scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery

Conference: NeurIPS 2025 arXiv: 2602.11609 Code: https://github.com/maitrix-org/scPilot Area: Interpretability Keywords: single-cell RNA-seq, LLM reasoning, omics-native reasoning, cell-type annotation, trajectory inference

TL;DR

This work proposes the scPilot framework and scBench benchmark, enabling LLMs to perform "omics-native reasoning" (ONR) directly on single-cell RNA-seq data—reading marker genes, forming hypotheses, invoking tools for verification, and iteratively refining conclusions—achieving an 11% improvement in cell-type annotation accuracy and a 30% reduction in trajectory inference graph-edit distance.

Background & Motivation

Background: Single-cell RNA-seq analysis relies on fixed pipelines (Scanpy, Seurat), with large amounts of implicit expert reasoning (e.g., differential genes → cell-type judgment) remaining unautomated. Existing LLM applications treat LLMs merely as "code generators" that invoke pre-existing tools.

Limitations of Prior Work: (a) Single-cell foundation models (e.g., scGPT) embed gene expression into vector spaces, sacrificing interpretability; (b) LLM code agents only wrap tools with default parameters without performing biological reasoning; (c) the biological logic underlying the analysis process is opaque.

Key Challenge: Single-cell analysis demands extensive expert reasoning (identifying cell types from marker genes, inferring developmental relationships from lineage trajectories), yet existing automated tools perform computation without reasoning.

Goal: To enable LLMs not merely to invoke tools, but to interpret data, formulate hypotheses, gather evidence, and iteratively refine conclusions in the manner of a biologist.

Key Insight: Define the omics-native reasoning (ONR) paradigm—LLMs receive textual summaries of single-cell data, perform explicit reasoning, invoke tools to obtain numerical evidence, and iterate until biological conclusions are reached.

Core Idea: Single-cell analysis is formalized as a natural language reasoning problem, where LLMs produce (claim, action) pairs at each step, constituting a dual-track "verbal + computational" proof.

Method

Overall Architecture

Three core components: (1) Problem-to-Text Converter \(\mathcal{C}\): compresses expression matrices of \(10^5\)–\(10^6\) cells into LLM-digestible textual summaries (e.g., cluster sizes, top-\(k\) marker genes); (2) Bio-Tool Library \(\mathcal{T}\): encapsulates Scanpy, Monocle, pySCENIC, and other tools as callable structured APIs; (3) LLM Reasoner \(\mathcal{R}_\phi\): uses reasoning LLMs such as o1/Gemini as the core, executing closed-loop reasoning \(\mathbf{X} \to \text{Prompt} \to \{(\text{Thought}_k, \text{Call}_k)\}_{k=1}^K \to \hat{y}\).

Key Designs

1. Omics-Native Reasoning (ONR) Formalization:

   • Function: Defines bioinformatics analysis tasks as a reasoning sequence \(\mathcal{R} = [(c_1,o_1), \ldots, (c_K,o_K)]\)
   • Mechanism: At each step, the LLM produces a natural language claim \(c_k\) (e.g., "cluster 5 highly expresses CD3D and CD3E, likely T cells") and an action \(o_k\) (e.g., "check NK cell marker genes"), with each action updating the data state \(S_k = o_k(S_{k-1})\)
   • Design Motivation: The key distinction from code agents—the reasoning trace constitutes an auditable biological argument, not merely code and output

2. Problem-to-Text Compression:

   • Function: Compresses million-scale cell matrices into text processable within an LLM context window
   • Mechanism: Task-specific compression strategies are designed: Leiden clustering + top-10 marker genes for cell annotation; PAGA graph + pseudotime for trajectory inference; top-150 TF–gene pairs for GRN
   • Design Motivation: Preserving biologically significant information while drastically reducing dimensionality, enabling LLMs to operate in the text domain

3. scBench Benchmark:

   • Function: Covers nine datasets across three major tasks (cell-type annotation, trajectory inference, gene regulatory network prediction)
   • Mechanism: Each task includes expert-verified ground truth and automated evaluation metrics (accuracy, graph-edit distance, AUROC)
   • Design Motivation: Existing single-cell benchmarks evaluate only embedding quality or numerical metrics, without assessing the biological validity of reasoning

Loss & Training

scPilot is a training-free framework that does not fine-tune LLMs. All reasoning capabilities derive from prompt engineering and iterative reasoning strategies. Core design principles: (a) biological context priority; (b) iterative reasoning; (c) minimal hand-crafted heuristics.

Key Experimental Results

Main Results

Task	Dataset	scPilot (o1)	Best Baseline	Gain
Cell Annotation	PBMC3k	~0.76	CellTypist 0.563	+35%
Cell Annotation	Liver	~0.50	CellTypist 0.464	+8%
Cell Annotation	Retina	~0.49	CellTypist 0.388	+26%
Trajectory Inference	Pancreas	GED reduced 30%	Traditional pipeline	Gemini-2.5-Pro best
GRN Prediction	Multi-organ	AUROC +0.03	pySCENIC direct output	Iterative reasoning gain

Ablation Study

Configuration	Effect	Note
Direct prompting (no iteration)	Baseline	Single-pass reasoning
Iterative reasoning (2–3 rounds)	+11% avg accuracy	Iterative hypothesis refinement
Without biological context	Significant drop	Species/tissue information is critical
LLM comparison	o1 best for annotation, Gemini best for trajectory	Task-specific LLM capability

Key Findings

• Iterative reasoning is critical—LLMs frequently err in the first round (e.g., confusing NK and T cells) but self-correct in the second round upon examining additional marker genes
• LLMs can identify potential issues in expert annotations—in some cases, scPilot's reasoning is more consistent than the original labels
• Different LLMs excel at different tasks: o1's strong reasoning suits annotation, while Gemini's large context window suits trajectory inference
• Reasoning traces are highly interpretable—biologists can audit the logic at every step

Highlights & Insights

• Paradigm Shift: From "LLM invoking tools" to "LLM performing biological reasoning." scPilot automates the expert thought process, not merely the pipeline
• Scientific Value of Reasoning Traces: Generated traces expose marker gene ambiguity, tissue-specific expression patterns, and other insights of independent analytical value to biologists
• Generalizable ONR Framework: The same "data → textual summary → LLM reasoning → tool verification" paradigm is transferable to proteomics, metabolomics, and other omics domains

Limitations & Future Work

• Performance depends on the quality of Problem-to-Text compression—information loss may introduce reasoning bias
• Current coverage is limited to three core tasks; spatial transcriptomics, multi-omics integration, and related areas are not addressed
• Full reliance on the LLM's biological knowledge may be insufficient for very novel or rare cell types
• Multi-round LLM reasoning per analysis incurs substantial computational cost (o1 API expenses)

Related Work & Insights

• vs. CellAgent: CellAgent instructs LLMs to write code invoking Scanpy but performs no biological reasoning; scPilot interprets differential genes and formulates hypotheses
• vs. scGPT/Geneformer: These models operate in vector space and produce no natural language reasoning; scPilot's traces constitute auditable arguments
• vs. GPTCellType: Addresses only cell annotation via direct prompting for a single task; scPilot covers multiple tasks with iterative reasoning

Rating

• Novelty: ⭐⭐⭐⭐⭐ The formal definition and systematic framework of "omics-native reasoning" is entirely novel, opening a new paradigm for LLMs in computational biology
• Experimental Thoroughness: ⭐⭐⭐⭐ Nine datasets, three tasks, and multiple LLMs and baselines, though gains on the GRN task are modest
• Writing Quality: ⭐⭐⭐⭐ Framework description is clear, though the mathematical formalization is somewhat over-notated
• Value: ⭐⭐⭐⭐⭐ Potentially transformative for the computational biology community—demonstrating the feasibility of LLMs as "scientific reasoning partners" rather than "code generators"

Related Papers

• [NeurIPS 2025] Latent Principle Discovery for Language Model Self-Improvement
• [NeurIPS 2025] Probabilistic Token Alignment for Large Language Model Fusion
• [NeurIPS 2025] Encoding and Understanding Astrophysical Information in Large Language Model-Generated Summaries
• [NeurIPS 2025] Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models
• [NeurIPS 2025] Curvature Tuning: Provable Training-free Model Steering From a Single Parameter