Emerging Chemical Patterns for Virtual Screening and Knowledge Discovery

Abstract

The adaptation and evaluation of contemporary data mining methods to chemical and biological problems is one of major areas of research in chemoinformatics. Currently, large databases containing millions of small organic compounds are publicly available, and the need for advanced methods to analyze these data increases. Most methods used in chemoinformatics, e.g. quantitative structure activity relationship (QSAR) modeling, decision trees and similarity searching, depend on the availability of large high-quality training data sets. However, in biological settings, the availability of these training sets is rather limited. This is especially true for early stages of drug discovery projects where typically only few active molecules are available. The ability of chemoinformatic methods to generalize from small training sets and accurately predict compound properties such as activity, ADME or toxicity is thus crucially important. Additionally, biological data such as results from high-throughput screening (HTS) campaigns is heavily biased towards inactive compounds. This bias presents an additional challenge for the adaptation of data mining methods and distinguishes chemoinformatics data from the standard benchmark scenarios in the data mining community.<br /> Even if a highly accurate classifier would be available, it is still necessary to evaluate the predictions experimentally. These experiments are both costly and time-consuming and the need to optimize resources has driven the development of integrated screening protocols which try to minimize experimental efforts but still reaching high hit rates of active compounds. This integration, termed “sequential screening” benefits from the complementary nature of experimental HTS and computational virtual screening (VS) methods.<br /> In this thesis, a current data mining framework based on class-specific nominal combinations of attributes (emerging patterns) is adapted to chemoinformatic problems and thoroughly evaluated. Combining emerging pattern methodology and the well-known notion of chemical descriptors, emerging chemical patterns (ECP) are defined as class- specific descriptor value range combinations. Each pattern can be thought of as a region in chemical space which is dominated by compounds from one class only. Based on chemical patterns, several experiments are presented which evaluate the performance of pattern-based knowledge mining, property prediction, compound ranking and sequential screening. ECP-based classification is implemented and evaluated on four activity classes for the prediction of compound potency levels. Compared to decision trees and a Bayesian binary QSAR method, ECP-based classification produces high accuracy in positive and negative classes even on the basis of very small training set, a result especially valuable to chemoinformatic problems.<br /> The simple nature of ECPs as class-specific descriptor value range combinations makes them easily interpretable. This is used to related ECPs to changes in the interaction network of protein-ligand complexes when the binding conformation is replaced by a computer-modeled conformation in a knowledge mining experiment. ECPs capture well-known energetic differences between binding and energy-minimized conformations and additionally present new insight into these differences on a class level analysis.<br /> Finally, the integration of ECPs and HTS is evaluated in simulated lead-optimization and sequential screening experiments. The high accuracy on very small training sets is exploited to design an iterative simulated lead optimization experiment based on experimental evaluation of randomly selected small training sets. In each iteration, all compounds predicted to be weakly active are removed and the remaining compound set is enriched with highly potent compounds. On this basis, a simulated sequential screening experiment shows that ECP-based ranking recovers 19% of available compounds while reducing the “experimental” effort to 0.2%. These findings illustrate the potential of sequential screening protocols and hopefully increase the popularity of this relatively new methodology.