Protein Link Augmentation for Functional Prediction: Combining Spectral and Machine Learning Techniques

Abstract

MS Thesis Defense

Recent advances in network-based methods for functional annotation of proteins have proved effective in well annotated species. However, network information is lacking in many species, so these methods cannot be effectively applied. Some methods, such as MUNK and MUNDO, attempt a co-embedding of well and lesser annotated species to boost performance. We explore the use of ML based approaches to augment these networks, with the intent of increasing the performance of co-embedding methods. A model that can predict interactions between two proteins was used to add potentially missing interactions or remove existing false interactions. We found that too much noise is added to the network to be useful. This does not rule out future endeavors in the area or specific applications of ML in instances where the model is known to perform well, only that generalized models are currently not capable of protein-protein interaction network augmentations.