Synthyra/DeNovo-Origin-Rigorous

Protein Sequence Origin and Engineering Provenance

Release status: compiled from the audited release manifest. This card describes the immutable, rights-reviewed release uploaded at the revision shown below; source receipts and split audits are retained in provenance/.

Dataset summary

This benchmark groups protein sequences by documented production history. It provides several aligned label views so that an experiment can ask whether a sequence is natural, designed or engineered, or a synthetic control, at the granularity supported by its provenance.

It is a dataset about recorded origin. It does not establish biological function, expression, folding, toxicity, pathogenicity, safety, or whether a sequence is harmful. A natural-origin label is not a safety certification. A design label does not imply that a sequence is functional or experimentally successful. A synthetic-control label describes the recorded control recipe; it is not a scientific finding that the sequence cannot function.

Release property	Value
Dataset revision	9883205aa889242f07ec4b6a00f4c3a88517f84f142e95f03b96703ee2510d6b
Compilation code revision and SHA-256	020d70089602031babc473f9a8712d3c64faf2d28bf0befdf22224ee76d6964f
Unique canonical sequences	40239
Source observations	40239
Derived sequence representations	40239
Train / validation / test	{"test": 1564, "train": 37019, "validation": 1656}
Unique sequence ceiling	1,000,000
Label views	Binary, three-way, four-way, documented subtype
Strict label coverage	{"test": {"de_novo": 11, "natural": 771, "parent_derived_engineering": 0, "synthetic_control": 782}, "train": {"de_novo": 9783, "natural": 8463, "parent_derived_engineering": 264, "synthetic_control": 18509}, "validation": {"de_novo": 112, "natural": 716, "parent_derived_engineering": 0, "synthetic_control": 828}}
MMseqs2 version and relation policy	{"policy": {"max_evalue": 0.001, "min_identity": 0.3, "min_query_coverage": 0.8, "min_target_coverage": 0.8}, "version": "14-7e284+ds-1+b2"}
Natural-reference release and hash	2026_03 / 4813d5bd72cf69fc19de11ed21897e00ca8c57ab874539c530ca7cb99677e578
Modal run receipts	provenance/run_manifest.json and provenance/*receipt.json files

The actual values above must come from the final immutable release manifest, not from a source paper's reported totals or an intermediate inventory.

The Modal compiler first writes a complete source ledger and a row-level review queue. It applies only the explicit curation JSONL described in CURATION_MANIFEST_SCHEMA.md. If that manifest is missing or incomplete, the run remains evidence-only and release_ready is false; no source method string or similarity hit is promoted to a training label.

Label views

All views refer to the same unique sequence rows and use the same train, validation, and test assignments. Rows without a single, evidence-supported origin receive -1 in numeric fields and unresolved in the subtype field; they are retained in the evidence ledger but excluded from strict supervised metrics.

Field	Values	Meaning
label_binary	0 natural; 1 non-natural	Natural versus the union of de novo, parent-derived engineering, and synthetic controls
label_three_way	0 natural; 1 designed or engineered; 2 synthetic control	Separates generated controls from documented designs and engineering
label_four_way	0 natural; 1 de novo; 2 parent-derived engineering; 3 synthetic control	Main origin task with the four evidence classes kept distinct
label_fine_name	<origin>:<subtype>	Retains the documented subtype; unresolved subtype is explicit
label	Same values as label_four_way	Convenience alias for the four-way task

parent_derived_engineering means that a specific parent sequence and a parent-based engineering event are documented. Subtypes distinguish, when evidence supports them, lead or optimization campaigns, directed evolution, deep mutational scanning, substitution or indel libraries, combinatorial mutagenesis, rational variants, and other documented variant-generation processes. The label does not imply that a variant was selected as a successful lead. Assay outcomes and design intent are separate fields.

de_novo means that the source describes a designed candidate and provides no documented sequence parent for that candidate. It does not mean that no natural homolog exists. A similarity hit cannot turn a documented design into natural origin or prove parentage. When the source documents a parent, the row belongs to the parent-derived engineering class even if its method name includes "design" or "optimization".

synthetic_control identifies a sequence generated as a control by the recorded procedure. The primary in-project controls call de_novo/random_gen.py; their records include its file hash, wrapper hash, parameters, per-sequence seed, branch, attempt, parent sequence, and parent source. External random-sequence archives are separate sources and are never relabeled as outputs of this generator.

The same sequence can have multiple histories. If source observations support conflicting origins, or any duplicate observation remains unreviewed, the sequence is marked ambiguous or pending and does not enter strict labeled evaluation. The complete set of observation IDs and claims remains available for audit.

A full-length exact match to a sequence in the frozen natural reference also withholds the strict origin label unless the natural observation is already represented in the evidence cohort. This is recorded as an external exact-match ambiguity; it is not converted into a natural label by similarity alone.

Source ledger and per-sequence traceability

Every unique sequence links to one or more source-observation rows. Every observation records, where supplied, the source and immutable revision, source record ID, artifact SHA-256, source file/member/row, evidence URL, source license, reported origin, reviewed origin, curation state, and reason codes. Duplicate observations are retained when sequence rows are deduplicated.

Designed and engineered records additionally link to the campaign and method family. Parent-derived variants carry the exact parent sequence digest and engineering subtype. Designs retain documented model, generator, target, campaign, and refinement information when the source provides it. The dataset does not infer missing parentage from sequence similarity.

Natural controls are reported with accession, reference release, search method, identity, aligned length, query and target coverage, bit score, E-value, and recorded evidence for natural origin. A top detected hit means the best match found in the specified searched database under the stated search settings. It does not mean a certified globally closest sequence unless the reference really supports that claim. Whole-sequence and local/domain comparisons are identified separately.

Raw sequence records with gaps, tags, ambiguous residues, stops, or other noncanonical symbols are retained in quarantine with their source metadata. The strict benchmark normalizes case and whitespace only. It does not silently strip tags or remove residues. A processed construct, when justified by source evidence, is stored as a derived representation linked to the raw observation.

Included sources and terms

The release manifest must list every included source with immutable revision, artifact hash, evidence URL, license or terms, included record count, and exclusion count. Set this dataset repository's license metadata only after confirming that the terms for every redistributed source and derived artifact permit the proposed distribution. If the sources cannot share one license, publish only a rights-cleared subset or provide a precise, source-scoped terms statement. The reviewed source candidates include:

Source	Role	Provenance and redistribution notes
ProteinBase snapshot dated 2026-01-28	Candidate design records and metadata	Collection-level terms may differ. Rows lacking method, parent, or campaign evidence remain unresolved.
ProtGPT2 released FASTA archive	Frozen generated-sequence examples and a separate published natural background	Dataset card declares MIT with upstream exceptions; retain per-artifact terms. The archive's random controls are not outputs of this project's generator.
Anthropic protein binder design release	Design and campaign-level provenance	Sequence tables are licensed CC BY 4.0 with third-party material exceptions. Include only the predeclared benign-provenance subset; do not infer biological activity from design provenance.
Somermeyer et al. GFP fitness-peak data	Parent-based GFP engineering and extant-protein examples	Repository is GPL-3.0; the data-specific redistribution terms require separate review. Rows are staging-only until cleared.
Tsuboyama et al. stability data	Parent-linked variant and mutagenesis observations plus other protein designs	Zenodo record declares CC BY 4.0. Assay background and WT_name do not by themselves establish natural origin. Measurements are not origin labels.

The table is a source-candidate inventory, not a claim that every listed row will be distributed. Records with unclear provenance, parentage, or redistribution rights are excluded from the labeled release or placed in a separately documented non-distributed audit ledger. Each upstream license and third-party exception remains attached to the individual observation.

Natural-reference matching

The frozen natural reference contains 573063 records from uniprot-swissprot-reviewed, release 2026_03, downloaded on ``, SHA-256 4813d5bd72cf69fc19de11ed21897e00ca8c57ab874539c530ca7cb99677e578. Its searchable scope and filters are UniProt reviewed:true, excluding taxonomy 32630; see run manifest.

Each designed or engineered query is searched against eligible reference records, including rows whose origin evidence is still pending or conflicting. The search receipt publishes tool and executable hashes, command, parameters, database hash, candidate cap, saturation counts, hit count, and logs. The matches/natural_matches.jsonl table retains ranked full-length matches and separately scoped local/domain matches with identity, query and target coverage, bit score, E-value, accession, sequence digest, source, and reference release. Every candidate query must be present in the audited query set, including queries with no detected match. A no-hit value means no eligible hit was observed under this search and threshold; it does not prove that no natural relative exists. Current-reference matches are not evidence that the same sequence was known before the design campaign.

The primary paired-control set and the independent natural background set are identified separately. Matching and source decisions are made before model evaluation. Length, composition, low complexity, terminal residues, family, organism, source, campaign, and representation are reported as possible predictive shortcuts.

Splits and homology audit

The released assignments target approximately 90/5/5 for train, validation, and test, subject to availability of independent connected components. Split ratios do not override evidence quality or component separation. If the source families cannot support a requested class in every split, the release reports that limitation rather than dividing a component.

Before assignment, the compiler connects exact duplicate sequences, observed homology relationships meeting the published operational criterion, variants sharing a parent, sequences in the same campaign or documented backbone group, matched natural-design pairs, and generated controls with their parents. The split algorithm assigns whole connected components. The release records the component graph and largest-component distribution.

The primary global search policy is {"max_evalue": 0.001, "min_identity": 0.3, "min_query_coverage": 0.8, "min_target_coverage": 0.8}. An independent post-allocation search compares validation and test sequences against the training sequences. Any qualifying cross-split relationship blocks release until the components are reassigned and the audit is repeated. A separate local/domain audit reports partial relationships. MMseqs2's prefilter and candidate cap can miss relationships; the benchmark therefore describes the exact observed-search audit and its limits instead of claiming no remote homology.

Additional campaign-held-out and method-family-held-out partitions are published only when there are enough independent campaigns or method families to support them. They are separate evaluations and must not be conflated with the primary sequence-relation-disjoint split.

Shortcut checks and recommended evaluation

Report per-split counts for each label view and subtype, unique source records, campaigns, parent roots, method families, homology components, natural-control coverage, and sequence-length bins. Report both sequence-level totals and independent group counts. A large mutagenesis library must not appear as thousands of independent campaigns.

Recommended diagnostics include label prediction from sequence length alone, amino-acid composition, terminal residues, low-complexity fraction, and simple k-mers; source/campaign-only prediction; duplicate and truncation collision checks; and performance broken down by source, campaign, parent, organism, method, and subtype. Any preprocessing, balancing, threshold selection, or calibration is fitted on training data only. Report uncertainty over independent groups. Good results on these checks do not prove that every possible shortcut has been removed.

Keep the final test set frozen. Do not choose inclusion thresholds, a split seed, or an input transformation based on final-test model performance. State whether metrics use all evidence-cleared rows or an explicitly defined subset. Ambiguous and weak-heuristic records may be analyzed as challenge sets, but they are not strict test truth.

Dataset schema and loading

The publication contains four related tables:

1. sequences: one row per unique canonical sequence, with sequence text, normalized-sequence hash, all label views, split, component ID, and pointers to its source observations, documented parent records, and natural-match counts by search scope.
2. observations: one row per source record, preserving the raw sequence, source fields, artifact and row provenance, evidence, license, and review status. Conflicting assertions are not collapsed.
3. relations: typed edges for exact identity, detected homology, campaign, parent, backbone, matched controls, and engineered variants.
4. matches/natural_matches.jsonl: ranked natural-reference hits, with full-length and local/domain search scopes kept distinct.

The exact manifest hash is reproducible from the manifest payload with the manifest_sha256 field omitted. The manifest contains hashes for every data file and the compiler source. Timestamps are kept in Modal run receipts rather than in the deterministic release manifest.

Schema version: 1. Column definitions and null/sentinel semantics are in manifest.json and sequence JSONL files. The split files use the same schema and release manifest: sequences/{train,validation,test}.jsonl.

from datasets import load_dataset

data = load_dataset("Synthyra/DeNovo-Origin-Rigorous", name="sequences")
train = data["train"]
sequence = train[0]["sequence"]
label = train[0]["label_four_way"]

To use another task, select label_binary, label_three_way, or label_fine_name. Exclude -1 labels for strict supervised training and preserve the supplied split. Do not re-split individual rows after download; all rows in a connected component must remain together.

Intended use and limitations

Intended use is research on protein-sequence provenance, design-history recognition, dataset leakage, and robustness to source and homology shift. The labels report evidence about how a record was produced. They are not guaranteed to identify the actual history of a sequence in all settings, especially when the sequence has multiple documented histories or source documentation is incomplete.

This dataset does not establish safety, harmfulness, function, folding, expression, novelty, intellectual-property status, or experimental success. Do not interpret predicted origin as a biological risk score or use it as a substitute for a domain-specific safety evaluation. This release does not contain per-sequence pathogenicity ground truth.

Known limitations include source and era imbalance, incomplete historical natural-reference coverage, literature reporting bias, uneven campaign documentation, possible residual source shortcuts, and incomplete sensitivity of heuristic homology search. The contribution of each source is reported in the release tables so users can evaluate these limitations directly.

Citation and reproducibility

Please cite this dataset as Synthyra/DeNovo-Origin-Rigorous at the immutable revision 9883205aa889242f07ec4b6a00f4c3a88517f84f142e95f03b96703ee2510d6b and cite the original source papers listed in Source citations and immutable URLs are retained in observations.jsonl. Compilation manifest SHA-256: 9883205aa889242f07ec4b6a00f4c3a88517f84f142e95f03b96703ee2510d6b. Full commands, software versions, source checksums, license review, curation decisions, split seed, control-generation seeds, and Modal resource limits are included in manifest.json, provenance/run_manifest.json, and hashed provenance receipts.

Compilation decisions that block publication

The release is not ready for publication until all of the following are populated and independently checked:

• Exact source counts, immutable revisions, artifact checksums, source-record exclusions, and per-observation redistribution terms.
• Sequence-level provenance adjudication, duplicate/conflict decisions, and parent joins for all included engineering examples.
• The selected benign Anthropic rows and documented target-based exclusions.
• The actual random_gen.py hash, parameters, seeds, branch counts, parent linkage, and final license decision for generated controls.
• Frozen natural-reference query, release, checksum, search receipts, and detected-hit limitations; complete full-length and local/domain query coverage; match-table hash; and explicit zero-saturation audit.
• Unique sequence ceiling, split/component counts, label support per split, exact and homology separation audits, and shortcut diagnostics.
• A clean reproducibility run from the recorded manifest, with the same release hashes and no unresolved template fields.