Evaluation and Improvement of Genome-Wide Human Protein-Protein Interaction PredictionDUNHAM, BRANDAN (2023) Evaluation and Improvement of Genome-Wide Human Protein-Protein Interaction Prediction. Doctoral Dissertation, University of Pittsburgh. (Unpublished)
AbstractBackground: Proteins are biological macromolecules that interact with each other, performing various functions and regulate many biological processes, making them vital to many types of biological research. However, as many protein-protein interactions (PPIs) remain unknown, and interacting protein pairs are rare among all protein pairs, it is important for researchers to find ways to predict novel interactors with high precision, reducing experimental costs by prioritizing likely interactors. Methods: In this thesis, we evaluate thirty-six previously published methods, and assess their suitability for predicting novel interactions. We analyze the ability of these methods to predict PPIs of proteins not used during training. This avoids a problem we hypothesized may exist in most methods, especially those that rely on protein sequence derived features. Similarly, we hypothesized removing this problem could yield better, more generalizable predictions when using annotation-based features for predicting interactions. Results: In our analyses, we found that most sequence-based models were unable to accurately predict interactions where the proteins were not in the training set. We obtained better results when using features that did not rely on primary sequence information, and showed that the models that performed well on unseen proteins were better at predicting proteome-wide interactions. Discussion: Our results show that models generated to maximize precision when predicting on protein pairs composed of proteins not used during training are better at making predictions proteome-wide. These models predict more validated PPIs from other data sources, and are less biased towards predicting hubs, than models trained in the traditional way. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}