Bio papers — 2026-09-14

Getting from raw single-cell RNA sequencing data to a testable biological idea is difficult because current AI tools do not directly interact with gene expression levels. To address this, ELISA was introduced as an interpretable framework that combines expression embeddings with semantic retrieval and large language models to allow interactive exploration of these datasets.

This system automatically sorts incoming questions into different pathways, which include scoring gene markers, finding semantic matches, or looking for combinations of both. These modules then perform tasks such as scoring pathway activity across more than sixty gene sets and predicting ligand-receptor interactions using over two hundred curated pairs.

The results demonstrate that ELISA is much better than classical methods like CellWhisperer when identifying cell types. It achieves a combined permutation test p value less than two times ten to the fifth for each type, which means it is significantly more accurate in finding cell populations. Furthermore, when asked questions about gene signatures, ELISA showed a very large Cohen's d of five point nine eight for MRR, indicating a strong ability to find relevant gene sets.

This capability allows the system to replicate existing biological findings with a mean composite score of zero point eight eight and then use reasoning from the large language model to generate candidate hypotheses based on the data. This process bridges the gap between simply looking at transcriptomic data and making real biological discoveries, while diffusion learning research is exploring how to map out viable parameter sets in complex dynamical systems using score-based diffusion models to sample these constrained regions.

Today's papers

• ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics Translating single-cell RNA sequencing data into mechanistic biological hypotheses remains a critical bottleneck, as agentic AI systems lack direct access to transcriptomic representations while expression foundation models remain opaque to natural language. [paper]
• Diffusion learning reveals viable parameter manifolds and compensation geometry in biological dynamical systems Models of complex systems often have many parameters, yet are constrained by far fewer experimentally accessible observables; consequently, similar activity can emerge from coordinated parameter changes. [paper]

The papers

• ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics —
• Diffusion learning reveals viable parameter manifolds and compensation geometry in biological dynamical systems —

Important terms

single-cell RNA sequencing data

This is raw data from sequencing individual cells to measure gene expression levels. The research focuses on how to turn this complex expression information into meaningful biological insights.

ELISA framework

A new interpretable system that combines gene expression embeddings with semantic retrieval and large language models. It helps explore datasets interactively.

expression embeddings

These are numerical representations of gene expression levels. They are used by the ELISA system to find semantic matches and understand the data better.

ligand-receptor interactions

These refer to how signaling molecules on one cell interact with receptors on another cell. The system predicts these interactions using curated pairs.

diffusion learning

A research area exploring how score-based diffusion models can map out viable parameter sets in complicated dynamical systems. This helps sample important regions.