acceptodds
Withdrawn from ICLR 2027

CellScientist: Autonomous Biological Discovery via Scientific Alignment and Multifaceted Biological Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in large language models have catalyzed growing interest in autonomous AI-driven scientific discovery, yet existing frameworks face a fundamental limitation when applied to knowledge-intensive domains such as cellular biology: they either produce scientifically invalid hypotheses that violate experimental constraints and physical laws, or generate methodologically sound but trivial findings that lack biological significance. To address this dual challenge, we observe that rigorous autonomous discovery requires both strict enforcement of validity constraints throughout the research process and systematic injection of domain knowledge to guide hypothesis search toward scientifically meaningful questions. We present CellScientist, an autonomous agent framework that implements this insight through a dual-grounding paradigm. First, Scientific Semantic Alignment enforces formal consistency among research questions, experimental protocols, statistical estimands, and biochemical plausibility, rejecting internally contradictory proposals before resource expenditure. Second, Comprehensive Biological Prior Injection integrates structured ontologies, automated literature retrieval, toolchain execution proficiencies, and empirical research heuristics across multiple decision points, steering the search toward biologically interpretable discoveries. Together, these mechanisms enable CellScientist to autonomously formulate research questions, design and execute multi-stage experiments, critically evaluate evidence through falsification tests, and synthesize findings into coherent scientific manuscripts. Evaluated across 20 single-cell and spatial omics research topics, CellScientist improves composite manuscript quality by 15–19% over the strongest general-purpose autonomous research agent and by 29–37% over prior AI Scientist systems, with the largest gains in reproducibility (+17–30%) and data-conclusion traceability (+24–35%). Running end to end without human intervention, CellScientist arrived at findings that revise prevailing expectations in the systems it studied: in post-influenza lung, it established that apparent myeloid transcriptional recovery is almost entirely demographic, with conditioning on cellular composition collapsing unique stage-associated variance from 72.7% to 0.80%; in human kidney, it showed that lineage-intrinsic transcriptional remodeling, rather than the compositional shifts long assumed to dominate, accounts for 92.09% of corticomedullary variance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Autonomous Scientific Discovery, AI for Science, Cell Biology, Scientific Semantic Alignment, Biological Priors, LLM Agents, Multi-source Knowledge Injection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2dDbOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents CellScientist, an autonomous agent framework for single-cell and spatial omics research. It combines Scientific Semantic Alignment, which enforces consistency among research questions, estimands, experimental designs, and conclusions, with Biological Prior Injection, which incorporates literature, domain knowledge, and analysis expertise to guide hypothesis generation. The system further performs iterative experiment search, validation, falsification, and manuscript generation.

CellScientist is evaluated on 20 real omics research topics against AI Scientist and OpenScientist. It achieves higher LLM-judged manuscript scores, particularly in reproducibility, research design coherence, and data–conclusion traceability. Ablation studies support the contributions of the two main components, and several biological case studies are provided to illustrate the generated findings.

Critical strengths

A key strength is the paper’s formulation of an important dilemma for autonomous scientific discovery: an AI scientist must avoid both scientifically invalid but seemingly interesting studies and methodologically sound but biologically trivial ones. This framing is meaningful because it goes beyond simply automating the research workflow and highlights two distinct requirements for useful scientific discovery: rigor and significance. The proposed separation between Scientific Semantic Alignment and Biological Prior Injection is a reasonable and intuitive response to this challenge.

Critical weaknesses
  1. The main evaluation does not directly validate the paper’s central claim. The paper is motivated by a dilemma between scientifically invalid but seemingly interesting studies and scientifically valid but biologically trivial ones. However, the primary evaluation reduces research quality to scores assigned by another LLM to AI-generated manuscripts. Because the evaluation rubric itself emphasizes properties such as rigor, reproducibility, coherence, and traceability—properties explicitly built into CellScientist—the reported gains are difficult to disentangle from alignment with the judging prompt. More importantly, aggregate scores do not show whether the framework actually prevents the concrete scientific failures it is designed to address.

  2. The paper lacks convincing case-level evidence demonstrating how the proposed mechanisms improve scientific discovery. The reported biological case studies mainly present final statistical findings, but do not clearly show how baseline agents fail, where Scientific Semantic Alignment or Biological Prior Injection intervenes, and how the resulting research question, analysis, or conclusion is materially improved. A small number of carefully analyzed paired examples would provide much stronger evidence for the claimed resolution of the rigor–significance dilemma than the current manuscript-level score comparisons.

Questions for authors

Clarification requests

  1. How sensitive are the reported manuscript-level scores to the exact evaluation prompt and rubric? In particular, have you tested alternative judge prompts, different evaluators, or human expert assessment to determine whether the observed gains persist beyond the current LLM-based scoring setup?

  2. Can you provide concrete paired examples in which a baseline system produces either a scientifically invalid study or a biologically trivial one, and show specifically how Scientific Semantic Alignment or Biological Prior Injection changes the research question, analysis, or conclusion? Such examples would materially strengthen the claim that the framework resolves the dilemma motivating the paper.

  3. For the biological case studies, how do you ensure that the reported findings are genuine discoveries rather than rediscoveries guided by retrieved literature? In particular, are the original source papers or closely related downstream studies excluded from the retrieval corpus during hypothesis generation?

Reviewer_adJ7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce CellScientist, an autonomous AI agent for biological discovery. The central premise is that existing AI scientist systems are often poorly grounded in biological knowledge and can generate scientifically inconsistent or biologically insignificant results. To address this, the authors introduce two components: scientific semantic alignment and comprehensive biological prior injection. The authors evaluate the system across 20 single-cell and spatial omics research topics, generating manuscripts that are then graded by Claude Opus 5. They find that Opus 5 grades the manuscripts produced by CellScientist more favorably than those generated by other AI scientist systems. They also present three biological case studies and ablation experiments supporting the contributions of their two components.

Critical strengths

The authors address an important problem. AI scientist systems can sometimes be poorly grounded in scientific knowledge and generate scientifically inconsistent results. Improving the scientific grounding of these systems is an important direction.

Critical weaknesses

First, there are already many existing works on autonomous AI for biological discovery, such as the Virtual Biotech from James Zou, AutoScientists from Marinka Zitnik, and Medea, which is an AI agent for analyzing omics data. These are just some examples of an already extensive literature developing AI agent systems that aim to address many of the same problems the authors are trying to address. It is unclear to me what CellScientist adds relative to these existing approaches, and for that reason, the novelty seems extremely limited.

Second, the benchmarking is problematic. The authors use Claude Opus 5 to grade generated manuscripts as a way to demonstrate that their method produces better-grounded scientific research. This raises a somewhat circular concern: if the central claim is that language models are not good at conducting scientifically grounded research, why should we trust a language model to reliably evaluate scientific grounding in the first place? Other benchmarks in this field use objective experimental outcomes, such as performance on actual data or improvements in model performance, to measure the quality of agent-generated research. The authors do perform computational analyses on real data and present biological case studies, but their central comparative claims still rely on LLM-based manuscript grading rather than independent validation of scientific correctness. On a related note reproducibility is often the axis by which the LLM says CellScientist does the best but this arises from a consideration of the reproducibility, the LLM does not try to reproduce the findings itself to qualify this. A manuscript can describe reproducible work without being reproducible.

Finally, I am not convinced by the central premise that current AI models necessarily suffer from the grounding problems described here to the extent the authors suggest. Earlier, less capable models may have exhibited these problems more frequently, but frontier models have improved substantially and can now be quite well grounded in scientific knowledge. Although the authors evaluate relatively recent model backbones, I think they need to establish that these models actually exhibit the particular problems they are trying to solve, and that these problems will not simply diminish as the underlying models continue to improve. Why do we need this additional method if stronger models may address the same issues through scaling?

Questions for authors

See Weaknesses

Reviewer_eXgnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CellScientist, an autonomous research framework for generating biologically meaningful questions and conducting verifiable single-cell and spatial-omics analyses. It combines Research Contracts for scientific semantic alignment with retrieval of biological and methodological priors, staged analysis execution, and evidence-linked manuscript generation. Experiments across 20 research topics, multiple backbones, and component ablations show improvements in automated manuscript scores. However, independent validation of scientific correctness, reproducibility, and biological novelty remains limited.

Critical strengths
  1. Research Contracts and staged artifact management address important scientific validity concerns, including inconsistent estimands, inappropriate inference units, and unsupported conclusions.

  2. Evaluations across multiple backbones and component ablations provide useful empirical evidence. Topic-level uncertainty estimates also account for dependencies among manuscripts generated from the same topic.

Critical weaknesses
  1. Automated manuscript scores lack reported calibration against expert assessments or independent reruns, making it difficult to distinguish reporting quality from scientific correctness. Unclear execution budgets and successful-run selection further limit the reliability of the comparisons.

  2. An apparent scaling error raises concerns about the claimed figure, evidence consistency. The biological case studies also lack sufficient analysis of composition effects and assay confounding, limiting the validity of their scientific conclusions.

Questions for authors
  1. What evidence did the automated judge inspect, and were its soundness and reproducibility ratings validated against expert assessments or independent reruns?

  2. What is the correct kidney bootstrap interval, and can the saved estimates confirm that the discrepancy is limited to plotting?

  3. How does the lung analysis distinguish cell-composition changes from within-lineage effects, and what ALS-associated effects remain identifiable after accounting for assay confounding?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.