Authors2027/iclr-reproducibility-bundle

Reproducibility Bundle — ICLR 2027 Submission

Evaluation datasets, scripts, and the model checkpoint for the anonymous ICLR 2027 submission.

Bundle Layout

Directory	Contents
expression_datasets/	Three .h5ad files (GEO-OmicsQA, Human Diseases, Tabula Sapiens)
embeddings/	Pre-computed OmicsLM input vectors (.npz)
geo_omics_qa/	GEO-OmicsQA question files (six .json variants, 3,000 questions)
omicslm_checkpoint/	OmicsLM model checkpoint in HuggingFace format
omicslm_evaluation_scripts/	Reproducible benchmark scripts and shared inference helpers
baselines_evaluation_scripts/	External baseline evaluation and PerturbQA scoring scripts
perturbqa/	PerturbQA inputs, control profiles, and asset downloader

expression_datasets/

All .h5ad files share a unified gene space of 19,260 protein-coding genes:

• adata.var_names — Ensembl Gene IDs
• adata.var['gene_symbol'] — HGNC gene symbols

Expression values are raw counts or Transcript Per Million (TPM) as inherited from the original source (see per-file notes below).

File	Task	Samples	Genes	Source	Licence
geo_qa_benchmark_expression_dataset.h5ad	GEO-OmicsQA (profiles)	6,729	19,260	GEO	NCBI/GEO source terms; Apache 2.0 (ARCHS4 pipeline)
human_diseases_cell_whisperer_benchmark.h5ad	Disease classification	14,112	19,260	GEO / SRA	CC BY 4.0
tabula_sapiens_geo_version.h5ad	Cell-type annotation	483,152	19,260	Tabula Sapiens	CC BY 4.0

geo_qa_benchmark_expression_dataset.h5ad

Expression profiles for 6,729 GEO samples, including those referenced by the GEO-OmicsQA question files. Each GSM identifier in the JSON questions maps to a sample via its geo_accession identifier. All samples come from GEO studies held out from training. Raw RNA-seq data was downloaded from NCBI GEO and processed using the open-source ARCHS4 pipeline (Apache License 2.0).

Expression: TPM Download source: NCBI GEO Licence: NCBI/GEO source terms / Apache 2.0 (ARCHS4 pipeline)

human_diseases_cell_whisperer_benchmark.h5ad

Bulk RNA-seq profiles from GEO/SRA representing a broad range of human diseases, assembled by Schaefer et al. (CellWhisperer). Used as a held-out disease-classification benchmark.

Expression: raw counts Download source: NCBI GEO / NCBI SRA Licence: CC BY 4.0

tabula_sapiens_geo_version.h5ad

Tabula Sapiens single-cell RNA-seq atlas (GEO version, GSE201048), used for cell-type annotation evaluation. This version matches the one used by CellWhisperer, enabling direct comparison. Excluded from training entirely.

Statistics: 483,152 cells · 24 tissues · 177 cell-ontology classes Expression: raw counts Download source: NCBI GEO GSE201048 Licence: CC BY 4.0

embeddings/

Pre-computed omics embeddings (*.npz), one file per dataset above with the same basename. Each archive holds an embeddings array. Human Diseases and Tabula Sapiens embeddings follow the corresponding .h5ad's obs row order; GEO-OmicsQA embeddings cover the samples referenced by the questions and are matched by GSM identifiers in obs_names.

geo_omics_qa/

Six JSON files comprising the 3,000-question GEO-OmicsQA benchmark, organised by question type and prompt variant. All questions reference GEO studies disjoint from the training corpus. Each question links to one or more expression profiles via geo_accession identifiers, which resolve to samples in geo_qa_benchmark_expression_dataset.h5ad.

File	Variant	Questions	Licence
geo_qa_binary.json	Binary QA — standard	500	CC BY-SA 4.0
geo_qa_binary_minimal.json	Binary QA — minimal	500	CC BY-SA 4.0
geo_qa_simple.json	Simple free-text — standard	500	CC BY-SA 4.0
geo_qa_simple_minimal.json	Simple free-text — minimal	500	CC BY-SA 4.0
geo_qa_context.json	Contextual free-text — standard	500	CC BY-SA 4.0
geo_qa_context_minimal.json	Contextual free-text — minimal	500	CC BY-SA 4.0

Binary QA (1,000 questions). Balanced yes/no questions testing whether expression profiles support a biological claim. The standard variant provides full biological context; the minimal variant reduces scaffolding to test whether models use the supplied profiles rather than textual priors.

Simple free-text (1,000 questions). Open-ended questions asking for concise biological explanations without extensive literature context.

Contextual free-text (1,000 questions). Long-form, literature-grounded questions requiring integration of biological background with multi-sample expression evidence.

perturbqa/

inputs.parquet contains 16,208 direction-evaluation questions and labels. profiles/ contains the K562, RPE1, HepG2, and Jurkat control profiles. download_assets.py downloads the official PerturbQA files and AROMA assets. See perturbqa/README.md for downloads and evaluation scripts.

omicslm_checkpoint/

The OmicsLM checkpoint is based on Qwen3.5-9B, whose base weights are licensed under Apache 2.0. It is a standard HuggingFace directory (config.json, six safetensors shards, tokenizer files). The custom OmicsLM classes are bundled (configuration.py, modeling.py) and registered via auto_map, so loading requires trust_remote_code=True.

Load with HuggingFace transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

CKPT = "omicslm_checkpoint"
tokenizer = AutoTokenizer.from_pretrained(CKPT, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    CKPT, trust_remote_code=True, torch_dtype=torch.bfloat16,
).cuda().eval()

Load with vLLM:

from vllm import LLM

# Install omicslm_evaluation_scripts/ first so the OmicsLM model class is
# registered with vLLM via the vllm.general_plugins entry point.
llm = LLM(
    model="omicslm_checkpoint",
    trust_remote_code=True,
    dtype="bfloat16",
    enable_mm_embeds=True,
    limit_mm_per_prompt={"image": 64},
)

See omicslm_evaluation_scripts/inference.py for end-to-end examples with the <omics> placeholder + omics-vector multimodal input.

omicslm_evaluation_scripts/

Reproducible benchmark scripts for the released model checkpoint.

Setup

Install the directory as a package so the OmicsLM model class is auto-registered with vLLM via the vllm.general_plugins entry point:

pip install ./omicslm_evaluation_scripts

Defaults expect: checkpoint at omicslm_checkpoint/, h5ad at expression_datasets/, npz at embeddings/, GEO-OmicsQA at geo_omics_qa/. Pass --checkpoint, --h5ad, --npz, --questions_dir to override.

GEO-OmicsQA and classification scripts write per-task JSONL + summary.json under logs/<benchmark>/. PerturbQA writes predictions and metrics to its specified output directory.

Determinism

Even with temperature=0, vLLM is not bit-deterministic. Results may vary between runs.

Benchmarks

#	Script	Metric
1	benchmark_geo_qa.py	accuracy (binary) / DeepEval (free-text)
2	benchmark_human_diseases.py	macro AUROC
3	benchmark_tabula_sapiens.py	macro AUROC
4	benchmark_perturbqa.py	mean per-gene AUROC, direction macro-F1

benchmark_common.py and perturbqa_scoring.py provide shared helpers.

The GEO-OmicsQA free-text splits are scored by Gemini via DeepEval — set GOOGLE_API_KEY in the environment before running benchmark #1.

baselines_evaluation_scripts/

Reproducibility scripts for the non-OmicsLM baselines compared in the paper. They evaluate external models (e.g. Gemini, GPT, Cell-o1, C2S-Scale, TxGemma, Qwen3) on GEO-OmicsQA and classification benchmarks, using top-expressed-gene "cell sentences" rendered as text where the model cannot consume omics embeddings directly.

Script	Benchmark
geo_qa_benchmark.py	GEO-OmicsQA (binary + free-text)
ts_hd_benchmark.py	Tabula Sapiens + Human Diseases annotation
cellwhisperer_lm.py	CellWhisperer Tabula Sapiens + Human Diseases classification
benchmark_aroma.py	PerturbQA AROMA inference and scoring
benchmark_perturbqa.py	PerturbQA published baseline scoring

environments/cellwhisperer/pyproject.toml defines the pinned dependencies for cellwhisperer_lm.py. The script downloads its upstream code and model checkpoints.

For C2S, task wording in ts_hd_benchmark.py follows the released templates accompanying the paper. Tissue uses one profile; disease subtype is a fixed adaptation because no native disease-classification template is provided. Candidate responses use a newline, lowercase label, and period. Cell-o1 prompts and model runtime settings are unchanged.

For geo_qa_benchmark.py and ts_hd_benchmark.py, pass --model <id> (HuggingFace ID, OpenAI / Gemini API name, or a local path) to select the backend; each script auto-detects whether to route through the OpenAI client, Gemini client, or local vLLM. Default expression datasets resolve to ../expression_datasets/<file>.h5ad. For geo_qa_benchmark.py also pass --questions <json...> pointing at the files in geo_omics_qa/. Free-text GEO-OmicsQA scoring requires GOOGLE_API_KEY in the environment, same as the OmicsLM scripts.

To rerun all six GEO-OmicsQA splits with up to 1,000 genes per profile:

uv run --no-sync bash baselines_evaluation_scripts/run_geoqa_dynamic.sh gemini-3.8-flash
CUDA_VISIBLE_DEVICES=0 uv run --no-sync bash baselines_evaluation_scripts/run_geoqa_dynamic.sh vandijklab/C2S-Scale-Gemma-2-27B

The same runner accepts gemini-3.1-pro-preview, gpt-6-sol, gpt-6-luna, and ncbi/Cell-o1. Run with the baseline dependencies installed; Gemini generation and judging support GOOGLE_API_KEY or configured Vertex AI credentials, and GPT requires OPENAI_API_KEY. The optional second argument sets the output directory. For local models, set GEOQA_TENSOR_PARALLEL_SIZE to the number of visible GPUs.

The runner starts with min(1000, (context_len - headroom) // num_profiles) genes per profile, then counts the complete prompt and shortens gene prefixes equally until it fits. It uses provider token-count APIs for Gemini/GPT and the model's chat template and tokenizer for local models. Questions and profiles are preserved. At least 1,000 tokens remain unused; local models also reserve their output cap (1,000 for C2S, 4,096 for Cell-o1), and GPT reserves 10,000. Both API backends retain 10,000-token output caps and default reasoning settings. C2S binary evaluation uses yes/no logits; Cell-o1 generates answers.

Predictions record gene counts and token budgets and are saved before judging. DeepEval disk caching is disabled for concurrent runs. Use the Python script's --dry_run to inspect prompts without generation or judging. Without --dynamic_genes, the original fixed 50-gene default remains available.

Licences

Licence	Applies to
Proprietary — All Rights Reserved	omicslm_evaluation_scripts/, perturbqa/download_assets.py
Proprietary Derivative — Research Use Only	omicslm_checkpoint/ (all files)
CC BY-SA 4.0	GEO-OmicsQA question files (geo_omics_qa/)
Apache 2.0	baselines_evaluation_scripts/ (all files)
CC BY 4.0	Human Diseases (CellWhisperer), Tabula Sapiens, perturbqa/inputs.parquet, perturbqa/profiles/k562.npy, perturbqa/profiles/rpe1.npy
GEO source terms	Nadig et al. (2025) profiles (perturbqa/profiles/hepg2.npy, perturbqa/profiles/jurkat.npy)
NCBI/GEO source terms / Apache 2.0 (ARCHS4 pipeline)	GEO QA expression file, GEO QA embeddings

Full per-file license terms are in LICENSE (this repository).