Last week, Dr. Lisa Chabrier successfully defended her PhD work on calculating SHAP values and using them to do differential network analysis. Whilst SHAP values are a popular framework for local explanations of machine learning predictions, they come with a serious computational cost. Lisa developed an approximation algorithm to avoid calculating lots of SHAP values that no-one cares about, and named it TopShap (see here). Next, she designed and implemented Re_actShap (see here), a modular pipeline with TopShap as one of its main components, to predict how differences between groups of cells can show as rewiring events in their regulatory networks.
She did a great job establishing herself at the interface of computer science and biology, and it was a pleasure to have worked with her as my first PhD student. I think that together with the other thesis advisors, Christophe Rigotti and Sergio Peignier, the four of us developed an exciting research line on (i) explainability in machine learning and (ii) its use to better understand regulatory networks inferred from single cell data.
Looking back at the 3-year project and also taking into account the Master’s students that I have advised over the years, I must say that guiding students to become independent researchers is a rewarding experience.
PS. TopShap and Re_actShap code, plus example usage in Jupyter notebooks, are available on our institute’s gitlab: https://gitlab.inria.fr/topshap.