The fixed backbone simplifies the size of the design problem by reducing it to a choice of residue types and orientations along a rigid frame. It also removes the reliance on the design force field to discriminate between competing backbone structures. Since core designs start with a wild-type protein template, the choice of backbone position is obvious and likely to be reasonable, as nature has already evolved toward it. This assumption breaks down, however, when the design template is not a natural protein and the initial backbone position may be suboptimal or impossible. It is also problematic when stabilizing residue choices require backbone shifts in order to pack into the protein core, which tends to be the rule rather than the exception.
One of the earliest attempts to incorporate backbone flexibility into computational design was the work of Harbury and coworkers on coiled coils (Harbury et al. 1993). This system has a high degree of symmetry and simple topology that allowed the use of algebraic parameterization to predict an ensemble of backbone conformations. A fixed number of backbone orientations and side-chain identities were searched. Dimer, trimer, and tetramer configurations were designed and characterized via circular dichroism, sedimentation equilibrium, and, in the case of the tetramer, crystallography. Each of the designs adopted the expected oligomerization state and the tetramer differed from the predicted structure by only 0.2 A. Though these designs were extraordinarily successful, this is a rare case in protein design, as most natural protein topologies are too complex to be parameterized in this fashion.
The first attempt at backbone flexibility in a more general protein design algorithm was the SoftROC program by Desjarlais and Handel (Desjarlais and Handel 1999). The intent of their work was to address the problem of compensating for unfavorable core mutations by relaxation of the backbone. Starting from a known crystallographi-cally determined structure, backbone torsion angles were randomly adjusted up to 3° from their original positions to generate a population of conformations (Figure 15.4).
Though the movement of each torsion angles was small, the compounded effect over many bonds led to a wide variety of backbones that preserved the overall topology of the protein.
SoftROC was tested on the bacteriophage protein 434 cro by forcing a disruptive point mutation into the core and having the program design stabilizing mutations to compensate for it (Desjarlais and Handel 1999). When compared to the results of its fixed-backbone counterpart, SoftROC produced new sequences that folded cooperatively and shared similar stability to the wild-type 434 cro protein. This suggested that there were false negatives in the context of the fixed backbone approach and additional solutions become available when a flexible backbone is permitted.
The real goal of flexible backbone design, however, was realized when Kuhlman and coworkers designed Top7, an a/p protein with a novel topology not yet observed in nature (Kuhlman et al. 2003). The initial backbone of the desired fold was built from secondary structure elements derived from existing structures and optimized by iterative rounds of sequence and backbone structure minimization. The predicted structure was verified by crystallography and deviated by only 1.17A from the actual coordinates while bearing no significant resemblance to any protein in the PDB (Figure 15.5. Also see color Figure 11.2).
The success of the Top7 design is promising for the potential of structural design, but it remains the only example of a fold being designed de novo via computational means. From the perspective of design being a validation of our knowledge of the forces involved in protein folding, there is also room for improvement. The current success requires knowledge-based potentials that use existing structural data to bias design results. Being able to achieve similar results using an ab initio potential
would solidify our understanding of protein structure and likely aid in our ability to predict folds from sequences.
Was this article helpful?