ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics

ELISA is an interpretable hybrid AI agent that bridges natural language and single-cell transcriptomic data by integrating scGPT embeddings with LLM reasoning to enable interactive, expression-grounded biological discovery, outperforming existing tools in cell type retrieval and hypothesis generation across diverse datasets.

The Gist

Imagine you have a massive, chaotic library containing the "instruction manuals" for every single cell in the human body. This library is called scRNA-seq (single-cell RNA sequencing). It's so big and complex that even the smartest human scientists get overwhelmed trying to find specific answers, like "Which cells are fighting this infection?" or "What genes are causing this cancer?"

For a long time, we had two separate tools to help us, but neither was perfect:

1. The "Gene List" Tool: Great if you know the exact names of the genes you are looking for, but useless if you just have a vague idea like "I'm looking for cells that are angry and inflamed."
2. The "Chatbot" Tool: Great if you ask in plain English, but it often gets lost in the details and misses the specific genetic evidence.

Enter ELISA.

ELISA (which stands for Embedding-Linked Interactive Single-cell Agent) is a new, super-smart AI detective designed to solve this problem. Think of it as a hybrid librarian that can speak both "Human" and "Genetic Code" fluently at the same time.

How ELISA Works: The "Smart Router" Analogy

Imagine you walk up to a help desk at a massive airport.

• If you say, "I need a flight to Paris," the agent knows you want a semantic search (looking for a destination).
• If you hand them a boarding pass with a specific flight number, they know you want a data search (looking for a specific record).

ELISA does the exact same thing with biology questions:

1. It listens to your question.
2. It decides how to answer:
   • If you ask, "Show me cells that look like angry macrophages," it uses its Language Brain (trained on millions of biology papers) to find the right cell types.
   • If you ask, "Find cells with genes A, B, and C," it switches to its Math Brain (trained on raw genetic data) to match those specific gene signatures.
   • If you mix them ("Find angry macrophages with genes A, B, and C"), it uses a fusion technique to combine both answers perfectly.

The "Magic Report" Generator

Once ELISA finds the right cells, it doesn't just spit out a list of numbers. It acts like a science journalist.

• It checks the "evidence" (the actual data).
• It compares it to "what we already know" (existing medical knowledge).
• It writes a clear, readable story for you.

It can tell you things like: "We found that in this patient's lung, the immune cells are talking to the tissue cells in a weird way, suggesting a new cause for the inflammation." It even flags these as "hypotheses" (educated guesses) so scientists know what to test next in the lab.

Why is this a Big Deal?

The paper tested ELISA against other AI tools (like a competitor called "CellWhisperer") using six different real-world medical datasets (ranging from cystic fibrosis to brain development).

• The Result: ELISA was significantly better at finding the right answers, especially when scientists asked questions using gene names (which is very common in research).
• The Analogy: If other AI tools were like a tourist who knows the general area but gets lost in the streets, ELISA is like a local taxi driver who knows every alleyway and can also explain the history of every building you pass.

The Bottom Line

ELISA bridges the gap between raw data (the messy, complex numbers) and biological discovery (the "Aha!" moment). It allows scientists to stop wrestling with spreadsheets and start having natural conversations with their data, leading to faster discoveries for diseases like cancer, lung disease, and more.

In short: ELISA is the translator that turns the chaotic language of genes into a clear story about how our bodies work.

Technical Summary

1. Problem Statement

The translation of single-cell RNA sequencing (scRNA-seq) data into mechanistic biological hypotheses faces a critical bottleneck. While expression foundation models (e.g., scGPT) can learn rich latent representations of transcriptomic data, they are often opaque to natural language queries. Conversely, agentic AI systems and Large Language Models (LLMs) excel at reasoning over text and structured databases but lack direct access to high-dimensional transcriptomic representations. Existing tools often rely on contrastive multimodal alignment (e.g., CLIP-style training), which tends to optimize for text-transcriptome correspondence at the expense of fine-grained gene-level specificity, leading to poor performance when users query via gene signatures rather than natural language descriptions.

2. Methodology: The ELISA Framework

ELISA (Embedding-Linked Interactive Single-cell Agent) is an interpretable framework that unifies transcriptomic embeddings, semantic retrieval, and LLM-mediated interpretation without retraining the underlying foundation models.

A. System Architecture

ELISA operates on a shared serialized PyTorch embedding file (.pt) per dataset, eliminating the need for access to the original count matrix during inference. The system comprises four core modules:

1. Data Preparation: Standard preprocessing (normalization, PCA, clustering) followed by the generation of two parallel representations:
   • Expression Embeddings: Cluster centroids derived from scGPT (generative pre-training on millions of cells).
   • Semantic Embeddings: Cluster descriptions (Cell Ontology names, top markers, GO/Reactome terms) encoded via BioBERT (768-dimensional).
2. Hybrid Retrieval Engine: An automatic Query Classifier routes inputs to the optimal pipeline based on token-level heuristics:
   • Gene Signature Queries: Routed to a Gene Marker Scoring pipeline. It scores clusters based on Differential Expression (DE) profiles using a weighted function of $|\log_2 \text{FC}|$ and expression specificity ($pct_{in} - pct_{out}$).
   • Natural Language Queries: Routed to a Semantic Retrieval pipeline using BioBERT cosine similarity, augmented by Cell Ontology name boosting and synonym expansion.
   • Mixed Queries: Resolved via Reciprocal Rank Fusion (RRF) of both pipelines.
   • Additive Union Strategy: For benchmarking, the system selects the higher-recall modality as primary and appends unique results from the secondary pipeline to maximize coverage.
3. Integrated Analytical Modules: These operate directly on the embedded data to perform:
   • Pathway Activity Scoring: Across 60+ curated gene sets (immune, metabolism, neuroscience, etc.).
   • Ligand-Receptor (L-R) Interaction Prediction: Using a curated database of 280+ pairs (CellChat, CellPhoneDB, NicheNet).
   • Condition-Aware Comparative Analysis: Identifying condition-biased gene expression and cell-type proportion shifts.
4. LLM Interpretation Layer: A LLaMA-3.1-8B model (via Groq API) generates structured biological reports. Crucially, the LLM is strictly grounded in the retrieved dataset evidence (cluster data, gene stats, pathway scores) to prevent hallucinations. It separates "dataset evidence" from "established biology" and proposes candidate hypotheses with probabilistic language.

B. Key Design Principle

Unlike systems that learn a single shared embedding space (implicit fusion), ELISA maintains separate representation spaces for text and expression, using explicit routing to leverage the strengths of each modality. This avoids the "compression artifacts" where text-supervised alignment sacrifices gene-level specificity.

3. Key Contributions

• Multimodal Discovery Agent: The first framework to integrate scGPT expression embeddings with BioBERT semantic retrieval and LLM reasoning for interactive single-cell discovery.
• Query-Adaptive Hybrid Retrieval: A novel architecture using automatic classification and dynamic routing (Gene scoring vs. Semantic matching vs. RRF) that outperforms single-modality approaches.
• Expression-Grounded Reasoning: Integrated analytical modules (pathway, L-R, proportion) that operate directly on embeddings, bridging the gap between data retrieval and biological interpretation.
• Benchmarking Framework: A quantitative evaluation strategy measuring an AI agent's ability to recover established biological findings (gene coverage, pathway alignment, interaction recovery) across diverse datasets.

4. Results

ELISA was validated on six diverse scRNA-seq datasets (Cystic Fibrosis, Neuroblastoma, Immune Checkpoint Blockade, Breast Tissue, Fetal Lung Organoids, and First-Trimester Brain).

A. Retrieval Performance

• Outperformance of CellWhisperer: ELISA significantly outperformed the state-of-the-art multimodal system, CellWhisperer, in cell type retrieval (Combined Permutation Test, $p < 0.001$).
• Gene Signature Queries: The advantage was most pronounced for expression queries, where ELISA achieved a Mean Reciprocal Rank (MRR) of 0.806 compared to CellWhisperer's 0.397 (Cohen's $d = 5.98$). This highlights the failure of contrastive text-transcriptome models to handle gene-list queries.
• Ontology Queries: ELISA matched or slightly exceeded CellWhisperer on natural language queries, demonstrating that explicit routing does not compromise semantic understanding.

B. Biological Replication

• Replication of Published Findings: ELISA achieved a mean composite score of 0.90 when compared against reference publications.
  • Pathway Alignment: 0.98 (near-perfect).
  • Theme Coverage: 0.98.
  • Gene Coverage: 0.85.
  • Interaction Recovery: 0.77.
• Case Studies:
  • Cystic Fibrosis: Correctly identified IFN-$\gamma$ signaling and the HLA-E/NKG2A axis.
  • Neuroblastoma: Detected the HB-EGF/ERBB4 paracrine axis and mTOR/MAPK pathways.
  • Breast Tissue: Resolved epithelial hierarchies with 0.96 gene coverage.

C. Discovery of Novel Hypotheses

ELISA's discovery mode generated plausible, non-obvious hypotheses not explicitly highlighted in the original papers, such as:

• CF Airways: Enrichment of the CALR–LRP1 phagocytic axis in macrophages.
• Fetal Lung: Upregulation of TRIM21/TRIM65 (ubiquitin ligases) in AT2 cells alongside ITCH.
• Neuroblastoma: Differential usage of AP-1 transcription factors (JUND at diagnosis vs. JUNB/FOS post-therapy).

5. Significance

• Architectural Insight: The study demonstrates that explicit modality routing is superior to implicit contrastive fusion for single-cell retrieval. It proves that text-supervised alignment often sacrifices the fine-grained transcriptomic structure required for gene-signature queries.
• Bridging the Gap: ELISA successfully bridges the divide between "data exploration" (finding relevant cells) and "biological discovery" (interpreting mechanisms), enabling researchers to move from raw atlas data to structured hypotheses in a single session.
• Reproducibility and Transparency: By enforcing strict grounding of LLM outputs in dataset evidence, ELISA mitigates the risk of hallucination, a critical requirement for scientific AI agents.
• Scalability: The system operates on cluster-level embeddings without requiring the original count matrix, making it computationally efficient and deployable on standard hardware.

In conclusion, ELISA represents a significant step forward in agentic AI for genomics, providing a robust, interpretable, and high-performance tool for translating complex single-cell data into actionable biological insights.