Surface similarity-based molecular query-retrieval

Predicting anatomic therapeutic chemical classification codes using tiered learning

BACKGROUND

The low success rate and high cost of drug discovery requires the development of new paradigms to identify molecules of therapeutic value. The Anatomical Therapeutic Chemical (ATC) Code System is a World Health Organization (WHO) proposed classification that assigns multi-level codes to compounds based on their therapeutic, pharmacological and chemical characteristics as well as the in-vivo sites(s) of activity. The ability to predict ATC codes of compounds can assist in creation of high-quality chemical libraries for drug screening and in applications such as drug repositioning. We propose a machine learning architecture called tiered learning for prediction of ATC codes that relies on the prediction results of the higher levels of the ATC code to simplify the predictions of the lower levels.

RESULTS

The proposed approach was validated using a number of compounds in both cross-validation and test setting. The validation experiments compared chemical descriptors, initialization methods and classification algorithms. The prediction accuracy obtained with tiered learning was found to be either comparable or better than that of established methods. Additionally, the experiments demonstrated the generalizability of the tiered learning architecture, in that its use was found to improve prediction rates for a majority of machine learning algorithms when compared to their stand-alone application.

CONCLUSION

The basis of our approach lies in the observation that anatomical-therapeutic biological activity of certain types typically precludes activities of many other types. Thus, there exists a characteristic distribution of the ATC codes, which can be leveraged to limit the search-space of possible codes that can be ascribed at a particular level once the codes at the preceding levels are known. Tiered learning utilizes this observation to constrain the learning space for ATC codes at a particular level based on the ATC code at higher levels. This simplifies the prediction and allows for improved accuracy.

Learning and Prediction of Complex Molecular Structure-Property Relationships

The problem of modeling and predicting complex structure-property relationships, such as the absorption, distribution, metabolism, and excretion of putative drug molecules is a fundamental one in contemporary drug discovery. An accurate model can not only be used to predict the behavior of a molecule and understand how structural variations may influence molecular property, but also to identify regions of molecular space that hold promise in context of a specific investigation. However, a variety of factors contribute to the difficulty of constructing robust structure activity models for such complex properties. These include conceptual issues related to how well the true bio-chemical property is accounted for by formulation of the specific learning strategy, algorithmic issues associated with determining the proper molecular descriptors, access to small quantities of data, possibly on tens of molecules only, due to the high cost and complexity of the experimental process, and the complex nature of bio-chemical phenomena underlying the data. This chapter attempts to address this problem from the rudiments: the authors first identify and discuss the salient computational issues that span (and complicate) structure-property modeling formulations and present a brief review of the state-of-the-art. The authors then consider a specific problem: that of modeling intestinal drug absorption, where many of the aforementioned factors play a role. In addressing them, their solution uses a novel characterization of molecular space based on the notion of surface-based molecular similarity. This is followed by identifying a statistically relevant set of molecular descriptors, which along with an appropriate machine learning technique, is used to build the structure-property model. The authors propose simultaneous use of both ratio and ordinal error-measures for model construction and validation. The applicability of the approach is demonstrated in a real world case study.

Using diffusion distances for flexible molecular shape comparison

Background

Many molecules are flexible and undergo significant shape deformation as part of their function, and yet most existing molecular shape comparison (MSC) methods treat them as rigid bodies, which may lead to incorrect shape recognition.

Results

In this paper, we present a new shape descriptor, named Diffusion Distance Shape Descriptor (DDSD), for comparing 3D shapes of flexible molecules. The diffusion distance in our work is considered as an average length of paths connecting two landmark points on the molecular shape in a sense of inner distances. The diffusion distance is robust to flexible shape deformation, in particular to topological changes, and it reflects well the molecular structure and deformation without explicit decomposition. Our DDSD is stored as a histogram which is a probability distribution of diffusion distances between all sample point pairs on the molecular surface. Finally, the problem of flexible MSC is reduced to comparison of DDSD histograms.

Conclusions

We illustrate that DDSD is insensitive to shape deformation of flexible molecules and more effective at capturing molecular structures than traditional shape descriptors. The presented algorithm is robust and does not require any prior knowledge of the flexible regions.

Large scale study of multiple-molecule queries

Background

In ligand-based screening, as well as in other chemoinformatics applications, one seeks to effectively search large repositories of molecules in order to retrieve molecules that are similar typically to a single molecule lead. However, in some case, multiple molecules from the same family are available to seed the query and search for other members of the same family.

Multiple-molecule query methods have been less studied than single-molecule query methods. Furthermore, the previous studies have relied on proprietary data and sometimes have not used proper cross-validation methods to assess the results. In contrast, here we develop and compare multiple-molecule query methods using several large publicly available data sets and background. We also create a framework based on a strict cross-validation protocol to allow unbiased benchmarking for direct comparison in future studies across several performance metrics.

Results

Fourteen different multiple-molecule query methods were defined and benchmarked using: (1) 41 publicly available data sets of related molecules with similar biological activity; and (2) publicly available background data sets consisting of up to 175,000 molecules randomly extracted from the ChemDB database and other sources. Eight of the fourteen methods were parameter free, and six of them fit one or two free parameters to the data using a careful cross-validation protocol. All the methods were assessed and compared for their ability to retrieve members of the same family against the background data set by using several performance metrics including the Area Under the Accumulation Curve (AUAC), Area Under the Curve (AUC), F1-measure, and BEDROC metrics.

Consistent with the previous literature, the best parameter-free methods are the MAX-SIM and MIN-RANK methods, which score a molecule to a family by the maximum similarity, or minimum ranking, obtained across the family. One new parameterized method introduced in this study and two previously defined methods, the Exponential Tanimoto Discriminant (ETD), the Tanimoto Power Discriminant (TPD), and the Binary Kernel Discriminant (BKD), outperform most other methods but are more complex, requiring one or two parameters to be fit to the data.

Conclusion

Fourteen methods for multiple-molecule querying of chemical databases, including novel methods, (ETD) and (TPD), are validated using publicly available data sets, standard cross-validation protocols, and established metrics. The best results are obtained with ETD, TPD, BKD, MAX-SIM, and MIN-RANK. These results can be replicated and compared with the results of future studies using data freely downloadable from http://cdb.ics.uci.edu/.

Proceedings of the 2006 International Workshop on Multiscale Biological Imaging, Data Mining and Informatics, Santa Barbara, USA (BII06)

The 2006 International Workshop on Multiscale Biological Imaging, Data Mining and Informatics was held at Santa Barbara, on Sept 7–8, 2006. Based on the presentations at the workshop, we selected and compiled this collection of research articles related to novel algorithms and enabling techniques for bio- and biomedical image analysis, mining, visualization, and biology applications.