Application of machine learning in systems biology

Doctoral thesis, 2020

Biological systems are composed of a large number of molecular components. Understanding their behavior as a result of the interactions between the individual components is one of the aims of systems biology. Computational modelling is a powerful tool commonly used in systems biology, which relies on mathematical models that capture the properties and interactions between molecular components to simulate the behavior of the whole system. However, in many biological systems, it becomes challenging to build reliable mathematical models due to the complexity and the poor understanding of the underlying mechanisms. With the breakthrough in big data technologies in biology, data-driven machine learning (ML) approaches offer a promising complement to traditional theory-based models in systems biology. Firstly, ML can be used to model the systems in which the relationships between the components and the system are too complex to be modelled with theory-based models. Two such examples of using ML to resolve the genotype-phenotype relationships are presented in this thesis: (i) predicting yeast phenotypes using genomic features and (ii) predicting the thermal niche of microorganisms based on the proteome features. Secondly, ML naturally complements theory-based models. By applying ML, I improved the performance of the genome-scale metabolic model in describing yeast thermotolerance. In this application, ML was used to estimate the thermal parameters by using a Bayesian statistical learning approach that trains regression models and performs uncertainty quantification and reduction. The predicted bottleneck genes were further validated by experiments in improving yeast thermotolerance.

In such applications, regression models are frequently used, and their performance relies on many factors, including but not limited to feature engineering and quality of response values. Manually engineering sufficient relevant features is particularly challenging in biology due to the lack of knowledge in certain areas. With the increasing volume of big data, deep-transfer learning enables us to learn a statistical summary of the samples from a big dataset which can be used as input to train other ML models. In the present thesis, I applied this approach to first learn a deep representation of enzyme thermal adaptation and then use it for the development of regression models for predicting enzyme optimal and protein melting temperatures. It was demonstrated that the transfer learning-based regression models outperform the classical ones trained on rationally engineered features in both cases. On the other hand, noisy response values are very common in biological datasets due to the variation in experimental measurements and they fundamentally restrict the performance attainable with regression models. I thereby addressed this challenge by deriving a theoretical upper bound for the coefficient of determination (R2) for regression models. This theoretical upper bound depends on the noise associated with the response variable and variance for a given dataset. It can thus be used to test whether the maximal performance has been reached on a particular dataset, or whether further model improvement is possible.