I started my career at INRIA (Institut National de Recherche en Informatique et Automatique) as a PhD researcher, under the supervision of François Coste. Together we introduced PPalign, a method for pairwise Markov Random Field alignment based on Integer Linear Programming, with application to protein sequences. Previous methods modelled protein sequence families using Hidden Markov Models (HMMs), looking at each position independently ("how frequent is this amino acid at this position?"). We proposed to use Potts models, which also capture dependencies between positions: not just individual frequencies but also how residues interact with each other. Aligning such models is significantly harder: the pairwise dependencies make the problem NP-hard. Our contribution was to formulate it as an ILP problem and solve it in tractable time, outperforming HMM baselines in alignment quality.

Early in my thesis, I also had the opportunity to work with Witold Dyrka from Politechnika Wrocławska (Wrocław, Poland) on training probabilistic context-free grammars on protein sequence families using structural contact constraints.

I then joined the National Museum of Natural History (MNHN Paris, Sorbonne Université) for a year as a postdoctoral researcher with Mathilde Carpentier. We experimented on alternative statistical approaches to build canonical, comparable models, and extended the PPalign solver to take into account the fact that insertions are more likely at some positions than others.

I then moved to Brussels to work with Dimitri Gilis at ULB (Université Libre de Bruxelles).

My first project there was a Transformer variational autoencoder trained to embed molecules into a structured, continuous latent space where distances reflect similarity, with application to molecular representation and optimization. The model was shipped as a user-friendly pip package.

I also collaborated with Matsvei Tsishyn on StructureDCA, a sparse Potts model inference method with structural constraints. The method, based on a C++ solver with a Python interface, achieves state-of-the-art performance in mutational landscape prediction.

Currently, I'm working on a framework for benchmark construction: given a dataset, it selects and partitions samples to control biases such as entities having strongly skewed label distributions or similar entities leaking between train and test, preventing machine learning from exploiting shortcuts, and allowing for a more reliable evaluation. Enforcing all these constraints simultaneously is a combinatorial problem which we solved using Integer Linear Programming, to be released as a pip package soon. The framework is already being applied to olfactory receptor activation prediction as part of a Master's thesis I am supervising.