jsoul/geo-perturbation-grouping-train

GEO Gene-Perturbation Grouping — Train / Validation

Supervised fine-tuning (SFT) data for teaching an LLM to read the sample metadata of a NCBI GEO gene-expression series and organize its samples into valid genetic-perturbation case/control experimental groups.

This repository holds the train and validation splits. The held-out test set lives in a separate private repo: jsoul/geo-perturbation-grouping-test.

What the task is

Given a GEO study (its title/summary/design plus a table of samples — GSM IDs with source_name, characteristics, treatment, description, etc.), the model must return the subset of samples that form clean perturbation experiments and group them as case vs. control. A valid group requires (among other rules): human samples only, a single cell line, a single allowed perturbation method (KD, KO, or OE), exactly one wild-type target gene (valid HGNC symbol), and ≥2 controls and ≥2 cases matched on cell line, method, and time point. Studies with no clean group must return No valid groups found. The full rulebook is provided verbatim as the system prompt in every example.

Data format

Each row is a single-turn chat in the standard messages layout, plus a small metadata struct:

field	type	description
messages	list[{role, content}]	Three messages, always in order system → user → assistant.
metadata.accession	string	The source GEO series ID (e.g. GSE62085). Unique per row.

• system — a fixed 7,091-character rulebook (identical across every example) defining eligibility criteria, name-interpretation heuristics, control rules, data-type rules, exceptions, and the exact required output format.
• user — the study's metadata: a worked example, then the target study's title/summary/design and a per-sample table. Length varies widely (~4.7k–113k characters; median ~6.7k).
• assistant — the target answer. Either one or more groups in the exact format below, or the literal string No valid groups found.

Target output format

Group <n>:
  Cell line: <cell line>
  Perturbation method: <KD | KO | OE>
  Target gene: <HGNC symbol>
  Control: <comma-separated GSM IDs>
  Case: <comma-separated GSM IDs>

Example (GSE62175, two groups):

Group 1:
  Cell line: HEK293
  Perturbation method: KD
  Target gene: MAML1
  Control: GSM1520991,GSM1520992
  Case: GSM1520993,GSM1520994
Group 2:
  Cell line: HEK293
  Perturbation method: OE
  Target gene: MAML1
  Control: GSM1520985,GSM1520986,GSM1520987
  Case: GSM1520988,GSM1520989,GSM1520990

Splits & statistics

split	rows	unique GEO accessions	No valid groups found
train	2,400	2,400	806 (33.6%)
validation	600	600	193 (32.2%)

Accessions are disjoint across train, validation, and the external test set — no study appears in more than one split.

Perturbation-method distribution (per group, train / validation):

method	train	validation
KD (knockdown: siRNA/shRNA/CRISPRi)	1,556	418
OE (overexpression)	369	122
KO (knockout)	221	53

Groups per example range from 0 (no valid groups) up to 8; the large majority of positive examples contain 1–2 groups.

Usage

from datasets import load_dataset

ds = load_dataset("jsoul/geo-perturbation-grouping-train")  # requires a token; private
train, val = ds["train"], ds["validation"]

row = train[0]
system = row["messages"][0]["content"]
user   = row["messages"][1]["content"]
target = row["messages"][2]["content"]
print(row["metadata"]["accession"])

The messages format plugs directly into chat-template SFT trainers (TRL SFTTrainer, Unsloth, Axolotl, etc.).

Licensing & provenance

Derived from publicly available metadata in NCBI GEO. The assistant targets are curated annotations produced for this project. Underlying GEO records are subject to NCBI's usage policies; individual studies retain their own attributions. Distributed here privately for research use — not an official NCBI product.

Citation

If you use this dataset, please cite this repository and acknowledge NCBI GEO as the metadata source.