https://bonkersworld.net/building-software

Back in 2016, I realized that I copy & pasted code snippets for my various microbiome analysis. Instead of coming up with a clean concept to export these snippets, I abused another repository (hence the strange name for mapping identifiers from GreenGenes) to at least re-use the same code: ggmap It once even had some unit tests, but in order to go fast, I ignored them failing over time. I never intended someone else to use this uggly part of software, but now realize that I torture my students by forcing them to do exactly this: use this badly documented, wobbly, partially deprecated code collection.

Should you have a slightly masochistic fun in refactoring my python code, I (and probably my students) would greatly appreachiate your efforts

The exoribonuclease-resistant RNA has some fascinating biological functions established through a very specific confirmation including a pseudoknot. This feature makes it hard a hard task to search for homologs. However, we should be able to design a hand tailored RNA folding algorithm for this pseudoknotted motif and search for compatible sequences in multiple genomes. (keywords: TDM for pknotsRG, alignment in Rfam: https://rfam.org/family/PLRV_xrRNA#tabview=tab9

Almost 25 years ago, Hidden Markov Models (HMM) were successfully used in the emerging new field of Bioinformatics to predict the topology of trans-membrane proteins (original papers "A hidden Markov model for predicting transmembrane helices in protein sequences" and "Predicting Transmembrane Protein Topology with a Hidden Markov Model: Application to Complete Genomes")

Authors of the program TM-HMM were excited about the tight coupling of mathematical modeling via HMMs and the modeled molecular biology. They used multiple algorithms to "score" amino acid sequences e.g. for being transmembrane proteins (Viterbi) or mark sub-sequences for being the helical regions (forward-backward) transecting the membrane. Labeling the many states of the HMM into 'inside', 'outside', 'transmembrane' or 'unlabeled' turns computation of the most probable "label" into an NP-hard problem - as proven much later by Broňa Brejová et al.. The TM-HMM program is a great example for real-world sized instances of dynamic programming; although a modern version is now based on deep learning - which, in principle, also uses forward-backward ideas.

Within a thesis, you would need to re-implement the HMM in Algebraic Dynamic Programming such that we can then use novel auto-generated algorithmic ideas to potentially improve prediction accuracy for this long standing problem.

The microbiome, i.e. the sum of bacteria, archaea, fungi and viruses, living in and on your body have an enormous impact on your health and well being, e.g. risk to develop autoimmune diseases, tendency for obesity, chronic gut diseases but also predisposition to depression or Parkinson's.

The majority of high profile studies of the microbiome are technically based on sequencing the 16S ribosomal RNA gene (i.e. amplicon sequencing) as a proxy to discern the different bacteria / archaea living in a microbiome. Recent algorithmic advances in amplicon-based microbiome studies enable the inference of exact amplicon sequence fragments to increase taxonomic resolution. However, these short (e.g., 150-nucleotide [nt]) DNA sequence fragments do not contain sufficient phylogenetic signal to reproduce a reasonable phylogenetic tree, introducing a barrier in the utilization of critical phylogenetically aware metrics such as Faith's PD or UniFrac.

Valid phylogenetic trees can be created by placing the sequenced fragments into a given high quality reference tree. This has been done for the Greengenes reference, but it is an open question how well this approach works for the popular SILVA reference tree: https://msystems.asm.org/content/3/3/e00021-18.

The aim of this project would be the evaluation of phylogenetic placements via SEPP into SILVA instead of Greengenes and the impact on the ability to detect biological signals in the samples that would be compared via metrics using this SILVA based tree. Furthermore, the use of a SILVA based insertion trees for taxonomic assignment shall be benchmarked within the TAX CREdiT framework.

RNAshapes studio front-end (B.Sc./M.Sc.)

Many RNA molecules do not encode proteins (mRNA), but by themselves exert important biological functions. Most functions are realized through their three dimensional structure - which is often times more conserved than the primary nucleotide sequence. Thus, sequence homology search with e.g. BLAST often does not return good results.

Secondary structure of RNA is the set of nucleotides that form base-pairs and functions as a scaffold for the final 3D structure. Since RNA folds hierarchically, prediction of secondary structure gives valuable insights about an RNA - while true 3D predictions are mostly computational intractable.

RNAshapes / pKiss / KnotInFrame and other software packages were developed at Bielefeld University and are popular tools to predict different aspects of secondary structures. They are based on the same algorithmic ideas (algebraic dynamic programming) and share the same code base back-end. The front end is written in Perl and lacks modern software engineering must-haves, like continuous testing, modularity or documentation.

Since our group is working on several algorithmic extensions of the back-end, we are looking for an encouraged student who re-implements the front-end in Python3, adds unit tests, creates bioconda packages, ... to allow for easy maintenance and distribution.

16S full length (M.Sc.)

Short read amplicon sequencing is still the state-of-the-art protocol to investigate large numbers of microbial samples. Although cheap and fast, phylogenetic resolution, i.e. the ability to discriminate different bacterial species / strains, is limited by the short read length of ~200 nucleotides. The other end of the spectrum is whole-metagenome shotgun sequencing. It is able to recover complete genomes of multiple bacterial species in a sample, but is expensive and needs high computational resources.

A compromise might be full length 16S rRNA sequencing. Third generation sequencing instruments like PacBio or Oxford Nanopore allow to span the full 16S gene (~1,800 bases) and thus capture potentially 10-fold more phylogenetic resolution. Benchmarks on mock communities (mixing ~20 bacterial isolates or rRNA genes) and on taxonomic assignments already exist, but it is unclear what the advantages and disadvantages are with respect to measure biological effect sizes in real experimental settings.

In cooperation with the University Clinic Düsseldorf, we sequenced ~500 samples with short and long read technology. Thus providing an ideal benchmarking scenario to develop recommendations for Vx / full-length / whole-metagenome sequencing to support future investigators to decide on the best suitable sequence strategy.

Benchmark genomic big data cancer processing pipelines (M.Sc.)

A single mutation in one of the three billion nucleotides that make up your genome might make the difference between a healthy life and suffering e.g. from cancer¹,². Our collaboration partner - the Department of Pediatric Oncology, Hematology and Clinical Immunology at Dusseldorf University - focuses on specific forms of childhood cancer and uses modern sequencing technologies to encypher and detect those mutations³,⁴,⁵. In fact, every human carries tens of thousands of mutations - compared to a given reference - and it requires elaborate computational pipelines to narrow down the list of potentially lethal mutations such that their wet lab team can concentrate on the most "promising" ones, i.e. those that might explain why a child has cancer - and ideally inform about possible treatments.

During the last six years, over 1,000 samples have been sequenced, producing 64 TB of raw data. The processing pipeline https://github.com/sjanssen2/spike encompasses over 50 individual computational steps which together took 25 CPU-years to produce meaningful mutation short lists.

However, during those years not only the computational tools improved, but also the human reference genome got updated with a more accurate one (hg38)⁶. Furthermore, we are going to employ a new sequencing technology, producing even more raw data. It is unclear how big the impact of those three dimensions on the filtration characteristics is. With those three measures, we aim to improve resolution such that more of the to date still unresolved cases can be explained. Unfortunately, it is not clear to what degree the changes impact positively or negatively on our capability to classify the important mutations⁷.

Thorough evaluation of those three changes would be the topic of a master thesis. Due to Duesseldorf's strong wet lab team, we are in the exceptional situation of having validation results (confirmation and rejection) of a high number of mutations, which would make up the gold standard for our evaluation.

If you love digging through really big data, are not afraid of commanding thousands of processors with a single key stroke from a command line, know how to program in our preferred language, e.g. Python, R, Java, and have an interest in boosting our cancer research, shoot a mail to stefan.janssen@computational.bio.uni-giessen.de or stop by at my office.

1. Genomic profiling of Acute lymphoblastic leukemia in ataxia telangiectasia patients reveals tight link between ATM mutations and chromothripsis.
   Ratnaparkhe M. et al. Leukemia 2017.
2. EBV Negative Lymphoma and Autoimmune Lymphoproliferative Syndrome Like Phenotype Extend the Clinical Spectrum of Primary Immunodeficiency Caused by STK4 Deficiency.
   Schipp C. et al. Front Immunology 2018.
3. Genomics and drug profiling of fatal TCF3-HLF-positive acute lymphoblastic leukemia identifies recurrent mutation patterns and therapeutic options.
   Fischer U. et al. Nature Genetics 2015.
4. Next-generation-sequencing of recurrent childhood high hyperdiploid acute lymphoblastic leukemia reveals mutations typically associated with high risk patients.
   Chen C. et al. Leukemia Research 2015.
5. Next-generation-sequencing-based risk stratification and identification of new genes involved in structural and sequence variations in near haploid lymphoblastic leukemia.
   Chen C. et al. Genes Chromosomes Cancer 2013.
6. Similarities and differences between variants called with human reference genome HG19 or HG38.
   Bohu Pan et al. BMC Bioinformatics, 2019.
7. Standards and Guidelines for Validating Next-Generation Sequencing Bioinformatics Pipeline.
   Somak Roy et al. Journal of Molecular Diagnostics, 2018.

Quantify folding ensemble differences (M.Sc.)

Stup: Bray-Curtis of dot-plot for related RNA sequences + UniFrac with tree that reflects neighborhood?

also good for FlowSoFine instead of gaussian weight?!

The American Gut Project is the world's largest crowd-sourced, citizen science microbiome research project. You can take and send in a sample, and you'll receive a report that shows a "snapshot" of the microbes found in your sample, along with comparisons to the rest of the population. In addition, you'll contribute to a dataset that can contribute to scientific research surrounding the gut.

To collect these data, we present a general survey to gather health and lifestyle information from participants (including a COVID-19 specific survey). The surveys at the moment are provided as single, very long, pages. A repeated concern from participants is the extent of "survey fatigue" they experience. From UI/UX research this past summer, it was noted that this fatigue may be reduced if questions were presented one-at-a-time with an indication of progress. The surveys themselves are described in JSON using a Vue.js compatible schema, which includes the question, its responses and response type (e.g., single choice, multiple choice, etc). Vue.js is used for rendering right now.

Join forces with the American Gut Developers and help improve the user experience to increase data quality and thus foster microbiome research! This programming project will focus on decomposing these structures to present individual questions. It should primarily orient around some Javascript modifications, with some HTML and Jinja2 templating.

UTF8 for RNA

Bellman's GAP cannot parse non-ASCII bases, i.e. base modifications. But those become ever more important, like pseudouridine in BioTechs Covid19 vaccine. There are a few thermodynamic parameters out there, that might be worth being integrated into Vienna's parameters + enable UTF8 parsing for Bellman's GAP.

What's in my sample

stup: machine learning on EMP dataset to guess metadata values for novel microbial samples, or "enrich" metadata.

Non-Terminal Report Algebra

Plotting algebra

Bioconda for FlowSoFine

Plasmid hunter

We found an antibiotic resistence gene, encoded on a plasmid in multiple clinical samples. The plasmid swapped bacterial hosts and might spread further. In order to assess necessary counter measures, we want to know if the very same plasmid was sequenced somewhere on the world before - probably as bycatch of metagenomic projects.

We need to develop a system that can crawl through the huge raw data of ENA metagenomic WGS samples (currently 39521 files totalling in 46.0 TB) and test if reads map to your plasmid. If so, we can create a world map of matches based off the geographic locations of the original sample sites.