Back

Can Lightweight LLM Agents Improve Spatial TranscriptomicsAnnotation?

Dip, S. A.; Zhang, L.

2025-11-10 genomics
10.1101/2025.11.08.687410 bioRxiv
Show abstract

Spatial transcriptomics (ST) links gene expression to tissue organization, yet automated annotation of spatial regions remains a persistent challenge. Recent studies have explored large language models (LLMs) for biological reasoning, but their applicability in low-compute, freetier settings is largely unexplored. We investigate whether lightweight LLM agents can improve ST annotation by integrating rule-based heuristics, prototype discovery, and multirole reasoning (Analyst-Consensus-Reviewer) within a unified agentic framework. Across six STARmap and MERFISH datasets, we benchmark single- and multi-agent variants using standard clustering and spatial coherence metrics (NMI, ARI, CHAOS, ASW). Our results show that small open-weight models such as llama3.2 and qwen3 match or slightly exceed deterministic baselines in cluster recovery, while producing more spatially consistent and interpretable predictions. These findings high-light the potential of modular LLM agents as resource-efficient components in future spatial omics annotation pipelines.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.