Research

• Database curation: At TU Wien, I curated an extensive database of validated, stereochemistry-corrected enzymatic reactions, enabling large performance improvements for bioretrosynthesis via deep learning.

• Similarity of molecules: At MIT, I developed a software package to compute the shape and electrostatic similarities of molecules called ESPsim, which was also featured in an RSC CICAG workshop.

• Deep learning: At MIT, I developed a machine learning algorithm to predict reaction properties based on message-passing neural nets on the condensed graph of reaction in the group of Prof. Green, outperforming state-of-the-art approaches in accuracy while being significantly smaller and thus faster to train. In addition, I actively contribute to the maintenance and development of Chemprop on Github, a graph neural network software for molecules.

• Machine learning of retrosynthesis: At MIT, I improved the automated extraction of templates for the machine-learning of retrosynthesis pathways of organic molecules, which largely boosts the performance of algorithms in template-based approaches such as ASKCOS.

• Enzymatic reaction prediction: At MIT, I successfully developed and implemented software solutions for the prediction of substrate ranges for enzymes based on molecular Hasse diagrams by extracting and scoring sets of reaction templates and relating them to their common substructures. The approach proved superior to previous approaches for the prediction of the feasibility of enzymatic reactions, as well as the proposal of co-substrates for multi-substrate reactions, and the correct treatment of regioselective reactions.

• Atomic polarizabilities: In a shared project between the London Imperial College, the University of Maryland and the University of Vienna, I developed a framework to compute atomic polarizabilities, a key parameter for polarizable force field development, for both neutral and charged species in arbitrary states and environments via quantum-mechanics. Previous approaches relied on the statistical dissection of experimental refractive indices, and could not account for non-bulk phase, or excited state polarizabilities. During a short term research stay at the University of Maryland, Baltimore, I furthermore implemented an automated system to determine partial charges and atomic polarizabilities for organic molecules via heuristics and machine learning in the group of Prof. Alexander MacKerell, which builds the foundation of their current work of developing a general polarizable force field (DGenFF).

• Solvation dynamics: At the University of Vienna, I achieved a quantitative agreement between experimental and computational time-dependent fluorescence signals in various systems by combining state-of-the-art non- equilibrium molecular dynamics simulations with polarizable force fields in the ground and excited state of the chromophore probe. I could gain insights into the underlying solvent motions of the observed solvation dynamics in different systems at atomic resolution, such as the cryoprotective properties of saccharides