Research

My research involves development of machine learning approaches, and application of these methods in close collaboration with experimental colleagues. Large amounts of data are being produced in the life sciences, at various levels (e.g. genomics, phenotypic, ecological), and much of this data is quite unstructured. Machine learning is often applied as a black box, which can be sufficient if all one cares about is good prediction performance. However, it can be very fruitful to learn from the machine learning model itself. This allows for example to generate testable hypotheses, leading in turn to improved understanding of the system of study. Importantly, this also gives rise to opportunities to modify the system in order to obtain specific desired behaviour.

Figure 1, Machine learning involves learning predictive patterns

Recent examples where I used machine learning in this way include PhD projects on the prediction of enzyme product specificity, the prediction of the pathogenic effect of protein coding variation, and the prediction of crossover occurrence in genomic locations for various plants. An important shared characteristic in these projects is that sufficient understanding of the application domain is essential to enable effective collaboration with domain scientists. Often, specific insights into the domain will be leading in how machine learning approaches can best be developed for specific purposes – standard solutions do not suffice.

Figure 2, A recent example of machine learning methodology involves using protein structures as input to predict functions and interactions (left). Using interpretable machine learning allows to find regions in a protein which are important for specific functions (right).