentiful) work in the development of suitable computer algo- rithms, systematic studies have been limited [Glll . Many of the available methods have been hampered by limitations in accuracy, speed and sensitivity. Here we present a new method for protein Structure ~ comparison that is accurate, fast and sensitive. Using this improved tool, we present the highlights of an ex- haustive comparison of all pairs of protein structures in the PDB. The discovery of a significant structural similarity between two well-studied protein families, the bacteriophage repressors and the globins, emphasizes the power of our method. Wjth its speed and sensitivity, it can aid the crystallographer and NMR spectroscopist in rapid identification of the relatedness of a newly de- termined structure to all previously reported ones. Such discoveries will in turn help to identify the rules chat govern higher order structural motifs. Results Aligning structures Our method aligns two protein structures by starting with an arbitrary equivalence of residues that are super- imposed in three-dimensions. A structural alignment matrix, which is calculated from distances behveen pairs of residues that are not in the same protein, is searched to achieve the optimal alignment. This gives I @ Current Biology 1993, VO~ 3 NO 3 141 i IFig. 1. A stereo view of the best structural alignment produced by our method between the Ca backbone of azurin (1AZU) and plasto- cyanin QPCV]. The matched residues of 1AZU are coloured green and other residues yellow; the matched residues of ZPCY are coloured red and other residues blue. For this alignment, the CRMS value is 2.80 A, for 89 matching Ca atoms. a new set of equivalent residues, which are again super- imposed, and the procedure is iterated to convergence (see Materials and methods). Known cases To test he accuracy of this method, we used the tech- nique to align some familiar cilses that are known to be at the extreme limit of detectability. Our alignment of rhe two copper-binding metalloproteins, azurin (PDB enuy IAZU) and phtocyanin (2PCY) (Fig. I), agrees with the findings of Taylor and Orengo [8], but dif- fered from the earlier alignment results of Chothia and Lesk [12] 'for 2PCY residues 45-65. In addiiton, our list of the mehe PDB enuies that are most related to hen egg-white Lysozyme (2LYM), as sorted by our structural alignment score (SAS), gustrates a sensible . - - ..-ea rank ordering of related structures in the database (Table 1). Sigdicandy, the seventh entry in our SAS list correctly identilies the structural similarity of T4 lysozyme (21zM) and hen egg-white lysozyme. This similarity, which is commonly believed to be at the limit of detection, can be considered as defining the ~ boundary between convergent and divergent evolution [13]. Using standard methods, the root mean square (Rh49 of equivalent Ca atoms (cRMS) for 2LZM and hen egg-white lysozyme is found to be 4.8211 over 89 n "2. .z.=. _. - ~. Table 1. Best matches to hen egg-white lysozyme (ZLYM 129). Protein N cRMS SA5 I% N,, n Biological name Source I_ 2HFC 126 2LZ2 119 3HFM 123 lLYM 122 lLZl 123 IALC 115 2LZM 89 BAD8 81 2PRK 98 2RUB 85 lGPl 81 lPPD 78 0.42 0.43 0.46 0.47 0.54 4.82 0.97 6.06 5.47 5.50 5.86 5.45 0.34 99.2 0.37 100.0 0.36 94.9 0.38 100.0 0.44 61.7 0.84 36.5 6.19 10.2 5.41 6.7 6.75 6.79 2.4 4.9 6.99 6.89 6.4 9.4 0 554 Fab HyHEL-S/lysozyme Mouse/chicken . 0 129 Lysozyme 0 558 Fab HyHEL-1Ollysozyme Turkey 0 258 Lysozyme Mouse/chicken 1 130 Lysozyme Hen egg-white, monoclin Human 4 122 a-Lactalbumin 8 164 Lysozyme Baboon 10 279 Proteinase K Tritiracl. alb. limb Bactetio hage 14 8 374 Alcohol dehydrogenase (apo) Horse liver 11 368 Glutathione peroxidase 9 882 Rubisco Bovine Rhodosp ir i 1 1 urn rubcum 7 212 Papain 0 Papaya The top 12 matches found by our method to the 129 residue hen egg-white lysozyme structure (ZLYM), using the 295 most order of decreasing similarity as measured by structural alignment scores (SASI. The PDB entry identifier of the structure is representative coordinates sets from the entire July 1991 release of the Brookhaven protein structure database. The list is in sequence identity lor