acceptodds
Under review as a conference paper at ICLR 2027

ADAPT: An Agent-Driven Autonomous Paradigm for Trans-Species Single-Cell Annotation

13 views

Abstract

Cross-species knowledge transfer provides a powerful strategy for single-cell RNA sequencing analysis in species with limited reference annotations, while offering insights into cellular diversity and evolutionary biology. However, cross-species transfer remains challenging due to species-specific expression programs, cellular differences, and complex gene homology. Existing methods often rely on predefined adaptation strategies that require manual trial-and-error for new transfer scenarios and lack closed-loop feedback to refine these strategies based on experimental outcomes. To address this challenge, we introduce ADAPT, a feedback-driven agentic framework for cross-species single-cell annotation that automatically searches and selects candidate adaptation approaches for transferring cell-type knowledge from reference to target species. Starting from frozen scGPT embeddings, ADAPT performs iterative hypothesis-driven exploration: an outer agent proposes scientific hypotheses and experimental plans, while an inner agent implements candidate adaptation models under controlled experimental protocols. The validation results are fed back to guide subsequent search rounds. We evaluate ADAPT on two cross-species transfer settings involving knowledge transfer from Tabula Sapiens to Tabula Muris and Macaca fascicularis. ADAPT achieves the highest average shared-class Macro-F1 among the compared methods, reaching 0.81 for human-to-mouse transfer and 0.73 for human-to-macaque transfer. These results suggest the potential of feedback-driven agentic search for cross-species adaptation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell RNA sequencing;Cross-species knowledge transfer;Agentic framework;Cell type annotation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Edx9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents ADAPT, an LLM-guided search framework for cross-species cell-type annotation using frozen scGPT embeddings. Two agents propose and implement adaptation models within a predefined experimental interface, using target-validation feedback to guide architecture, objective, and optimizer search. Experiments report mean shared-class Macro-F1 scores of 0.81 and 0.73 on human-to-mouse and human-to-macaque transfer.

Critical strengths
  • Structured agentic search framework. ADAPT provides a reasonably systematic closed-loop workflow that connects hypothesis generation, candidate implementation, experimental evaluation, and feedback-driven planning. This goes beyond using an LLM solely for conventional hyper-parameter tuning.

  • Careful evaluation protocol. The separation of validation, monitor, and sealed test sets, together with controlled per-round comparisons, makes the reported results relatively credible and reduces concerns about direct test-set overfitting.

Critical weaknesses
  • Limited connection to cross-species transfer. The current results mainly demonstrate task-specific improvements from agent-driven model search, but provide limited evidence that the discovered strategies capture insights specific to cross-species prediction. It is also unclear whether the search experience can transfer across settings. For example, whether findings from human-to-mouse transfer can provide a better starting point or reduce the search cost for human-to-macaque transfer.

  • The methodological advance over existing agentic AutoML is insufficiently established. Section 3 describes iterative generation and evaluation of architectures, objectives, and optimizer configurations. Although candidates can contain generated code, Appendix A.3 already supplies substantial design guidance. The paper does not clearly establish what its search procedure contributes beyond existing LLM-based AutoML methods.

Questions for authors
  • Why is scGPT kept frozen throughout the search? Since the paper aims to show that agents can discover effective strategies for cross-species adaptation, I would expect the search to potentially uncover mechanisms with clearer biological structure. However, the agent only operates on frozen 512-dim scGPT embeddings and cannot modify the representation itself. Does this constrain the search to adapting the embedding distributions and, consequently, to the biological structure already encoded by scGPT?

  • Can the search discover an efficient evolutionary trajectory? Beyond finding a strong final configuration, does the feedback from previous rounds lead to a progressively more effective search strategy? For example, compared with restarting the search, can intermediate discoveries provide better starting points or reduce the number of experiments needed to reach strong solutions?

Reviewer_rnwZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces ADAPT, an agent-driven framework for cross-species single-cell annotation. Starting from frozen scGPT embeddings, an outer LLM agent proposes adaptation strategies while an inner agent implements candidate architectures, objectives, and optimizers. The framework iteratively uses target-validation feedback to guide model search. Experiments on human-to-mouse and human-to-macaque transfer demonstrate improvements over several existing cross-species annotation methods.

Critical strengths

Automating model adaptation for cross-species single-cell annotation is an interesting and practically important direction.

Critical weaknesses
  1. First of all, the choice of scGPT as the backbone for cross-species matching is not justified. scGPT is trained on human data only and does not model gene homology / cross species transfer explicitly. More fundamentally, the paper does not describe how genes are mapped across species before scGPT inference, including the orthology mapping procedure, handling of unmatched genes, or vocabulary coverage.
  2. Related to last point, the work can be better fitted using natively cross species model such as SATURN, UCE, and Transcriptformer, but miss the opportunity to include them as ablations to the agent input or as cross species alignment baselines
  3. Further, the agent optimizes adaptation strategies using extensive feedback derived from labeled target cells (Section 4.1), which is not representative of many practical cross-species annotation settings,
  4. Also the paper provides very limited insight/explainations on what the agent actually discovered, it could be very interesting to systematically analyze / understand the final architectures, objectives, or adaptation mechanisms selected across tissues and species
Questions for authors

See the weakness

Reviewer_AjbqOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ADAPT, a feedback-driven agentic framework for cross-species single-cell cell-type annotation. Built on frozen scGPT embeddings, the framework adopts a nested two-agent architecture: an outer planning agent formulates scientific hypotheses and designs iterative experimental rounds, while an inner implementation agent generates executable candidate adaptation models spanning architecture, training objective, and optimizer configurations. Through closed-loop validation feedback, the system automatically searches for effective transfer strategies without manual trial-and-error. Evaluated on two cross-species settings (human-to-mouse and human-to-macaque), ADAPT achieves higher shared-class Macro-F1 scores than classical baseline methods including Seurat, SAMap, CAME, and scANVI.

Critical strengths

S1. This work presents an innovative feedback‑driven two‑agent paradigm to automatically search for cross‑species adaptation strategies, moving past manually fixed pipelines for single‑cell annotation.

S2. It achieves consistent performance gains over established baselines across human‑to‑mouse and human‑to‑macaque transfer experiments on multiple real tissue datasets.

Critical weaknesses

W1. The paper fails to establish that LLM‑agentic search provides tangible benefits over conventional automated search methods, making it unclear whether performance gains stem from the agent design or simply broader exploration of model configurations.

W2. Critical practical evaluations are missing, including computational cost profiling, low‑target‑data testing, and broader cross‑species generalization, which limit real‑world applicability and weaken support for the core contribution.

Questions for authors

Q1. Could you provide results for additional foundation‑model‑relevant baselines (e.g., DANN, CORAL, simple fine‑tuning) built on the identical frozen scGPT embeddings? These controls are needed to separate performance improvements brought by your agentic search loop from gains originating purely from strong pre‑trained representations paired with basic adaptation modules.

Q2. Can you present direct comparisons between your LLM‑driven agent and conventional automated search approaches (random search, Bayesian optimization, grid search) over your exact architecture‑objective‑optimizer design space, to verify whether the LLM agent delivers better or more sample‑efficient search outcomes?

Q3. What is the concrete computational overhead (GPU‑hours, wall‑clock time, total training runs) for completing the full 50‑round ADAPT search pipeline on your experimental datasets? These figures are essential for assessing practical usability for typical research laboratories.

Q4. Beyond engineering‑level implementation of an LLM agent wrapper, what fundamental conceptual or algorithmic novelty does ADAPT introduce for cross‑species single‑cell domain adaptation? Could you also provide supporting analysis explaining under what conditions agent‑driven search is expected to outperform hand‑designed adaptation strategies?

Q5. Why was human exclusively used as the source species, with evaluations limited to only two phylogenetically close target species? Could you explain the biological or experimental rationale for not including reverse transfer, distant‑species transfer, or cross‑tissue setups, and further elaborate the key factors driving the large cross‑tissue performance heterogeneity observed in your results?

Q6. Can you report convergence properties, typical failure modes of your outer planning agent, and performance sensitivity when swapping the DeepSeek‑V4‑Flash backend for alternative LLMs? Without these results, it remains unclear how generalizable your agent‑based findings are.

Q7. Could you give a thorough justification for adopting the source‑only logistic‑regression probe plus multi‑way train/validation/monitor/test data splitting protocol? Please also address potential overfitting risks introduced by the monitor split used for final‑model selection, even though this split is isolated from agent planners.

Q8. What amount of unlabeled target data is minimally required for ADAPT to yield reliable annotation performance? Can you share results from low‑unlabeled‑target‑data regimes, which represent the key real‑world use‑case motivating cross‑species annotation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.