Neural networks have almost always been treated as black boxes, as models that are able to capture the highly nonlinear relationships that exists among inputs that lead to a correct output with no clear idea about how to get an interpretation of how such relationship are learnt. However, feature importance and input correlation are crucial in certain domains, such as bioinformatics and medicine, in which researchers are not only interested to get a correct output, but also to understand the reason why the network output such results. In this spirit, the aim of this project is to explore techniques for neural network interpretability and to apply those techniques in the domain of bioinformatics, particularly in genomics, hopefully leading to new insights both in neural networks and bioinformatics domains, showing that deep learning can be an effective tool for genomics and that the obtained results can be interpreted, fulfilling the needs of medical doctors and geneticists.
The successful outcome of the research can lead to advancements in two scientific sectors: neural networks and bioinformatics.
Since neural networks are mostly considered as black boxes, having demonstrated that the interpretability methods are effective even in a specific field, takes us a step closer to open the not-so-black box of neural network.
In bioinformatics, if this research is able to find a neural network model able to deliver a high accuracy, it can help into the detection of diseases in people and consequently lead to therapies that can enhance people's life.
The interpretation of the models and the understanding of the most important features, and thus the genetic variants related to the disease, can make the work of medical doctors faster; they can concentrate their medical studies on the variants that the network considers to be more important, and thus more likely to be related to the trait under study. This can lead to the discovery of new variants that were previously not known to be trait-related.
Furthermore, the ability of neural networks to perform feature selection will convey a speedup to the current way of studying the association between genetic variants and diseases; instead of performing regression on single variants to find correlations with a trait, such relevant variants will be found by the network itself and uncovered through interpretabily.
In conclusion, the success of this research can represent a breakthrough in two important and hot topics nowadays: neural network interpretability and bioinformatics. Using a neural network approach in bioinformatics (in which feature selection is crucial to understand which genetic variants are related to a disease), can lead to new important contributions in opening the neural networks black box while finding answer to bioinformatics problems that can improve the health of the people.
Due to the lack of gold standards nowadays for deep learning models and interpretability techniques in bioinformatics, this research could help into both defining standards for the application of neural networks in this field and understanding which are the most effective techniques for interpretability to use in such specific domain.