Horizons in Molecular Biology 2017: Applications of Artificial Intelligence to the Prediction of Gene Function

One of the challenges of the post genomic era is to predict the function of a gene or protein given its sequence. Most automated function prediction methods rely upon identifying well annotated sequence and structural homologues to transfer annotations to uncharacterized proteins. Sequence similarity based methods are relatively successful at annotating closely related proteins, however, they are not applicable to annotating orphan proteins or proteins whose relatives are themselves not functionally characterised. Currently, around 30% of the human genome cannot be functionally identified by standard bioinformatics techniques based on homology-transfer, which highlights the urgent need for new function prediction methods that are able to work on these hard cases.

Given the complexity of both describing and predicting gene function, it is not surprising that machine learning based approaches have gained wide attention recently, owing to the fact that substantial knowledge about protein function has been steadily accumulated in public databases, and machine learning technology has been advancing in leaps and bounds over the past few years, especially in the area of so-called “deep learning”. To help monitor progress in the field, there have also been three worldwide Critical Assessment of Function Annotation (CAFA) challenges to evaluate protein function prediction methods, and I will discuss what we have learned from our participation in these challenges, along with a more general overview of my lab’s work on trying to solve this problem and a brief look to future developments.

Free registration at www.horizons.uni-goettingen.de