Presenting LLMBDC at BIOKDD 2026 (KDD Workshop)

I’m happy to share that our paper, “LLMBDC: Language Model for Biological Domains Oriented Clustering of Gene Ontology,” has been accepted and presented at the 25th International Workshop on Data Mining in Bioinformatics (BIOKDD’26), held in conjunction with KDD 2026!

The Problem

Gene Ontology (GO) is a foundational resource in bioinformatics — a large, hierarchical vocabulary of biological terms covering molecular functions, biological processes, and cellular components. But GO contains tens of thousands of terms, making it difficult to summarize or interpret a set of GO terms associated with a disease or experiment.

Existing clustering approaches group GO terms primarily by graph topology or co-occurrence statistics. These methods can miss the semantic nuance that distinguishes, say, “synaptic transmission” from “glutamate receptor signaling” even when both are statistically co-enriched.

Our Approach

LLMBDC takes a different route: we use a large language model in a zero-shot, training-free setting to cluster GO terms based on their semantic meaning. Rather than learning from labeled examples, the model reasons about biological relevance using its pre-trained knowledge, then assigns confidence scores to each clustering decision.

On Alzheimer’s disease and Fragile X syndrome datasets, LLMBDC outperformed six existing GO clustering approaches on standard clustering metrics while preserving biologically meaningful groupings.

Why Training-Free Matters

One recurring challenge in computational biology is that high-quality labeled datasets are scarce and expensive to produce. A method that requires no task-specific training can be applied immediately to any disease or organism without data collection overhead — and the interpretability of LLM-generated cluster labels is a practical bonus for downstream biological analysis.

What’s Next

We’re exploring extensions to other ontologies (HPO, ChEBI) and richer multi-ontology clustering scenarios. The preprint is available on arXiv.

— Ximing