Predicting VNN resistance in European sea bass using machine learning on high dimensional low sample size data

Aquaculture is a rapidly growing sector in the global food production chain as a recognized fundamental source of high-quality proteins. One of the crucial tasks in aquaculture is phenotype prediction. While machine learning research has mainly focused on classification tasks on Big Data, in many bioinformatics applications, including aquaculture, the real challenge behind prediction problems is dealing with small sample and high-dimensional data. In such contexts, it is in fact common that the number of genetic features (such as SNPs) far exceeds the sample size. As a test case, this study focuses on the prediction of resistance to Viral Nervous Necrosis(VNN) from a population of European sea bass. We explore a range of machine learning techniques, from established methods such as Support Vector Machines and Gradient Boosting, to increasingly popular Deep Learning Approaches, also including a variant of image-based classification based on Chaos Game Representation. Besides standard training-test partitioning, we also considered a more challenging partition of the dataset that maximize the genomic distance among training and testing set to better reflect the kind of generalization problem encountered in breeding practice due to data scarcity typical of non-model species. Although all the animals belong to the same population, this approach offered the most appropriate way to ensure the procedure was sufficiently challenging given the available data. We assessed the performance of learning approaches in different scenarios, reducing the data dimensionality by selecting SNPs on the basis of functional information. Our experiments confirmed the difficult nature of this association task. However, each tested tool showed promising results in at least one scenario. While predicting disease susceptibility remains a challenging task for breeding programs, within the boundaries of the tested scenarios, our results show that machine learning approaches, combined with a controlled amount of additional functional information, can help mitigate the issues arising from high dimensional, low sample size datasets typical in the study of non-model species.