julkuhn/HIS3InterspeciesEpistasis

--- license: cc-by-4.0 task_categories: - other language: - en tags: - biology - protein - deep-mutational-scanning - fitness - ProteinGym pretty_name: HIS7_YEAST Pokusaeva2019 - Segmentwise Split --- # Dataset Card for HIS3 Interspecies Epistasis ## Dataset Description - **Repository:** (https://github.com/julkuhn/fopra.git) - **Point of Contact:** [e-mail] ## Dataset Summary A re-split of the [ProteinGym](https://proteingym.org/) 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](https://www.uniprot.org/uniprotkb/P06633/entry)). 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 ```python 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` ```python 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: ```python 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](https://proteingym.org/) 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](https://doi.org/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](https://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: ```bibtex @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} } ```