Large Language Models in Medicine

In my introductory article on computational linguistics, I mentioned that I had chosen an interdisciplinary topic for my master’s thesis, linking large language models with bioinformatics and genetics. I will describe what exactly lies behind this and what I have been working on for the past few months.

In my thesis, I focused on the automatic interpretation of genetic variant mentions in biomedical literature using large language models (LLMs). I developed an open-source web application that helps molecular geneticists streamline their daily workflow when assessing the pathogenicity of genetic mutations in patients.

From Finding a Problem to a Functional Web Application

Just like with my volunteering activities in Hnutí Brontosaurus (more here), where it is essential for me to see a tangible result behind my work, I wanted to create something practical for my thesis as well.

I chose the field of medicine and bioinformatics because it makes sense to me to use technology where it can truly help someone.

I reached the final topic gradually, and I must admit I was quite surprised by how specialized it turned out to be, but I took it as a challenge.

During consultations with experts from CEITEC, we defined how I could help them. Molecular geneticists receive tables of mutations found in patients, and for the lesser-known ones, they must manually go through scientific articles to determine whether the mutations are dangerous. This process is tedious and demanding because there are simply too many papers. the resulting web application accelerates and streamlines article browsing.

What It Taught Me

The master’s thesis was a test for me in several areas. It showed me that:

  • I can drive a project to completion independently: From the initial user requirements analysis through architecture design, dealing with missing API documentation, all the way to deploying a functional prototype as a web application.

  • I acquired new technical skills: I learned to validate model outputs using pydantic-ai, deployed the application via Docker to Koyeb, and practiced prompt engineering alongside XML and JSON parsing.

  • I can orient myself in an unfamiliar field: I did not know what HGVS nomenclature or ACMG/AMP guidelines were, or what the difference was between PubMed and MEDLINE – but I looked it up and managed to put it to use.

I look forward to leveraging the experience gained in my future professional practice!

The text of the thesis is available here, and the full code can be found on my GitHub here.