V.
LogD Comparisons, continued
Figures 2(a) and (b) show the correlation between theory and experiment
graphically for the generic set of parameters and the user-trained set. It is evident that user training yields a
narrower distribution that is closer to the y = x ideal agreement.
Figure 2. Comparison of logDexp vs logDcalc calculated
by the ACD/Labs method using A. generic parameters and B. user-trained parameters for 18 compounds from Ref.[2].
A similar situation holds for the other case of logD training.
Table 5 summarizes the logD calculations for the other set of compounds, the 2,3,4-substituted analogs of compound (VI).
The calculated logD value has an average discrepancy from experiment of 1.39 logD units.
The presence of fused ring heterocycles may be related to the unusually high error in these calculations.
This case becomes even more of a challenge for the system training, and in fact its effect is even more striking
than in the previous example: the average discrepancy has been reduced to 0.33 logD
units.
Table 5. Experimental and calculated logD values at pH = 7.4 for
substituted analogs of structure (VI).
| N |
-R1 |
-R2 |
-R3 |
logDexpa) |
logDcalcb) |
ΔLogD |
logDcalcc)
(after
training) |
ΔLogDtr |
| 1 |
-H |
-H |
-H |
-0.97 |
0.15 |
-1.12 |
-0.97 (in
training) |
0.00 |
| 2 |
-Me |
-H |
-H |
-0.83 |
0.61 |
-1.44 |
-0.51 |
-0.32 |
| 3 |
-F |
-H |
-H |
-1.19 |
0.00 |
-1.19 |
-0.12 |
-0.07 |
| 4 |
-H |
-H |
-Me |
-0.73 |
0.61 |
-1.34 |
-0.51 |
-0.22 |
| 5 |
-H |
-H |
-F |
-1.18 |
0.44 |
-1.62 |
-0.68 |
-0.50 |
| 6 |
-Me |
-H |
-Me |
-0.07 |
1.07 |
-1.14 |
-0.05 |
-0.02 |
| 7 |
-Me |
-H |
-F |
-0.43 |
0.90 |
-1.33 |
-0.22 |
-0.21 |
| 8 |
-F |
-H |
-F |
-0.48 |
0.33 |
-0.81 |
-0.79 |
0.31 |
| 9 |
-Cl |
-H |
-Cl |
-0.55 |
1.40 |
-1.95 |
0.27 |
-0.82 |
| 10 |
-F |
-Me |
-F |
-1.12 |
0.79 |
-1.91 |
-0.33 |
-0.79 |
| Average Δ |
1.39 |
|
0.33 |
a) Experimental Data from Reference 3.
b) Calculated values obtained without training.
c) Calculated values obtained after training by structure (VIII).
|
