Case-control genetics datasets evolved to be epistatic

This repository contains the data for the following two studies:

1. Evolving hard problems: Generating human genetics datasets with a complex etiology (Prefered Citation)
   Daniel S. Himmelstein, Casey S. Greene, Jason H. Moore
   BioData Mining (2011) DOI: 10.1186/1756-0381-4-21 · PMID: 21736753 · PMCID: PMC3154150
   

2. A Model Free Method to Generate Human Genetics Datasets with Complex Gene-Disease Relationships
   Casey S. Greene, Daniel S. Himmelstein, Jason H. Moore
   Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics (2010) DOI: 10.1007/978-3-642-12211-8_7

Previously, this data was hosted at http://discovery.dartmouth.edu/model_free_data/. In July 2016, the data was migrated to this repository and modernized. The legacy site is no longer supported, use this site instead. The data contents are equivalent, although directory structures, file extensions, and text-file header rows have been refined.

Introduction

This repository contains genetics datasets simulated to be complex. Each dataset contains 3,000 observations (rows), which represent samples/subjects for a genetic association study. Subjects are classified as cases (i.e. diseased when Class column is 1) or controls (i.e. healthy when Class column is 0). The remaining columns (e.g. X1, X2, X3) represent biallelic SNPs and are coded as 0 for homozygous, 1 for heterozygous, and 2 for homozygous of the alternate allele. Datasets were created with 3, 4, and 5 SNPs. For example, a subset of a 5 SNP dataset looks like:

X1	X2	X3	X4	X5	Class
0	1	1	1	1	0
0	0	0	1	0	0
2	1	1	2	0	0
1	1	0	1	0	1
1	1	1	1	2	1
1	2	2	0	0	1

Datasets were created using an evolution strategy. Each run used a population size of 1,000 (number of datasets) and evolved for 2,000 generations. Each generation consisted of introducing mutations and selecting to survive datasets that were optimal for at least one attribute. Only the Pareto-optimal datasets from the final generation of a run were retained. From the Pareto-optimal datasets yielded by a run, a single "best" dataset was chosen.

We created datasets with following attributes:

• n-way: the number of SNPs in the dataset, ranging from 3–5. The joint-predictiveness of all SNPs was maximized, while the predictiveness of lower order SNP-combinations was minimized.
• NoLow. All runs were optimized for having no one-way (marginal) or two-way (pairwise epistatic) associations. NoLow refers to whether, in addition to minimizing 1 and 2-way effects, all lower order effects were minimized. For example, fivewayNoLow maximized the 5-way effect, while minimizing 1, 2, 3, and 4-way effects.
• HWE: whether SNPs were optimized to maintain Hardy-Weinberg equilibrium.

In total, eight types of datasets were created, which combined the attributes above. For each dataset type, 100 runs were performed resulting in 100 best datasets per type. Most users will be interested in only the best datasets, since multiple datasets from the same run (set) may not be independent.

Access datasets

The following table links to repository location with the datasets for a given attribute combination. SNPs refers to the number of SNPs in each dataset. 1-Way through 5-Way indicate whether associations of that order exist: No means datasets were evolved not to have that effect; Yes means datasets were evolved to have that effect; and Possibly means datasets were not optimized for that effect.

SNPs	1-Way	2-Way	3-Way	4-Way	5-Way	HWE	Name
3	No	No	Yes			No	threeway
3	No	No	Yes			Yes	HWthreeway
4	No	No	Possibly	Yes		No	fourway
4	No	No	Possibly	Yes		Yes	HWfourway
4	No	No	No	Yes		No	fourwayNoLow
5	No	No	Possibly	Possibly	Yes	No	fiveway
5	No	No	Possibly	Possibly	Yes	Yes	HWfiveway
5	No	No	No	No	Yes	No	fivewayNoLow

Need help?

If you have any questions or feedback, please submit an Issue on GitHub.

License

No rights reserved. This repository including its datasets are released under the CC0 Public Domain Dedication (see LICENSE.md).