Nicos Angelopoulos

Bayesian learning of Bayesian networks with informative priors

This paper presents and evaluates an approach to Bayesian model averaging where the models are Bayesian nets (BNs). A comprehensive study of the literature on structural priors for BNs is conducted. A number of prior distributions are defined using stochastic logic programs and the MCMC Metropolis-Hastings algorithm is used to (approximately) sample from the posterior. We use proposals which are tightly coupled to the priors which give rise to cheaply computable acceptance probabilities. Experiments using data generated from known BNs have been conducted to evaluate the method. The experiments used 6 different BNs and varied: the structural prior, the parameter prior, the Metropolis-Hasting proposal and the data size. Each experiment was repeated three times with different random seeds to test the robustness of the MCMC-produced results. Our results show that with effective priors (i) robust results are produced and (ii) informative priors improve results significantly.

Bayesian model averaging for ligand discovery.

High-throughput screening (HTS) is now a standard approach used in the pharmaceutical industry to identify potential drug-like lead molecules. The analysis linking biological data with molecular properties is a major goal in both academic and pharmaceutical research. This paper presents a Bayesian analysis of high-dimensional descriptor data using Markov chain Monte Carlo (MCMC) simulations for learning classification trees as a novel method for pharmacophore and ligand discovery. We use experimentally determined binding affinity data with the protein pyruvate kinase to train and assess our model averaging algorithm and then apply it to a large database of over 3.7 million molecules. We compare the results of a number of variations on the central Bayesian theme to that of two Neural Network (NN) architectures and that of Support Vector Machines (SVM). The main Bayesian algorithm, in addition to achieving high specificity and sensitivity, also lends itself naturally to classifying test sets with missing data and providing a ranking for the classified compounds. The approach has been used to select and rank potential biologically active compounds and could provide a powerful tool in compound testing.

Tempering for Bayesian C&RT

This paper concerns the experimental assessment of tempering as a technique for improving Bayesian inference for C&RT models. Full Bayesian inference requires the computation of a posterior over all possible trees. Since exact computation is not possible Markov chain Monte Carlo (MCMC) methods are used to produce an approximation. C&RT posteriors have many local modes: tempering aims to prevent the Markov chain getting stuck in these modes. Our results show that a clear improvement is achieved using tempering.

Prolog issues and experimental results of an MCMC algorithm

We present a Markov chain Monte Carlo algorithm that operates on generic model structures that are represented by terms found in the computed answers produced by stochastic logic programs. The objective of this paper is threefold (a) to show that SLD-trees are an elegant means for describing prior distributions over model structures (b) to sketch an implementation of the MCMC algorithm in Prolog, and (c) to provide insights on desirable properties for SLPs

Selective chemical intervention in the proteome of Caenorhabditis elegans.

We present the first study of protein regulation by ligands in Caenorhabditis elegans. The ligands were peptidyl-prolyl isomerase inhibitors of cyclophilins. Up-regulation is observed for several heat shock proteins and one ligand in particular caused a greater than 2-fold enhancement of cyclophilin CYN-5. Additionally, several metabolic enzymes display elevated levels. This approach, using label-free relative quantification, provides an extremely attractive way of measuring the effect of ligands on an entire proteome, with minimal sample pretreatment, which could be applicable to large-scale studies. In this initial study, which compares the effect of three ligands, 54 unique proteins have been identified that are up- (51) or down- (3) regulated in the presence of a given ligand. A total of 431 C. elegans proteins were identified. Our methodology provides an intriguing new direction for in vivo screening of the effects of novel and untested ligands at the whole organism level.

Probabilistic space partitioning in constraint logic programming

We present a language for integrating probabilistic reasoning and logic programming. The key idea is to use constraints based techniques such as the constraints store and finite domain variables. First we show how these techniques can be used to integrate a number of probabilistic inference algorithms with logic programming. We then proceed to detail a language which effects conditioning by probabilistically partitioning the constraint store. We elucidate the kinds of reasoning effected by the introduced language by means of two well known probabilistic problems: the three prisoners and Monty Hall. In particular we show how the syntax of the language can be used to avoid the pitfalls normally associated with the two problems. An elimination algorithm for computing the probability of a query in a given store is presented.