Variational Markov Chain Monte Carlo for Bayesian smoothing of non-linear niffusions

In this paper we develop set of novel Markov chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. Flexible blocking strategies are introduced to further improve mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample, applications the algorithm is accurate except in the presence of large observation errors and low observation densities, which lead to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient.

Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions

In this paper we develop set of novel Markov Chain Monte Carlo algorithms for Bayesian smoothing of partially observed non-linear diffusion processes. The sampling algorithms developed herein use a deterministic approximation to the posterior distribution over paths as the proposal distribution for a mixture of an independence and a random walk sampler. The approximating distribution is sampled by simulating an optimized time-dependent linear diffusion process derived from the recently developed variational Gaussian process approximation method. The novel diffusion bridge proposal derived from the variational approximation allows the use of a flexible blocking strategy that further improves mixing, and thus the efficiency, of the sampling algorithms. The algorithms are tested on two diffusion processes: one with double-well potential drift and another with SINE drift. The new algorithm's accuracy and efficiency is compared with state-of-the-art hybrid Monte Carlo based path sampling. It is shown that in practical, finite sample applications the algorithm is accurate except in the presence of large observation errors and low to a multi-modal structure in the posterior distribution over paths. More importantly, the variational approximation assisted sampling algorithm outperforms hybrid Monte Carlo in terms of computational efficiency, except when the diffusion process is densely observed with small errors in which case both algorithms are equally efficient. © 2011 Springer-Verlag.

Anisotropic seismic tomography with the reversible jump Markov chain Monte Carlo

The established isotropic tomographic models show the features of subduction zones in terms of seismic velocity anomalies, but they are generally subjected to the generation of artifacts due to the lack of anisotropy in forward modelling. There is evidence for the significant influence of seismic anisotropy in the mid-upper mantle, especially for boundary layers like subducting slabs. As consequence, in isotropic models artifacts may be misinterpreted as compositional or thermal heterogeneities. In this thesis project the application of a trans-dimensional Metropolis-Hastings method is investigated in the context of anisotropic seismic tomography. This choice arises as a response to the important limitations introduced by traditional inversion methods which use iterative procedures of optimization of a function object of the inversion. On the basis of a first implementation of the Bayesian sampling algorithm, the code is tested with some cartesian two-dimensional models, and then extended to polar coordinates and dimensions typical of subduction zones, the main focus proposed for this method. Synthetic experiments with increasing complexity are realized to test the performance of the method and the precautions for multiple contexts, taking into account also the possibility to apply seismic ray-tracing iteratively. The code developed is tested mainly for 2D inversions, future extensions will allow the anisotropic inversion of seismological data to provide more realistic imaging of real subduction zones, less subjected to generation of artifacts.

Independent doubly Adaptive Rejection Metropolis Sampling

Adaptive Rejection Metropolis Sampling (ARMS) is a wellknown MCMC scheme for generating samples from onedimensional target distributions. ARMS is widely used within Gibbs sampling, where automatic and fast samplers are often needed to draw from univariate full-conditional densities. In this work, we propose an alternative adaptive algorithm (IA2RMS) that overcomes the main drawback of ARMS (an uncomplete adaptation of the proposal in some cases), speeding up the convergence of the chain to the target. Numerical results show that IA2RMS outperforms the standard ARMS, providing a correlation among samples close to zero.

ALTERNATIVE SAMPLING STRATEGY FOR A RANDOM OPTIMIZATION ALGORITHM

Alternative sampling procedures are compared to the pure random search method. It is shown that the efficiency of the algorithm can be improved with respect to the expected number of steps to reach an epsilon-neighborhood of the optimal point.

Sampling phylogenetic tree space with the generalized Gibbs sampler

The generalized Gibbs sampler (GGS) is a recently developed Markov chain Monte Carlo (MCMC) technique that enables Gibbs-like sampling of state spaces that lack a convenient representation in terms of a fixed coordinate system. This paper describes a new sampler, called the tree sampler, which uses the GGS to sample from a state space consisting of phylogenetic trees. The tree sampler is useful for a wide range of phylogenetic applications, including Bayesian, maximum likelihood, and maximum parsimony methods. A fast new algorithm to search for a maximum parsimony phylogeny is presented, using the tree sampler in the context of simulated annealing. The mathematics underlying the algorithm is explained and its time complexity is analyzed. The method is tested on two large data sets consisting of 123 sequences and 500 sequences, respectively. The new algorithm is shown to compare very favorably in terms of speed and accuracy to the program DNAPARS from the PHYLIP package.

A retrospective analysis of mycophenolic acid and cyclosporin concentrations with acute rejection in renal transplant recipients

Objectives: Although monitoring of cyclosporin (CsA) is standard clinical practice postrenal transplantation. mycophenolic acid (MPA) concentrations are not routinely measured. There is evidence that a relationship exists between MPA area under the concentration-time curve (AUC) and rejection. In this study, a retrospective analysis was undertaken of 27 adult renal transplant recipients. Methods: Patients received CsA and MPA therapy and had a four-point MPA AUC investigation. The relationship between MPA AUC performed in the first week after transplantation, as well as median trough cyclosporin concentrations, and clinical outcomes in the first month posttransplant were evaluated. Results: A total of 12 patients experienced biopsy proven rejection (44.4%) and 4 patients had gastrointestinal adverse events (14.8%). A statistically significant relationship was observed between the incidence of biopsy proven rejection and both MPA AUC (p = 0.02) and median trough CsA concentration (p = 0.008). No relationship between trough MPA concentration and rejection was observed (p = 0.21). Only 3 of 11 (27%) patients with an MPA AUC > 30 mg.h/L and a median trough CsA > 175 mug/L experienced acute rejection, compared with a 56% incidence of rejection for the remaining 16 patients. Patients who experienced adverse gastrointestinal events had significantly lower MPA AUC (p = 0.04), but median trough CsA concentrations were not significantly different (p = 0.24). Further, 3 of these 4 patients had rejection episodes. Conclusions: in addition to standard CsA monitoring, we propose further investigation of the use of a 4-point sampling strategy to predict MPA AUC in the first week posttransplant, which may facilitate optimization of mycophenolate mofetil dose at a rime when patients are most vulnerable to acute rejection. (C) 2001 The Canadian Society of Clinical Chemists. All rights reserved.

Active Learning of Very-High Resolution Optical Imagery with SVM: Entropy vs Margin Sampling

An active learning method is proposed for the semi-automatic selection of training sets in remote sensing image classification. The method adds iteratively to the current training set the unlabeled pixels for which the prediction of an ensemble of classifiers based on bagged training sets show maximum entropy. This way, the algorithm selects the pixels that are the most uncertain and that will improve the model if added in the training set. The user is asked to label such pixels at each iteration. Experiments using support vector machines (SVM) on an 8 classes QuickBird image show the excellent performances of the methods, that equals accuracies of both a model trained with ten times more pixels and a model whose training set has been built using a state-of-the-art SVM specific active learning method

Irregular to regular sampling, denoising and deconvolution

We propose a restoration algorithm for band limited images that considers irregular(perturbed) sampling, denoising, and deconvolution. We explore the application of a family ofregularizers that allow to control the spectral behavior of the solution combined with the irregular toregular sampling algorithms proposed by H.G. Feichtinger, K. Gr¨ochenig, M. Rauth and T. Strohmer.Moreover, the constraints given by the image acquisition model are incorporated as a set of localconstraints. And the analysis of such constraints leads to an early stopping rule meant to improvethe speed of the algorithm. Finally we present experiments focused on the restoration of satellite images, where the micro-vibrations are responsible of the type of distortions we are considering here. We will compare results of the proposed method with previous methods and show an extension tozoom.

EADock : design of a new molecular docking algorithm and some of its applications

3 Summary 3. 1 English The pharmaceutical industry has been facing several challenges during the last years, and the optimization of their drug discovery pipeline is believed to be the only viable solution. High-throughput techniques do participate actively to this optimization, especially when complemented by computational approaches aiming at rationalizing the enormous amount of information that they can produce. In siiico techniques, such as virtual screening or rational drug design, are now routinely used to guide drug discovery. Both heavily rely on the prediction of the molecular interaction (docking) occurring between drug-like molecules and a therapeutically relevant target. Several softwares are available to this end, but despite the very promising picture drawn in most benchmarks, they still hold several hidden weaknesses. As pointed out in several recent reviews, the docking problem is far from being solved, and there is now a need for methods able to identify binding modes with a high accuracy, which is essential to reliably compute the binding free energy of the ligand. This quantity is directly linked to its affinity and can be related to its biological activity. Accurate docking algorithms are thus critical for both the discovery and the rational optimization of new drugs. In this thesis, a new docking software aiming at this goal is presented, EADock. It uses a hybrid evolutionary algorithm with two fitness functions, in combination with a sophisticated management of the diversity. EADock is interfaced with .the CHARMM package for energy calculations and coordinate handling. A validation was carried out on 37 crystallized protein-ligand complexes featuring 11 different proteins. The search space was defined as a sphere of 15 R around the center of mass of the ligand position in the crystal structure, and conversely to other benchmarks, our algorithms was fed with optimized ligand positions up to 10 A root mean square deviation 2MSD) from the crystal structure. This validation illustrates the efficiency of our sampling heuristic, as correct binding modes, defined by a RMSD to the crystal structure lower than 2 A, were identified and ranked first for 68% of the complexes. The success rate increases to 78% when considering the five best-ranked clusters, and 92% when all clusters present in the last generation are taken into account. Most failures in this benchmark could be explained by the presence of crystal contacts in the experimental structure. EADock has been used to understand molecular interactions involved in the regulation of the Na,K ATPase, and in the activation of the nuclear hormone peroxisome proliferatoractivated receptors a (PPARa). It also helped to understand the action of common pollutants (phthalates) on PPARy, and the impact of biotransformations of the anticancer drug Imatinib (Gleevec®) on its binding mode to the Bcr-Abl tyrosine kinase. Finally, a fragment-based rational drug design approach using EADock was developed, and led to the successful design of new peptidic ligands for the a5ß1 integrin, and for the human PPARa. In both cases, the designed peptides presented activities comparable to that of well-established ligands such as the anticancer drug Cilengitide and Wy14,643, respectively. 3.2 French Les récentes difficultés de l'industrie pharmaceutique ne semblent pouvoir se résoudre que par l'optimisation de leur processus de développement de médicaments. Cette dernière implique de plus en plus. de techniques dites "haut-débit", particulièrement efficaces lorsqu'elles sont couplées aux outils informatiques permettant de gérer la masse de données produite. Désormais, les approches in silico telles que le criblage virtuel ou la conception rationnelle de nouvelles molécules sont utilisées couramment. Toutes deux reposent sur la capacité à prédire les détails de l'interaction moléculaire entre une molécule ressemblant à un principe actif (PA) et une protéine cible ayant un intérêt thérapeutique. Les comparatifs de logiciels s'attaquant à cette prédiction sont flatteurs, mais plusieurs problèmes subsistent. La littérature récente tend à remettre en cause leur fiabilité, affirmant l'émergence .d'un besoin pour des approches plus précises du mode d'interaction. Cette précision est essentielle au calcul de l'énergie libre de liaison, qui est directement liée à l'affinité du PA potentiel pour la protéine cible, et indirectement liée à son activité biologique. Une prédiction précise est d'une importance toute particulière pour la découverte et l'optimisation de nouvelles molécules actives. Cette thèse présente un nouveau logiciel, EADock, mettant en avant une telle précision. Cet algorithme évolutionnaire hybride utilise deux pressions de sélections, combinées à une gestion de la diversité sophistiquée. EADock repose sur CHARMM pour les calculs d'énergie et la gestion des coordonnées atomiques. Sa validation a été effectuée sur 37 complexes protéine-ligand cristallisés, incluant 11 protéines différentes. L'espace de recherche a été étendu à une sphère de 151 de rayon autour du centre de masse du ligand cristallisé, et contrairement aux comparatifs habituels, l'algorithme est parti de solutions optimisées présentant un RMSD jusqu'à 10 R par rapport à la structure cristalline. Cette validation a permis de mettre en évidence l'efficacité de notre heuristique de recherche car des modes d'interactions présentant un RMSD inférieur à 2 R par rapport à la structure cristalline ont été classés premier pour 68% des complexes. Lorsque les cinq meilleures solutions sont prises en compte, le taux de succès grimpe à 78%, et 92% lorsque la totalité de la dernière génération est prise en compte. La plupart des erreurs de prédiction sont imputables à la présence de contacts cristallins. Depuis, EADock a été utilisé pour comprendre les mécanismes moléculaires impliqués dans la régulation de la Na,K ATPase et dans l'activation du peroxisome proliferatoractivated receptor a (PPARa). Il a également permis de décrire l'interaction de polluants couramment rencontrés sur PPARy, ainsi que l'influence de la métabolisation de l'Imatinib (PA anticancéreux) sur la fixation à la kinase Bcr-Abl. Une approche basée sur la prédiction des interactions de fragments moléculaires avec protéine cible est également proposée. Elle a permis la découverte de nouveaux ligands peptidiques de PPARa et de l'intégrine a5ß1. Dans les deux cas, l'activité de ces nouveaux peptides est comparable à celles de ligands bien établis, comme le Wy14,643 pour le premier, et le Cilengitide (PA anticancéreux) pour la seconde.

Free-breathing T2 mapping at 3T for the monitoring of cardiac allograft rejection: initial results

rejection can lead to loss of function. Histological reading of endomyocardial biopsy remains the "gold standard" for guiding immunosuppression, despite its methodological limitations (sampling error and interobserver variability). The measurement of the T2 relaxation time has been suggested for detection of allograft rejection, on the pathophysiological basis that the T2 relaxation time prolongs with local edema resulting from acute allograft rejection. Using breath-held cardiac magnetic resonance T2 mapping at 1.5 T, Usman et al. (CircCardiovascImaging2012) detected moderate allograft rejection (grade 2R, ISHLT 2004). With modern immunosuppression grade 2R rejection has become a rare event, but the need remains for a technique that permits the discrimination of absent (grade 0R) and mild rejection (grade 1R). We therefore investigated whether an increase of magnetic field strength to 3T and the use of real-time navigator-gated respiration compensation allow for an increase in the sensitivity of T2 relaxation time detection that is necessary to achieve this discrimination. Methods: Eighteen patients received EMB (Tan et al., ArchPatholLabMed2007) and cardiac T2 mapping on the same day. Reading of T2 maps was blinded to the histological results. For final analysis, 3 cases with known 2R rejection at the time of T2 mapping were added, yielding 21 T2 mapping sessions. A respiration-navigator-gated radial gradient-recalled-echo pulse sequence (resolution 1.17 mm2, matrix 2562, trigger time 3 heartbeats, T2 preparation duration TET2 Prep = 60/30/0 ms) was applied to obtain 3 short-axis T2 maps (van Heeswijk et al., JACCCardiovascImaging2012), which were segmented according to AHA guidelines (Cerqueira et al, Circulation2001). The highest segmental T2 values were grouped according to histological rejection grade and differences were analyzed by Student's t-test, except for the non-blinded cases with 2R rejection. The degree of discrimination was determined using the Spearman's ranked correlation test. Results: The high-quality T2 maps allowed for visual differentiation of the rejection degrees (Figure 1), and the correlation of T2 mapping with the histological grade of acute cellular rejection was significant (Spearman's r = 0.56, p = 0.007). The 0R (n = 15) and 1R (n = 3) degrees demonstrated significantly different T2 values (46.9 ± 5.0 and 54.3 ± 3.0 ms, p = 0.02, Figure 2). Cases with 2R rejection showed clear T2 elevation (T2 = 60.3 ± 16.2 ms). Conclusions: This pilot study demonstrates that non-invasive free-breathing cardiac T2 mapping at 3T discriminates between no and mild cardiac allograft rejection. Confirmation of these encouraging results in a larger cohort should consider a study able to show equivalency or superiority of T2 mapping.

Bio-inspired optimization & sampling technique for side-chain packing in MCCE

The prediction of proteins' conformation helps to understand their exhibited functions, allows for modeling and allows for the possible synthesis of the studied protein. Our research is focused on a sub-problem of protein folding known as side-chain packing. Its computational complexity has been proven to be NP-Hard. The motivation behind our study is to offer the scientific community a means to obtain faster conformation approximations for small to large proteins over currently available methods. As the size of proteins increases, current techniques become unusable due to the exponential nature of the problem. We investigated the capabilities of a hybrid genetic algorithm / simulated annealing technique to predict the low-energy conformational states of various sized proteins and to generate statistical distributions of the studied proteins' molecular ensemble for pKa predictions. Our algorithm produced errors to experimental results within .acceptable margins and offered considerable speed up depending on the protein and on the rotameric states' resolution used.

Improving sampling, optimization and feature extraction in Boltzmann machines

L’apprentissage supervisé de réseaux hiérarchiques à grande échelle connaît présentement un succès fulgurant. Malgré cette effervescence, l’apprentissage non-supervisé représente toujours, selon plusieurs chercheurs, un élément clé de l’Intelligence Artificielle, où les agents doivent apprendre à partir d’un nombre potentiellement limité de données. Cette thèse s’inscrit dans cette pensée et aborde divers sujets de recherche liés au problème d’estimation de densité par l’entremise des machines de Boltzmann (BM), modèles graphiques probabilistes au coeur de l’apprentissage profond. Nos contributions touchent les domaines de l’échantillonnage, l’estimation de fonctions de partition, l’optimisation ainsi que l’apprentissage de représentations invariantes. Cette thèse débute par l’exposition d’un nouvel algorithme d'échantillonnage adaptatif, qui ajuste (de fa ̧con automatique) la température des chaînes de Markov sous simulation, afin de maintenir une vitesse de convergence élevée tout au long de l’apprentissage. Lorsqu’utilisé dans le contexte de l’apprentissage par maximum de vraisemblance stochastique (SML), notre algorithme engendre une robustesse accrue face à la sélection du taux d’apprentissage, ainsi qu’une meilleure vitesse de convergence. Nos résultats sont présent ́es dans le domaine des BMs, mais la méthode est générale et applicable à l’apprentissage de tout modèle probabiliste exploitant l’échantillonnage par chaînes de Markov. Tandis que le gradient du maximum de vraisemblance peut-être approximé par échantillonnage, l’évaluation de la log-vraisemblance nécessite un estimé de la fonction de partition. Contrairement aux approches traditionnelles qui considèrent un modèle donné comme une boîte noire, nous proposons plutôt d’exploiter la dynamique de l’apprentissage en estimant les changements successifs de log-partition encourus à chaque mise à jour des paramètres. Le problème d’estimation est reformulé comme un problème d’inférence similaire au filtre de Kalman, mais sur un graphe bi-dimensionnel, où les dimensions correspondent aux axes du temps et au paramètre de température. Sur le thème de l’optimisation, nous présentons également un algorithme permettant d’appliquer, de manière efficace, le gradient naturel à des machines de Boltzmann comportant des milliers d’unités. Jusqu’à présent, son adoption était limitée par son haut coût computationel ainsi que sa demande en mémoire. Notre algorithme, Metric-Free Natural Gradient (MFNG), permet d’éviter le calcul explicite de la matrice d’information de Fisher (et son inverse) en exploitant un solveur linéaire combiné à un produit matrice-vecteur efficace. L’algorithme est prometteur: en terme du nombre d’évaluations de fonctions, MFNG converge plus rapidement que SML. Son implémentation demeure malheureusement inefficace en temps de calcul. Ces travaux explorent également les mécanismes sous-jacents à l’apprentissage de représentations invariantes. À cette fin, nous utilisons la famille de machines de Boltzmann restreintes “spike & slab” (ssRBM), que nous modifions afin de pouvoir modéliser des distributions binaires et parcimonieuses. Les variables latentes binaires de la ssRBM peuvent être rendues invariantes à un sous-espace vectoriel, en associant à chacune d’elles, un vecteur de variables latentes continues (dénommées “slabs”). Ceci se traduit par une invariance accrue au niveau de la représentation et un meilleur taux de classification lorsque peu de données étiquetées sont disponibles. Nous terminons cette thèse sur un sujet ambitieux: l’apprentissage de représentations pouvant séparer les facteurs de variations présents dans le signal d’entrée. Nous proposons une solution à base de ssRBM bilinéaire (avec deux groupes de facteurs latents) et formulons le problème comme l’un de “pooling” dans des sous-espaces vectoriels complémentaires.

Policy Improvement for POMDPs Using Normalized Importance Sampling

We present a new method for estimating the expected return of a POMDP from experience. The estimator does not assume any knowle ge of the POMDP and allows the experience to be gathered with an arbitrary set of policies. The return is estimated for any new policy of the POMDP. We motivate the estimator from function-approximation and importance sampling points-of-view and derive its theoretical properties. Although the estimator is biased, it has low variance and the bias is often irrelevant when the estimator is used for pair-wise comparisons.We conclude by extending the estimator to policies with memory and compare its performance in a greedy search algorithm to the REINFORCE algorithm showing an order of magnitude reduction in the number of trials required.