Reliable and Performant Identification of Low-Energy Conformers in the Gas Phase and Water.

Prediction of compound properties from structure via quantitative structure-activity relationship and machine-learning approaches is an important computational chemistry task in small-molecule drug research. Though many such properties are dependent on three-dimensional structures or even conformer ensembles, the majority of models are based on descriptors derived from two-dimensional structures. Here we present results from a thorough benchmark study of force field, semiempirical, and density functional methods for the calculation of conformer energies in the gas phase and water solvation as a foundation for the correct identification of relevant low-energy conformers. We find that the tight-binding ansatz GFN-xTB shows the lowest error metrics and highest correlation to the benchmark PBE0-D3(BJ)/def2-TZVP in the gas phase for the computationally fast methods and that in solvent OPLS3 becomes comparable in performance. MMFF94, AM1, and DFTB+ perform worse, whereas the performance-optimized but far more expensive functional PBEh-3c yields energies almost perfectly correlated to the benchmark and should be used whenever affordable. On the basis of our findings, we have implemented a reliable and fast protocol for the identification of low-energy conformers of drug-like molecules in water that can be used for the quantification of strain energy and entropy contributions to target binding as well as for the derivation of conformer-ensemble-dependent molecular descriptors.

Prediction errors in learning drug response from gene expression data - influence of labeling, sample size, and machine learning algorithm.

Model-based prediction is dependent on many choices ranging from the sample collection and prediction endpoint to the choice of algorithm and its parameters. Here we studied the effects of such choices, exemplified by predicting sensitivity (as IC50) of cancer cell lines towards a variety of compounds. For this, we used three independent sample collections and applied several machine learning algorithms for predicting a variety of endpoints for drug response. We compared all possible models for combinations of sample collections, algorithm, drug, and labeling to an identically generated null model. The predictability of treatment effects varies among compounds, i.e. response could be predicted for some but not for all. The choice of sample collection plays a major role towards lowering the prediction error, as does sample size. However, we found that no algorithm was able to consistently outperform the other and there was no significant difference between regression and two- or three class predictors in this experimental setting. These results indicate that response-modeling projects should direct efforts mainly towards sample collection and data quality, rather than method adjustment.

Using expression data for quantification of active processes in physiologically based pharmacokinetic modeling.

Active processes involved in drug metabolization and distribution mediated by enzymes, transporters, or binding partners mostly occur simultaneously in various organs. However, a quantitative description of active processes is difficult because of limited experimental accessibility of tissue-specific protein activity in vivo. In this work, we present a novel approach to estimate in vivo activity of such enzymes or transporters that have an influence on drug pharmacokinetics. Tissue-specific mRNA expression is used as a surrogate for protein abundance and activity and is integrated into physiologically based pharmacokinetic (PBPK) models that already represent detailed anatomical and physiological information. The new approach was evaluated using three publicly available databases: whole-genome expression microarrays from ArrayExpress, reverse transcription-polymerase chain reaction-derived gene expression estimates collected from the literature, and expressed sequence tags from UniGene. Expression data were preprocessed and stored in a customized database that was then used to build PBPK models for pravastatin in humans. These models represented drug uptake by organic anion-transporting polypeptide 1B1 and organic anion transporter 3, active efflux by multidrug resistance protein 2, and metabolization by sulfotransferases in liver, kidney, and/or intestine. Benchmarking of PBPK models based on gene expression data against alternative models with either a less complex model structure or randomly assigned gene expression values clearly demonstrated the superior model performance of the former. Besides accurate prediction of drug pharmacokinetics, integration of relative gene expression data in PBPK models offers the unique possibility to simultaneously investigate drug-drug interactions in all relevant organs because of the physiological representation of protein-mediated processes.

Quantifying stability in gene list ranking across microarray derived clinical biomarkers.

BACKGROUND: Identifying stable gene lists for diagnosis, prognosis prediction, and treatment guidance of tumors remains a major challenge in cancer research. Microarrays measuring differential gene expression are widely used and should be versatile predictors of disease and other phenotypic data. However, gene expression profile studies and predictive biomarkers are often of low power, requiring numerous samples for a sound statistic, or vary between studies. Given the inconsistency of results across similar studies, methods that identify robust biomarkers from microarray data are needed to relay true biological information. Here we present a method to demonstrate that gene list stability and predictive power depends not only on the size of studies, but also on the clinical phenotype. RESULTS: Our method projects genomic tumor expression data to a lower dimensional space representing the main variation in the data. Some information regarding the phenotype resides in this low dimensional space, while some information resides in the residuum. We then introduce an information ratio (IR) as a metric defined by the partition between projected and residual space. Upon grouping phenotypes such as tumor tissue, histological grades, relapse, or aging, we show that higher IR values correlated with phenotypes that yield less robust biomarkers whereas lower IR values showed higher transferability across studies. Our results indicate that the IR is correlated with predictive accuracy. When tested across different published datasets, the IR can identify information-rich data characterizing clinical phenotypes and stable biomarkers. CONCLUSIONS: The IR presents a quantitative metric to estimate the information content of gene expression data with respect to particular phenotypes.