In drug discovery, thousands of compounds are assayed to detect activity against a biological target. The goal of drug discovery is to identify compounds that are active against the target (e.g. inhibit a virus). Statistical learning in drug discovery seeks to build a model that uses descriptors characterizing molecular structure to predict biological activity. However, the characteristics of drug discovery data can make it difficult to model the relationship between molecular descriptors and biological activity. Among these characteristics are the rarity of active compounds, the large volume of compounds tested by high-throughput screening, and the complexity of molecular structure and its relationship to activity. This thesis focuses on the design of statistical learning algorithms/models and their applications to drug discovery. The two main parts of the thesis are: an algorithm-based statistical method and a more formal model-based approach. Both approaches can facilitate and accelerate the process of developing new drugs. A unifying theme is the use of unsupervised methods as components of supervised learning algorithms/models. In the first part of the thesis, we explore a sequential screening approach, Cluster Structure-Activity Relationship Analysis (CSARA). Sequential screening integrates High Throughput Screening with mathematical modeling to sequentially select the best compounds. CSARA is a cluster-based and algorithm driven method. To gain further insight into this method, we use three carefully designed experiments to compare predictive accuracy with Recursive Partitioning, a popular structureactivity relationship analysis method. The experiments show that CSARA outperforms Recursive Partitioning. Comparisons include problems with many descriptor sets and situations in which many descriptors are not important for activity. In the second part of the thesis, we propose and develop constrained mixture discriminant analysis (CMDA), a model-based method. The main idea of CMDA is to model the distribution of the observations given the class label (e.g. active or inactive class) as a constrained mixture distribution, and then use Bayes’ rule to predict the probability of being active for each observation in the testing set. Constraints are used to deal with the otherwise explosive growth of the number of parameters with increasing dimensionality. CMDA is designed to solve several challenges in modeling drug data sets, such as multiple mechanisms, the rare target problem (i.e. imbalanced classes), and the identification of relevant subspaces of descriptors (i.e. variable selection). We focus on the CMDA1 model, in which univariate densities form the building blocks of the mixture components. Due to the unboundedness of the CMDA1 log likelihood function, it is easy for the EM algorithm to converge to degenerate solutions. A special Multi-Step EM algorithm is therefore developed and explored via several experimental comparisons. Using the multi-step EM algorithm, the CMDA1 model is compared to model-based clustering discriminant analysis (MclustDA). The CMDA1 model is either superior to or competitive with the MclustDA model, depending on which model generates the data. The CMDA1 model has better performance than the MclustDA model when the data are high-dimensional and unbalanced, an essential feature of the drug discovery problem! An alternate approach to the problem of degeneracy is penalized estimation. By introducing a group of simple penalty functions, we consider penalized maximum likelihood estimation of the CMDA1 and CMDA2 models. This strategy improves the convergence of the conventional EM algorithm, and helps avoid degenerate solutions. Extending techniques from Chen et al. (2007), we prove that the PMLE’s of the two-dimensional CMDA1 model can be asymptotically consistent.

In drug discovery, thousands of compounds are assayed to detect activity against a biological target. The goal of drug discovery is to identify compounds that are active against the target (e.g. inhibit a virus). Statistical learning in drug discovery seeks to build a model that uses descriptors characterizing molecular structure to predict biological activity. However, the characteristics of drug discovery data can make it difficult to model the relationship between molecular descriptors and biological activity. Among these characteristics are the rarity of active compounds, the large volume of compounds tested by high-throughput screening, and the complexity of molecular structure and its relationship to activity. This thesis focuses on the design of statistical learning algorithms/models and their applications to drug discovery. The two main parts of the thesis are: an algorithm-based statistical method and a more formal model-based approach. Both approaches can facilitate and accelerate the process of developing new drugs. A unifying theme is the use of unsupervised methods as components of supervised learning algorithms/models. In the first part of the thesis, we explore a sequential screening approach, Cluster Structure-Activity Relationship Analysis (CSARA). Sequential screening integrates High Throughput Screening with mathematical modeling to sequentially select the best compounds. CSARA is a cluster-based and algorithm driven method. To gain further insight into this method, we use three carefully designed experiments to compare predictive accuracy with Recursive Partitioning, a popular structureactivity relationship analysis method. The experiments show that CSARA outperforms Recursive Partitioning. Comparisons include problems with many descriptor sets and situations in which many descriptors are not important for activity. In the second part of the thesis, we propose and develop constrained mixture discriminant analysis (CMDA), a model-based method. The main idea of CMDA is to model the distribution of the observations given the class label (e.g. active or inactive class) as a constrained mixture distribution, and then use Bayes’ rule to predict the probability of being active for each observation in the testing set. Constraints are used to deal with the otherwise explosive growth of the number of parameters with increasing dimensionality. CMDA is designed to solve several challenges in modeling drug data sets, such as multiple mechanisms, the rare target problem (i.e. imbalanced classes), and the identification of relevant subspaces of descriptors (i.e. variable selection). We focus on the CMDA1 model, in which univariate densities form the building blocks of the mixture components. Due to the unboundedness of the CMDA1 log likelihood function, it is easy for the EM algorithm to converge to degenerate solutions. A special Multi-Step EM algorithm is therefore developed and explored via several experimental comparisons. Using the multi-step EM algorithm, the CMDA1 model is compared to model-based clustering discriminant analysis (MclustDA). The CMDA1 model is either superior to or competitive with the MclustDA model, depending on which model generates the data. The CMDA1 model has better performance than the MclustDA model when the data are high-dimensional and unbalanced, an essential feature of the drug discovery problem! An alternate approach to the problem of degeneracy is penalized estimation. By introducing a group of simple penalty functions, we consider penalized maximum likelihood estimation of the CMDA1 and CMDA2 models. This strategy improves the convergence of the conventional EM algorithm, and helps avoid degenerate solutions. Extending techniques from Chen et al. (2007), we prove that the PMLE’s of the two-dimensional CMDA1 model can be asymptotically consistent.

京都大学 / 新制・課程博士 / 博士(理学) / 甲第23699号 / 理博第4789号 / 新制||理||1686(附属図書館) / 京都大学大学院理学研究科物理学・宇宙物理学専攻 / (主査)教授 永江 知文, 准教授 銭廣 十三, 教授 中家 剛 / 学位規則第4条第1項該当 / Doctor of Science / Kyoto University / DGAM

京都大学 / 0048 / 新制・課程博士 / 博士(理学) / 甲第20897号 / 理博第4349号 / 新制||理||1624(附属図書館) / 京都大学大学院理学研究科物理学・宇宙物理学専攻 / (主査)准教授 川畑 貴裕, 教授 永江 知文, 教授 鶴 剛 / 学位規則第4条第1項該当 / Doctor of Science / Kyoto University / DGAM

O modelo de cluster-alfa é aplicado aos núcleos de massa intermediária 90Sr, 92Zr, 94Mo, 96Ru e 98Pd com a Abordagem de Potencial Local. As bandas do estado fundamental dos respectivos sistemas alfa + caroço são calculadas com um único parâmetro variável, fornecendo uma boa descrição geral dos níveis experimentais. Mostra-se que o potencial alfa + caroço é fracamente dependente do momento angular L, e que tal dependência pode ser descrita de forma simples e padronizada para os cinco núcleos. O comportamento do parâmetro radial R do potencial alfa + caroço é discutido em relação ao raio do núcleo total e a soma dos raios do cluster-alfa e do caroço. As taxas de transição B(E2) reproduzem corretamente as ordens de grandeza de quase todos os dados experimentais sem o uso de cargas efetivas. A análise das separações intercluster rms e das larguras-alfa reduzidas nas bandas do estado fundamental sugere que há uma redução da intensidade de aglomeração-alfa com o aumento do spin. Uma análise complementar das bandas de estado fundamental dos núcleos 20Ne, 44Ti e 212Po aponta uma diminuição da intensidade de aglomeração-alfa com o aumento da massa nuclear, e mostra uma condição mais fraca de aglomeração-alfa para os núcleos da região do Mo em comparação com os núcleos leves. Bandas de paridade negativa são calculadas para os núcleos 92Zr, 94Mo, 96Ru e 98Pd e informações não relatadas anteriormente são comparadas a níveis experimentais disponíveis. A avaliação geral dos resultados indica que os núcleos com A par e N=52 na região do Mo possuem estruturas alfa + caroço com características semelhantes. / The alpha-cluster model is applied to the intermediate mass nuclei 90Sr, 92Zr, 94Mo, 96Ru and 98Pd with the Local Potential Approach. The ground state bands of the respective alpha + core systems are calculated with only one variable parameter, giving a good general description of the experimental data. It is shown that the alpha + core potential is weakly dependent on the angular momentum L and such dependence may be described in a simple and standardized form for the five nuclei. The behavior of the radial parameter R of the alpha + core potential is discussed in relation to the radius of the total nucleus and the sum of the radii of the alpha-cluster and the core. The calculated B(E2) transition rates reproduce correctly the orders of magnitude of almost all experimental data without the use of effective charges. The analysis of the rms intercluster separations and the reduced alpha-widths for the ground state bands suggests a reduction of the alpha-cluster intensity with the increasing spin. A complementary analysis of the ground state bands of the nuclei 20Ne, 44Ti and 212Po points to a decrease of the alpha-cluster intensity with the increasing nuclear mass, and shows a weaker alpha-cluster condition for the nuclei of the Mo region in comparison with the light nuclei. Negative parity bands are calculated for the nuclei 92Zr, 94Mo, 96Ru and 98Pd and previously not reported information are compared to available experimental levels. The general evaluation of the results indicates that the N=52 even-A nuclei in the Mo region have alpha + core structures with similar features.

O modelo de cluster-alfa é aplicado aos núcleos de massa intermediária 90Sr, 92Zr, 94Mo, 96Ru e 98Pd com a Abordagem de Potencial Local. As bandas do estado fundamental dos respectivos sistemas alfa + caroço são calculadas com um único parâmetro variável, fornecendo uma boa descrição geral dos níveis experimentais. Mostra-se que o potencial alfa + caroço é fracamente dependente do momento angular L, e que tal dependência pode ser descrita de forma simples e padronizada para os cinco núcleos. O comportamento do parâmetro radial R do potencial alfa + caroço é discutido em relação ao raio do núcleo total e a soma dos raios do cluster-alfa e do caroço. As taxas de transição B(E2) reproduzem corretamente as ordens de grandeza de quase todos os dados experimentais sem o uso de cargas efetivas. A análise das separações intercluster rms e das larguras-alfa reduzidas nas bandas do estado fundamental sugere que há uma redução da intensidade de aglomeração-alfa com o aumento do spin. Uma análise complementar das bandas de estado fundamental dos núcleos 20Ne, 44Ti e 212Po aponta uma diminuição da intensidade de aglomeração-alfa com o aumento da massa nuclear, e mostra uma condição mais fraca de aglomeração-alfa para os núcleos da região do Mo em comparação com os núcleos leves. Bandas de paridade negativa são calculadas para os núcleos 92Zr, 94Mo, 96Ru e 98Pd e informações não relatadas anteriormente são comparadas a níveis experimentais disponíveis. A avaliação geral dos resultados indica que os núcleos com A par e N=52 na região do Mo possuem estruturas alfa + caroço com características semelhantes. / The alpha-cluster model is applied to the intermediate mass nuclei 90Sr, 92Zr, 94Mo, 96Ru and 98Pd with the Local Potential Approach. The ground state bands of the respective alpha + core systems are calculated with only one variable parameter, giving a good general description of the experimental data. It is shown that the alpha + core potential is weakly dependent on the angular momentum L and such dependence may be described in a simple and standardized form for the five nuclei. The behavior of the radial parameter R of the alpha + core potential is discussed in relation to the radius of the total nucleus and the sum of the radii of the alpha-cluster and the core. The calculated B(E2) transition rates reproduce correctly the orders of magnitude of almost all experimental data without the use of effective charges. The analysis of the rms intercluster separations and the reduced alpha-widths for the ground state bands suggests a reduction of the alpha-cluster intensity with the increasing spin. A complementary analysis of the ground state bands of the nuclei 20Ne, 44Ti and 212Po points to a decrease of the alpha-cluster intensity with the increasing nuclear mass, and shows a weaker alpha-cluster condition for the nuclei of the Mo region in comparison with the light nuclei. Negative parity bands are calculated for the nuclei 92Zr, 94Mo, 96Ru and 98Pd and previously not reported information are compared to available experimental levels. The general evaluation of the results indicates that the N=52 even-A nuclei in the Mo region have alpha + core structures with similar features.

[...] Mes travaux portent sur les clusters structurant le réseau et l'innovation car 1) le cluster impacte collectivement plutôt qu’individuellement la sortie du réseau, 2) les couplages intra et inter-cluster représentent la structure même des clusters mais ils influencent différemment l'innovation ou la croissance du cluster, 3) un certain compromis reste à définir entre la structure dense et éparse des différents réseaux. Un cluster est de façon générale défini comme un groupe de choses similaires ou de personnes qui travaillent sur des sujets analogues. Selon le domaine auquel il s’applique, même si l’idée reste la même, la définition s’affine. En sciences des organisations, un cluster représente un regroupement d’entreprises et d’institutions qui interagissent entre-elles par le biais de contrats, d’opérations formelles ou informelles et de réunions occasionnelles afin de contribuer collectivement à un résultat innovant. [...] La thèse est structurée comme suit. Dans l'introduction générale, nous passons en revue la littérature des connaissances existantes qui sert de base pour le cadre conceptuel des documents. Nous définissons ensuite certains concepts utilisés dans les trois articles présentés tels que la structure de réseau complexe (utilisée dans le premier article), l'innovation et les liens de réseau (utilisés principalement dans le deuxième article), et la gestion des connaissances utilisées (dans le troisième article). Dans le premier article, nous discutons les différents mécanismes de formation de liens dictés par les réseaux dirigés permettant de distinguer la distribution des degrés. Dans le deuxième article, nous abordons l'impact de la dynamique de groupe sur l'innovation du groupe de projet OSS. Dans le troisième article, nous nous attachons à l'impact du transfert des connaissances à l'intérieur des groupes sur le transfert des connaissances entre les groupes. L'annexe A permettra de discuter la modélisation analytique de la croissance des réseaux sociaux en utilisant la projection de réseaux multicouches ; l'annexe B sera l’occasion de présenter statistiquement le lien entre les relations intragroupe et les relations intergroupe. / [...] However, there is a gap in the literature with regard to the analysis of cluster or group structure as an input and cluster or group innovation as an output, e.g. “impact of network cluster structure on cluster innovation and growth”, i.e. how intra- and inter-cluster coupling, structural holes and tie strength impact cluster innovation and growth; and how intra-cluster density affects inter-cluster coupling; that I address in my thesis.Therefore, I focus on the cluster (or group of individuals) rather than the individual to analyze both network structure and innovation, because 1) clusters represent collective impact on network output rather than individuals’ impact, 2) intra and inter cluster couplings both represent cluster structure but have different impacts on cluster innovation and growth, 3) trade-offs among dense and sparse network cluster structures are different from those associated with networks of individuals. [...] The thesis is structured as follows. In the general introduction, I review the literature of existing knowledge in the field, which serves as a basis for the conceptual framework for the papers. I then define certain concepts used in the papers, such as complex network structure used in the first paper, innovation and network ties mainly used in the second paper, and knowledge management used in the third paper. In the first paper I discuss directed networks’ different link formation mechanisms causing degree distribution distinction. In the second paper, I discuss the impact of group dynamics on OSS project group innovation. In the third paper, I discuss impact of knowledge transfer inside groups onto knowledge transfer between groups. In appendix A, I discuss analytical modeling of social network growth using multilayer network projection; and in appendix B, I discuss statistically how intragroup ties and intergroup ties are related.