Characterising the influence of rule-based knowledge representations in biological knowledge extraction from transcriptomics data

Currently, there is a wealth of biotechnologies (e.g. sequencing, proteomics, lipidomics) able to generate a broad range of data types out of biological samples. However, the knowledge gained from such data sources is constrained by the limitations of the analytics techniques. The state-of-the-art machine learning algorithms are able to capture complex patterns with high prediction capacity. However, often it is very difficult if not impossible to extract human-understandable knowledge out of these patterns. In recent years evolutionary machine learning techniques have shown that they are competent methods for biological/biomedical data analytics. They are able to generate interpretable prediction models and, beyond just prediction models, they are able to extract useful knowledge in the form of biomarkers or biological networks.

This talk presents our efforts in thoroughly characterising the impact that a core component of the evolutionary machine learning process, its knowledge representations, has in the process of extracting biologically-useful knowledge out of transcriptomics datasets. Using the FuNeL evolutionary machine learning-based network inference method, we evaluate several variants of rule knowledge representations on a range of transcriptomics datasets to quantify the volume and complementarity of the knowledge that each of them can extract. Overall we show that knowledge representations, often considered a minor detail, greatly impact on the downstream biological knowledge extraction process.