Structural Basis for Partial Redundancy in a Class of Transcription Factors, the LIM Homeodomain Proteins, in Neural Cell Type Specification*

Background: Lhx and Isl proteins contribute to genetic control in developing neurons. Results: The Lhx3/4-binding motif in Isl2 was identified, and the structures of Lhx-Isl complexes were characterized and compared. Conclusion: There are minor differences in the structures of Lhx3/4 binding Isl1/2 reflected by mutational and biophysical analyses. Significance: Redundant sets of interactions conserve function in developing neurons while allowing divergence in other contexts. Combinations of LIM homeodomain proteins form a transcriptional “LIM code” to direct the specification of neural cell types. Two paralogous pairs of LIM homeodomain proteins, LIM homeobox protein 3/4 (Lhx3/Lhx4) and Islet-1/2 (Isl1/Isl2), are expressed in developing ventral motor neurons. Lhx3 and Isl1 interact within a well characterized transcriptional complex that triggers motor neuron development, but it was not known whether Lhx4 and Isl2 could participate in equivalent complexes. We have identified an Lhx3-binding domain (LBD) in Isl2 based on sequence homology with the Isl1LBD and show that both Isl2LBD and Isl1LBD can bind each of Lhx3 and Lhx4. X-ray crystal- and small-angle x-ray scattering-derived solution structures of an Lhx4·Isl2 complex exhibit many similarities with that of Lhx3·Isl1; however, structural differences supported by mutagenic studies reveal differences in the mechanisms of binding. Differences in binding have implications for the mode of exchange of protein partners in transcriptional complexes and indicate a divergence in functions of Lhx3/4 and Isl1/2. The formation of weaker Lhx·Isl complexes would likely be masked by the availability of the other Lhx·Isl complexes in postmitotic motor neurons.

LIM domains, named for the first three proteins in which the motif was identified (Lin-11/Islet-1/Mec-3), are zinc fingers that coordinate two zinc ions and mediate protein-protein interactions (reviewed in Ref. 1). Two related families of LIMcontaining proteins, LIM homeodomain and LIM-only proteins, are expressed in a combinatorial manner along with their protein-binding partners to form the "LIM code," a transcriptional code that helps specify cell type during the development of the central nervous system and many other tissues and organs (e.g. Refs. 2 and 3). The 12 mammalian LIM homeodomain proteins comprise six pairs of paralogues (4), which often have overlapping expression patterns and functions. LIM homeobox protein 3 and 4 (Lhx3 and Lhx4) 5 and Islet-1 and -2 (Isl1 and Isl2) are two such paralogous pairs. All four proteins are functional in developing motor neurons (3,5). Mice in which these genes have been individually disrupted are all embryonic or perinatal lethal, but the phenotypes differ, showing that the pairs of paralogues are not genetically redundant. For example, both Lhx3 Ϫ/Ϫ and Lhx4 Ϫ/Ϫ mice are perinatal lethal but show primary defects in pituitary and lung development, respectively (6,7). Isl1 Ϫ/Ϫ mice die about halfway through embryonic development (embryonic day 11.5) with abnormal heart and pancreas development and an absence of motor neurons (8), whereas Isl2 Ϫ/Ϫ mice are perinatal lethal, probably due to disrupted breathing resulting from defects in motor neuron differentiation in the thoracic levels of the spinal cord (9).
In differentiating motor neurons, the paralogues show some redundancy. For example, during chick ventral motor neuron development, Lhx3 and Lhx4 must both be knocked out to specify dorsal rather than ventral motor neurons (5). In mice and zebrafish, exogenous Isl1 or Isl2 can trigger motor neuron differentiation, but both must be expressed to maintain proper motor neuron cell fate (9 -11).
Lhx3 and Isl1 form a simple and well characterized example of the LIM code associated with the development of adjacent cell types (V2 interneurons and motor neurons) in the ventral spinal cord. Lhx3 is expressed in developing V2 interneurons, whereas both Lhx3 and Isl1 are expressed in postmitotic motor neurons. These LIM homeodomain proteins are strong drivers of cell specification. The addition of Lhx3 to dorsal spinal cord cells results in the formation of V2 interneurons, whereas further addition of Isl1 switches their identity to motor neuron cells (12).
Lhx3 and Isl1 form cell-specific transcriptional complexes that also contain LIM domain-binding protein 1 (Ldb1), a widely expressed transcriptional regulator that binds to, and is essential for the activity of, members of the LIM homeodomain and LIMonly families (13)(14)(15). Lhx3 binds Ldb1 to generate an active binary complex in V2 interneurons (Fig. 1A). In motor neurons, Isl1 disrupts this interaction to form a ternary complex in which Isl1 displaces Lhx3 as the binding partner of Ldb1 and forms additional contacts with Lhx3 ( Fig. 1B) (12). The contacts between Lhx3 or Isl1 and Ldb1 are made by the tandem LIM domains from the LIM homeodomain proteins and the ϳ30-residue LIM interaction domain from Ldb1 (Ldb1 LID ) (Fig. 1C) (13,14,16). An essentially identical interaction is formed between the LIM domains from Lhx3 and an Ldb1 LID -like sequence near the C terminus of Isl1, the Lhx3-binding domain (Isl1 LBD ) (17).
This study explores the possibility that Isl2 and Lhx4 may be capable of substituting for Isl1 and Lhx3, respectively, in the ternary complex. We identify a new Lhx3-binding domain in Isl2 (Isl2 LBD ), which binds both Lhx3 and Lhx4, and show that Isl1 LBD also binds Lhx4. The x-ray crystal structure of Lhx4 LIM1ϩ2 ⅐Isl2 LBD , 6 combined with biophysical and muta-genic analysis of the series of pairwise interactions, shows that although the structures of Lhx3/4⅐Isl1/2 pairwise complexes are very similar, there are differences in the nature of the binding interactions between these LIM homeodomain family proteins that likely contribute to diverging biological functions.
X-ray Structure Determination and Refinement-The crystallization, collection, and processing of data for Lhx4-Isl2 were described previously (19). The program SOLVE (22) was used with anomalous data to 3.0 Å resolution to identify the zinc atom positions and generate initial phases. Statistical density modification and local pattern matching were carried out using RESOLVE (23). The model was built and revised manually with Coot (24). Refinement of the diffraction data was carried out to 2.16 Å using REFMAC5 and PHENIX with TLS refinement (25)(26)(27). The final model was refined using PHENIX. The MOLPROBITY server was used to identify steric clashes and unconventional geometry for validation of the structure (28).
Small-angle X-ray Scattering (SAXS) Data Acquisition and Analysis-Solution SAXS data (I(q) versus q, where q ϭ (4sin)/, 2 is the scattering angle, and the x-ray wavelength) were measured at 10°C using a SAXSess (Anton Paar) line collimation instrument (10-mm slit width; Ref. 29). For details of sample environments, see supplemental data 3. Data reduction included corrections for beam geometry, sample absorbance, and detector sensitivity as described previously (29). Data were placed on an absolute scale, and the forward scattering at zero angle (I(0)) was used to evaluate the molecular weight of the scattering particles (30) using contrast values (⌬ M ) and partial specific volumes () derived from the pro- A, the proteins Ldb1 (gray) and Lhx3 (black) assemble into a binary complex. B, with the addition of Isl1 (white), the ternary complex is preferentially formed (12). Note that the self-association domain of Ldb1 forms trimers in vitro (21) and that the homeodomains from Lhx3 and Isl1 bind DNA. The spacing between the homeodomain-binding sites (dashed lines) is not defined and may represent close or distant binding sites. C, details of domain contacts within the binary and ternary complexes. The LIM interaction domain (LID) from Ldb1 makes contacts with the LIM domains (LIM) from Lhx3 in the binary complex, but in the ternary complex, Ldb1 LID contacts the LIM domains of Isl1, and the LBD from Isl1 binds the LIM domains from Lhx3. SD, self-association domain; HD, homeodomain.
gram MULCh (31). Guinier analysis of the desmeared data was performed using PRIMUS (32). GIFT (31) was used to calculate probable atom pair distance distributions (P(r) versus r) from which the maximum dimension (D max ), radius of gyration (R g ), and I(0) of each tethered complex were determined. Ab initio shape restorations of the molecular envelopes for each complex were performed using DAMMIF with the outputs from GIFT. Further rigid body modeling was performed using BUNCH (33) (see supplemental data 3 for details). The fits of the refined BUNCH models against the SAXS data were assessed using CRYSOL, which also was used to calculate scattering profiles and fits of the original crystal structures to the data. The P(r) versus r of the crystal or BUNCH structures were determined as described in Ref. 34 using GNOM.
Yeast Two-hybrid Analysis-Inserts were cloned into pGAD10 and modified pGBT9 plasmids and co-transformed into AH109 cells (Clontech), as described previously (35). Transformed cells were selected by growth on leucine-and tryptophan-deficient media. For the detection of an interaction, the following selection media were used as indicated (all selection media lacked leucine and tryptophan). For low stringency (ϪH ϩ X-␣-gal), the medium was additionally deficient in histidine (ϪH) and supplemented with 40 g ml Ϫ1 X-␣-gal; for moderate stringency (ϪH ϩ X-␣-gal ϩ 3-AT), the medium was as for low stringency but further supplemented with 1 mM 3-amino-1,2,4-triazole (3-AT); for high stringency, the medium was deficient in both histidine and adenine (ϪH-A). Numbers of yeast cells were normalized such that A 600 nm ϭ 0.2 and were deposited in 2-l drops at the dilutions indicated.
Chemical Denaturation-Stability studies using guanidine hydrochloride as the denaturant were conducted as described previously (36). Tryptophan fluorescence following excitation at 295 nm was recorded in the range 320 -380 nm. The fluorescence emission wavelength maximum ( max ) was used as an indication of the unfolding of the tethered protein complex. The max was obtained by fitting of a Gaussian curve through the emission spectra using Origin 6.1 (OriginLab).

RESULTS
Identification of the Lhx3-binding Domain in Isl2-Isl1 LBD spans residues 262-291 ( Fig. 2A) (17). Comparison of the sequences of Isl1 and Isl2 suggested that a similar region in Isl2, residues 272-301 (Isl2 LBD ), might mediate binding to Lhx3 ( Fig. 2A). We used yeast two-hybrid analysis to test the ability of Isl1 LBD and the putative Isl2 LBD to interact with the LIM domains of Lhx3 and Lhx4. When the Isl LBDs were fused to the GAL4 DNA-binding domain (GAL4 DBD ), all pairs of proteins exhibited growth on high stringency (ϪH-A) selection plates, indicating that all four pairwise interactions among these proteins take place (Fig. 2B). When the Lhx3/4 LIM1ϩ2 constructs were fused to the GAL4 DBD , the Lhx3/4 LIM1ϩ2 ⅐Isl1/2 LBD interactions were all evident under moderate stringency conditions (ϪH ϩ X-␣-gal ϩ 3-AT), and only the interaction between Lhx4 LIM1ϩ2 and Isl1 LBD was detected under high stringency conditions (Fig. 2C). Thus, Isl2 LBD does mediate interactions with Lhx3, and both Lhx3 and Lhx4 can interact with equivalent domains in Isl1 and Isl2.

Generation of Stable Lhx-Isl Constructs-When expressed in
bacteria, the LIM domains from Lhx3 and Lhx4, like those from other LIM-only/LIM homeodomain proteins, tend to aggregate and are largely insoluble (37,38). Tethering an interaction partner such as Isl1 LBD or Ldb1 LID to the LIM domains protects them from aggregation, allowing soluble tethered complexes to be produced (e.g. Refs. 18, 38, and 39). As shown for Lhx4⅐Isl2 in Fig. 3, A and B, we generated constructs of the form Lhx3/ 4 LIM1ϩ2 -linker-Isl1/2 LBD , where the 11-residue linker was designed to be flexible by the inclusion of glycine and serine residues (18,19). These constructs are subsequently referred to as Lhx3-Isl1, Lhx3-Isl2, Lhx4-Isl1, and Lhx4-Isl2.
Crystal Structure of Lhx4-Isl2-The protein Lhx4-Isl2 was crystallized as described previously (19). The structure of Lhx4-Isl2 was determined using single-wavelength anomalous dispersion data recorded at the zinc absorption edge ( ϭ 1.2819 Å) to a resolution of 2.16 Å ( Table 1). The R work and R free are consistent with an x-ray crystal structure of this resolution with some disordered regions (e.g. the linker between Lhx4 LIM1ϩ2 and Isl2 LBD ).
Lhx4 contains two typical LIM domains, each comprising two zinc-coordinating modules made up of two sequential ␤-hairpins and a short segment of ␣-helix (Fig. 4A, orange ribbon). Isl2 LBD (purple ribbon) binds across the length of both LIM domains, forming a "head-to-tail" complex. Isl2 LBD forms small stretches of ␤-strand structure along the interface, packing against the second ␤-hairpin of all four zinc-coordinating modules to make small, three-stranded antiparallel ␤-sheets.
The two molecules in the asymmetric unit (chains A and B, chain B is shown in Fig. 4A) are highly similar (root mean square deviation ϳ0.28 Å for the backbone atoms). Fourteen residues are missing from both chains, an additional residue is found only in chain A, and another four are found only in chain B. The missing residues are from the linker and termini of the protein (supplemental Table S2A). Of the residues modeled, ϳ20% have missing side chains (see supplemental Table S2B), most commonly in the region between the two LIM domains (the "hinge") and the adjacent region in Isl2 LBD (the "spacer"). In both chains, residue Asp-66 has poorly defined electron density and lies in the disallowed region of the Ramachandran plot. Apart from these minor differences, both molecules have the same basic structure.
The overall conformation of the structure is unexpectedly compact when compared with previously determined structures of Lhx3-Isl1 and a related LIM-only complex (e.g. Fig. 4B) (17,35). The Lhx4-Isl2 complex has a sharp bend of ϳ100°at the hinge/spacer region between the two domains. As a result, the two LIM domains of Lhx4 lie almost perpendicular to one another.
Solution Properties of Lhx4-Isl2-SAXS data were measured to probe the structure of Lhx4-Isl2 in solution (supplemental Fig. S3). The linearity of the Guinier plot and the molecular weight of Lhx4-Isl2 estimated from the extrapolated zero-angle scattering (I(0)) indicate that under these conditions, the solution contains non-interacting, monodisperse Lhx4-Isl2 molecules ( Table 2 and supplemental Fig. S3). The calculated scattering curve from CRYSOL (40) using the coordinates of the Lhx4-Isl2 crystal structure shows deviations from the shape of the experimental curve in the very low-q and mid-q regimes (supplemental Fig. S3). These differences become more pronounced when comparing the interatomic distance distribution functions, P(r), which are very sensitive to symmetry and domain structure and can provide more intuitive information about the shape of a molecule (Fig. 5A). The larger radius of gyration, R g , and longer maximum dimension, D max , of the molecule for the solution data (R g ϭ 23.5 Å; D max ϭ 75-80 Å) when compared with the crystal structure (R g ϭ 20.5 Å; D max ϭ 70 -75 Å) suggest that the solution structure is more elongated than that in the crystal form.
This type of tandem LIM-binding peptide complex is thought to have some flexibility at the hinge/spacer (17,41). A rigid body refinement using the crystal structure and the SAXS data was carried out in which the two LIM peptide halves of the crystal structure were allowed to flex at the hinge/spacer. The BUNCH models support a more elongated conformation in solution with an average angle at the hinge/spacer of ϳ120° (  Fig. 5B). The more extreme bend in the Lhx4-Isl2 crystal structure likely represents a rarely populated conformation that had a preference for crystallization under the conditions used.
Solution Properties of Lhx3-Isl1 and Lhx3-Isl2-SAXS data were also collected for Lhx3-Isl1, Lhx3-Isl2, and Lhx4-Isl1. The Lhx4-Isl1 sample showed signs of aggregation, and data were not further analyzed (data not shown), but Lhx3-Isl1 and Lhx3-Isl2 were monomers in solution as evidenced by molecular mass estimates from I(0) ( Table 2). Lhx3-Isl1 and Lhx3-Isl2 have similar asymmetric P(r) profiles, indicating that the complexes adopted extended conformations in solution (Fig. 5C). However, the R g and D max for Lhx3-Isl2 are shorter (R g ϭ 22.3 Å; D max ϳ75 Å) than Lhx3-Isl1(R g ϭ 26.9 Å; D max ϳ90 Å), indicating that Lhx3-Isl2 is more compact than its Lhx3-Isl1 counterpart, as seen in the BUNCH models (Fig. 5, D and E). The solution structural parameters are essentially identical for Lhx3-Isl2 and Lhx4-Isl2 (Fig. 5, B and E, and supplemental Table S3), suggesting that binding to Isl2 LBD causes both Lhx3 and Lhx4 to adopt a more compact structure than does binding to Isl1 LBD .

TABLE 1 Refinement statistics for Lhx4-Isl2
Values for the highest resolution shell are given in parentheses. Overall these solution scattering studies (see supplemental data 3 for a detailed analysis) indicate that tethered Lhx3/4-Isl1/2 complexes are likely to exist as an ensemble of elongated complexes with some flexibility at the hinge/spacer. Although the molecular details of each LIM domain peptide module are unlikely to be affected by flexion, details concerning the hinge/ spacer regions will likely vary between members of the ensembles, and analyses of high resolution structure should be careful not to overinterpret details from these regions.
Identifying Binding Hotspots in Lhx3/4⅐Isl1/2 Complexes-We used yeast two-hybrid analysis to compare the ability of the individual LIM domains from Lhx3 and Lhx4 to interact with Isl1 LBD , Isl2 LBD , and Ldb1 LID (supplemental Table S4) (17). The LIM2 domains from Lhx3 and Lhx4 were both capable of independently interacting with Isl1 LBD and Ldb1 LID , whereas the LIM1 domains could bind weakly to Ldb1 LID (some yeast growth was evident only under low stringency selection conditions), but not to Isl1 LBD . No individual LIM domain from either Lhx3 or Lhx4 showed evidence of an interaction with Isl2 LBD in this assay.
The Lhx3-binding domains of Isl1 and Isl2 were subjected to mutagenesis to identify the key residues in those domains that mediate binding with Lhx3 and Lhx4. Residues were mutated to alanine (or glycine if the native residue was alanine). Initially sequential sets of three consecutive residues were mutated and tested for their ability to interact with the LIM domains from Lhx3 and Lhx4 using yeast two-hybrid analysis. For those triple mutants that showed significantly reduced levels of binding, a second round of single point mutants was generated and tested in the same manner. The results from these mutagenic screens are reported in supplemental data 4 and are summarized in Fig. 6A. Finally, two point mutants, Isl1 M265L and Isl2 L275M , were used to test binding specificity ( Fig. 6B and supplemental data 4).
Mutations in the N-terminal (and not the C-terminal) half of Isl1 LBD significantly disrupted binding to both Lhx3 and Lhx4 (Fig. 6A) (17), which is consistent with only the LIM2 domains from the Lhx proteins being able to independently bind Isl1 LBD . In contrast, mutations in both the N-terminal and the C-terminal halves of Isl2 LBD abrogated binding to Lhx3 and Lhx4 (Fig.  6A), consistent with both LIM domains of the Lhx proteins being required to mediate an interaction with Isl2 LBD .
Among the residues identified as being most important for the interactions, Isl1 265-267 , and Isl2 275-277/293-295 , the only sequence difference between the two paralogues is a conservative switch from Met-265 in Isl1 to Leu-275 in Isl2. Interaction data for Isl1 M265L and Isl2 L275M suggest that although Lhx3 has no apparent preference for either residue at Isl1 265 /Isl2 275 , Lhx4 binds with higher affinity when there is a methionine at this position (Fig.  6B, supplemental Fig. S4, and supplemental Table S4).
Relative Stabilities of Lhx-Isl Complexes-We used an indirect approach to probe the relative binding affinities of Lhx3 and Lhx4 for each of Isl1 and Isl2. Stable forms of isolated LIM domains from Lhx3 and Lhx4 cannot be produced in a recom-   DECEMBER 16, 2011 • VOLUME 286 • NUMBER 50 binant fashion, preventing the use of standard binding experiments (such as ELISA, isothermal titration calorimetry, or surface plasmon resonance). We instead compared the relative resistance of the Lhx3/4-Isl1/2 complexes to chemical denaturation, an approach previously used to gauge the relative stability of tethered Lhx3-Isl1 and Lhx3 LIM1ϩ2 -Ldb1 LID complexes (17). The LBD/LID peptides are largely unstructured, and if we only compare identical folded LIM domains, the differences in resistance to chemical denaturation should report on differences in binding affinity between the LIM domains and the peptides. As the sequences of Lhx3 LIM1ϩ2 and Lhx4 LIM1ϩ2 differ, we only compared the apparent stability of Lhx3-Isl1 with Lhx3-Isl2, or Lhx4-Isl1 with Lhx4-Isl2, and did not compare Lhx3 constructs with Lhx4 constructs. Note that these proteins do not exhibit reversible unfolding, meaning that the data cannot be used to calculate differences in free energy of unfolding and/or actual binding affinities.

Structure of an Lhx4⅐Isl2 Complex
The resistance of the constructs to denaturation was measured by monitoring the change in max of tryptophan fluorescence as a function of guanidine hydrochloride concentration. Lhx3-Isl1 and Lhx3-Isl2 tethered complexes had virtually identical unfolding curves, with a midpoint of unfolding at ϳ2 M guanidine hydrochloride, suggesting that the Lhx3-Isl1 and Lhx3-Isl2 complexes have the same apparent stability (Fig. 6C). In contrast, Lhx4-Isl1 showed a midpoint of unfolding at ϳ2.5 M when compared with ϳ2 M for Lhx4-Isl2, suggesting that the Lhx4-Isl1 tethered complex is probably more stable than the Lhx4-Isl2 complex (Fig. 6D).

DISCUSSION
We have now shown that an Lhx3-binding domain exists in Isl2 as well as Isl1 and that both of these domains can bind to each of Lhx3 and Lhx4. Overall Lhx3 and Lhx4 share 66% sequence identity, whereas Isl1 and Isl2 share 74% sequence FIGURE 5. The solution structures of Lhx3/4-Isl1/2. A, P(r) profiles from experimental scattering data for Lhx4-Isl2 (green squares) and calculated scattering profiles from the Lhx4-Isl2 crystal structure (purple line) and the generated BUNCH model (green line). B, alignment using the atoms of Lhx4 LIM2 of the Lhx4-Isl2 crystal structure (purple) with the best-fit BUNCH model (green) and ab initio DAMMIF reconstruction (transparent green surface). C, P(r) profiles from experimental scattering data for Lhx3-Isl1 (blue triangles) and Lhx3-Isl2 (orange squares) and calculated scattering profiles from the Lhx3-Isl1 crystal structure (red dashed line), the Lhx3-Isl1 BUNCH model (dark blue line), and the Lhx3-Isl2 BUNCH model (dark orange line). D, alignment using the atoms of Lhx3 LIM2 of the Lhx3-Isl1 crystal structure (red) with the best-fit BUNCH model (blue) and ab initio DAMMIF reconstruction (transparent blue surface). E, the best-fit BUNCH model (orange) and ab initio DAMMIF reconstruction (transparent orange surface) for Lhx3-Isl2. LIM domains are shown as ribbons, and LBDs as are shown as spheres.
identity, which is highest in the homeodomains (95 and 98% sequence identity, respectively) and LIM domains (83 and 82% identity, respectively). Isl1 LBD and Isl2 LBD have lower sequence identity (60%). The structure of the Lhx4-Isl2 complex reported here is most divergent in sequence from the existing Lhx3-Isl1 complex structure (17), allowing a structure-guided interpretation of sequence, binding, and mutational data for Lhx3/Lhx4 and Isl1/Isl2 interactions.
The Lhx3-Isl1 and Lhx4-Isl2 Structures Have Similar Overall Structures-The overall shapes of Lhx3-Isl1 and Lhx4-Isl2 from the crystal structures differ by the angles at the hinge/spacer (ϳ100°for Lhx4-Isl2 and ϳ165°for Lhx3-Isl1; Fig. 4B). The solution structural parameters derived from SAXS data indicate that, although this difference is exaggerated in the crystal forms, in solution there is a difference in the hinge/spacer angles between the two complexes such that Lhx4-Isl2 and Lhx3-Isl2 are more compact than Lhx3-Isl1 (have a smaller angle). At the domain level, the folds of the LIM domains of Lhx3 and Lhx4 are highly conserved. An alignment of the back-bone atoms from chain B of the Lhx4-Isl2 structure and the Lhx3-Isl1 structure (Protein Data Bank (PDB) accession code 2RGT, chain B; Ref. 17) gives an r.m.s.d. of 0.77 Å for the LIM1 domains (Fig. 7A) and 0.59 Å for the LIM2 domains (Fig. 7B).
Isl1 LBD and Isl2 LBD have a backbone r.m.s.d. of 8.7 Å, which is not surprising given the different shapes of the complexes. If we consider the LIM1-and LIM2-binding regions of the LBDs separately, the r.m.s.d. drops to 0.56 Å for Isl1 279 -287 /Isl2 289 -297 and 0.38 Å for Isl1 262-273 /Isl2 272-283 . Isl2 LBD forms four well defined ␤-strands at the interface with Lhx4, and Isl1 LBD shows an almost identical binding topology, with three well defined strands and evidence of a fourth short ␤-strand from two interdomain backbone-backbone hydrogen bonds over the third zinc-coordinating module of Lhx3. The program PISA (42) reveals similar surface areas buried at the interfaces of the two complexes, ϳ1490 Å 2 for Lhx3 LIM1ϩ2 ⅐Isl1 LBD and ϳ1540 Å 2 for Lhx4 LIM1ϩ2 ⅐Isl2 LBD .
Structure-guided Interpretation of Binding Data-A comparison of the interfaces in the two structures shows that there is a Colored boxes indicate where mutation had a strong (red), moderate (orange), or minor effect (yellow). White boxes indicate residues mutated in triple-alanine constructs that had a minor effect on growth of yeast. Gray boxes indicate no effect on yeast growth. B, comparison of wild-type and mutant LBDs from Isl1/Isl2 binding to the LIM domains of Lhx3/Lhx4 by yeast two-hybrid analysis. AH109 yeast cells co-transformed with pGBT9/pGAD10 vectors shown on the right were tested for growth under different selection conditions (ϪH ϩ X-␣-gal ϩ 3-AT) or (ϪHϪA); 0 indicates no dilution of yeast cells (A 600 nm ϭ 0.2), 1 indicates a 1:10 dilution (A 600 nm ϭ 0.02), and 2 indicates a 1:100 dilution (A 600 nm ϭ 0.002). C, resistance of Lhx3-Isl1/2 complexes to denaturation by guanidine hydrochloride (Gdn.HCl). Ⅺ, Lhx3-Isl1; f, Lhx3-Isl2. D, resistance of Lhx4-Isl1/2 complexes to denaturation by guanidine hydrochloride. E, Lhx4-Isl1; •, Lhx4-Isl2. For C and D, max reports maximum emission wavelength in the range 320 -380 nm with excitation at 295 nm.

Structure of an Lhx4⅐Isl2 Complex
high sequence similarity in the stretches of the LBDs that contact each LIM domain (the binding motifs) but poor sequence conservation in and around the spacer (Fig. 6A). The specific intermolecular contacts of the two complexes are almost identical at the LIM2/N-terminal LBD interfaces but are more varied at the LIM1/C-terminal LBD interfaces (Fig. 7, C and D). In the latter region, there are minor variations in the side-chain conformations of some residues at the interface when compared with those at the LIM2/N-terminal LBD interfaces (Fig.  7, E and F). These differences in structure are reflected by different contributions to binding of the C-terminal halves of Isl1 LBD and Isl2 LBD and the effects of mutation on binding. Isl2 E293A/V294A/Q295A abrogated binding to both Lhx3 and Lhx4 (Fig. 6A, supplemental Fig. S4, and supplemental Table  S4). Isl2 E293-Q295 is equivalent to Isl1 E283-Q285 in Isl1, but the side chains of those residues form predominantly different sets of contacts in the two complexes (Fig. 7, C and D), and the backbone alignment of Lhx3 and Lhx4 in this region shows more deviation than elsewhere in the molecules (Fig. 7A, marked by an asterisk). The side-chain conformations of Isl2 E293 and Isl2 Q295 and the residues that they contact, Lhx4 F41 and Lhx4 K44 , also differ from those of the corresponding residues in the Lhx3-Isl1 structure (Fig. 7E). Although these differences are mostly subtle, the changes appear significant when compared with the N-terminal halves of the Isl1 LBD and Isl2 LBD , where sets of contacts and side-chain conformations are almost identical (e.g. Fig. 7, C, D, and F), even in non-conserved residues.
The essentially identical solution parameters for Lhx3-Isl2 and Lhx4-Isl2, coupled with the same patterns of binding of Isl2 LBD for Lhx3 and Lhx4 (which differ from the Isl1 LBD patterns for the same targets), imply that the binding of Isl2 LBD results in a different conformation in Lhx3 and Lhx4 than does Isl1 LBD . That is, although the mutation to alanine of residues in the LBD spacers does not appear to significantly affect binding, sequence variation in the spacers and the minor differences in the binding motifs induce different angles at the hinge/spacer. These differences might influence the formation of higher order complexes involving these protein pairs to regulate different activities in cells by creating or obscuring binding sites.
The chemical denaturation data for Lhx4 tethered complexes indicate that Lhx3-Isl1 and Lhx3-Isl2 have the same resistance to denaturation, but Lhx4-Isl2 is less resistant than Lhx4-Isl1. This difference may reflect an increased binding affinity of Lhx4⅐Isl1 over Lhx4⅐Isl2; however, thermodynamic binding data, which have thus far not been possible to obtain, would be required to establish whether this is the case. A small increase in the apparent affinity of Lhx4 (but not Lhx3) for the equivalent Isl1 residue and a decreased apparent affinity for the Isl2 residue at position Isl1 M265 /Isl2 L275 (Fig. 6B, supplemental  Fig. S4, and supplemental Table S4) implies that although Lhx3⅐Isl1 and Lhx3⅐Isl2 are interchangeable in terms of binding specificity, Lhx4 appears to have a small preference for binding Isl1 over Isl2.
Redundancy and Divergence of LIM Homeodomain Proteins-Paralogous pairs of genes are often thought to have arisen via genome duplication events. As demonstrated by their different knock-out phenotypes in mice, the pairs Lhx3/Lhx4 and Isl1/ Isl2 have diverged in some of their functions, but all four proteins are expressed in postmitotic ventral motor neurons (43). Lhx3 and Lhx4 are apparently redundant in those cells as mice embryos in which either Lhx3 or Lhx4 is knocked out have normal ventral motor neurons (5). In addition, human Lhx3 mutations that cause combined pituitary hormone deficiency syndrome do not cause motor neuron abnormalities (44). It is only when both Lhx3 and Lhx4 are knocked out that ventral motor neurons fail to develop (5). Although at least some expression of Isl2 is required for all of the motor neuron subtypes to develop, Isl2 appears to act by increasing levels of Islet activity rather than providing a separate function (9,11). Indeed, either Isl1 or Isl2 can effectively promote ectopic motor neuron differentiation when co-expressed with Lhx3, suggesting that the Islet proteins do have the same molecular function in this context (9 -12). If Lhx4 does bind Isl2 less strongly than it does Isl1, co-expression of Isl2 with Lhx4 rather than Lhx3 should show a reduction in the efficiency of motor neuron development.
It seems likely that conservation of function of Lhx3/Lhx4 and Isl1/Isl2 in motor neurons represents a co-evolutionary process; high levels of Lhx3/Lhx4 must be matched by high levels of Islet proteins to achieve the correct balance of cellspecific transcription complexes. However, the presence of four possible pairwise complexes indicates a highly redundant system that could accommodate evolutionary divergence; simultaneous disruptions to both Lhx3 and Isl1 would probably be required to severely compromise motor neuron development. Teleost fish have higher numbers of LIM homeodomain (and many other) genes than other vertebrates, most likely from additional genome duplication events. The sequences of LBDs from the Islet proteins of these fish suggest the presence of a single Isl1 paralogue, with two or more proteins that are Isl2like (or more divergent in sequence), suggesting that the duplicated genes are acquiring more diverse functions.
Divergence is also evident in apparent differences in the way that Isl1 and Isl2 mediate binding to the Lhx3 and Lhx4 targets. Solution scattering data indicate that Isl2 induces a more compact structure (smaller angle at the hinge/spacer) in Lhx3 and Lhx4 than does Isl1 (Fig. 5), probably due to differences in con- FIGURE 7. Comparison of Lhx3/4⅐Isl1/2 interaction interfaces. A and B, backbone alignment of Lhx3 (red) and Lhx4 (orange) using the LIM1 (A) and LIM2 (B) domains. The backbones of Isl1 (blue) and Isl2 (purple) are also shown. An asterisk indicates loops in Lhx3/Lhx4 that show the most variation in the backbone alignment. C and D, interaction maps indicating the residues from Lhx3-Isl1 (C) and Lhx4-Isl2 (D) that make contacts as identified by LIGPLOT (47). Gray boxes are residues of Isl1 LBD or Isl2 LBD . Colored boxes for the Lhx residues define the type of interaction as indicated; solid lines represent conserved contacts between the two complexes, and dashed lines indicate non-conserved contacts between the two complexes. Note that the numbering between Lhx3 and Lhx4 differs by four (e.g. Lhx4 F130 corresponds to Lhx3 F134 ) and by 10 between Isl1 and Isl2 (e.g. Isl1 M265 corresponds to Isl2 L275 ). E and F, the key residues in the two complexes from Lhx3 and Lhx4 LIM1 domains (E) and Lhx3 and Lhx4 LIM2 domains (F). The side chains of critical residues from Isl1 (blue) and Isl2 (purple) identified using yeast two-hybrid analysis and the residues they contact are shown (Lhx3 in red and Lhx4 in orange), as well as the backbone atoms of non-critical residues from Isl1 LBD and Isl2 LBD . Non-Isl-binding residues from Lhx4 LIM1 and Lhx4 LIM2 are shown in surface representation (white). The residues Isl1 M265 and Isl2 L275 adopt equivalent rotamers, mtp and mt, respectively. tacts between hinge and spacer residues (Fig. 7, C and D). The C-terminal half of Isl2 LBD appears to make a larger contribution to binding than the same region of Isl1. Despite these different binding patterns, Isl1 LBD and Isl2 LBD appear to have equivalent affinities for Lhx3 (Fig. 6B), suggesting that the N-terminal half of Isl2 makes a correspondingly weaker contribution to binding. However, although these complexes combine modular interactions between the LIM domains and cognate peptide sequences to result in high affinity complexes, the interaction affinities of the modules may not be additive, as suggested by data for LIM-only protein 2 (41,45). Differences in binding at the two domains could play a role in the kinetics of exchange of partners between different transcription complexes that contain LIM homeodomain and LIM-only proteins (17).
In conclusion, our studies have shown that although the pairwise interactions between the paralogous pairs of LIM homeodomain proteins Lhx3/Lhx4 and Isl1/Isl2 have many similarities that explain high levels of redundancy in motor neuron development, minor differences are evident that reflect evolutionary divergence for these proteins in other developmental pathways.