scPilot: LLM-Enabled Single-Cell Analysis

• scPilot is a framework that automates single-cell RNA-seq analysis by integrating omics-native reasoning through large language models.
• It converts raw expression data into semantic sketches, enabling iterative, auditable reasoning that links biological claims to computational evidence.
• Paired with scBench, scPilot enhances cell-type annotation, trajectory inference, and GRN prediction using transparent, tool-guided workflows.

Searching arXiv for the scPilot paper and closely related single-cell reasoning context. scPilot is a framework for automated single-cell RNA-seq analysis that uses LLMs to perform what its authors call omics-native reasoning: the model converses in natural language while inspecting single-cell data summaries derived from raw expression matrices, invoking bioinformatics tools on demand, and producing a transparent reasoning trace that links biological claims to computational evidence (Gao et al., 12 Feb 2026). The system targets three core tasks—cell-type annotation, developmental trajectory reconstruction, and transcription-factor targeting—and is paired with scBench, a benchmark suite of 9 curated datasets intended to evaluate this style of reasoning systematically (Gao et al., 12 Feb 2026).

1. Conceptual basis and problem formulation

The framework is motivated by the claim that single-cell analysis remains bottlenecked by human-only reasoning even when mature computational pipelines are available. In the formulation used by the paper, conventional workflows can be written as

$y = g_{\text{tool}}(X; \theta),$

where $X \in \mathbb{R}^{G \times N}$ is the expression matrix, $g_{\text{tool}}$ denotes a tool such as Scanpy or Monocle, and $\theta$ denotes manually tuned hyperparameters (Gao et al., 12 Feb 2026). The criticism is not that such tools are ineffective, but that the biological assumptions encoded in parameter choices and downstream interpretation are often implicit rather than auditable.

The paper contrasts this with code-generating tool agents, expressed as

$$y = g_{\text{src}_K}\big(g_{\text{src}_{K-1}}(\cdots g_{\text{src}_1}(X)\cdots)\big),$$

arguing that in those systems the causal chain between data evidence and biological conclusion is often fragmented across code, logs, and dialogue (Gao et al., 12 Feb 2026). scPilot is presented as a response to this gap: not merely a tool runner, and not merely a summarizer, but a framework in which reasoning over omics evidence is the primary computational object.

The central formalism is omics-native reasoning (ONR). The initial state is the raw data,

$S_0 = X.$

At reasoning step $k$, the model emits a pair $(c_k, o_k)$, where $c_k$ is a natural-language claim or justification and $o_k$ is a primitive omics operator acting on the current state. State evolution is

$X \in \mathbb{R}^{G \times N}$0

and the full trace is

$X \in \mathbb{R}^{G \times N}$1

The final answer is obtained from the terminal state $X \in \mathbb{R}^{G \times N}$2 via $X \in \mathbb{R}^{G \times N}$3 (Gao et al., 12 Feb 2026). The paper characterizes this as a “verbal + computational proof.” This suggests that scPilot’s main innovation is methodological: it reframes single-cell analysis as a sequence of inspectable claim–evidence transitions rather than as a one-shot prediction or opaque pipeline.

2. System architecture and workflow

scPilot comprises three principal components: a Problem-to-Text Converter $X \in \mathbb{R}^{G \times N}$4, a Bio-Tool Library $X \in \mathbb{R}^{G \times N}$5, and an LLM Reasoner $X \in \mathbb{R}^{G \times N}$6 (Gao et al., 12 Feb 2026). The converter exists because full expression matrices are too large for direct inclusion in an LLM context window. It constructs a task-specific semantic sketch

$X \in \mathbb{R}^{G \times N}$7

where $X \in \mathbb{R}^{G \times N}$8 is a compressed textual abstraction of the dataset tailored to query $X \in \mathbb{R}^{G \times N}$9 (Gao et al., 12 Feb 2026). The paper stresses that this conversion is algorithmic rather than learned.

Depending on the task, the semantic sketch may include cluster sizes, top marker genes, dotplot summaries, pseudotime or lineage graph structure, TF–gene candidate scores, and Gene Ontology overlap (Gao et al., 12 Feb 2026). This is meant to preserve the information that a human analyst would inspect while discarding the full matrix.

The Bio-Tool Library wraps established single-cell software, including Scanpy, Seurat via Reticulate, Monocle 3, pySCENIC, and plotting utilities (Gao et al., 12 Feb 2026). Each tool returns structured JSON outputs together with short textual descriptions. This design choice is technically important because it constrains the LLM to reason over machine-readable evidence rather than arbitrary unstructured logs.

The reasoner is instantiated with off-the-shelf models such as o1 and Gemini-2.5-Pro (Gao et al., 12 Feb 2026). The end-to-end workflow is summarized in the paper as

$g_{\text{tool}}$0

Operationally, a natural-language query is converted into a reasoning loop in which the model proposes hypotheses, requests tool outputs, revises prior beliefs if necessary, and emits a final answer with an accompanying rationale (Gao et al., 12 Feb 2026).

The paper emphasizes three design principles across tasks: biological context is included explicitly in prompts; reasoning is iterative rather than one-shot; and gains are obtained without task-specific weight fine-tuning, relying instead on prompting structure and evidence integration (Gao et al., 12 Feb 2026).

3. Task design and reasoning procedures

scPilot supports three task families, each converted into an ONR workflow rather than treated as direct prediction.

For cell-type annotation, the inputs include cluster assignments, top-$g_{\text{tool}}$1 marker genes per cluster with default $g_{\text{tool}}$2, metadata such as tissue and species, and prior iteration outputs when applicable (Gao et al., 12 Feb 2026). The textual representation used in scBench is “Scanpy-Leiden clusters + top-$g_{\text{tool}}$3 marker genes per cluster ($g_{\text{tool}}$4)” (Gao et al., 12 Feb 2026). Annotation proceeds through a propose–filter–solve loop. The model first generates cell-type hypotheses from top differentially expressed genes and context, then proposes additional marker genes for unresolved clusters, then evaluates tool-generated dotplot-style evidence to determine which genes are informative or ambiguous, and which clusters should be stabilized or revisited (Gao et al., 12 Feb 2026). The paper states that annotation uses a maximum of three reasoning iterations.

A canonical biological example is the PBMC case in which NKG7 alone is treated as insufficiently specific, whereas the joint pattern CD3D + GNLY + NKG7 is used to distinguish NK cells from cytotoxic T cells, especially when CD8A is weak; plasma-cell hypotheses are excluded when SDC1 is absent (Gao et al., 12 Feb 2026). This is used to illustrate that scPilot attempts multi-marker reasoning rather than single-marker lookup.

For developmental trajectory reconstruction, the inputs include clustered scRNA-seq data, top marker genes, developmental timepoint proportions per cluster, biological context, and PAGA- or Monocle-derived trajectory evidence (Gao et al., 12 Feb 2026). The workflow first annotates clusters using top genes and timepoint distributions, then identifies the most likely root cluster, then constructs a directional tree over remaining clusters, and then performs a Monocle-guided self-audit in which missing progenitor relations, annotation mismatches, hierarchy errors, and implausible transitions are explicitly reconsidered (Gao et al., 12 Feb 2026). The output is a nested trajectory tree together with refined cluster annotations and a rationale for root and branch assignments.

For transcription-factor targeting / GRN prediction, the inputs include a TF–gene pair, tissue context, pySCENIC or GRNdb candidate edges with motif support, and Gene Ontology overlap between TF and gene (Gao et al., 12 Feb 2026). The model is prompted to recall TF biology, compare the candidate gene with known targets, and return both a reasoning string and a possibility score in $g_{\text{tool}}$5 (Gao et al., 12 Feb 2026). The paper uses examples such as Stat1 $g_{\text{tool}}$6 Irf7 and Klf4 $g_{\text{tool}}$7 Muc5ac as successful predictions, and Usf2 $g_{\text{tool}}$8 Pigr and Fos $g_{\text{tool}}$9 Hmox1 as informative failures (Gao et al., 12 Feb 2026). These cases are presented as evidence that ontology overlap and tissue context can both help and mislead the model.

This task structure suggests that scPilot’s contribution is not only the use of tools, but the explicit decomposition of biological inference into revisable subclaims.

4. Benchmark construction and evaluation methodology

The companion benchmark scBench contains 9 expertly curated datasets: 3 for cell-type annotation, 3 for trajectory inference, and 3 for GRN prediction (Gao et al., 12 Feb 2026). The benchmark is intended to measure what the paper terms omics-native reasoning rather than only generic LLM competence.

For cell-type annotation, the datasets are PBMC3k (2,638 cells $\theta$0 13,714 genes, 8 cell types), Liver (41,000 cells $\theta$1 2,000 HVGs, 31 cell types), and Retina (20,091 cells $\theta$2 19,719 genes, 9 cell types) (Gao et al., 12 Feb 2026). Ground truth comes from author-provided labels. Evaluation is ontology-aware: predicted names are standardized to Cell Ontology IDs and scored as $\theta$3 for exact matches, $\theta$4 for partial matches via parent–child overlap, and $\theta$5 otherwise (Gao et al., 12 Feb 2026).

For trajectory inference, the datasets are Pancreas (36,351 cells, 17,327 genes, 4 timepoints, 14 trajectory nodes), Liver (44,010 cells, 2,000 HVGs, 4 timepoints, 15 nodes), and Neocortex (33,976 cells, 35,543 genes, 2 timepoints, 16 nodes) (Gao et al., 12 Feb 2026). Ground truth is a manually curated lineage tree from the original study. The metrics are node Jaccard similarity,

$\theta$6

graph edit distance,

$\theta$7

and spectral distance,

$\theta$8

(Gao et al., 12 Feb 2026).

For GRN prediction, the tissues are Stomach (23 verified TF–gene edges, 46 total questions), Liver (71 verified edges, 142 questions), and Kidney (49 verified edges, 98 questions) (Gao et al., 12 Feb 2026). Positives are validated against TRRUST v2, while negatives are sampled to avoid validated or SCENIC-supported edges (Gao et al., 12 Feb 2026). The main metric is AUROC,

$\theta$9

with

$$y = g_{\text{src}_K}\big(g_{\text{src}_{K-1}}(\cdots g_{\text{src}_1}(X)\cdots)\big),$$0

The paper also specifies the confusion-matrix form

$$y = g_{\text{src}_K}\big(g_{\text{src}_{K-1}}(\cdots g_{\text{src}_1}(X)\cdots)\big),$$1

(Gao et al., 12 Feb 2026).

The benchmark uses pre-specified termination conditions rather than allowing models to decide when to stop reasoning, which the authors frame as a reproducibility measure (Gao et al., 12 Feb 2026).

5. Empirical results

The paper reports that iterative omics-native reasoning improves average cell-type annotation accuracy by about 11% over direct prompting (Gao et al., 12 Feb 2026). In Table 2, scPilot (o1) achieves 0.518 on Liver, 0.792 on PBMC, and 0.728 on Retina, compared with Direct (o1) at 0.560, 0.667, and 0.474 respectively (Gao et al., 12 Feb 2026). Table 3 reports that the best scPilot values include 0.792 for PBMC3k with o1 and also 0.792 with Gemini-2.0-Pro, 0.763 for Retina with Gemini-2.0-Pro, and 0.518 for Liver with o1 (Gao et al., 12 Feb 2026). The largest median gain is reported on Retina at +0.180, with smaller gains on PBMC3k (+0.042) and Liver (+0.024) (Gao et al., 12 Feb 2026).

For trajectory inference, the paper states that Gemini-2.5-Pro under scPilot cuts trajectory graph-edit distance by 30% versus one-shot prompting, and elsewhere summarizes an average reduction of 26% (Gao et al., 12 Feb 2026). Table 4 gives representative improvements. For Gemini-2.5-Pro, GED-nx improves from 8.33 to 5.00 on Pancreas, 8.00 to 3.33 on Liver, and 13.33 to 9.50 on Neocortex; Jaccard improves to 1.000 on all three; spectral distance improves from 0.453 to 0.310 on Pancreas and 0.388 to 0.199 on Liver, with a slight worsening on Neocortex from 0.977 to 1.052 (Gao et al., 12 Feb 2026). The paper reports median improvements of $$y = g_{\text{src}_K}\big(g_{\text{src}_{K-1}}(\cdots g_{\text{src}_1}(X)\cdots)\big),$$2 in GED and $$y = g_{\text{src}_K}\big(g_{\text{src}_{K-1}}(\cdots g_{\text{src}_1}(X)\cdots)\big),$$3 in spectral distance across model–metric pairs (Gao et al., 12 Feb 2026).

For GRN prediction, the contribution summary reports an AUROC improvement of 0.03 over direct prompting, while the task-specific section reports an average +0.098 across tissues and models (Gao et al., 12 Feb 2026). Table 5 shows, for o1, Stomach improving from 0.827 \pm 0.002 to 0.873 \pm 0.004, Liver from 0.753 \pm 0.001 to 0.760 \pm 0.000, and Kidney from 0.777 \pm 0.000 to 0.797 \pm 0.001 (Gao et al., 12 Feb 2026). The paper notes that GPT-4o shows the largest relative improvement, with average AUROC gain +0.162 (Gao et al., 12 Feb 2026). On the Stomach subset, Table 7 reports scPilot (o1) at 0.873 \pm 0.004, exceeding graph-based baselines and direct o1 at 0.827 \pm 0.002 (Gao et al., 12 Feb 2026).

Across the full benchmark, the paper states that scPilot wins in 87 of 108 total comparisons (Gao et al., 12 Feb 2026). This suggests broad but not universal improvement. The authors are explicit that scPilot does not dominate every dataset or model.

6. Interpretation, limitations, and future directions

A central claim of the framework is that it improves auditability by tying each conclusion to prompt stages, tool outputs, and explicit natural-language justifications (Gao et al., 12 Feb 2026). The paper repeatedly contrasts this with pipelines that either produce only outputs or hide the rationale in code and intermediate files. In scPilot, the reasoning trace is treated as a first-class artifact. A plausible implication is that the framework is designed as much for expert inspection and debugging as for automation.

The paper also presents several ablations. Removing dataset-level metadata from PBMC3k reduces annotation accuracy from 0.792 to 0.688 for o1, 0.646 to 0.583 for GPT-4o, and 0.604 to 0.416 for GPT-4o-mini (Gao et al., 12 Feb 2026). Shuffling GO annotations in GRN prediction lowers Stomach AUROC from 0.873 to 0.813 for o1, 0.800 to 0.710 for GPT-4o, and 0.697 to 0.617 for GPT-4o-mini (Gao et al., 12 Feb 2026). Corrupting Monocle inputs worsens Liver trajectory metrics for o1, with Jaccard falling from 1.000 to 0.933, GED-nx rising from 8.00 to 11.33, and spectral distance rising from 0.567 to 0.593 (Gao et al., 12 Feb 2026). These experiments are used to argue that the system is genuinely using biological context and tool outputs rather than ignoring them.

The framework’s limitations are described explicitly. Compression into semantic sketches may discard subtle rare-cell signals (Gao et al., 12 Feb 2026). On complex datasets, deeper reasoning can lead to “overthinking”; the Liver annotation result with o1—0.518 for scPilot versus 0.560 for direct prompting—is cited as a case where iterative reasoning amplified ambiguity among related developmental states (Gao et al., 12 Feb 2026). Performance also depends on the quality of upstream tools, as shown by the Monocle and GO perturbation studies (Gao et al., 12 Feb 2026). The paper notes that plausible but incorrect biological narratives remain possible, especially in GRN tasks when GO overlap is overinterpreted (Gao et al., 12 Feb 2026).

The open-source model results are presented as another practical limitation. Gemma-3-27B is reported as substantially slower and weaker than proprietary models; PBMC3k inference takes 135.7 s on four A100 80GB GPUs, compared with 8.8 s for a comparable GPT-4o task (Gao et al., 12 Feb 2026). The authors conclude that fully on-premise ONR is not yet practical for most laboratories. They also report per-call runtimes for TF–gene prediction ranging from 3.75 s for GPT-4o-mini to 30.61 s for Gemini-2.0-Pro, with o1 at 11.42 s (Gao et al., 12 Feb 2026).

Future directions proposed in the paper include improved data compression for rare populations, larger-context ONR workflows, retrieval-augmented biological reasoning, stronger hallucination mitigation, and the incorporation of wet-lab feedback for experimental validation (Gao et al., 12 Feb 2026). This suggests that scPilot is intended not only as an automation layer over existing pipelines, but as an initial template for a broader class of LLM-guided, evidence-grounded bioinformatics systems.

In the paper’s own positioning, scPilot is not merely an LLM wrapper for single-cell software. It is a framework that formalizes single-cell analysis as a sequence of inspectable biological claims and tool-grounded revisions, with scBench serving as the corresponding evaluation substrate (Gao et al., 12 Feb 2026).