Different potential energy functions have been used in protein dynamics simulations, protein design calculations, and protein structure prediction. Clearly, the same physics applies in all three cases, so the variation in potential energy functions reflects differences in how the calculations are performed. With improvements in computer power and algorithms, the same potential energy function should be applicable to all three problems.

Here we show that a standard molecular-mechanics potential energy function without any modifications can be used to engineer protein-ligand binding. A molecular-mechanics potential is used to reconstruct the coordinates of various binding sites with an average root mean square error of 0.61 Å, and to reproduce known ligand-induced side-chain conformational shifts. Within a series of 34 mutants, the calculation can always distinguish weak ( K_d > 1 mM) and tight (K_d < 10 μM) binding sequences. Starting from partial coordinates of the ribose binding protein lacking the ligand and the ten primary contact residues, the molecular-mechanics potential is used to redesign a ribose binding site. Out of a search space of 2 × 10¹² sequences, the calculation selects a point mutant of the native protein as the top solution (experimental Kd = 17 μM), and the native protein as the second best solution (experimental K_d = 210 nM). The quality of the predictions depends on the accuracy of the generalized Born electrostatics model, treatment of protonation equilibria, high resolution rotamer sampling, a final local energy minimization step, and explicit modeling of the bound, unbound, and unfolded states.

After this initial proof of principle experiment, we next used a standard molecular mechanics potential energy function to redesign ribose binding protein to bind a series of ligands: L-arabinose, D-xylose, indole-3-acetic acid, and estradiol. The resulting proteins have 5–10 mutations from the native, are stable, the predicted structures have good hydrogen bonds and shape complementarity, and they use motifs similar to natural binding proteins. All of the designed proteins bind to their target ligands with measurable but weak affinity. The affinity was improved by random mutagenesis and screening.

The application of unmodified molecular-mechanics potentials to protein design links two fields in a mutually beneficial way. Design provides a new avenue to test molecular-mechanics energy functions, and future improvements in these energy functions will presumably lead to more accurate design results.

This is the first time a single model has been used to predict structures, binding constants, and to design new small-molecule binding sites. Using a standard model should improve the generality of protein design, which could enable the creation of custom proteins for a wide variety of applications, including sensors, enzymes, and protein therapeutics.