Deep Learning and Complex Network Analysis for uncovering epistatic interactions underlying complex phenotypic traits

The integration of Darwinian evolution and Mendelian genetics in the early 20th century provided the conceptual bases for the study of how specific genetic differences, known as genotypes, are linked to observable traits or characteristics, known as phenotypes. However, advances in our understanding of this relationship were limited by genotyping techniques that could only examine a small subset of known genetic markers. The emergence of next generation sequencing in the early 2000's led to so-called genome-wide studies focused on identifying individual variants associated with specific phenotypes. Such studies helped identify several variants associated with important phenotypic traits in plant and animal breeds. It also uncovered many variants linked to disease in humans. However much of the variation underlying these traits remains unexplained. Moreover, the increase in size of genomic datasets and complexity of traits being studied represent important challenges to the statistical frameworks in use. It is now clear that complex phenotypic traits may be determined not only by many genes of small effect but also by so-called epistatic interactions among them. Some progress has been made in detecting interactions among a small number of variants but the role of high-order epistatic interactions still needs to be addressed. Thus, the challenge today is to develop new methods of analysis that can scale up to modern population genomics databases and uncover interactions between many genetic variants. Our project addresses these challenges by harnessing the power of deep learning (DL) methods and complex network analysis (CNA) to develop an end-to-end computational tool to associate causal genetic variants to a phenotype of interest and also detect underlying epistatic interactions. Our approach will go beyond pairwise gene-to-gene interactions and study higher-order interactions. We will implement DL models that scale up to high-dimensional input and learn complex nonlinear interaction patterns, which can then be unveiled using the latest advances in explainable and interpretable machine learning approaches. Once important variants and potential low dimensional interactions are identified, CNA techniques will allow us to explore higher-order interactions using an enormous range of new analysis methods that are unavailable in lower-order settings. Our approach will help identify essential genes (as network hubs), gene clusters with similar functionalities, and genes with suppressing and augmenting effects for a specific phenotype. Our proposed framework can reveal how genes drive biological functions and contribute to diversity, health, and disease and will be applicable in a wide range of domains. In evolutionary biology, it will provide new insights into how certain traits evolved and adapted in response to environmental pressures. This new knowledge is crucial for the design of efficient animal and plant breeding programs aimed at increasing disease resistance and adaptation to global change; as well as for conservation efforts and for predicting how populations may respond to changing environments. It will also help to uncover the genetic basis of complex human diseases, which may lead to new diagnostic tools, personalised medicine, and targeted therapies. We aim to develop proof-of-concept and identify the right techniques. This, in turn, will allow us in the near future to develop an open-sourced library for geneticists and evolutionary biologists interested in understanding the genetic architecture of complex traits, which will facilitate new genetic discoveries.