Results

A summary of the experimental design is shown in Fig. 1. The amine groups of Lys and the N terminus react with the homobifunctional cross-linker BS³, producing either single residue modifications (⁺C₈H₁₂O₃) or interresidue cross-links (⁺C₈H₁₀O₂). Under the cross-linking conditions used, approximately one Lys-Lys cross-link per protein molecule was observed (data not shown), a level that is unlikely to distort tertiary structure (6). Masses of tryptic peptides derived from FGF-2, including unmodified, modified, and cross-linked, were assigned from the mass spectra by using asap. Trypsin, which normally cleaves at lysines and arginines, will not cleave at BS³-modified lysines, aiding in the identification of cross-linked peptides.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Experimental Design. (Top) Cross-linking of FGF-2. (Middle) Purification of monomeric FGF-2. (Bottom) Proteolytic digestion, HPLC separation, and mass spectrometry.

Eighteen masses corresponding to cross-linked peptides are listed in Table 1. For example (Fig. 2A), the peak at m/z 2059 was assigned as the single peptide E45-R60, cross-linked through internal lysines 46 and 52. The peak at m/z 2465 was assigned as a cross-link between lysines 26 and 46 in peptides L23–R33 and E45–K52. These two assignments, along with those made for six additional cross-linked peptides, were confirmed by MALDI-PSD of the parent ions (Fig. 2 B and C), which contained immonium ions (⁺NH₂⩵CHR) and sequence information consistent with the expected peptides and cross-link position (Table 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Mass spectrometry. (A) MALDI-TOF spectrum from tryptic digest of BS³ cross-linked FGF-2. Cross-linked peptides are identified by using the program asap and are denoted with an asterisk (9). (B) MALDI-PSD spectrum of cross-linked peptide E45–R60 (M + H⁺ = m/z 2059.08). (C) MALDI-PSD spectrum of cross-linked peptides L23–R33 and E45–K52 (M + H⁺ = m/z 2465.25). Fragments from peptide L23–R33 are labeled with an α subscript; those from peptide E45–K52 with a β subscript. Fragments due to amide bond cleavages (b-ions, y-ions, and immonium ions) are labeled according to their single letter abbreviations. The most abundant fragments arose from cleavages at cross-linked lysines and aspartic acid. Peaks at m/z 696 and 1771 correspond to fragmentation at y_6β and/or y_6α, and b_2β and/or b_5α, respectively.

Of the total of 18 cross-links identified, 3 (cross-linked peptides 46–52, 119–125, and 125–129) were within 7 amino acids of each other in the FGF sequence. Because BS³ conjugated to two lysines can span up to 7 amino acids (the total span is ≈24 Å), these 3 cross-links provided little or no three-dimensional structure information. The remaining 15 cross-links provided new and very valuable throughspace distance information (Fig. 3). It is this type of throughspace distance constraint that is the key to determining the fold of a protein.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

The 15 nonlocal throughspace distance constraints generated by the chemical cross-links (yellow dashed lines) superimposed on the average NMR structure of FGF-2 (1BLA). The 14 lysines of FGF-2 are shown in red.

Protein fold identification was carried out in a two step process. First, we used sequence threading (program 123D, ref. 15) of FGF-2 (FGF2_BOVIN) to find the best 20 structural models out of a database of 635 proteins sharing <30% sequence identity (16). Three β-trefoil proteins are ranked 1st (FGF-2: 4FGF), 5th (IL-1β: 1I1B), and 12th (hisactophilin: 1HCE). Second, we reranked the models by their agreement with the 18 cross-link-derived distance constraints (Table 2). The threading models were scored by using the equation: where n is the number of constraints; d_i is the C^α–C^α distance in the model for the two residues in constraint i; and d₀ is 24 Å, the maximum C^α–C^α throughspace distance between BS³-cross-linked lysines. Two models were discarded that had >50% of the constraints undefined because of unresolved structural regions or gaps in the sequence alignment (2PHY, 2PLDA). The three β-trefoil proteins are now ranked 1st, 2nd, and 4th.

FGF-2 is clearly identified as a member of the β-trefoil family and is correctly predicted to share the same fold as IL-1β even though their sequence identity is <13%. Like the β-trefoils, gastrotropin (1EAL), ranking third, has a β-barrel fold with a meander motif (17) and is partially alignable to FGF-2 by using the program dali (18). Without the distance constraints, the threading algorithm would suggest that FGF-2 belongs to the unrelated β-clip fold family (D-UTPase).

Given the robust identification of FGF-2 as a β-trefoil, we built a homology model based on IL-1β and the threading alignment (Fig. 4). In the model, the core β-strands are positioned correctly. Sequence alignment and modeling errors occur mainly in the loop regions (19) and the 20 N-terminal amino acids. The backbone rms deviation of the model to the FGF-2 average NMR structure (1BLA) is 4.8 Å over 101 residues (14).

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

FGF-2 homology model based on IL-1β (cyan) aligned to the NMR structure (white) of FGF-2. Insertions of FGF-2 relative to IL-1β were not modeled. Structural gaps were closed with 100 steps of energy minimization by using Tripos (St. Louis) sybyl 6.4. The backbone rms deviation is 4.8 Å over 101 residues.

Discussion

The determination of the three-dimensional structure of a molecule can be represented as a problem in obtaining sufficient information, in the form of distance constraints, to fix the positions of the atoms or residues to a certain accuracy, or resolution (4). The number and precision of the constraints define the types of structural questions that can be answered. At one extreme, the placement of a pair of rigid domains with respect to each other (6 degrees of freedom) can be carried out with as few as three constraints per domain (20, 21). At the other, the determination of a macromolecular structure to residue resolution, using strictly distance constraints, would require ca. 3N constraints, where N is the number of residues. The work reported here shows that assigning the tertiary fold of a protein domain can be accomplished with few constraints (approximately N/10) if the fold of interest has been previously observed.

FGF-2 proved a good choice for the initial study because it contains a high concentration of dispersed lysine residues and is a member of a well-catalogued fold family. For proteins with lower lysine content, or in cases where more distance information is needed, additional cross-linkers could be used. Because the number of easily modified (i.e., polar) surface-exposed residues in a typical protein greatly exceeds N/10 (22), this number of identifiable cross-links should be readily achievable. There are commercially available cross-linkers that react specifically with other polar groups besides lysine, have different spacer arm lengths, or have different flexibility. Alternatively, nonspecific homo- or heterobifunctional photocross-linkers can be used. This methodology as implemented has another dependency besides the requirement that ≈N/10 constraints be generated; the scoring function of the threading algorithm must be able to select a set of plausible models where one has the correct fold. A threading-based approach, in turn, requires that the correct fold has at least one representative in the sequence-unique structural database. For sequences for which cross-linking data are inconsistent with all threading models in the sequence-unique structural database, one might explore de novo structure prediction methods to construct models consistent with distance constraint information (23–28). By using simulated data sets, such methods have been shown to generate low resolution structures with, again, N/10 distance constraints (29).

Once the general methodologies were developed, the distance constraints used in this study were derived from three experiments, each taking a week and using approximately 100 μg of protein. Given that commercial purification/protease mapping HPLC workstations are available, the time needed to perform an experiment could be reduced to 2 days, and the amount of protein required to less than 10 μg. Throughput can be significantly increased by performing experiments in parallel by using automated systems to perform multiple experiments simultaneously. The purity of the protein analyte is not critical, because the only peaks of interest are those consistent with cross-linked peptides of known sequence. Therefore, by using a standard baculovirus expression system for a His-tagged protein, the necessary amount of protein at the required purity could be prepared in 2 days. By using the methodology we have described here, and with high throughput systems in place, we estimate the time to determine the fold family of a protein to be on the order of 1 week.

Our studies show that it is feasible to obtain multiple cross-links per experiment and that even a small number of distance constraints can assist in the determination of a protein structure. Protein models constructed by using this technique may provide sufficient resolution to assist in elucidating function, macromolecular interactions, and multiple conformational states, as well as in the design of protein mimetics. This method should be broadly applicable to the field of structure determination. It is many times faster than the current standard techniques, it can be performed on much less material, automation is eminently possible, and the means already exist to apply it more generally in the field of protein structure analysis.