Roberto Tagliaferri

Discovery of drug mode of action and drug repositioning from transcriptional responses

A bottleneck in drug discovery is the identification of the molecular targets of a compound (mode of action, MoA) and of its off-target effects. Previous approaches to elucidate drug MoA include analysis of chemical structures, transcriptional responses following treatment, and text mining. Methods based on transcriptional responses require the least amount of information and can be quickly applied to new compounds. Available methods are inefficient and are not able to support network pharmacology. We developed an automatic and robust approach that exploits similarity in gene expression profiles following drug treatment, across multiple cell lines and dosages, to predict similarities in drug effect and MoA. We constructed a “drug network” of 1,302 nodes (drugs) and 41,047 edges (indicating similarities between pair of drugs). We applied network theory, partitioning drugs into groups of densely interconnected nodes (i.e., communities). These communities are significantly enriched for compounds with similar MoA, or acting on the same pathway, and can be used to identify the compound-targeted biological pathways. New compounds can be integrated into the network to predict their therapeutic and off-target effects. Using this network, we correctly predicted the MoA for nine anticancer compounds, and we were able to discover an unreported effect for a well-known drug. We verified an unexpected similarity between cyclin-dependent kinase 2 inhibitors and Topoisomerase inhibitors. We discovered that Fasudil (a Rho-kinase inhibitor) might be “repositioned” as an enhancer of cellular autophagy, potentially applicable to several neurodegenerative disorders. Our approach was implemented in a tool (Mode of Action by NeTwoRk Analysis, MANTRA, http://mantra.tigem.it).

Drug repositioning: a machine-learning approach through data integration

Existing computational methods for drug repositioning either rely only on the gene expression response of cell lines after treatment, or on drug-to-disease relationships, merging several information levels. However, the noisy nature of the gene expression and the scarcity of genomic data for many diseases are important limitations to such approaches. Here we focused on a drug-centered approach by predicting the therapeutic class of FDA-approved compounds, not considering data concerning the diseases. We propose a novel computational approach to predict drug repositioning based on state-of-the-art machine-learning algorithms. We have integrated multiple layers of information: i) on the distances of the drugs based on how similar are their chemical structures, ii) on how close are their targets within the protein-protein interaction network, and iii) on how correlated are the gene expression patterns after treatment. Our classifier reaches high accuracy levels (78%), allowing us to re-interpret the top misclassifications as re-classifications, after rigorous statistical evaluation. Efficient drug repurposing has the potential to significantly impact the whole field of drug development. The results presented here can significantly accelerate the translation into the clinics of known compounds for novel therapeutic uses.

Identifying Network of Drug Mode of Action by Gene Expression Profiling

Drug mode of action (MOA) of novel compounds has been predicted using phenotypic features or, more recently, comparing side effect similarities. Attempts to use gene expression data in mammalian systems have so far met limited success. Here, we built a drug similarity network starting from a public reference dataset containing genome-wide gene expression profiles (GEPs) following treatments with more than a thousand compounds. In this network, drugs sharing a subset of molecular targets are connected by an edge or lie in the same community. Our approach is based on a novel similarity distance between two compounds. The distance is computed by combining GEPs via an original rank-aggregation method, followed by a gene set enrichment analysis (GSEA) to compute similarity between pair of drugs. The network is obtained by considering each compound as a node, and adding an edge between two compounds if their similarity distance is below a given significance threshold. We show that, despite the complexity and the variety of the experimental conditions, our approach is able to identify similarities in drug mode of action from GEPs. Our approach can also be used for the identification of the MOA of new compounds.

A multifrequency analysis of radio variability of blazars

We have carried out a multifrequency analysis of the radio variability of blazars, exploiting the data obtained during the extensive monitoring programs carried out at the University of Michigan Radio Astronomy Observatory (UMRAO, at 4.8, 8, and 14.5 GHz) and at the Metsahovi Radio Observatory (22 and 37 GHz). Two different techniques detect, in the Metsahovi light curves, evidence of periodicity at both frequencies for 5 sources (0224 + 671, 0945 + 408, 1226 + 023, 2200 + 420, and 2251 + 158). For the last three sources, consistent periods are found also at the three UMRAO frequencies and the Scargle (1982) method yields an extremely low false-alarm probability. On the other hand, the 22 and 37 GHz periodicities of 0224+671 and 0945 + 408 (which were less extensively monitored at Metsahovi and for which we get a significant false-alarm probability) are not confirmed by the UMRAO database, where some indications of ill-defined periods of about a factor of two longer are retrieved. We have also investigated the variability index, the structure function, and the distribution of intensity variations of the most extensively monitored sources. We find a statistically significant difference in the distribution of the variability index for BL Lac objects compared to flat-spectrum radio quasars (FSRQs), in the sense that the former objects are more variable. For both populations the variability index steadily increases with increasing frequency. The distribution of intensity variations also broadens with increasing frequency, and approaches a log-normal shape at the highest frequencies. We find that variability enhances by 20-30% the high frequency counts of extragalactic radio-sources at bright flux densities, such as those of the WMAP and PLANCK surveys. In all objects with detected periodicity we find evidence for the existence of impulsive signals superimposed on the periodic component.

Wide field imaging – I. Applications of neural networks to object detection and star/galaxy classification

Astronomical wide-field imaging performed with new large-format CCD detectors poses data reduction problems of unprecedented scale, which are difficult to deal with using traditional interactive tools. We present here NExt (Neural Extractor), a new neural network (NN) based package capable of detecting objects and performing both deblending and star/galaxy classification in an automatic way. Traditionally, in astronomical images, objects are first distinguished from the noisy background by searching for sets of connected pixels having brightnesses above a given threshold; they are then classified as stars or as galaxies through diagnostic diagrams having variables chosen according to the astronomers taste and experience. In the extraction step, assuming that images are well sampled, NExt requires only the simplest a priori definition of ‘what an object is’ (i.e. it keeps all structures composed of more than one pixel) and performs the detection via an unsupervised NN, approaching detection as a clustering problem that has been thoroughly studied in the artificial intelligence literature. The first part of the NExt procedure consists of an optimal compression of the redundant information contained in the pixels via a mapping from pixel intensities to a subspace individualized through principal component analysis. At magnitudes fainter than the completeness limit, stars are usually almost indistinguishable from galaxies, and therefore the parameters characterizing the two classes do not lie in disconnected subspaces, thus preventing the use of unsupervised methods. We therefore adopted a supervised NN (i.e. a NN that first finds the rules to classify objects from examples and then applies them to the whole data set). In practice, each object is classified depending on its membership of the regions mapping the input feature space in the training set. In order to obtain an objective and reliable classification, instead of using an arbitrarily defined set of features we use a NN to select the most significant features among the large number of measured ones, and then we use these selected features to perform the classification task. In order to optimize the performance of the system, we implemented and tested several different models of NN. The comparison of the NExt performance with that of the best detection and classification package known to the authors (SExtractor) shows that NExt is at least as effective as the best traditional packages.

A novel neural network-based survival analysis model

A feedforward neural network architecture aimed at survival probability estimation is presented which generalizes the standard, usually linear, models described in literature. The network builds an approximation to the survival probability of a system at a given time, conditional on the system features. The resulting model is described in a hierarchical Bayesian framework. Experiments with synthetic and real world data compare the performance of this model with the commonly used standard ones.

Mining the SDSS archive. I. Photometric redshifts in the nearby universe

In this paper we present a supervised neural network approach to the determination of photometric redshifts. The method, even though of general validity, was fine tuned to match the characteristics of the Sloan Digital Sky Survey (SDSS) and as base of ’a priori’ knowledge, it exploits the rich wealth of spectroscopic redshifts provided by this unique survey. In order to train, validate and test the networks, we used two galaxy samples drawn from the SDSS spectroscopic dataset, namely: the general galaxy sample (GG) and the luminous red galaxies subsample (LRG). Due to the uneven distribution of measured redshifts in the SDSS spectroscopic subsample, the method consists of a two steps approach. In the first step, objects are classified in nearby (z < 0.25) and distant (0.25 < z < 0.50), with an accuracy estimated in 97.52%. In the second step two different networks are separately trained on objects belonging to the two redshift ranges. Using a standard Multi Layer Perceptron operated in a Bayesian framework, the optimal architectures were found to require 1 hidden layer of 24 (24) and 24 (25) neurons for the GG (LRG) sample. The presence of systematic deviations was then corrected by interpolating the resulting redshifts. The final results on the GG dataset give a robust σz ≃ 0.0208 over the redshift range [0.01, 0.48] and σz ≃ 0.0197 and σz ≃ 0.0238 for the nearby and distant Department of Physical Sciences, University of Napoli Federico II, via Cinthia 9, 80126 Napoli, ITALY INAF-Italian National Institute of Astrophysics, via del Parco Mellini, Rome, I INFN Napoli Unit, Dept. of Physical Sciences, via Cinthia 9, 80126, Napoli, ITALY Department of Mathematics and Applications, University of Salerno, Fisciano, ITALY Department of Applied Science, University of Napoli ”Parthenope”, via A. De Gasperi 5, 80133 Napoli, ITALY Institute of Astronomy, University of Cambridge, Madingley Rd, Cambridge CB4 0HA, UK

Neural networks in astronomy

In the last decade, the use of neural networks (NN) and of other soft computing methods has begun to spread also in the astronomical community which, due to the required accuracy of the measurements, is usually reluctant to use automatic tools to perform even the most common tasks of data reduction and data mining. The federation of heterogeneous large astronomical databases which is foreseen in the framework of the astrophysical virtual observatory and national virtual observatory projects, is, however, posing unprecedented data mining and visualization problems which will find a rather natural and user friendly answer in artificial intelligence tools based on NNs, fuzzy sets or genetic algorithms. This review is aimed to both astronomers (who often have little knowledge of the methodological background) and computer scientists (who often know little about potentially interesting applications), and therefore will be structured as follows: after giving a short introduction to the subject, we shall summarize the methodological background and focus our attention on some of the most interesting fields of application, namely: object extraction and classification, time series analysis, noise identification, and data mining. Most of the original work described in the paper has been performed in the framework of the AstroNeural collaboration (Napoli-Salerno).

Data mining in cancer research

This article is not intended as a comprehensive survey of data mining applications in cancer. Rather, it provides starting points for further, more targeted, literature searches, by embarking on a guided tour of computational intelligence applications in cancer medicine, structured in increasing order of the physical scales of biological processes.This article is not intended as a comprehensive survey of data mining applications in cancer. Rather, it provides starting points for further, more targeted, literature searches, by embarking on a guided tour of computational intelligence applications in cancer medicine, structured in increasing order of the physical scales of biological processes.

Neural networks for blind-source separation of Stromboli explosion quakes

Independent component analysis (ICA) is used to analyze the seismic signals produced by explosions of the Stromboli volcano. It has been experimentally proved that it is possible to extract the most significant components from seismometer recorders. In particular, the signal, eventually thought as generated by the source, is corresponding to the higher power spectrum, isolated by our analysis. Furthermore, the amplitude of the source signals has been found by using a simple trick and so overcoming, for this specific case, the classical problem of ICA regarding the amplitude loss of the separated signals.