Research
Embedding-based retrieval of oligonucleotide binding sites.
Where
Research Intern at the Medical Bioinformatics group of Prof. Alexander Schliep, BTU Cottbus-Senftenberg, since November 2026, guided by a PhD student in the group. The lab works on AI for antisense oligonucleotide (ASO) drug design.
The problem
Given a short query oligonucleotide of about 15–25 nucleotides, find every site in a very large sequence database that it could bind, meaning sites where the binding free energy ΔG falls below a threshold. This matters for drug design, because an oligo that also binds unintended sites causes off-target effects.
Computing ΔG exactly for every pair is far too slow at that scale. The approach is filter-and-verify: embed every target once, use nearest-neighbour search to pull a small candidate set for a query, then compute exact ΔG only on those candidates. The idea comes from FlashPPI, which does this for protein interactions, and from the NeuroSEED framework for embedding biological sequences. The difference for DNA is that binding depends on complementarity rather than similarity.
What I've established so far
- It's a range query, not top-K.We want all sites below the energy threshold, and the number of hits varies per query, so recall is measured against that whole set.
- Mismatch count is a weak proxy for binding energy.Measured on sequences with varied GC content, edit distance only loosely tracks ΔG, which leaves real room for a learned embedding to help.
- It needs two separate encoders.A sequence binds its reverse complement far more strongly than itself, so binding is not a symmetric metric on one set. That rules out a single shared encoder and forces a query tower and a target tower.
- A first learned encoder generalises.It places sequences it has never seen, and the bar to beat is a cheap reverse-complement-aware k-mer baseline that needs no training at all. Closing that gap is the research question.
Measurements and figures stay with the lab; this page describes the shape of the work rather than its results.
Methods and tools
Python, NumPy and primer3 for thermodynamics · PyTorch for the encoders · contrastive learning with InfoNCE · FAISS and approximate nearest-neighbour indexes · DNA language models (DNABERT-2, HyenaDNA, Nucleotide Transformer) under evaluation. Related reading is on my papershelf.
Background
I also gave an MSc seminar talk on Hist-Tree and learned index structures, arguing that classical indexes can match learned ones on read-only sorted data without the build and retraining cost. I read the three learned indexes it is measured against — RMI, the PGM-index and RadixSpline — alongside it. Indexing and retrieval is the thread connecting that talk, miniDB and this project.