suite · audited · verified 2026-07-22

TAPE

A five-task benchmark for protein representation learning spanning secondary structure, residue contacts, remote homology, fluorescence, and stability.

+2 more

Benchmark definition

What is counted

Version: original-2019
Total: 5 (supervised downstream benchmark tasks)
Task formats: sequence classification; per-residue classification; residue-pair classification; sequence regression
Capabilities: PredictionClassificationRegression
Modalities: Protein sequence

Version history

Version	Status	Release / as-of	Total	Formal tracks
original-2019 `tape-original-2019`	current	2019-06-19	5 (supervised downstream benchmark tasks)	None registered

Tracks and subsets

ID	Count	Basis	Partition?	Notes
Structure prediction tasks `tape-structure-tasks`	2	supervised downstream tasks	Exclusive & exhaustive	Secondary structure and contact prediction.
Evolutionary understanding task `tape-evolution-task`	1	supervised downstream tasks	Exclusive & exhaustive	Remote homology detection.
Protein engineering tasks `tape-engineering-tasks`	2	supervised downstream tasks	Exclusive & exhaustive	Fluorescence and stability landscape prediction.

Scientific Task Atlas

Scientific task classification

complete for original-2019. The five supervised downstream tasks are explicitly defined and grouped in the creator paper.

Scientific task	Coverage	Count	Mapping	Evidence
Protein secondary-structure prediction	explicitly-in-scope	1 tasks Supervised downstream benchmark tasks.	official-taxonomy high confidence	`tape-paper-definition-evidence` Per-residue three-state and eight-state secondary-structure prediction constitute one benchmark task.
Protein contact-map prediction	explicitly-in-scope	1 tasks Supervised downstream benchmark tasks.	official-taxonomy high confidence	`tape-paper-definition-evidence` Residue-pair contact prediction is the second structure task; it is not relabeled as full 3D folding.
Protein remote-homology detection	explicitly-in-scope	1 tasks Supervised downstream benchmark tasks.	official-taxonomy high confidence	`tape-paper-definition-evidence` Fold-level remote-homology classification.
Protein fluorescence prediction	explicitly-in-scope	1 tasks Supervised downstream benchmark tasks.	official-taxonomy high confidence	`tape-paper-definition-evidence` Sequence-to-fluorescence regression; this is prediction, not sequence generation.
Protein stability prediction	explicitly-in-scope	1 tasks Supervised downstream benchmark tasks.	official-taxonomy high confidence	`tape-paper-definition-evidence` Sequence-to-stability regression; this is prediction, not sequence generation.

Scientific coverage notes

Domain	Coverage	Count	Interpretation
Protein structure	explicitly-in-scope	2	Secondary-structure and contact-map prediction are distinct tasks; TAPE is not a full 3D folding benchmark.
Protein design	explicitly-in-scope	2	The paper calls fluorescence and stability protein-engineering tasks, but they evaluate prediction rather than sequence generation.

Evaluation registry

Works and run settings

A setting change—scope, prompt, tools, budget, grader, or repeats—creates a separate run. Charts never cross a comparability group.

Evaluation run

tape-creator-full

From Evaluating Protein Transfer Learning with TAPE

Scopefull · n=5

ShotsNot applicable

TurnsNot applicable

System prompt publicNot applicable

Reasoning / effortNot applicable

BrowserNot applicable

InternetNot applicable

DatabasesNot applicable

Code executionNot applicable

ContainerNot reported

External toolstask-specific supervised heads over frozen or fine-tuned protein representations

Token budgetNot applicable

Time / cost budgetNot reported

TemperatureNot applicable

SeedNot reported

RepeatsNot reported

Graderdeterministic task-specific scorer · human review: no

StatisticsTask-level evaluation only; no cross-task normalized aggregate.

ContaminationSequence-identity filtering and biologically motivated held-out splits.

Metrics, results, and full protocol

Metrics

Metric	Kind / baseline	Unit	Aggregation	Threshold / tolerance
Per-amino-acid accuracy	absolute	proportion	across labeled residues	Not reported
L/5 medium and long-range precision	absolute	proportion	per protein then task summary	Not reported
Fold-level accuracy	absolute	proportion	held-out test examples	Not reported
Spearman's rho	absolute	correlation	held-out test examples	Not reported

No numeric result rows are published yet; the verified protocol remains useful.

Evidence

section: Sections 4.2-5 and Table 2; Appendix A (Defines all five task splits, architectures, training procedures, native metrics, and creator comparison table.) — supports /scope, /protocol, /metrics

Evidence and change history

Source locators remain visible; expand an item to inspect the exact Registry fields it supports.

Evaluating Protein Transfer Learning with TAPE · section: Abstract; Sections 4.2 and 5; pp. 4-7, Table 2 (Defines the five tasks, their three biological groups, splits, primary metrics, and creator evaluation.) · Supports 27 fields

Open source →

/name
/aliases
/summary
/kind
/organizations
/release_date
/latest_version
/domains
/capabilities
/modalities
/task_formats
/task_counts/total
/task_counts/basis
/task_counts/subsets
/coverage_notes
/access/level
/access/license
/resources
/versions/0/release_date
/versions/0/task_counts/total
/versions/0/task_counts/basis
/versions/0/task_counts/subsets
/scientific_task_classification/entries/0
/scientific_task_classification/entries/1
/scientific_task_classification/entries/2
/scientific_task_classification/entries/3
/scientific_task_classification/entries/4

tape-original-repository-resource · repository-path: README.md; tape/tasks; tape/data_utils at dad242d0341379255213cb8715d33b57aa9369bb (Confirms the five downloadable supervised datasets, evaluators, model list, and MIT code license.) · Supports 7 fields

Open source →

/access/tasks
/access/artifacts
/access/grader
/access/license
/resources
/implementations
/versions/0/formal_tracks

View source-level modification history on GitHub →

Access & reproducibility

Access: Fully open
License: MIT for the original implementation; benchmark datasets retain their source-specific terms and citation requirements.
Grader: Deterministic task-specific accuracy, precision, and rank-correlation scorers.
Implementations: 1 recorded · official