Are AM1 ligand-protein binding enthalpies good enough for use in the rational design of new drugs?

We have examined the performance of semiempirical quantum mechanical methods in solving the problem of accurately predicting protein-ligand binding energies and geometries. Firstly, AM1 and PM3 geometries and binding enthalpies between small molecules that simulate typical ligand-protein interactions were compared with high level quantum mechanical techniques that include electronic correlation (e.g., MP2 or B3LYP). Species studied include alkanes, aromatic systems, molecules including groups with hypervalent sulfur or with donor or acceptor hydrogen bonding capability, as well as ammonium or carboxylate ions. B3LYP/6-311+G(2d,p) binding energies correlated very well with the BSSE corrected MP2/6-31G(d) values. AM1 binding enthalpies also showed good correlation with MP2 values, and their systematic deviation is acceptable when enthalpies are used for the comparison of interaction energies between ligands and a target. PM3 otherwise gave erratic energy differences in comparison to the B3LYP or MP2 approaches. As one would expect, the geometries of the binding complexes showed the known limitations of the semiempirical and DFT methods. AM1 calculations were subsequently applied to a test set consisting of "real" protein active site-ligand complexes. Preliminary results indicate that AM1 could be a valuable tool for the design of new drugs using proteins as templates. This approach also has a reasonable computational cost. The ligand-protein X-ray structures were reasonably reproduced by AM1 calculations and the corresponding AM1 binding enthalpies are in agreement with the results from the "small molecules" test set.