This paper is focused on modern approaches to machine learning, most of which are as yet used infrequently or not at all in chemoinformatics. Machine learning methods are characterized in terms of the “modes of statistical inference” and “modeling levels” nomenclature and by considering different facets of the modeling with respect to input/ouput matching, data types, models duality, and models inference. Particular attention is paid to new approaches and concepts that may provide efficient solutions of common problems in chemoinformatics: improvement of predictive performance of structure–property (activity) models, generation of structures possessing desirable properties, model applicability domain, modeling of properties with functional endpoints (e.g., phase diagrams and dose–response curves), and accounting for multiple molecular species (e.g., conformers or tautomers).
J. Chem. Inf. Model., 2012, 52 (6), pp 1413–1437

Comparative assessment of nine different scoring functions (OpenEye and Tripos implementation) applied to structure-based virtual screening based on rigid docking of the pregenerated conformations library of glycogen synthase kinase 3β (GSK-3β) inhibitors has been carried out. The functions studied belong to the following types: Gaussian (Chemgauss3, Shapegauss), empirical (Chemscore, OEChemscore, Piecewise Linear Potential, Screenscore), force field-based (D_score and G_score), and potential of mean force (PMF_score). Overall enrichment of the large true inhibitors set against the set of true non-inhibitors, Directory of Useful Decoys (DUD), cyclin-dependent kinase 2 subset, and NCI Diversity Set was evaluated by means of ROC (receiver operating characteristic) method. According to this analysis, scoring function Chemscore leads to the best enrichment of the inhibitors whereas the best early enrichment of the actives may be obtained with the help of Chemgauss3 function as estimated by BEDROC (Boltzmann-enhanced discrimination of ROC) metrics.
Volume 78, Issue 3, pages 378–390, September 2011

It is proved that any molecular graph invariant (that is any topological index) can be uniquely represented as (1) a linear combination of occurrence numbers of some substructures (fragments), both connected and disconnected, or (2) a polynomial on occurrence numbers of connected substructures of corresponding molecular graph. Besides, any (0,l)-valued molecular graph invariant can be uniquely represented as a linear combination (in the terms of logic operations) of some basic (0, 1)-valued invariants indicating the presence of some substructures in the chemical structure. Thus, the occurrence numbers of substructures in a structure (or numbers indicating the presence or absence of substructures in a structure for the case of (0,l)-valued invariants) are shown to constitute the basis of invariants of labeled molecular graphs. A possibility to use these results for the mathematical justification of substructures-based methods in the “structure-property” problem is also discussed.
J. Chem. Inf. Comput. Sci., 1995, V. 35, No. 3, P. 527-531; DOI: 10.1021/ci00025a021

Two inductive knowledge transfer approaches - multitask learning (MTL) and Feature Net (FN) - have been used to build predictive neural networks (ASNN) and PLS models for 11 types of tissue-air partition coefficients (TAPC). Unlike conventional single-task learning (STL) modeling focused only on a single target property without any relations to other properties, in the framework of inductive transfer approach, the individual models are viewed as nodes in the network of interrelated models built in parallel (MTL) or sequentially (FN). It has been demonstrated that MTL and FN techniques are extremely useful in structure-property modeling on small and structurally diverse data sets, when conventional STL modeling is unable to produce any predictive model. The predictive STL individual models were obtained for 4 out of 11 TAPC whereas application of inductive knowledge transfer techniques resulted in models for 9 TAPC. Differences in prediction performances of the models as a function of the machine-learning method, and of the number of properties simultaneously involved in the learning, has been discussed.
J. Chem. Inf. Model., 2009, V. 49, No. 1, P. 133-144. DOI: 10.1021/ci8002914