Functional characterization of single amino acis variats
- Autores: Víctor López Ferrando
- Directores de la Tesis: Josep Lluis Gelpi Buchaca (dir. tes.), Modesto Orozco López (codir. tes.)
- Lectura: En la Universitat de Barcelona ( España ) en 2019
- Idioma: inglés
- Tribunal Calificador de la Tesis: Patrick Aloy Calaf (presid.), Fátima Al-Shahrour (secret.), María Jesús Martín Fernández (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Single amino acid variants (SAVs) are one of the main causes of Mendelian disorders, and play an important role in the development of many complex diseases. At the same time, they are the most common kind of variation affecting coding DNA, without generally presenting any damaging effect. With the advent of next generation sequencing technologies, the detection of these variants in patients and the general population is easier than ever, but the characterization of the functional effects of each variant remains an open challenge. It is our objective in this work to tackle this problem by developing machine learning based in silico SAVs pathology predictors. Having the PMut classic predictor as a starting point, we have rethought the entire supervised learning pipeline, elaborating new training sets, features and classifiers. PMut2017 is the first result of these efforts, a new general-purpose predictor based on SwissVar and trained on 12 different conservation scores. Its performance, evaluated bothby cross-validation and different blind tests, was in line with the best predictors published to date. Continuing our efforts in search for more accurate predictors, especially for those cases were general predictors tend to fail, we developed PMut-S, a suite of 215 protein-specific predictors. Similar to PMut in nature, PMut-S introduced the use of co-evolution conservation features and balanced training sets, and showed improved performance, specially for those proteins that were more commonly misclassified by PMut. Comparing PMut-S to other specific predictors we proved that it is possible to train specific predictors using a unique automated pipeline and match the results of most gene specific predictors released to date. The implementation of the machine learning pipeline of both PMut and PMut-S was released as an open source Python module: PyMut, which bundles functions implementing the features computation and selection, classifier train- ing and evaluation, plots drawing, among others. Their predictions were also made available in a rich web portal, which includes a precomputed repository with analyses of more than 700 million variants on over 100,000 human pro- teins, together with relevant contextual information such as 3D visualizations of protein structures, links to databases, functional annotations, and more.
