Moomboh/mutafitup-datasets

Leakage-free protein function prediction benchmarks for multi-task protein language model (pLM) fine-tuning.

This dataset contains 16 prediction tasks drawn from three independent data families, resplit to eliminate cross-dataset sequence similarity-based data leakage between training, validation, and test sets.

Datasets

Dataset	Task type	Labels	Metric	Provenance
secstr	per-residue classification	3	accuracy	DSSP assignments from NetSurfP-2.0 (Klausen et al., 2019), test set from ProtTrans NEW364 (Elnaggar et al., 2022)
secstr8	per-residue classification	8	accuracy	DSSP assignments from NetSurfP-2.0 (Klausen et al., 2019), test set from ProtTrans NEW364 (Elnaggar et al., 2022)
disorder	per-residue regression	continuous	spearman	CheZOD database (Dass et al., 2020) -> SETH (Ilzhofer et al., 2022) -> Schmirler et al., 2024
rsa	per-residue regression	continuous	spearman	Isolated-chain RSA from NetSurfP-2.0 (Klausen et al., 2019), test set CB513
meltome	per-protein regression	continuous	spearman	Meltome Atlas (Jarzab et al., 2020) -> FLIP (Dallago et al., 2021) -> Schmirler et al., 2024
subloc	per-protein classification	10	accuracy	Light Attention (Stärk et al., 2021) -> Schmirler et al., 2024
gpsite_dna	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_rna	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_pep	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_pro	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_atp	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_hem	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_zn	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_ca	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_mg	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)
gpsite_mn	per-residue classification	2	accuracy	GPSite (Yuan et al., 2024)

Loading

from datasets import load_dataset

# Load a single dataset (e.g. meltome)
ds = load_dataset("Moomboh/mutafitup-datasets", "meltome")

# Access splits
train = ds["train"]
valid = ds["validation"]
test = ds["test"]

Resplit methodology

The original datasets were collected from three independent sources (see provenance chain below). Because these sources share overlapping protein sequences, naively combining their original train/test splits introduces cross-dataset data leakage: a protein in one dataset's training set may appear in another dataset's test set.

To eliminate this leakage, all sequences across all 16 datasets were pooled, deduplicated, and clustered using mmseqs at 20.0% minimum sequence identity. New training, validation, and test splits were then assigned at the cluster level so that no two splits share sequence-similar proteins:

1. Cluster all unique sequences across datasets using mmseqs at 20.0% minimum sequence identity.
2. Merge original train + valid pools per dataset.
3. Remove within-dataset test-similar sequences from the pool.
4. Reassign cross-dataset test-contaminated cluster members to validation.
5. Top up validation to at least 10.0% of the non-test data (random seed: 42).

Shared protein groups: (secstr, secstr8, rsa) -- datasets sharing the same underlying proteins are coordinated during split assignment.

Test set reconstruction: meltome -- test sets are reconstructed from original test sequences with cluster-aware decontamination.

Duplicate aggregation: meltome -- duplicate sequences within a split are aggregated (mean strategy for regression scores).

The metadata/ directory contains the full clustering and split assignment artifacts for reproducibility.

Parquet schemas

Per-residue classification

Used by: secstr, secstr8, gpsite_*

Column	Type	Description
sequence	str	Amino acid sequence
label	list[int]	Per-residue integer class labels
resolved	list[int]	Per-residue binary mask (1 = ordered, 0 = disordered)

Per-residue regression

Used by: disorder, rsa

Column	Type	Description
sequence	str	Amino acid sequence
score	list[float]	Per-residue float scores (rsa uses 999.0 sentinel for unresolved residues)

Per-protein regression

Used by: meltome

Column	Type	Description
sequence	str	Amino acid sequence
score	float	Scalar score (melting temperature in degrees Celsius)

Per-protein classification

Used by: subloc

Column	Type	Description
sequence	str	Amino acid sequence
label	str	Class name (subcellular localization compartment)
label_numeric	int	Integer-encoded label

Data provenance

Schmirler 2024 (meltome, subloc, disorder)

Training data from Schmirler, Heinzinger & Rost (2024), downloaded from the companion GitHub repository RSchmirler/data-repo_plm-finetune-eval, also archived at Zenodo. Original train/valid/test splits were constructed by the authors using MMseqs2 clustering at >20% pairwise sequence identity.

Full provenance chains:

• meltome: Experimental thermal proteome profiling from the Meltome Atlas (Jarzab et al., 2020) -- standardized into benchmark splits by FLIP (Dallago et al., 2021) -- adopted with added validation set by Schmirler et al., 2024.
• subloc: Curated subcellular localization annotations from Light Attention (Starck et al., 2021) -- included directly by Schmirler et al., 2024.
• disorder: NMR-derived per-residue disorder Z-scores from the CheZOD database / ODiNPred (Dass et al., 2020) -- used for pLM embedding prediction by SETH (Ilzhofer et al., 2022) -- re-split by Schmirler et al., 2024.

NetSurfP-2.0 (secstr, secstr8, rsa)

Structure-derived annotations from NetSurfP-2.0 (Klausen et al., 2019), hosted at DTU Health Tech. Secondary structure labels are DSSP-based categorical assignments; RSA targets are solvent-accessibility values for resolved residues.

The secondary-structure test set (NEW364) comes from the ProtTrans project (Elnaggar et al., 2022). The RSA test set is CB513.

GPSite (gpsite_*)

Residue-level binding-site annotations from GPSite (Yuan et al., 2024), pinned to commit 58cfa4e. Labels correspond to experimentally resolved protein-ligand interactions from the PDB across 10 ligand types (DNA, RNA, peptide, protein, ATP, heme, Zn, Ca, Mg, Mn).

References

• Dallago, C. et al. (2021). FLIP: Benchmark tasks in fitness landscape inference for proteins. bioRxiv 2021.11.09.467890. doi:10.1101/2021.11.09.467890
• Dass, R., Mulder, F.A.A. & Nielsen, J.T. (2020). ODiNPred: comprehensive prediction of protein order and disorder. Sci. Rep. 10, 14780. doi:10.1038/s41598-020-71716-1
• Elnaggar, A. et al. (2022). ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning. IEEE TPAMI 44(10), 7112--7127. doi:10.1109/TPAMI.2021.3095381
• Ilzhofer, D., Heinzinger, M. & Rost, B. (2022). SETH predicts nuances of residue disorder from protein embeddings. Front. Bioinform. 2, 1019597. doi:10.3389/fbinf.2022.1019597
• Jarzab, A. et al. (2020). Meltome atlas -- thermal proteome stability across the tree of life. Nat. Methods 17, 495--503. doi:10.1038/s41592-020-0801-4
• Klausen, M.S. et al. (2019). NetSurfP-2.0: Improved prediction of protein structural features by integrated deep learning. Proteins 87(6), 520--527. doi:10.1002/prot.25674
• Schmirler, R., Heinzinger, M. & Rost, B. (2024). Fine-tuning protein language models boosts predictions across diverse tasks. Nat. Commun. 15, 7407. doi:10.1038/s41467-024-51844-2
• Starck, H. et al. (2021). Light attention predicts protein location from the language of life. Bioinformatics Advances 1(1), vbab035. doi:10.1093/bioadv/vbab035
• Steinegger, M. & Soding, J. (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026--1028. doi:10.1038/nbt.3988
• Yuan, Q., Tian, C. & Yang, Y. (2024). Genome-scale annotation of protein binding sites via language model and geometric deep learning. eLife 13, e93695. doi:10.7554/eLife.93695