## Scoring results with different reference states

Figure 3 shows derived PMF using reference states Ref l and Ref3. The curves appear to be very similar, Refl generates slightly more negative potentials than Ref3 does. The scoring conditions are chosen as in Muegge and Martin [27], that is, carbon-carbon interactions are scored with a 6 A, cut-off radius and all other interactions with a 9 A, cut-offradius. The PMF are derived using a reference sphere radius of 12 A Table 2 shows the correlation between PMF

Table 2. Correlation between PMF score and measured binding affinity for four test setsa

No. Test set

Serine protease

2 Metalloprotease

2a Metalloprotease w/o 1 outlier

Diverse setl

Diverse set2

2a Metalloprotease w/o 1 outlier

No. of |
Reference |
R2 |
SD |

complexes |
state | ||

16 |
Refl |
0.91 |
0.79 |

Ref2 |
0.86 |
1.10 | |

Ref3 |
0.91 |
0.77 | |

15 |
Refl |
0.59 |
1.84 |

Ref2 |
0.55 |
1.88 | |

Ref3 |
0.53 |
1.94 | |

14 |
Refl |
0.77 |
1.39 |

Ref2 |
0.61 |
1.82 | |

Ref3 |
0.76 |
1.42 | |

17 |
Refl |
0.68 |
1.55 |

Ref2 |
0.42 |
2.00 | |

Ref3 |
0.52 |
1.75 | |

63 |
Refl |
0.54 |
1.80 |

Ref2 |
0.45 |
1.95 | |

Ref3 |
0.39 |
2.07 |

a The standard deviations (SD) between the binding affinities and a linear regression line, calculated by using M-estimates [42], are given in log Ki values.

scores and binding affinities for the four test sets described above. Figures 47 show the correlation between the PMF scores and the binding affinities for the four test sets and the three different reference states. From Table 2 one can see that PMF/Refl generally performs best. For test sets of protein-ligand complexes from the same protein class (sets 1 and 2) the differences between the reference states are only small. In set 1, Ref2 leads to an inability of distinguishing between low binding complexes. The most notable difference between Ref2 and Refl/3 in set 2 is that the outlier, the hydroxamate binding complex 1mnc, disappears in Ref2. However, the overall correlation between PMF score and log K is significant for all cases, particularly after removing the 1mnc outlier in set 2. Larger differences are found in test sets 3 and 4. Set 3 shows significant correlation only for PMF/Refl and PMF/Ref3. PMF/Ref2 performs worst. However, for a larger set of diverse protein-ligand complexes (set 4), Ref2 performs better than Ref3. Here, only PMF/Refl leads to a significant correlation between experiment and calculated score. The most

notable shortcoming of PMF/Ref2 is the finding that compounds with the highest binding affinities are no longer given the best scores (Figure 7). The results suggest that a reference state that is specific for the set of protein-ligand complexes used to derive the PMF and that most comprehensively accounts for solvation effects should be used to derive the PMF. This is the more important the more diverse the protein-ligand complexes are that are scored by the PMF scoring function.

Figure 6. PMF score of 17 diverse protein-ligand complexes (set 3) as function ofobserved binding affinities calculated by using Refl Ref2 and Ref3 respectively.

## Post a comment