Data and modelling

Built from administrative records I worked with as school staff, in a public school in Lisbon. The source data holds personal information about minors, so the first thing I wrote was the anonymisation pipeline: identity documents, tax and social-security numbers and emergency contacts removed outright, names replaced by sequential pseudonyms, dates of birth reduced to age, free-text health and dietary notes collapsed into broad categories, and any small group folded into a larger one so nobody could be identified indirectly. Only the anonymised output is versioned.

The analysis and an interactive dashboard went to the school’s management. The results are not published here: the data belongs to the school.

Between 2015 and 2019 I asked what the microRNA miR-34c-5p actually does inside an activated human CD4+ T cell. The answer was a logical network model, built in GINsim, predicting what should happen downstream and guiding the wet-lab work. The unfinished thesis then sat on a hard drive for six years.

I rebuilt a piece of it in Python: a parser for the thesis’s own notation, a multi-valued evaluator, and an asynchronous simulator running 400 trajectories per condition. The 76 nodes and 89 logical formulas are transcribed verbatim from Table 5 of the thesis draft, not reconstructed from memory.

miR-155-5p turns out to be the dominant driver: on its own it suppresses the Th2, Tfh, Th22 and Th9 regulators while nudging Th1 and iTreg upward. miR-34c-5p alone barely moves the baseline — but together the two suppress that branch further than miR-155-5p does alone. The effect is cooperative, not additive.

That is a related but not identical story to the 2019 discussion, which leaned towards a Th17/iTreg shift. I would rather report the divergence than tune the model until it agrees: the simulation method differs, and two formulas in the original table are ambiguous as published and were evaluated exactly as written. Both causes are documented in the repository.