Synthyra/SwissProt-Latent-Separation

Swiss-Prot Latent-Separation Benchmark

This benchmark contains 2,000 reviewed Swiss-Prot proteins selected from distinct 30% sequence-identity clusters. The train and test splits each contain 1,000 proteins. Labels are ternary: true and false denote evidence-backed classes, while null denotes insufficient evidence or an example outside that task's matched subset. Every task has at least 100 positive and 100 negative examples, with each class comprising at least 40% of its labeled subset in each split.

Sequences are 30 to 2,048 residues long and contain only the 20 standard amino-acid letters. Biochemical tasks use experimentally supported positives and evidence-rich negatives. Positive and negative examples are approximately matched jointly by taxonomic domain and sequence-length bin.

For GO tasks, an alternative experimental annotation in the same ontology aspect is an evidence-rich contrast, not a claim that the two biological properties are mutually exclusive under GO's open-world semantics. Exact per-task negative rules are recorded in tasks.json.

How the dataset was built

1. Pin the sources. Download UniProtKB/Swiss-Prot release 2026_02 (575,503 entries), its DAT and FASTA files, UniProt GOA annotations, and GO basic OBO. Record source URLs, dates, sizes, and hashes in source_manifest.json.
2. Filter canonical proteins. Require a reviewed, non-fragment entry; an exact DAT/FASTA sequence match; a taxon ID; length 30–2,048 residues; and only the 20 canonical amino acids.
3. Control homology. Cluster the complete Swiss-Prot FASTA with MMseqs2 18-8cc5c using local alignment, 30% minimum sequence identity, and 50% minimum coverage. Among eligible proteins, keep one quality-ranked representative per cluster, prioritizing experimental evidence, annotation support, publications, and accession as a deterministic tie-breaker.
4. Select the corpus hierarchically. Use an exact HiGHS mixed-integer solve to choose 2,000 cluster- and sequence-distinct representatives while requiring soluble, enzyme, go__go_0003677, go__go_0002376. Optimize evidence quality and taxonomic diversity, then partition the fixed corpus into 1,000 train and 1,000 test proteins.
5. Add task-specific subsets. Scan the fixed corpus for every additional experimentally supported GO task that is feasible, nonduplicate, noncomplementary, not an auxiliary ancestor/descendant duplicate, and below the label-correlation threshold. Taxonomy tasks are intentionally excluded.
6. Balance each task. In each split require at least 100 positives and 100 negatives; keep each class at least 40% of labeled examples; and keep each class at least 30% within joint taxonomic-domain and sequence-length strata. Rows outside a task's matched subset remain null.
7. Validate and package. Verify split sizes, accession/sequence/cluster uniqueness, evidence rules, partial-label balance, stratum balance, label redundancy/correlation, Parquet schema, deterministic hashes, and Hub reload. The build uses distinct MMseqs cluster IDs as its homology boundary and does not claim an exhaustive post-selection pairwise alignment audit.

Dataset structure

Each row contains the Swiss-Prot accession and full sequence; MMseqs cluster and split; length, organism, taxon, lineage, and domain; gene names; experimental GO and EC annotations; supporting PMIDs; selected UniProt annotation text; and evidence-quality counts. Dynamic label__* columns contain true, false, or null, while positive_task_ids and negative_task_ids list the evidence-backed assignments for that row.

How to read the labels

• true: the protein meets the task's experimental positive definition.
• false: the protein has the stated evidence-backed contrasting annotation; it is not merely unannotated.
• null: available evidence is insufficient, contradictory, or the protein was not selected for that task's matched subset.
• GO positives include experimentally annotated descendant terms. GO negatives use experimental NOT evidence or an experimental alternative in the same GO aspect.
• enzyme contrasts experimental catalytic evidence with experimental non-catalytic molecular-function evidence. soluble contrasts curated soluble-compartment localization with experimental membrane localization.

Reference ESMC-300 separability baseline

These values provide a rough reference for initial linear and shallow-classifier separability, not a tuned benchmark target. A frozen ESMC-300 encoder processed every sequence in evaluation mode, without gradients, after truncation to at most 2,048 residues. For each protein, the final per-residue hidden states were pooled by taking the mean and variance along the residue axis after excluding padding and special tokens; concatenating them produced a feature vector of width twice the encoder hidden size.

Each task was fit independently using only its true and false rows; null rows were excluded. The labeled subsets contained approximately 252-556 train and 265-895 test proteins per task. Four classifiers were fit on the fixed train split: standardized logistic regression (balanced class weights, C=1.0, liblinear), standardized linear SVM (balanced class weights, C=1.0), standardized 15-neighbor KNN (distance weighting), and XGBoost (100 trees, depth 3, learning rate 0.1, histogram algorithm). The 80 fits ran with four CPU workers. There was no validation split, hyperparameter or threshold search, or calibration against test labels.

The per-label value below is the unweighted mean test Matthews correlation coefficient (MCC) across the four classifiers. The overall score gives every label equal weight, regardless of labeled-subset size. Across the 20 labels, the unweighted macro-average was 0.3881. The strongest initial results were Nucleus (0.7145), DNA binding (0.6333), Enzyme (0.6232).

Validation

The publication command validates the local Parquet files and reloads the uploaded DatasetDict before reporting success. It checks exact split sizes, label balance, accession/sequence/cluster disjointness, and experimental evidence requirements. Homology separation is defined by distinct MMseqs2 cluster IDs; no exhaustive post-selection pairwise audit is claimed.

Validated content SHA-256: 642a222bcdde021cbb0bcd02eb99b3406bbf746429aa8a5e62139d00d519c10f Retained label count: 20

Label definitions

Label	Name	Concise biological interpretation	Reference test MCC
label__go__go_0003677	DNA binding	Captures proteins that bind DNA selectively and non-covalently.	0.6333
label__enzyme	Enzyme	Distinguishes experimentally supported enzymes from proteins with experimentally supported non-catalytic molecular functions.	0.6232
label__go__go_0002376	Immune system process	Captures proteins that participate in immune-system development, activation, or defense.	0.3209
label__soluble	Soluble localization	Distinguishes proteins experimentally localized to curated soluble compartments from experimentally membrane-localized proteins.	0.3494
label__go__go_0016043	Cellular component organization	Captures proteins involved in forming, arranging, or dismantling cellular structures.	0.2739
label__go__go_0005737	Cytoplasm	Captures proteins experimentally localized to the cytoplasm.	0.3529
label__go__go_0006955	Immune response	Captures proteins that execute or support an immune response.	0.3706
label__go__go_0043232	Intracellular membraneless organelle	Captures proteins localized to an intracellular structure without a surrounding membrane.	0.4476
label__go__go_0043170	Macromolecule metabolic process	Captures proteins involved in reactions and pathways that transform macromolecules.	0.2825
label__go__go_0010605	Negative regulation of macromolecule metabolic process	Captures proteins that decrease metabolic processing of cellular macromolecules.	0.3934
label__go__go_0005634	Nucleus	Captures proteins experimentally localized to the cell nucleus.	0.7145
label__go__go_0010557	Positive regulation of macromolecule biosynthetic process	Captures proteins that increase the synthesis of cellular macromolecules.	0.4791
label__go__go_0048584	Positive regulation of response to stimulus	Captures proteins that increase the strength, rate, or duration of a response to a stimulus.	0.2040
label__go__go_0044238	Primary metabolic process	Captures proteins involved in core metabolic reactions needed for normal cellular activity.	0.2695
label__go__go_0002682	Regulation of immune system process	Captures proteins that increase, decrease, or otherwise tune an immune-system process.	0.2539
label__go__go_0051239	Regulation of multicellular organismal process	Captures proteins that tune a biological process occurring at the multicellular-organism level.	0.2564
label__go__go_0009966	Regulation of signal transduction	Captures proteins that increase, decrease, or otherwise tune signal transduction.	0.3916
label__go__go_0006357	Regulation of transcription by RNA polymerase II	Captures proteins that control transcription carried out by RNA polymerase II.	0.5866
label__go__go_0051707	Response to other organism	Captures proteins involved in responding to the presence or activity of another organism.	0.2663
label__go__go_0007165	Signal transduction	Captures proteins that convert an internal or external signal into a cellular response.	0.2924

Included artifacts

The repository includes tasks.json, source and build manifests, balance diagnostics, rejected-task reasons, and the selected FASTA file. audit_report.json is empty because the workflow uses MMseqs2 cluster membership directly.

License

The benchmark is released under CC BY 4.0. UniProt source records remain subject to UniProt attribution requirements.