Design of diverse and druglike libraries

An important advantage of performing library design in product space is the ability to optimise the properties of individual molecules within the library simultaneously with the diversity of the library. SELECT has been designed with a multi-component fitness function whereby the physicochemical property profiles of the libraries can be optimised with respect to the profile of the same property in some reference collection, for example a 'drug-like' profile as found in WDI. In the following experiments we have compared the physicochemical property profiles of diverse libraries selected by analysing reactant space with the profiles of the same physicochemical properties in libraries selected from product space that are optimised on property and diversity, simultaneously. The property profiles we have investigated are the distribution of rotatable bonds and molecular weight, although the methods are applicable to any rapidly computable molecular property. In each case, the profile is recorded in a series of 20 bins where each bin represents the percentage of compounds in the library having a given number of rotatable bonds or having molecule weight within a given range. In the case of rotatable bond profiles the bins represent the occurrence of 0, 1, 2, . . . >19 rotatable

Table 7. The RMSD between molecular weight profiles of WDI and reactant-based libraries is compared with the RMSD achieved by optimising the profile in the product libraries. Diversity is measured using Daylight fingerprints and SUMOT. The product-based libraries have much better profiles and are also more diverse than the reactant-based libraries

Library

Size

Reactants

Diversity Amw

Products

Diversity Amw

Amide

0.564 0.561

0.573 0.565

Thiazoline-2-imine

0.394 0.362

9.47

0.415 0.390

Table 8. The RMSD between rotatable bond profiles of WDI and reactant-based libraries is compared with the RMSD achieved by optimising the profile in the product libraries. Diversity is measured using Daylight fingerprints and SUM„s. The product-based libraries have better profiles and are also more diverse than the reactant-based libraries

Library

Size

Reactants

Diversity Arb

Products

Diversity Arb

Amide

0.564 0.561

0.574 0.594

Thiazoline-2-imine

0.394 0.362

0.410 0.376

bonds. In the case of molecular weight profiles the bins cover the following ranges: 0. . .49,50. . .99, . . . <950. Distributions of the properties in WDI are used as the reference distributions and SELECT attempts to minimise the root mean squared difference (RMSD) between the profile of the designed library and the reference profile.

The profiles of rotatable bonds and molecular weights were calculated for the subsets consisting of 900 compounds that were generated by selecting diverse reactants followed by enumeration, using Daylight fingerprints and SUMC0J. SELECT was then run to choose combinatorial subsets in product space that are optimised firstly on diversity and rotatable bond profile and secondly on diversity and molecular weight profile. In each case, the fitness

Figure 7. The physical property profiles of amide libraries designed using reactant-based selection (in white) are compared with libraries that are optimsed in product space (grey) and with the property profiles found in WDI (black). (a) Rotatable bond profiles; (b) molecular weight profiles.

Figure 8. The physical property profiles of thiazoline-2-imine libraries designed using re-actant-based selection (in white) are compared with libraries that are optimised in product space (grey) and with the property profiles found in WDI (black). (a) Rotatable bond profiles; (b) molecular weight profiles.

Figure 8. The physical property profiles of thiazoline-2-imine libraries designed using re-actant-based selection (in white) are compared with libraries that are optimised in product space (grey) and with the property profiles found in WDI (black). (a) Rotatable bond profiles; (b) molecular weight profiles.

function consisted of the sum of two weighted terms, the diversity term and the relevant property term. The property was included in the fitness function as the RMSD between the distribution of the property in the library represented in a chromosome and the distribution of the property in WDI, where the distributions are given as percentages. The weight assigned to diversity was 1.0 and the weight assigned to the RMSD of the property was 0.1, these weights being chosen so that the RMSD property values were approximately in the same range of values as diversity.

The results for the amide and thiazoline-2-imine libraries, shown graphically in Figures 7 and 8, respectively, demonstrate that reactant-based selection often results in libraries with poor physicochemical property profiles. By performing the analysis in product space it is possible to design libraries with optimised physicochemical property profiles. The profiles of the libraries designed in product space have much more 'WDI-like' profiles and are therefore likely to contain more compounds with the potential to be bioactive.

0 0

Post a comment