Publicação
Machine learning methods for quantitative structure-property relationship modeling
| Resumo: | Due to the high rate of new compounds discovered each day and the morosity/cost of experimental measurements there will always be a significant gap between the number of known chemical compounds and the amount of chemical compounds for which experimental properties are available. This research work is motivated by the fact that the development of new methods for predicting properties and organize huge collections of molecules to reveal certain chemical categories/patterns and select diverse/representative samples for exploratory experiments are becoming essential. This work aims to increase the capability to predict physical, chemical and biological properties, using data mining methods applied to complex non-homogeneous data (chemical structures), for large information repositories. In the first phase of this work, current methodologies in quantitative structure-property modelling were studied. These methodologies attempt to relate a set of selected structure-derived features of a compound to its property using model-based learning. This work focused on solving major issues identified when predicting properties of chemical compounds and on the solutions explored using different molecular representations, feature selection techniques and data mining approaches. In this context, an innovative hybrid approach was proposed in order to improve the prediction power and comprehensibility of QSPR/QSAR problems using Random Forests for feature selection. It is acknowledged that, in general, similar molecules tend to have similar properties; therefore, on the second phase of this work, an instance-based machine learning methodology for predicting properties of compounds using the similarity-based molecular space was developed. However, this type of methodology requires the quantification of structural similarity between molecules, which is often subjective, ambiguous and relies upon comparative judgements, and consequently, there is currently no absolute standard of molecular similarity. In this context, a new similarity method was developed, the non-contiguous atom matching (NAMS), based on the optimal atom alignment using pairwise matching algorithms that take into account both topological profiles and atoms/bonds characteristics. NAMS can then be used for property inference over the molecular metric space using ordinary kriging in order to obtain robust and interpretable predictive results, providing a better understanding of the underlying relationship structure-property. |
|---|---|
| Autores principais: | Teixeira, Ana Isabel Lino, 1986- |
| Assunto: | Bioinformática Estrutura molecular Aprendizagem automática Dados - informação Teses de doutoramento - 2014 |
| Ano: | 2014 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Due to the high rate of new compounds discovered each day and the morosity/cost of experimental measurements there will always be a significant gap between the number of known chemical compounds and the amount of chemical compounds for which experimental properties are available. This research work is motivated by the fact that the development of new methods for predicting properties and organize huge collections of molecules to reveal certain chemical categories/patterns and select diverse/representative samples for exploratory experiments are becoming essential. This work aims to increase the capability to predict physical, chemical and biological properties, using data mining methods applied to complex non-homogeneous data (chemical structures), for large information repositories. In the first phase of this work, current methodologies in quantitative structure-property modelling were studied. These methodologies attempt to relate a set of selected structure-derived features of a compound to its property using model-based learning. This work focused on solving major issues identified when predicting properties of chemical compounds and on the solutions explored using different molecular representations, feature selection techniques and data mining approaches. In this context, an innovative hybrid approach was proposed in order to improve the prediction power and comprehensibility of QSPR/QSAR problems using Random Forests for feature selection. It is acknowledged that, in general, similar molecules tend to have similar properties; therefore, on the second phase of this work, an instance-based machine learning methodology for predicting properties of compounds using the similarity-based molecular space was developed. However, this type of methodology requires the quantification of structural similarity between molecules, which is often subjective, ambiguous and relies upon comparative judgements, and consequently, there is currently no absolute standard of molecular similarity. In this context, a new similarity method was developed, the non-contiguous atom matching (NAMS), based on the optimal atom alignment using pairwise matching algorithms that take into account both topological profiles and atoms/bonds characteristics. NAMS can then be used for property inference over the molecular metric space using ordinary kriging in order to obtain robust and interpretable predictive results, providing a better understanding of the underlying relationship structure-property. |
|---|