PertReasonQA

Dataset repository of "PertReason: A Knowledge-Grounded Benchmark and Framework for Cell-State-Conditioned Mechanistic Reasoning of Perturbation Effects"

Abstract

Evaluating machine learning in scientific domains requires separating correct predictions from correct reasons under realistic distribution shifts. We introduce PertReason, a knowledge-grounded benchmark and framework suite for cell-state--conditioned reasoning about perturbation effects. At its core, PertReasonQA is a benchmark that tests whether models can generate mechanistically faithful explanations while remaining robust to complex shifts, such as new cells, unseen perturbations, cross-modal or combinatorial extrapolations. PertReasonQA combines single-cell genetic and chemical perturbation data across multiple cellular contexts with knowledge graphs, and dynamically conditions pathways on cell-specific basal states to avoid generic memorization. Evaluations on state-of-the-art models reveal systematic gaps between predictive accuracy and mechanistic reasoning. Specifically, these models exhibit failure modes largely invisible to standard benchmarks, such as deriving correct answers through flawed logic, ignoring cellular context, and generating directionally inconsistent mechanisms. As a reference probe of the benchmark, we present PertReasonLM, a large language model trained to align outcome predictions with context-specific mechanistic reasoning. Our model targets the identified failure modes by grounding rationales in context-specific pathways and tightening agreement between outcomes and mechanisms. Together, we provide a diagnostic framework for exposing and mitigating failures in faithful reasoning in data-rich scientific systems.

Repository Structure

• data_outcome_only_jsonl/: Outcome-only training and validation records with task metadata, three-way expression labels, and basal expression values.
• data_reasoning_jsonl/: Reasoning records for training, validation, and the ten test folds, including reference and candidate pathways. Training records also include reference reasoning text.
• prompt_cell_cond_path_jsonl/: Evaluation records with model-ready prompts containing discretized basal expression states and retrieved cell-conditioned pathways.
• prompt_no_path_jsonl/: Evaluation records with model-ready prompts containing basal expression context without retrieved pathway edges.

JSON Schema

Each JSONL line contains one QA record.

Top-Level Keys

Key	Description
id	QA identifier based on cell type, perturbation, and effect gene.
input	Cell type, perturbation, effect gene, and split metadata.
basal_context	Gene-to-value mapping of log-normalized expression in the unperturbed control state.
output	Reference pathways and reasoning; absent from outcome-only records.
prompt	Model-ready question with basal context and instructions.
label	Ground-truth expression label in prompt records.
pert_type	Perturbation modality: chemical or genetic.
data_type	reasoning in prompt records.

input Keys

Key	Description
cell_type	Cell line or immune-cell type.
pert	Compound name, target gene, or list of target genes.
gene	Effect gene whose expression change is predicted.
task	Prediction task: 3way.
split_cell	Cell-context split: id or ood_test.
split_pert	Perturbation split: train, valid, id_test, or ood_test.
label	Ground-truth expression label: up, down, or unchanged.
pert_type	Perturbation category or transfer-fold identifier.
is_combo	Whether multiple genes are perturbed together.
pert_splits	Split assignment for each gene in pert.

output Keys

Pathways are lists of [source, relation, target] triplets.

Key	Description
triplets_ref	Reference pathways consistent with the observed outcome.
triplets_uncond	Pathways retrieved without cell-state conditioning.
triplets_cell_cond	Candidate pathways conditioned on the basal cell state.
reasoning_text	Reference reasoning text generated by Qwen3-4B; training records only.
reasoning_id	Reference reasoning index.

Citation

@inproceedings{kim2026pertreason,
  title={PertReason: A Knowledge-Grounded Benchmark and Framework for Cell-State-Conditioned Mechanistic Reasoning of Perturbation Effects},
  author={Kim, Dongkwan and Gao, Yiming and Yang, Yining and Shen, Yang},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026},
  url={https://arxiv.org/abs/2607.18777}
}