Word cloud summary of diverse topics associated with QSAR modeling that are discussed in this review.

The aim of virtual screening (VS) is to identify bioactive compounds through computational means, by employing knowledge about the protein target (structure-based VS) or known bioactive ligands (ligand-based VS). In VS, a large number of molecules are ranked according to their likelihood to be bioactive compounds, with the aim to enrich the top fraction of the resulting list (which can be tested in bioassays afterward). At its core, VS attempts to improve the odds of identifying bioactive molecules by maximizing the true positive rate, that is, by ranking the truly active molecules as high as possible (and, correspondingly, the truly inactive ones as low as possible). In choosing the right approach, the researcher is faced with many questions: where does the optimal balance between efficiency and accuracy lie when evaluating a particular algorithm; do some methods perform better than others and in what particular situations; and what do retrospective results tell us about the prospective utility of a particular method? Given the multitude of settings, parameters, and data sets the practitioner can choose from, there are many pitfalls that lurk along the way which might render VS less efficient or downright useless. This review attempts to catalogue published and unpublished problems, shortcomings, failures, and technical traps of VS methods with the aim to avoid pitfalls by making the user aware of them in the first place.

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.

Background and Purpose Because robotic devices record the kinematics and kinetics of human movements with high resolution, we hypothesized that robotic measures collected longitudinally in patients after stroke would bear a significant relationship to standard clinical outcome measures and, therefore, might provide superior biomarkers. Methods In patients with moderate-to-severe acute ischemic stroke, we used clinical scales and robotic devices to measure arm movement 7, 14, 21, 30, and 90 days after the event at 2 clinical sites. The robots are interactive devices that measure speed, position, and force so that calculated kinematic and kinetic parameters could be compared with clinical assessments. Results Among 208 patients, robotic measures predicted well the clinical measures (cross-validated R2 of modified Rankin scale=0.60; National Institutes of Health Stroke Scale=0.63; Fugl-Meyer=0.73; Motor Power=0.75). When suitably scaled and combined by an artificial neural network, the robotic measures demonstrated greater sensitivity in measuring the recovery of patients from day 7 to day 90 (increased standardized effect=1.47). Conclusions These results demonstrate that robotic measures of motor performance will more than adequately capture outcome, and the altered effect size will reduce the required sample size. Reducing sample size will likely improve study efficiency.