davidnovak9000 at gmail dot com
I am a bioinformatician and PhD researcher at Saeys lab, Center for Inflammation Research of the Flemish Institute of Biotechnology and Ghent University in Ghent, Belgium. I am currently working on my PhD, having secured an FWO Strategic Basic research fellowship.
My interests include machine learning and (including) deep learning applied in biomedical contexts. My favourite biological data is single-cell.
Check out my full CV for more about my work experience and education. See my Google Scholar profile for publication activity.
Some of my publicly available project repositories, including collaborations, are listed below.
ViVAE is a dimensionality reduction algorithm implemented in Python, aimed at creating useful embeddings of cytometry and scRNA-seq data. Using variational autoencoders and a stochastic multi-dimensional scaling (MDS) loss (adapted version of idea from SQuadMDS), we achieve a better balance of local and global structure preservation. I am the first author of a paper now submitted we are now getting published with my co-authors for Ghent University and UCLouvain (read pre-print here).
ViScore is a collection of evaluation metrics for dimensionality reduction. Since previously published evaluation methods suffer from severe drawbacks, together with collaborators from UCLouvain we have released a battery of both unsupervised and supervised evalation metrics. They build on RNX curves and the Neighbouhood Proportion Error to provide embedding-level and population-level scores. The current version is described in our ViVAE pre-print.
GroupEnc is a project I presented at the BNAIC/BeNeLearn 2023 machine learning conference. Implemented in Python, this work is similar in spirit to ViVAE in that it uses a VAE architecture for generating embeddings with an MDS scaling criterion. This work is more conceptual and allows for varying more hyperparameters, as well as dispensing with the decoder to create a faster-training encoder-only model to produce embeddings with significantly better global structure preservation than a VAE does. The KL-divergence term from VAEs is kept, in order to obtain an interpretable latent space.
tviblindi is a semi-supervised trajectory inference tool for single-cell data that uses topological data analysis (TDA) and persistent homology to be able to work will the full, high-dimensional data, mitigating loss of information due to dimensionality reduction. For my master thesis, I implemented parts of the TDA pipeline in C++ and created a method for clustering trajectories based on persistent homology, as well a GUI for semi-supervised analyses using that clustering, in R & Shiny.
hloss is a work I presented at the ABLS 2022 bioinformatics conference, implemented in R. This project tackles the issue of evaluating a cell type classification in single-cell data while considering a known hierarchical structure of phenotypes. I introduce a novel scoring approach, which incorporates the biological prior to assess error based on degrees of relatedness.
npefaster is an optimised and extended version of the Neighbouhood Proportion Error.
SingleBench is a benchmarking framework for high-dimensional data transformations and clustering in R. Relying on HDF5 and recycling of intermediate results, it enables extensive hyperparameter sweep and easy interpretation of results. It is designed to be highly extensible. I presented this work at CYTO 2021.
tidycell is a differential expression analysis tool written in R for cytometry data that I developed when working at Childhood Leukaemia Investigation Prague. It combines two pipelines: one based on FlowSOM clustering and one based on extraction of learned convolutional filters from CellCNN.
qctoy is an R package for simulating aberrances in flow cytometry measurements that are relevant in designing QC tools and pipelines.
auto_compensate is an automated pipeline for cytometry data compensation I designed at Childhood Leukaemia Investigation Prague.
hidden is a neat hidden Markov model simulator in R. Why not.
CommandLineParser is a C#/.NET API I co-wrote with Kačka Břicháčková.
avl_tree is an Adelson-Velsky and Landis tree implementation in Pascal.
RCondaRun is a tiny package for switching between Conda environments within a single R session when interfacing with Python.