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

We present scPilot, the first systematic framework to practice omics-native reasoning: a large language model (LLM) converses in natural language while directly inspecting single-cell RNA-seq data and on-demand bioinformatics tools. scPilot converts core single-cell analyses, i.e., cell-type annotation, developmental-trajectory reconstruction, and transcription-factor targeting, into step-by-step reasoning problems that the model must solve, justify, and, when needed, revise with new evidence. To measure progress, we release scBench, a suite of 9 expertly curated datasets and graders that faithfully evaluate the omics-native reasoning capability of scPilot w.r.t various LLMs. Experiments with o1 show that iterative omics-native reasoning lifts average accuracy by 11% for cell-type annotation and Gemini-2.5-Pro cuts trajectory graph-edit distance by 30% versus one-shot prompting, while generating transparent reasoning traces explain marker gene ambiguity and regulatory logic. By grounding LLMs in raw omics data, scPilot enables auditable, interpretable, and diagnostically informative single-cell analyses. Code, data, and package are available at https://github.com/maitrix-org/scPilot

💡 Research Summary

The paper introduces scPilot, the first systematic framework that brings large language models (LLMs) into the core of single‑cell RNA‑seq analysis through a paradigm called omics‑native reasoning (ONR). Traditional single‑cell pipelines rely on human experts to interpret high‑dimensional expression matrices, choose tools, set parameters, and draw biological conclusions. Existing LLM applications in bioinformatics either use the model as a natural‑language front‑end for fixed tools or embed the data into opaque vector spaces, offering little interpretability. scPilot changes this by requiring the LLM to (i) receive a concise textual sketch of the raw matrix, (ii) formulate biological hypotheses in natural language, (iii) invoke targeted bioinformatics operators (e.g., Scanpy clustering, Monocle trajectory inference, pySCENIC GRN inference) directly on the data, (iv) evaluate the numerical output, and (v) iteratively refine its claims until a biologically coherent answer is reached.

Formally, the reasoning proceeds as a sequence of claim‑operator pairs (cₖ, oₖ). Each claim cₖ is a natural‑language statement (hypothesis, justification, decision) and each operator oₖ is a primitive, verifiable action on the current data state (filtering, clustering, scoring, lookup). The chain R = {(c₁,o₁)…(c_K,o_K)} constitutes a verbal‑computational proof; the final state S_K is mapped to a prediction ŷ that answers the user query. This explicit trace makes the analysis auditable and allows human experts to follow, critique, or correct any step.

The architecture consists of three components: (1) a problem‑to‑text converter Φ_q that algorithmically compresses a massive G × N expression matrix into a digestible summary s_q (cluster sizes, top‑k markers, lineage connections, TF‑target scores); (2) a curated bio‑tool library T providing primitive operators from established packages (Scanpy, Seurat via Reticulate, Monocle 3, pySCENIC) that return structured JSON together with a short natural‑language description; (3) an LLM reasoner R_ϕ (e.g., o1, Gemini‑2.5‑Pro) that receives s_q, the user query, and a reasoning scaffold, then generates iterative “thought” and “call” steps. Design principles emphasize (a) biological context first – prompts always embed species, tissue, and experimental protocol; (b) iterative reasoning – each tool call is followed by reflection and possible hypothesis revision; (c) minimal manual heuristics – only high‑level prompts are supplied, performance gains arise from richer evidence and better prompting.

To evaluate scPilot, the authors release scBench, a benchmark suite of nine expertly curated single‑cell datasets covering three canonical tasks: cell‑type annotation (PBMC3k, Liver, Retina), developmental trajectory inference (Pancreas, Liver, Neocortex), and gene‑regulatory network (GRN) edge prediction (stomach, liver, kidney). For each task, scBench provides a compressed textual representation, ground‑truth labels (author‑provided or curated lineage trees), and task‑specific metrics (cluster‑level accuracy, graph‑edit distance, AUROC). Automatic termination conditions are pre‑specified to prevent uncontrolled LLM looping.

Experimental results show that ONR substantially improves performance over one‑shot prompting and over existing specialized tools. Using o1 as the LLM, scPilot raises average cell‑type annotation accuracy by 11 % (e.g., from 0.56 to 0.79 on Liver) and reduces trajectory graph‑edit distance by roughly 30 % compared with Gemini‑2.5‑Pro in a one‑shot setting. In GRN prediction, scPilot improves AUROC by 0.03 over direct prompting. Crucially, scPilot generates transparent reasoning traces that expose marker‑gene ambiguities, lineage inconsistencies, and tissue‑specific regulatory logic, enabling diagnostic inspection and expert validation.

Limitations include the dependence on the quality of the Φ_q summarizer (LLM context windows still constrain how much raw data can be presented), the lack of robust automatic error‑recovery when a tool call fails, and the inherent risk of LLM hallucinations, which necessitates expert oversight. Future work is suggested to integrate long‑memory architectures, multimodal inputs (e.g., spatial transcriptomics), and confidence‑estimation modules, as well as to expand the tool library to cover perturbation analysis and multi‑omics integration.

In summary, scPilot demonstrates that large language models can be co‑pilots for single‑cell analysis, delivering automation, interpretability, and auditability across the full analytical workflow. The accompanying scBench benchmark provides a reproducible platform for measuring progress and for comparing future LLM‑tool integration strategies in the rapidly evolving field of single‑cell genomics.