On the use of neural network ensembles in QSAR and QSPR

Despite their growing popularity among neural network practitioners, ensemble methods have not been widely adopted in structure-activity and structure-property correlation. Neural networks are inherently unstable, in that small changes in the training set and/or training parameters can lead to large changes in their generalization performance. Recent research has shown that by capitalizing on the diversity of the individual models, ensemble techniques can minimize uncertainty and produce more stable and accurate predictors. In this work, we present a critical assessment of the most common ensemble technique known as bootstrap aggregation, or bagging, as applied to QSAR and QSPR. Although aggregation does offer definitive advantages, we demonstrate that bagging may not be the best possible choice and that simpler techniques such as retraining with the full sample can often produce superior results. These findings are rationalized using Krogh and Vedelsby's decomposition of the generalization error into a term that measures the average generalization performance of the individual networks and a term that measures the diversity among them. For networks that are designed to resist over-fitting, the benefits of aggregation are clear but not overwhelming.     

Solvolysis of para-substituted cumyl chlorides. Brown and Okamoto's electrophilic substituent constants revisited using continuum solvent models

Brown and Okamoto (J. Am. Chem. Soc. 1958, 80, 4979) derived their electrophilic substitutent constants, sigma(p)+, from the relative rates of solvolysis of ring-substituted cumyl chlorides in an acetone/water solvent mixture. Application of the Hammett equation to the rates for the meta-substituted cumyl chlorides, where there could be no resonance interaction with the developing carbocation, gave a slope, rho(+) = -4.54 ( identical with 6.2 kcal/mol free energy). Rates for the para-substituted chlorides were then used to obtain sigma(p)+ values. We have calculated gas-phase C-Cl heterolytic bond dissociation enthalpy differences, Delta BDE(het) (= BDE(het)(4-YC(6)H(4)CMe(2)Cl) - BDE(het)(C(6)H(5)CMe(2)Cl)), for 16 of the 4-Y substituents employed by Brown and Okamoto. The plot of Delta BDE(het) vs sigma(p)+ gave rho(+) (SD) = 16.3 (2.3) kcal/mol, i.e., a rho(+) value roughly 2.5 times greater than experiment. Inclusion of solvation (water) energies, calculated using three continuum solvent models, reduced rho(+) and SD. The computationally least expensive model used, SM5.42R (Li et al. Theor. Chem. Acc. 1999, 103, 9) gave the best agreement with experiment. This model yielded rho(+) (SD) = 7.7 (0.9) kcal/mol, i.e., a rho(+) value that is only 24% larger than experiment.     

QSPR modeling of decomposition temperature of energetic cocrystals using artificial neural network

Journal of Thermal Analysis and Calorimetry - The quantitative structure–property relationship for the decomposition temperature (Td) of energetic cocrystals was investigated. The artificial...     

Taft substituent constants for arene chromium dicarbonyl phosphine and arsine derivatives

A series of biphenyl, meta- and para-fluorobiphenyl complexes of chromiumtricarbonyl, chromiumdicarbonyltriphenylphosphine and chromiumdicarbonyltriphenylarsine have been prepared and their spectral and physical properties recorded. Taft substituent constants have been calculated for the coordinated arenes and the consequences of substituting a Group V ligand for a carbonyl are discussed.     

QSPR study for the prediction of half-wave potentials of benzoxazines by heuristic method and radial basis function neural network

The half-wave potential (E_1/2) is an important electrochemical property of organic compounds. In this work, a quantitative structure-property relationship (QSPR) analysis has been conducted on the half-wave reduction potential (E_1/2) of 40 substituted benzoxazines by means of both a heuristic method (HM) and a non-linear radial basis function neural network (RBFNN) modeling method. The statistical parameters provided by the HM model (R² =0.946; F=152.576; RMSCV=0.0141) and the RBFNN model (R²=0.982; F=1034.171 and RMS =0.0209) indicated satisfactory stability and predictive ability. The obtained models showed that benzoxazines with larger Min valency of a S atom (MVSA), lower Relative number of H atom (RNHA) and Min n-n repulsion for a C-H bond (MnnRCHB) and Minimal Electrophilic Reactivity Index for a C atom (MERICA) can be more easily reduced. This QSPR approach can contribute to a better understanding of structural factors of the organic compounds that contribute to the E_1/2, and can be useful in predicting the E_1/2 of other compounds.      

Comparative studies on some metrics for external validation of QSPR models

Quantitative structure-property relationship (QSPR) models used for prediction of property of untested chemicals can be utilized for prioritization plan of synthesis and experimental testing of new compounds. Validation of QSPR models plays a crucial role for judgment of the reliability of predictions of such models. In the QSPR literature, serious attention is now given to external validation for checking reliability of QSPR models, and predictive quality is in the most cases judged based on the quality of predictions of property of a single test set as reflected in one or more external validation metrics. Here, we have shown that a single QSPR model may show a variable degree of prediction quality as reflected in some variants of external validation metrics like Q2(F1), Q2(F2), Q2(F3), CCC, and r2(m) (all of which are differently modified forms of predicted variance, which theoretically may attain a maximum value of 1), depending on the test set composition and test set size. Thus, this report questions the appropriateness of the common practice of the "classic" approach of external validation based on a single test set and thereby derives a conclusion about predictive quality of a model on the basis of a particular validation metric. The present work further demonstrates that among the considered external validation metrics, r2(m) shows statistically significantly different numerical values from others among which CCC is the most optimistic or less stringent. Furthermore, at a given level of threshold value of acceptance for external validation metrics, r2(m) provides the most stringent criterion (especially with Δr2(m) at highest tolerated value of 0.2) of external validation, which may be adopted in the case of regulatory decision support processes.     

Correlation of ESCA shifts and Hammett substituent constants in substituted benzene derivatives

ESCA shifts of substituted benzene derivatives are shown to correlate with Hammett σ constants. An explanation of this correlation is given in terms of the thermodynamic model for core electron binding energy shifts, and it is shown that ESCA shifts in benzene derivatives can be treated as a linear free energy relationship.     

Accurate calculation of N1s and C1s core electron binding energies of substituted pyridines. Correlation with basicity and with Hammett substituent constants

Substituent shifts of the energetics of four related ionization processes of pyridines and benzoic acids (Fig. 1) were investigated. The first process is core-electron ionization of gas-phase pyridines (Fig. 1A), while the second concerns gas-phase acid-base reaction between a substituted pyridine and a conjugated acid (Fig. 1B), and the third and fourth processes are the acid dissociation of substituted benzoic acids in aqueous solution (Fig. 1C) and in vacuum (Fig. 1D), respectively. Core-electron binding energies for the first process were calculated using density-functional theory with the scheme ΔE_KS (PW86x-PW91c/TZP+C_rel)//HF/6-31G*. Average absolute deviation of calculated core electron binding energy shifts at N atom in substituted pyridines from experiment was 0.08 eV. The shift at N coincides highly with that at a ring carbon atom. The four shifts corresponding to the four processes shown in Figs. 1A–D correlate strongly with one another, with numerical values fairly close to each other when expressed in unit of electron volts.