Hi there, I'm Jason and I'm interested in Data Science and Computational Biology. I studied bioinformatics at the University of Queensland and, more recently, a masters of Data Science at LMU in Munich. I'm particularly interested in incorporating thermodynamics into metabolic models and hybrid modelling, that is, the synthesis of ML and mechanistic models to hopefully combine the rigidity, specificty and interpretability of mechanistic models with the accuracy of ML models. I'm currently working on a new statistical model for improving the covariance estimates of the group contribution (and component contribution) methods.
Unfortunately, my current projects are still private until we get closer to publishing, but I have a handful here that might be interesting.
In this project I was tasked with integrating height data into an existing Faster-RCNN object detection network to detect conifer seedlings from drone imagery. Interestingly, it turned out that if you introduced the height data directly as a fourth image channel, the performance of the network decreased. However, pooling the height data and incorporating it after the backbone and before the region proposal network improved performance. It goes to show that it's important to think about how to incorporate new information when using transfer learning.
In this project I wanted to integrate thermodynamics into steady-state flux sampling for my master's thesis. The requirement that the fluxes are steady-state introduces a number of constraints into the model that need to be adressed. I solved this by making explicit the linear relations between the models parameters and removing superfluous parameters. This project also involved improving existing estimates of formation energy covariance to address prediction collinearity.