23rd Discussion-20 Nov 2025

A knowledge sharing session on the usage of LLMs in Bioinfo cores hosted by Sivarajan Karunanithi (Siva) and James Gilbert.

Below is an AI generated summary of the [Meeting Notes](https://docs.google.com/document/d/1zdZvvpbsnktxlg-FbvhLLQ3JfXrvbX4X4r4krYi528g/edit?usp=sharing)

Summary:

* LLMs can generate code and assist with tasks like QC or documentation, but they perform poorly on complex analyses (e.g., multiomics, scRNA-seq), often producing plausible but incorrect results.
* Agentic LLM systems (Cell Voyager, Biomni) are difficult to debug, inconsistent, and prone to choosing wrong tests or outdated methods, making their analyses unreliable.
* While LLMs help non-experts write substantial code, users often lack the background knowledge to detect errors, increasing risk and reducing incentive to learn fundamentals.
* Teaching must adapt by emphasizing AI literacy, safe use, critical thinking, and maintaining foundational coding skills despite reduced student interest in learning basics.
* Institutions are exploring secure/open-source LLM deployments, with AI useful for snippets and documentation but not yet a threat to bioinformatics facilities due to its unreliability in advanced analysis.
* LLMs can create a false sense of competence, enabling inexperienced users to perform advanced analyses without understanding assumptions, which can lead to confidently wrong scientific conclusions.
* There’s no consensus on how to acknowledge AI use, raising questions about responsibility, transparency, and how AI-assisted work should be reported in scientific outputs.