Stochastic versus Stepwise Strategies for Quantitative Structure−Activity Relationship GenerationHow Much Effort May the Mining for Successful QSAR Models Take?

Abstract

Descriptor selection in QSAR typically relies on a set of upfront working hypotheses in order to boil down the initial descriptor set to a tractable size. Stepwise regression, computationally cheap and therefore widely used in spite of its potential caveats, is most aggressive in reducing the effectively explored problem space by adopting a greedy variable pick strategy. This work explores an antipodal approach, incarnated by an original Genetic Algorithm (GA)-based Stochastic QSAR Sampler (SQS) that favors unbiased model search over computational cost. Independent of a priori descriptor filtering and, most important, not limited to linear models only, it was benchmarked against the ISIDA Stepwise Regression (SR) tool. SQS was run under various premises, varying the training/validation set splitting scheme, the nonlinearity policy, and the used descriptors. With the considered three anti-HIV compound sets, repeated SQS runs generate sometimes poorly overlapping but nevertheless equally well validating model sets. Enabling SQS to apply nonlinear descriptor transformations increases the problem space: nevertheless, nonlinear models tend to be more robust validators. Model validation benchmarking showed SQS to match the performance of SR or outperform it in cases when the upfront simplifications of SR "backfire", even though the robust SR got trapped in local minima only once in six cases. Consensus models from large SQS model sets validate well--but not outstandingly better than SR consensus equations. SQS is thus a robust QSAR building tool according to standard validation tests against external sets of compounds (of same families as used for training), but many of its benefits/drawbacks may yet not be revealed by such tests. SQS results are a challenge to the traditional way to interpret and exploit QSAR: how to deal with thousands of well validating models, nonetheless providing potentially diverging applicability ranges and predicted values for external compounds. SR does not impose such burden on the user, but is "betting" on a single equation or a narrow consensus model to behave properly in virtual screening a sound strategy? By posing these questions, this article will hopefully act as an incentive for the long-haul studies needed to get them answered.

Fuzzy Tricentric Pharmacophore Fingerprints. 1. Topological Fuzzy Pharmacophore Triplets and Adapted Molecular Similarity Scoring Schemes

“…Nonlinear approaches occupy the top position of the best validating model in eight out of 13 cases, an observation reinforcing the already reported 25 trend of improving model validation propensity upon allowing SQS to employ predefined nonlinear transformations.…”

“…In principle, due to the stochastic nature of the model builder, it is risky to extrapolate the intrinsic quality of descriptors from validation score differences of single models. However, the analysis of the 39 duplicated SQS simulations for all 13 data sets, performed with D, D-R, and D-S descriptors, respectively, showed that duplicate simulations generate significantly diverging representative sets 25 a 1 -Validation criteria for the globally optimal, best validating 2D-FPT models: RMSPE (plain text), validation set R 2 V (italics), and percentage of correctly classified inactives (bold) in an additional inactive validation set, except for (a), reporting the overall correct classification rate of both validation set actives and inactives. 2 -2D-FPT setup and nonlinearity policy (linear, nonlinear) leading to results (1).…”

“…Linear simulations with D, D-R, and D-S descriptors have been duplicated, and representative model pools were merged in order to allow the estimation of the average validation correlation coefficient shifts attributable to switching from one descriptor variant to the other, according to a previously outlined formalism 25 (also see below). Eventually, a second round of SQS simulations mining for fully nonlinear models was also carried out for all compound sets using all descriptor versions, leading to a total of 5 × 13 (first round, linear) + 3 × 13 (linear model generation duplicates: D, D-R, and D-S descriptors only) + 5 × 13 (nonlinear) ) 169 different SQS runs using various Linux and IRIX workstations of the laboratory.…”

“…These results reinforce the similar trend reported in earlier work. Table 1 of the previous publication 25 for more details about the predefined nonlinear transformations in SQS.] Given the standard 25 average <HD4-Hp6-PC4> and variance σ(HD4-Hp6-PC4) population levels of 1.2 and 6.4, respectively, and knowing that HD4-Hp6-PC4 is not populated in either of the (b) and (c) molecules from Figure 8, the absence of such a triplet contributes 11.1×zexp3(0))+10 to predicted pK i values.…”

“…Duplicate SQS runs were performed with both default 2D-FPT, the rule-based version D-R, and the nonfuzzy version D-S, in order to generate extended representative model sets, allowing a comparison of average validation propensities according to a previously described approach. 25 Comparison of QSARs based on triplet versions D, O, and C relies on the validation statistics of best validating models, like in literature model benchmarking.…”