julkuhn/HIS3InterspeciesEpistasis

Dataset Card for HIS3 Interspecies Epistasis

Dataset Summary

A re-split of the ProteinGym deep mutational scanning (DMS) dataset HIS7_YEAST_Pokusaeva_2019 for the HIS3 gene (IGP dehydratase, imidazoleglycerol-phosphate dehydratase) in Saccharomyces cerevisiae, an enzyme in the histidine biosynthesis pathway (UniProt P06633).

The original dataset contains 483,609 combinatorial multi-mutant variants scored for growth fitness. This version introduces a segment-wise train/validation/test split to benchmark out-of-distribution (OOD) generalization across evolutionary sequence space.

Segment-wise splitting

Pokusaeva et al. divided the HIS3 protein into segments with their extant amino acid states. The split is applied independently per segment: within each segment, variants are grouped by the distance of their nearest orthologue in a multiple sequence alignment (MSA). Train, validation, and test sets therefore represent progressively more distant evolutionary contexts.

This design makes the benchmark sensitive to interspecies epistasis: fitness effects that only manifests in sequence backgrounds diverged from the reference, which standard splits do not to capture.

Dataset Structure

Data Fields

Field	Type	Description
mutant	string	Amino acid substitutions from the wildtype in {WT}{pos}{mut} format, colon-separated for multiple mutations (e.g. A42G:L103V)
mutated_sequence	string	Full protein sequence of this variant
DMS_score	float	DMS fitness score of this variant
DMS_score_bin	int	Binary fitness label: functional = 1, non-functional = 0
n_mut	int	Number of mutations relative to wildtype
mut_pos	string	List of mutated positions (1-indexed)
segment	string	Protein segment the variant is assigned to (S02–S12)
phylo_dist_to_ref	float	Normalised Hamming distance of the nearest orthologue to the reference sequence within the segment (0 = identical, 1 = fully diverged)
nearest_orthologue	string	Header of the nearest orthologue in the MSA used for cluster assignment

Data Splits

The protein sequence is divided into structural/functional segments (S02–S12). For each segment, an MSA of homologous sequences is clustered by Hamming distance (UPGMA linkage). Each DMS variant is assigned to the cluster of its nearest orthologue in the alignment. Clusters are then assigned to splits such that:

• Test: clusters that are phylogenetically distant from the reference (out-of-distribution)
• Validation: clusters at an intermediate phylogenetic distance
• Train: the remaining reference-proximal clusters

Split	Variants	Share
train	254,109	52.5 %
validation	63,531	13.1 %
test	165,969	34.3 %
total	483,609

Segments covered: S02, S03, S04, S05, S06, S07, S08, S12 (S01, S09, S11 excluded due to insufficient data.)

Usage

Load the full dataset

from datasets import load_dataset

ds = load_dataset("julkuhn/HIS3InterspeciesEpistasis")

train = ds["train"]
val   = ds["validation"]
test  = ds["test"]

Filter to a single segment

Segments available: S02, S03, S04, S05, S06, S07, S08, S12

from datasets import load_dataset

ds = load_dataset("julkuhn/HIS3InterspeciesEpistasis")

segment = "S07"
train_s07 = ds["train"].filter(lambda x: x["segment"] == segment)
val_s07   = ds["validation"].filter(lambda x: x["segment"] == segment)
test_s07  = ds["test"].filter(lambda x: x["segment"] == segment)

Or with pandas:

import pandas as pd

segment = "S07"
df_train = pd.read_csv("hf://datasets/julkuhn/HIS3InterspeciesEpistasis/data/train.csv")
df_train_s07 = df_train[df_train["segment"] == segment]

Source Data

The original data was downloaded from the ProteinGym download page.

ProteinGym Filename	Dataset Name
HIS7_YEAST_Pokusaeva_2019	HIS3 Interspecies Epistasis

• Original DMS study: Pokusaeva et al. (2019). An experimental assay of the interactions of amino acids from orthologous sequences shaping a complex fitness landscape. PLOS Genetics. doi:10.1371/journal.pgen.1008079
• ProteinGym benchmark: Notin et al. (2023). ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. NeurIPS 2023. proteingym.org
• MSA source: Supplementary file pgen.1008079.s010.fas from the original publication.

Licensing Information

Creative Commons Attribution 4.0 International

Citation

If you use this split, please cite the original DMS study and ProteinGym:

@article{pokusaeva_experimental_2019,
    title = {An experimental assay of the interactions of amino acids from orthologous sequences shaping a complex fitness landscape.},
    volume = {15},
    issn = {1553-7404 1553-7390},
    doi = {10.1371/journal.pgen.1008079},
    journaltitle = {{PLoS} genetics},
    shortjournal = {{PLoS} Genet},
    author = {Pokusaeva, Victoria O. and Usmanova, Dinara R. and Putintseva, Ekaterina V. and Espinar, Lorena and Sarkisyan, Karen S. and Mishin, Alexander S. and Bogatyreva, Natalya S. and Ivankov, Dmitry N. and Akopyan, Arseniy V. and Avvakumov, Sergey Ya and Povolotskaya, Inna S. and Filion, Guillaume J. and Carey, Lucas B. and Kondrashov, Fyodor A.},
    date = {2019-04},
    pmid = {30969963},
}
@inproceedings{NEURIPS2023_cac723e5,
 author = {Notin, Pascal and Kollasch, Aaron and Ritter, Daniel and van Niekerk, Lood and Paul, Steffanie and Spinner, Han and Rollins, Nathan and Shaw, Ada and Orenbuch, Rose and Weitzman, Ruben and Frazer, Jonathan and Dias, Mafalda and Franceschi, Dinko and Gal, Yarin and Marks, Debora},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {A. Oh and T. Neumann and A. Globerson and K. Saenko and M. Hardt and S. Levine},
 pages = {64331--64379},
 publisher = {Curran Associates, Inc.},
 title = {ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design},
 url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/cac723e5ff29f65e3fcbb0739ae91bee-Paper-Datasets_and_Benchmarks.pdf},
 volume = {36},
 year = {2023}
}