Many apologies for the interval between updates. And, thank you for your patience and computer cycles as we move into the final stages of this challenging project.

Phase 2 has been fully optimized, tested, and launched. The initial results of our large-scale free energy of binding calculations, launched during the past year on test systems, have supported our expectations that these calculations would provide improved enrichment rates and reduced false positive rates relative to virtual screening based on traditional docking programs (in this case, AutoDock). More details will be provided below.

While we continue to benchmark test cases (based on the systems developed by Shoichet and Irwin laboratories [http://dud.docking.org/ ]) to better understand the strengths and limitations of our phased approach, we have launched Phase 2 production runs for the dengue and West Nile virus proteases, and are readying hepatitis C virus protease for Phase 2 free energy calculations. Each Phase 2 production run is expected to take ~4 months of wall-clock time (based on the turn-around times we are seeing with the test systems).

For completeness, we finished laboratory testing (i.e., using protease inhibition assays) of ~50 compounds predicted from the Phase 1 simulations to be potential inhibitors of the dengue and West Nile proteases. Not surprisingly (actually, quite expected), none of the tested Phase 1 “hits” showed activities that were more promising than our already discovered dengue and West Nile protease inhibitors (see, for instance, Tomlinson and Watowich, Anthracene-based inhibitors of dengue virus NS2B-NS3 protease. 2011. Antiviral Res. 89, 127-35). It is the Phase 2 simulations that are expected to find compounds that will be good protease inhibitors (and, thus, good antivirals).

For those interested, here are some details about the initial tests that were done to validate our ongoing free energy calculations. A small-scale virtual screening experiment, using a collection of 16 binders and 14 non-binders to the lysozyme protein, showed that free energy of binding calculations could accurately discriminate between binders and non-binders. The success rates of the free energy of binding predictions closely followed an ideal curve, and were a marked improvement over the Autodock predictions. These results demonstrated that our two phase approach could reduce false positives in virtual screening experiments.

A recently completed medium-scale virtual screening experiment (using our Phase 1 and Phase 2 strategy) examined the estrogen receptor agonist system available from Shoichet and Irwin’s “Directory of Useful Decoys” (DUD ). This test system contained ~70 known small molecules binders and ~2,300 chemically similar ligands that were assumed to be non-binders (i.e., decoys). From a practical point of view, we are most interested in the early success rate of virtual screening calculations, since this will dictate our success at validating predicted inhibitors in the laboratory. Because of time and budget constraints, it is only practical to test a few dozen to a few hundred of the best computer predictions. Thus, it is important that the computer’s best predictions correlate strongly to true binders. We compared the early success rates for Autodock and free energy calculations. On average, the lowest energy Autodock-predicted “hits” had an success rate of ~15% while the lowest free-energy-predicted “hits” had a success rate of ~40%. The success rates varied as a function of number of top predicted “hits” that were examined, but typically the free energy calculations identified ~3 times more binders than the Autodock calculations.

We are investigating methods to further improve the success rate of the free energy calculations, since we would like early success rates that exceed 80%. Moreover, we are completing Phase 2 calculations on additional DUD test systems to determine if the observed free energy-based improvements to the success rate are applicable to different target proteins. This is necessary, since virtual screening success rates vary dramatically as a function of the target protein. If one is interested, within the free virtual screening portal that we constructed we have posted results obtained from running several virtual screening programs against the 40 protein targets (and libraries) that are part of the DUD database.

Finally, we are honored to have been awarded a grant from the IBM International Foundation to support expanded testing of compounds predicted by World Community Grid calculations. The IBM International Foundation grant is based in part on the cash prize received from the recent Jeopardy! win by the IBM’s Watson supercomputer.

Thanks,
Stan