Predicting single-cell perturbations response with prior biological knowledge graphs
Visualitza/Obre
Estadístiques de LA Referencia / Recolecta
Inclou dades d'ús des de 2022
Cita com:
hdl:2117/372832
Realitzat a/ambUniversity of Oxford
Tipus de documentProjecte Final de Màster Oficial
Data2022-06-28
Condicions d'accésAccés obert
Tots els drets reservats. Aquesta obra està protegida pels drets de propietat intel·lectual i
industrial corresponents. Sense perjudici de les exempcions legals existents, queda prohibida la seva
reproducció, distribució, comunicació pública o transformació sense l'autorització del titular dels drets
Abstract
The recent advances in single-cell experimental technologies have opened the door to a broad study of cell perturbations such as drugs or gene knock- outs. Knowing how a cell would respond to a certain perturbation can boost drug discovery field accelerating the development of new drugs and therapies. Nevertheless, the perturbation search space is so large that an exhaustive classical search is not feasible. For this reason, computational methods to predict the perturbation response and guide the search must be developed. Those computational methods must be able to work out-of- distribution and predict the behaviour of cells under unknown perturbations, since not possible combinations of perturbations can be seen in training time. We hypothesise that prior biological existing knowledge can help current Deep Learning systems to perform better OOD and generalize. To do so, in this thesis we present a system that incorporates prior biological knowledge into Deep Learning systems structuring the data using Gene Regulatory Networks (GRN) and feeding these data to Graph Neural Networks (GNN). We explore, in a biological problem, the idea of using prior existing knowledge about the nature of the system to regularize the models in such a way that their OOD performance improves. We propose different architectures: from ones that trust exclusively in prior knowledge graph structured data to others that merge prior knowledge-driven embeddings and tabular data embeddings. We show that, unfortunately, leveraging GRN to encode the data in such a way that prior knowledge is exploited is useful for doing in-distribution predictions but it is not for OOD settings. Finally, we point out at the current state of the existing prior knowledge as the main bottleneck of the performance of the system.
TitulacióMÀSTER UNIVERSITARI EN INNOVACIÓ I RECERCA EN INFORMÀTICA (Pla 2012)
Col·leccions
Fitxers | Descripció | Mida | Format | Visualitza |
---|---|---|---|---|
169967.pdf | 7,102Mb | Visualitza/Obre |