Identificación de blancos moleculares para el tratamiento de alzheimer mediante interacciones proteína-proteína (ppi) y algoritmos de deep learning.
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La enfermedad de Alzheimer (EA) carece de terapias capaces de detener o revertir su progresión, por lo que priorizar blancos moleculares a partir de redes de interacción proteína–proteína (PPI) resulta clave. Este trabajo tuvo por objetivo detectar módulos funcionales en una red PPI asociada a EA e integrar dicha estructura con anotaciones de Gene Ontology (GO) para proponer candidatos a blanco. Para ello, se desarrolló un pipeline basado en redes neuronales del tipo Graph Neural Network (GNN) aplicado a una red de 5,390 proteínas y 430,206 interacciones, anotada con 6,928 términos GO.
La metodología consistió en la construcción de representaciones de nodos y aristas que permitieron entrenar un modelo basado en redes neuronales gráficas. Posteriormente, se generaron embeddings que fueron analizados mediante estrategias de reducción de dimensionalidad y agrupamiento, lo que posibilitó la identificación de módulos funcionales y su posterior caracterización biológica.
En cuanto a los resultados, el modelo seleccionado alcanzó un AUC de 0.9596 en el conjunto de prueba y permitió la detección de 14 módulos funcionales (denominados clusters) consistentes. Adicionalmente, se calcularon métricas de conectividad intra- e interclúster junto con el Connectivity Score. A partir de estas métricas se identificaron las cinco proteínas con los valores más altos en cada cluster, conformando así el conjunto final de 145 proteínas priorizadas para su estudio posterior.
En conclusión, el enfoque propuesto constituye una estrategia computacional para la integración de información topológica y funcional en redes PPI asociadas a la EA. Los hallazgos no sólo generan una lista priorizada de posibles blancos moleculares, sino que también establecen una base metodológica extensible a otras patologías, reforzando el potencial del aprendizaje profundo aplicado a problemas biológicos.
Alzheimer’s disease (AD) lacks therapies capable of halting or reversing its progression, making the prioritization of molecular targets from protein–protein interaction (PPI) networks a key step. The objective of this work was to detect functional modules in a PPI network associated with AD and to integrate this structure with Gene Ontology (GO) annotations in order to propose candidate targets. To this end, a pipeline was developed based on Graph Neural Networks (GNN), applied to a network of 5,390 proteins and 430,206 interactions, annotated with 6,928 GO terms. The methodology consisted of building node and edge representations to train a graph neural network model. Subsequently, embeddings were generated and analyzed through dimensionality reduction and clustering strategies, enabling the detection of functional modules and their subsequent biological characterization. Regarding the results, the selected model achieved an AUC of 0.9596 on the test set and allowed the detection of 14 consistent functional modules (denominated clusters). Additionally, intra- and inter-cluster connectivity metrics were computed together with the Connectivity Score. Based on these metrics, the five proteins with the highest values in each case were identified, thus constituting the final set of 145 prioritized proteins for further study. In conclusion, the proposed approach represents a robust computational strategy for integrating topological and functional information in PPI networks associated with AD. The findings not only provide a prioritized list of potential molecular targets, but also establish a methodological foundation that can be extended to other pathologies, reinforcing the potential of deep learning applied to biological problems.
Alzheimer’s disease (AD) lacks therapies capable of halting or reversing its progression, making the prioritization of molecular targets from protein–protein interaction (PPI) networks a key step. The objective of this work was to detect functional modules in a PPI network associated with AD and to integrate this structure with Gene Ontology (GO) annotations in order to propose candidate targets. To this end, a pipeline was developed based on Graph Neural Networks (GNN), applied to a network of 5,390 proteins and 430,206 interactions, annotated with 6,928 GO terms. The methodology consisted of building node and edge representations to train a graph neural network model. Subsequently, embeddings were generated and analyzed through dimensionality reduction and clustering strategies, enabling the detection of functional modules and their subsequent biological characterization. Regarding the results, the selected model achieved an AUC of 0.9596 on the test set and allowed the detection of 14 consistent functional modules (denominated clusters). Additionally, intra- and inter-cluster connectivity metrics were computed together with the Connectivity Score. Based on these metrics, the five proteins with the highest values in each case were identified, thus constituting the final set of 145 prioritized proteins for further study. In conclusion, the proposed approach represents a robust computational strategy for integrating topological and functional information in PPI networks associated with AD. The findings not only provide a prioritized list of potential molecular targets, but also establish a methodological foundation that can be extended to other pathologies, reinforcing the potential of deep learning applied to biological problems.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Informático/a.
Keywords
Enfermedad de Alzheimer, Algoritmos, Deep learning (Machine learning)