- Molecular similarity for machine learning in drug development : poster presentation (2008)
- Poster presentation In pharmaceutical research and drug development, machine learning methods play an important role in virtual screening and ADME/Tox prediction. For the application of such methods, a formal measure of similarity between molecules is essential. Such a measure, in turn, depends on the underlying molecular representation. Input samples have traditionally been modeled as vectors. Consequently, molecules are represented to machine learning algorithms in a vectorized form using molecular descriptors. While this approach is straightforward, it has its shortcomings. Amongst others, the interpretation of the learned model can be difficult, e.g. when using fingerprints or hashing. Structured representations of the input constitute an alternative to vector based representations, a trend in machine learning over the last years. For molecules, there is a rich choice of such representations. Popular examples include the molecular graph, molecular shape and the electrostatic field. We have developed a molecular similarity measure defined directly on the (annotated) molecular graph, a long-standing established topological model for molecules. It is based on the concepts of optimal atom assignments and iterative graph similarity. In the latter, two atoms are considered similar if their neighbors are similar. This recursive definition leads to a non-linear system of equations. We show how to iteratively solve these equations and give bounds on the computational complexity of the procedure. Advantages of our similarity measure include interpretability (atoms of two molecules are assigned to each other, each pair with a score expressing local similarity; this can be visualized to show similar regions of two molecules and the degree of their similarity) and the possibility to introduce knowledge about the target where available. We retrospectively tested our similarity measure using support vector machines for virtual screening on several pharmaceutical and toxicological datasets, with encouraging results. Prospective studies are under way.
- Ideenschmiede mit Praxisbezug : fünf Jahre Beilstein-Stiftungsprofessur für Chemieinformatik (2007)
- Eine Stiftungsprofessur ermöglicht die konzentrierte Forschung auf einem speziellen Fachgebiet und schafft den notwendigen Freiraum, Neues zu erproben. Insbesondere kann sie dazu dienen, Brücken zwischen Disziplinen zu errichten. Mit diesem Ziel wurde vor fünf Jahren die Beilstein-Stiftungsprofessur für Chemieinformatik an der Johann Wolfgang Goethe-Universität eingerichtet. Gefördert von dem in Frankfurt am Main ansässigen Beilstein-Institut zur Förderung der Chemischen Wissenschaften, wurde sie in enger Zusammenarbeit mit dem Institut für Organische Chemie und Chemische Biologie unter der Federführung von Prof. Dr. Michael Göbel konzipiert. Nachdem die Förderperiode von fünf Jahren im März 2007 ausgelaufen war, ist die Stiftungsprofessur nahtlos in den ordentlichen Universitätsbetrieb übernommen worden. Dies gibt Anlass, ein Fazit zu ziehen.
- Kernel learning for ligand-based virtual screening:discovery of a new PPARgamma agonist (2010)
- Poster presentation at 5th German Conference on Cheminformatics: 23. CIC-Workshop Goslar, Germany. 8-10 November 2009 We demonstrate the theoretical and practical application of modern kernel-based machine learning methods to ligand-based virtual screening by successful prospective screening for novel agonists of the peroxisome proliferator-activated receptor gamma (PPARgamma) . PPARgamma is a nuclear receptor involved in lipid and glucose metabolism, and related to type-2 diabetes and dyslipidemia. Applied methods included a graph kernel designed for molecular similarity analysis , kernel principle component analysis , multiple kernel learning , and, Gaussian process regression . In the machine learning approach to ligand-based virtual screening, one uses the similarity principle  to identify potentially active compounds based on their similarity to known reference ligands. Kernel-based machine learning  uses the "kernel trick", a systematic approach to the derivation of non-linear versions of linear algorithms like separating hyperplanes and regression. Prerequisites for kernel learning are similarity measures with the mathematical property of positive semidefiniteness (kernels). The iterative similarity optimal assignment graph kernel (ISOAK)  is defined directly on the annotated structure graph, and was designed specifically for the comparison of small molecules. In our virtual screening study, its use improved results, e.g., in principle component analysis-based visualization and Gaussian process regression. Following a thorough retrospective validation using a data set of 176 published PPARgamma agonists , we screened a vendor library for novel agonists. Subsequent testing of 15 compounds in a cell-based transactivation assay  yielded four active compounds. The most interesting hit, a natural product derivative with cyclobutane scaffold, is a full selective PPARgamma agonist (EC50 = 10 ± 0.2 microM, inactive on PPARalpha and PPARbeta/delta at 10 microM). We demonstrate how the interplay of several modern kernel-based machine learning approaches can successfully improve ligand-based virtual screening results.
- Predicting olfactory receptor neuron responses from odorant structure (2007)
- Background Olfactory receptors work at the interface between the chemical world of volatile molecules and the perception of scent in the brain. Their main purpose is to translate chemical space into information that can be processed by neural circuits. Assuming that these receptors have evolved to cope with this task, the analysis of their coding strategy promises to yield valuable insight in how to encode chemical information in an efficient way. Results We mimicked olfactory coding by modeling responses of primary olfactory neurons to small molecules using a large set of physicochemical molecular descriptors and artificial neural networks. We then tested these models by recording in vivo receptor neuron responses to a new set of odorants and successfully predicted the responses of five out of seven receptor neurons. Correlation coefficients ranged from 0.66 to 0.85, demonstrating the applicability of our approach for the analysis of olfactory receptor activation data. The molecular descriptors that are best-suited for response prediction vary for different receptor neurons, implying that each receptor neuron detects a different aspect of chemical space. Finally, we demonstrate that receptor responses themselves can be used as descriptors in a predictive model of neuron activation. Conclusions The chemical meaning of molecular descriptors helps understand structure-response relationships for olfactory receptors and their 'receptive fields'. Moreover, it is possible to predict receptor neuron activation from chemical structure using machine-learning techniques, although this is still complicated by a lack of training data.
- PocketPicker: analysis of ligand binding-sites with shape descriptors (2007)
- Background Identification and evaluation of surface binding-pockets and occluded cavities are initial steps in protein structure-based drug design. Characterizing the active site's shape as well as the distribution of surrounding residues plays an important role for a variety of applications such as automated ligand docking or in situ modeling. Comparing the shape similarity of binding site geometries of related proteins provides further insights into the mechanisms of ligand binding. Results We present PocketPicker, an automated grid-based technique for the prediction of protein binding pockets that specifies the shape of a potential binding-site with regard to its buriedness. The method was applied to a representative set of protein-ligand complexes and their corresponding apo-protein structures to evaluate the quality of binding-site predictions. The performance of the pocket detection routine was compared to results achieved with the existing methods CAST, LIGSITE, LIGSITEcs, PASS and SURFNET. Success rates PocketPicker were comparable to those of LIGSITEcs and outperformed the other tools. We introduce a descriptor that translates the arrangement of grid points delineating a detected binding-site into a correlation vector. We show that this shape descriptor is suited for comparative analyses of similar binding-site geometry by examining induced-fit phenomena in aldose reductase. This new method uses information derived from calculations of the buriedness of potential binding-sites. Conclusions The pocket prediction routine of PocketPicker is a useful tool for identification of potential protein binding-pockets. It produces a convenient representation of binding-site shapes including an intuitive description of their accessibility. The shape-descriptor for automated classification of binding-site geometries can be used as an additional tool complementing elaborate manual inspections.
- Optimized Particle Swarm Optimization (OPSO) and its application to artificial neural network training (2006)
- Background: Particle Swarm Optimization (PSO) is an established method for parameter optimization. It represents a population-based adaptive optimization technique that is influenced by several "strategy parameters". Choosing reasonable parameter values for the PSO is crucial for its convergence behavior, and depends on the optimization task. We present a method for parameter meta-optimization based on PSO and its application to neural network training. The concept of the Optimized Particle Swarm Optimization (OPSO) is to optimize the free parameters of the PSO by having swarms within a swarm. We assessed the performance of the OPSO method on a set of five artificial fitness functions and compared it to the performance of two popular PSO implementations. Results: Our results indicate that PSO performance can be improved if meta-optimized parameter sets are applied. In addition, we could improve optimization speed and quality on the other PSO methods in the majority of our experiments. We applied the OPSO method to neural network training with the aim to build a quantitative model for predicting blood-brain barrier permeation of small organic molecules. On average, training time decreased by a factor of four and two in comparison to the other PSO methods, respectively. By applying the OPSO method, a prediction model showing good correlation with training-, test- and validation data was obtained. Conclusion: Optimizing the free parameters of the PSO method can result in performance gain. The OPSO approach yields parameter combinations improving overall optimization performance. Its conceptual simplicity makes implementing the method a straightforward task.
- The plasmodium export element revisited (2008)
- We performed a bioinformatical analysis of protein export elements (PEXEL) in the putative proteome of the malaria parasite Plasmodium falciparum. A protein family-specific conservation of physicochemical residue profiles was found for PEXEL-flanking sequence regions. We demonstrate that the family members can be clustered based on the flanking regions only and display characteristic hydrophobicity patterns. This raises the possibility that the flanking regions may contain additional information for a family-specific role of PEXEL. We further show that signal peptide cleavage results in a positional alignment of PEXEL from both proteins with, and without, a signal peptide.
- Prediction of extracellular proteases of the human pathogen Helicobacter pylori reveals proteolytic activity of the Hp1018/19 protein HtrA (2008)
- Exported proteases of Helicobacter pylori (H. pylori) are potentially involved in pathogen-associated disorders leading to gastric inflammation and neoplasia. By comprehensive sequence screening of the H. pylori proteome for predicted secreted proteases, we retrieved several candidate genes. We detected caseinolytic activities of several such proteases, which are released independently from the H. pylori type IV secretion system encoded by the cag pathogenicity island (cagPAI). Among these, we found the predicted serine protease HtrA (Hp1019), which was previously identified in the bacterial secretome of H. pylori. Importantly, we further found that the H. pylori genes hp1018 and hp1019 represent a single gene likely coding for an exported protein. Here, we directly verified proteolytic activity of HtrA in vitro and identified the HtrA protease in zymograms by mass spectrometry. Overexpressed and purified HtrA exhibited pronounced proteolytic activity, which is inactivated after mutation of Ser205 to alanine in the predicted active center of HtrA. These data demonstrate that H. pylori secretes HtrA as an active protease, which might represent a novel candidate target for therapeutic intervention strategies.
- Domain organization of long signal peptides of single-pass integral membrane proteins reveals multiple functional capacity (2008)
- Targeting signals direct proteins to their extra- or intracellular destination such as the plasma membrane or cellular organelles. Here we investigated the structure and function of exceptionally long signal peptides encompassing at least 40 amino acid residues. We discovered a two-domain organization ("NtraC model") in many long signals from vertebrate precursor proteins. Accordingly, long signal peptides may contain an N-terminal domain (N-domain) and a C-terminal domain (C-domain) with different signal or targeting capabilities, separable by a presumably turn-rich transition area (tra). Individual domain functions were probed by cellular targeting experiments with fusion proteins containing parts of the long signal peptide of human membrane protein shrew-1 and secreted alkaline phosphatase as a reporter protein. As predicted, the N-domain of the fusion protein alone was shown to act as a mitochondrial targeting signal, whereas the C-domain alone functions as an export signal. Selective disruption of the transition area in the signal peptide impairs the export efficiency of the reporter protein. Altogether, the results of cellular targeting studies provide a proof-of-principle for our NtraC model and highlight the particular functional importance of the predicted transition area, which critically affects the rate of protein export. In conclusion, the NtraC approach enables the systematic detection and prediction of cryptic targeting signals present in one coherent sequence, and provides a structurally motivated basis for decoding the functional complexity of long protein targeting signals.