Hanna Geppert

Scaffold hopping using two-dimensional fingerprints: true potential, black magic, or a hopeless endeavor? Guidelines for virtual screening.

The scaffold hopping potential of popular 2D fingerprints has been thoroughly investigated. We have found that these types of fingerprints have at least limited scaffold hopping ability including early enrichment of small numbers of active scaffolds at high database ranks. However, it has not been possible to derive Tanimoto coefficient value ranges for individual fingerprints that are generally preferred for scaffold hopping. For selected fingerprints, similarity threshold values have been identified that yield small database selection sets having a high probability to contain a few active scaffolds. Furthermore, essentially all tested fingerprints have shown the ability to enrich scaffold hops in approximately 1% of a screening database. For the test cases reported herein, selecting 0.5-1% of the screening database yields approximately 25% of the available scaffolds. On the basis of our findings, practical guidelines for virtual screening using different types of 2D fingerprints have been formulated.

Integrating structure- and ligand-based virtual screening: comparison of individual, parallel, and fused molecular docking and similarity search calculations on multiple targets.

Similarity searching is often used to preselect compounds for docking, thereby decreasing the size of screening databases. However, integrated structure‐ and ligand‐based screening schemes are rare at present. Docking and similarity search calculations using 2D fingerprints were carried out in a comparative manner on nine target enzymes, for which significant numbers of diverse inhibitors could be obtained. In the absence of knowledge‐based docking constraints and target‐directed parameter optimisation, fingerprint searching displayed a clear preference over docking calculations. Alternative combinations of docking and similarity search results were investigated and found to further increase compound recall of individual methods in a number of instances. When the results of similarity searching and docking were combined, parallel selection of candidate compounds from individual rankings was generally superior to rank fusion. We suggest that complementary results from docking and similarity searching can be captured by integrated compound selection schemes.

Ligand Prediction from Protein Sequence and Small Molecule Information Using Support Vector Machines and Fingerprint Descriptors

Support vector machine (SVM) database search strategies are presented that aim at the identification of small molecule ligands for targets for which no ligand information is currently available. In pharmaceutical research and chemical biology, this situation is faced, for example, when studying orphan targets or newly identified members of protein families. To investigate methods for de novo ligand identification in the absence of known three-dimensional target structures or active molecules, we have focused on combining sequence and ligand information for closely and distantly related proteins. To provide a basis for these investigations, a set of 11 protease targets from different families was assembled together with more than 2000 inhibitors directed against individual proteases. We have compared SVM approaches that combine protein sequence and ligand information in different ways and utilize 2D fingerprints as ligand descriptors. These methodologies were applied to search for inhibitors of individual proteases not taken into account during learning. A target sequence-ligand kernel and, in particular, a linear combination of multiple target-directed SVMs consistently identified inhibitors with high accuracy including test cases where homology-based similarity searching using data fusion and conventional SVM ranking nearly or completely failed. The SVM linear combination and target-ligand kernel methods described herein are intuitive and straightforward to adopt for ligand prediction against other targets.

Ligand Prediction for Orphan Targets Using Support Vector Machines and Various Target-Ligand Kernels Is Dominated by Nearest Neighbor Effects

Support vector machine (SVM) calculations combining protein and small molecule information have been applied to identify ligands for simulated orphan targets (i.e., targets for which no ligands were available). The combination of protein and ligand information was facilitated through the design of target-ligand kernel functions that account for pairwise ligand and target similarity. The design and biological information content of such kernel functions was expected to play a major role for target-directed ligand prediction. Therefore, a variety of target-ligand kernels were implemented to capture different types of target information including sequence, secondary structure, tertiary structure, biophysical properties, ontologies, or structural taxonomy. These kernels were tested in ligand predictions for simulated orphan targets in two target protein systems characterized by the presence of different intertarget relationships. Surprisingly, although there were target- and set-specific differences in prediction rates for alternative target-ligand kernels, the performance of these kernels was overall similar and also similar to SVM linear combinations. Test calculations designed to better understand possible reasons for these observations revealed that ligand information provided by nearest neighbors of orphan targets significantly influenced SVM performance, much more so than the inclusion of protein information. As long as ligands of closely related neighbors of orphan targets were available for SVM learning, orphan target ligands could be well predicted, regardless of the type and sophistication of the kernel function that was used. These findings suggest simplified strategies for SVM-based ligand prediction for orphan targets.

Searching for Target-Selective Compounds Using Different Combinations of Multiclass Support Vector Machine Ranking Methods, Kernel Functions, and Fingerprint Descriptors

The identification of small chemical compounds that are selective for a target protein over one or more closely related members of the same family is of high relevance for applications in chemical biology. Conventional 2D similarity searching using known selective molecules as templates has recently been found to preferentially detect selective over non-selective and inactive database compounds. To improve the initially observed search performance, we have attempted to use 2D fingerprints as descriptors for support vector machine (SVM)-based selectivity searching. Different from typically applied binary SVM compound classification, SVM analysis has been adapted here for multiclass predictions and compound ranking to distinguish between selective, active but non-selective, and inactive compounds. In systematic database search calculations, we tested combinations of four alternative SVM ranking schemes, four different kernel functions, and four fingerprints and were able to further improve selectivity search performance by effectively removing non-selective molecules from high ranking positions while retaining high recall of selective compounds.

Support-vector-machine-based ranking significantly improves the effectiveness of similarity searching using 2D fingerprints and multiple reference compounds.

Similarity searching using molecular fingerprints is computationally efficient and a surprisingly effective virtual screening tool. In this study, we have compared ranking methods for similarity searching using multiple active reference molecules. Different 2D fingerprints were used as search tools and also as descriptors for a support vector machine (SVM) algorithm. In systematic database search calculations, a SVM-based ranking scheme consistently outperformed nearest neighbor and centroid approaches, regardless of the fingerprints that were tested, even if only very small training sets were used for SVM learning. The superiority of SVM-based ranking over conventional fingerprint methods is ascribed to the fact that SVM makes use of information about database molecules, in addition to known active compounds, during the learning phase.

Targeting multifunctional proteins by virtual screening: structurally diverse cytohesin inhibitors with differentiated biological functions

Virtual screening (VS) of chemical libraries formatted in silico provides an alternative to experimental high-throughput screening (HTS) for the identification of small molecule modulators of protein function. We have tailored a VS approach combining fingerprint similarity searching and support vector machine modeling toward the identification of small molecular probes for the study of cytohesins, a family of cytoplasmic regulator proteins with multiple cellular functions. A total of 40 new structurally diverse inhibitors were identified, and 26 of these compounds were more active than the primary VS template, a single known inhibitory chemotype, in at least one of three different assays (guanine nucleotide exchange, Drosophila insulin signaling, and human leukocyte cell adhesion). Moreover, these inhibitors displayed differential inhibitory profiles. Our findings demonstrate that, at least for the cytohesins, computational extrapolation from known active compounds was capable of identifying small molecular probes with highly diversified functional profiles.

Methods for Computer-Aided Chemical Biology. Part 3: Analysis of Structure–Selectivity Relationships through Single- or Dual-Step Selectivity Searching and Bayesian Classification

The identification of small molecules that are selective for individual targets within target families is an important task in chemical biology. We aim at the development of computational approaches for the study of structure–selectivity relationships and prediction of target‐selective ligands. In previous studies, we have introduced the concept of selectivity searching. Here we study compound selectivity on the basis of 18 selectivity sets that are designed to contain target‐selective molecules and compounds that are comparably active against related targets. These sets consist of a total of 432 compounds and focus on eight targets belonging to four target families. This compound source has enabled us to evaluate different computational approaches to search for target‐selective compounds in large databases. These investigations have revealed a preferred search strategy to enrich database selection sets with target‐selective compounds. The selectivity sets reported here are made publicly available to support the development of other computational tools for applications in chemical biology and medicinal chemistry.

Development of a Method To Consistently Quantify the Structural Distance between Scaffolds and To Assess Scaffold Hopping Potential

We introduce a method to determine a structural distance between any pair of molecular scaffolds. The development of this approach was motivated by the need to accurately evaluate scaffold hopping studies in virtual screening and medicinal chemistry and assess the degree of difficulty involved in facilitating a transition from one structure to another. In order to consistently derive structural distances, scaffolds of different composition and topology are subjected to molecular editing procedures that abstract from original scaffolds in a defined manner until compositional and topological equivalence can be established. Pairs of corresponding scaffold representations are transformed into one-dimensional atom sequences that are aligned using approaches adapted from biological sequence comparison. From best scoring atom sequence alignments, interscaffold distances are derived. The algorithm is evaluated at different levels including the analysis of a series of model scaffolds with defined chemical changes, a scaffold library, and scaffolds from reference compounds and hits of successful virtual screening applications. It is demonstrated that chemically intuitive scaffold distances are obtained for pairs of scaffolds with varying composition and topology. Distance threshold values for close and remote structural relationships between scaffolds are also determined. The methodology is made publicly available in order to provide a basis for a consistent assessment of scaffold hopping ability and to aid in the evaluation and comparison of virtual screening methods.

Shannon entropy-based fingerprint similarity search strategy.

For fingerprint searching using multiple active reference compounds, an information entropy-based similarity method is introduced as an alternative to conventional similarity coefficients and search strategies. The approach involves the determination of the fingerprint bit pattern entropy of a compound reference set and recalculation of the entropy following the addition of individual test compounds. If a database compound shares similar bit patterns with reference set molecules, adding this compound to the reference set only produces a small change in system entropy. By contrast, inclusion of a compound having a dissimilar fingerprint leads to a notable increase in entropy. Thus, database compounds can be screened for candidate molecules that do not cause significant changes in reference set fingerprint entropy. Compared to nearest neighbor methods, this approach has the computational advantage that it extracts reference set information only once prior to similarity searching. Test calculations on different compound data sets, fingerprints, and screening databases reveal that the ability of our entropy-based method to detect active compounds is often superior to data fusion techniques and Tanimoto similarity calculations.