Computational biology researcher and MSc student in Bioinformatics & Systems Biology (UvA/VU). I build mechanistic and inference models of RNA Pol II transcription, and apply deep learning to multi-omics imaging.
BSc thesis (transcription simulation, University of Crete): 10.00/10. MSc (Bioinformatics & Systems Biology, UvA/VU): 8.05, in progress.
Stack across these projects: Python, Cython, stochastic/Markov-chain modeling, and vision transformers (DINO ViT) for image analysis.
Van Steensel group, Division of Gene Regulation. Extending PARM, a
deep-learning promoter-activity CNN published in Nature (doi:
10.1038/s41586-025-10093-z), to endogenous transcription in K562, benchmarked
against Activity-by-Contact. Daily supervision by Oscar Garcia Blay,
co-supervised by Vinicius Franceschini-Santos. The analysis pipeline
(metron, private) already runs the raw qPRO-seq mapping, GENCODE-resolved
window counting, and a PARM-reproduction step ahead of the official start. It
already carries the project's reproducibility scaffolding: every input traced
to a checksummed manifest, every threshold preregistered ahead of the
placement's data.
Under Matthieu Lavigne, building a four-part framework that simulates, solves, and infers RNA Pol II transcription kinetics end to end - continuing a BSc thesis that modeled transcription in silico (University of Crete):
- POLYARIS: stochastic simulation of Pol II via a TASEP (Totally Asymmetric Simple Exclusion Process) model with spatial exclusion, pausing, and obstacle interactions, run as a continuous-time Markov chain (Gillespie algorithm), Cython-accelerated.
- KINEMA: closed-form Markov-chain solutions for Pol II initiation kinetics - the analytical counterpart to POLYARIS's simulation - including "cascade" transitions for packed-polymerase trains. Co-authored with Michalis Loulakis (NTUA).
- GENEFIT: recovers transcription-kinetic parameters from ChIP-seq/PRO-seq/ mNET-seq features by nested brentq root inversion plus joint Poisson profile MLE, no stochastic search.
- SCAN: downstream statistical analysis and figure generation across the suite.
POLYARIS, KINEMA, and GENEFIT each carry an automated pytest suite run on every change; GENEFIT is additionally validated against synthetic ground-truth data before being applied to real sequencing data.
This research is ongoing; the code is private, with
Private_Projects_Showcase (pinned below) marking that it exists. Happy to
talk through it directly on request.
Ahmed Mahfouz's group, Human Genetics. Cell segmentation and vision transformers on tissue imaging:
- vitseg: unsupervised tissue-region discovery in kidney microscopy, using DINO ViT-S/16 self-supervised features clustered with MiniBatchKMeans.
- ATLAS-Cellpose: adaptive tiled segmentation wrapper around Cellpose, merging results across tiles via a four-step algorithm.
Co-author, Manzato et al., "Image-guided alignment of consecutive multi-modal tissue slides" (bioRxiv, 12/2025).