3.3. Model accuracy
We assessed the CASP14 predictions of TSP4-N by individual domains because the flexible interdomain linkers may adopt different conformations than those seen in the crystal structures. Several groups predicted the correct folds of domains XD1-XD3, with different level of accuracy. Overall, the AlphaFold2 predictions (DeepMind team, group 427) were the most accurate with respect to the XD domains. Superposition of the crystal structure and the AlphaFold2 model 1 using PyMOL43 yielded root mean squared deviation (RMSD) values for aligned Cα atoms of 0.38 Å for XD1 (90 of 100 superposed amino acid residues), 0.36 Å for XD2 (52 of 60 superposed amino acid residues) and 0.63 Å for XD3 (63 of 65 superposed amino acid residues). Several other CASP14 participants predicted the structures of these domains successfully, typically with twice or more the RMSD values. Fig. 5B shows the superposition of the XD2 domain, illustrating the remarkable similarity between the experimental and Alphafold2 structures, and also an excellent structure similarity produced by another group, even though not as good as the AlphaFold2 structure (ZhangTBM server, group 226, RMSD = 0.7-0.8 Å).
None of the structure predictions of the AD domain resembled the entire triple β-helix region. Nevertheless, the AlphaFold2 model of the AD subunit contains a meandered polypeptide chain covering residues 20-50 that resembles the β-helix trace seen in the crystal structure, with RMSD value for 28 of the 31 aligned Cα atoms of 1.7 Å. Moreover, the predicted ensuing 3-stranded antiparallel β-sheet that forms the trimeric antiparallel β-prism II (residues 51-75) is quite accurate, with RMSD value for all 25 aligned Cα atoms of 0.65 Å. In contrast, the AlphaFold2 residues 1-19 diverge from the experimental structure and the five deposited models exhibit a wide range of extended polypeptide chain orientations. Fig. 5C illustrates this by superposing only the closely related 3-stranded antiparallel β-sheet regions of the X-ray structure and the AlphaFold2 model 1. Considering that the AD triple β-helix polypeptide lacks significant amino acid sequence homology to those of known protein structures, and that there are no intra subunit interactions in triple β-helices, it is surprising that the fold calculated by the AI methods resembles at all the actual fold.
AlphaFold2 models enable solving crystal structure of Af1503 transmembrane receptor (CASP: T1100) that withstood experimental approaches for years – by MDH, RA and ANL.From email to the CASP Prediction Center: I cannot overstate my excitement at the fact that Marcus Hartmann solved the structure of Af1503 by molecular replacement with the models of group g427. Andrei Lupas
Brief description of the target
Our department has a long-standing interest in coiled coils and their role in transmembrane signal transduction. Coiled coils are bundles of α-helices with a specific regular and repetitive packing44; they are found in innumerable structural contexts in essentially all aspects of cell biology 45. While their structural and functional roles are well understood in many contexts, their role in transmembrane signal transduction is still debated. Many transmembrane receptors are homo-dimeric proteins in which a membrane-spanning coiled coil connects extracellular sensor domains to intracellular effector domains, such that signals have to be passed along the coiled-coil segment. To study this process, we have been working on the minimalistic putative receptor Af1503 fromArchaeoglobus fulgidus - fortuitously, we had already entered its genomic neighbor, Af1502, into the CASP 11 experiment46,47. Sequence analysis suggested that Af1503 forms a homo-dimer merely consisting of an extracellular PAS domain connected to an intracellular HAMP domain via an antiparallel tetrameric coiled coil. While we conducted several structural studies on the isolated HAMP domain48,49and on chimeric fusion proteins in which we fused the Af1503 HAMP domain to other coiled coil-based signaling domains50,51, we were so far unable to determine the structure of the full receptor52.
How AlphaFold2 models helped to solve the structure
Our problems in obtaining the structure of the full receptor did not lie in the behavior of the protein. The protein was very well behaved, stable, and readily crystallized in a range of conditions. However, crystal quality was very erratic, could not be improved systematically, and diffraction was generally strongly anisotropic and not to high resolution. This led to the failure of experimental phasing approaches, despite several different strategies employed. On the other hand, molecular replacement (MR) was not successful, as we only had the structure of the HAMP domain as an available search model, and as the approach was further complicated by the presence of translational non-crystallographic symmetry. To aid MR, we decided to tackle a truncated construct covering the extracellular PAS domain, but this construct failed to crystallize. In contrast, we succeeded with an NMR analysis of this construct, revealing the fold of the PAS domain, but the structural models derived from the NMR data were too far from the actual crystal structure to succeed in MR attempts.
Finally, years later, we easily managed to solve the crystal structure using the AlphaFold2 models. As the predictions were modeled as monomers, without constraints for the homo-dimeric state, they were not fully compatible with the dimeric state along the whole chain, and a very first, naive MR attempt employing a full model did not succeed. However, in the second attempt, only employing a single PAS domain with a short coiled-coil segment as a search model, the structure was essentially solved using MOLREP with standard parameters53; after the correct placement and initial refinement of the PAS domains, the electron density for the rest of the protein was clearly traceable.