rewirebio.iobenchmarks
Source

RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 · Project source

Original source

What this source contributed

Records in this release that cite this source. A record can cite several sources.

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Model or configurationMetricValueDatasetEvaluation
AlphaProtein-novo (Codesign), Atomic Motif EnzymePassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: protein-backbone RMSD at most 2 A, required motif all-atom RMSD at most 1.5 A, and AF3 minimum interface PAE below 1.5 A)26.02 count-per-hourFrozen project benchmark Atomic Motif Enzyme target set, 37 tasksAlphaProtein-novo (Codesign) on Frozen project benchmark Atomic Motif Enzyme, fixed ligand: distinct passers per GPU-hour
BoltzGen, frozen project benchmark recipePassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 minimum interface PAE at most 1.5 A and DNA-target-aligned binder backbone RMSD at most 8 A, aligning on all target DNA atoms in both the fixed and unfixed modes)9.46 count-per-hourFrozen project benchmark DNA-binding target set, 3 targetsBoltzGen (fixed DNA) on Frozen project benchmark DNA binding, fixed DNA: distinct passers per GPU-hour
BoltzGen, frozen project benchmark recipePassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 minimum interface PAE at most 1.5 A, chain-of-interest ipTM at least 80 out of 100, protein-backbone RMSD at most 1.5 A, and protein-backbone-aligned ligand RMSD at most 5 A (the sm_binder criterion, not the looser historical disco criterion))15.48 count-per-hourFrozen project benchmark small-molecule binding target set, 77 targetsBoltzGen on Frozen project benchmark small-molecule binding, unfixed ligand: distinct passers per GPU-hour
Proteina-Complexa (V1), protein binding, all-atom conditioningPassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 binder pTM at least 80 on the native 0 to 100 scale, minimum interface PAE at most 1.5 A, and target-backbone-aligned binder backbone RMSD at most 2.5 A, with valid required target mapping and conditioning)13.24 count-per-hourFrozen project benchmark protein-binding target set, 24 targetsProteina-Complexa on Frozen project benchmark protein binding, all-atom conditioning: distinct passers per GPU-hour
RFD4-Proteina (high thinking), previous release pairPassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: protein-backbone RMSD at most 2 A, required motif all-atom RMSD at most 1.5 A, and AF3 minimum interface PAE below 1.5 A)91.18 count-per-hourFrozen project benchmark Atomic Motif Enzyme target set, 37 tasksRFD4-Proteina, high thinking on Frozen project benchmark Atomic Motif Enzyme, fixed ligand: distinct passers per GPU-hour
RFD4-Proteina (high thinking), previous release pairPassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 minimum interface PAE at most 1.5 A and DNA-target-aligned binder backbone RMSD at most 8 A, aligning on all target DNA atoms in both the fixed and unfixed modes)41.97 count-per-hourFrozen project benchmark DNA-binding target set, 3 targetsRFD4-Proteina, high thinking on Frozen project benchmark DNA binding, unfixed DNA: distinct passers per GPU-hour
RFD4-Proteina (high thinking), previous release pairPassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 binder pTM at least 80 on the native 0 to 100 scale, minimum interface PAE at most 1.5 A, and target-backbone-aligned binder backbone RMSD at most 2.5 A, with valid required target mapping and conditioning)71.60 count-per-hourFrozen project benchmark protein-binding target set, 24 targetsRFD4-Proteina, high thinking on Frozen project benchmark protein binding, backbone conditioning: distinct passers per GPU-hour
RFD4-Proteina (high thinking), previous release pairPassing design clusters per accelerator hour (mean over targets of structurally distinct computational passers per GPU-hour, counting generation and common AlphaFold3 evaluation time, with 200 original attempts per target and failed attempts included; passing designs clustered within each model, protocol and target by Foldseek at TM 0.5; pass criterion: AF3 minimum interface PAE at most 1.5 A, chain-of-interest ipTM at least 80 out of 100, protein-backbone RMSD at most 1.5 A, and protein-backbone-aligned ligand RMSD at most 5 A (the sm_binder criterion, not the looser historical disco criterion))112.25 count-per-hourFrozen project benchmark small-molecule binding target set, 77 targetsRFD4-Proteina, high thinking on Frozen project benchmark small-molecule binding, unfixed ligand: distinct passers per GPU-hour

Claims

Statements this source supports, each with where in the source it was checked.

  • RFD4-Proteina (high thinking), previous release pair: checkpoint identity of the pair the efficiency table was measured on: Checkpoint identity The efficiency table's scope note names only 'the previous release pair'. The card's checkpoint table gives a 'Previous release' column with flow checkpoint chk_epoch_00000169_step_000000340000-ema, autoencoder 2026_09_04_release_ae_z16_400k-ema, latent dimension 16, 33 residue classes and 434,348,737 stored parameters. The same card separately says the 2026-10-19 pair supersedes a 2026-10-15 pair. Neither pair is tied to the table, and the benchmark methods supplement says each cohort's exact schedule and checkpoint come from its own native config, which the packet does not publish. No checkpoint filename is assigned to the table's rows in this batch. (Hugging Face model card, 'Measured design efficiency' blockquote above the table, and 'Checkpoints in this release' table column 'Previous release'; benchmarks/METHODS.md, 'Model recipes and adaptations' row RFD4-Proteina)
  • RFD4-Proteina frozen project benchmark, October 9 2026 supplement: attempt accounting: Attempt accounting Each model, protocol and target has 200 original attempts, including native failures and unmatched required motifs. Rejected originals are not replaced and outputs are not selected with the common AlphaFold3 evaluation. Failed attempts consume time and stay included in the efficiency denominator, while setup, compilation, CPU scoring and idle allocation time are excluded. Values are means of target-level rates, not pooled totals or single-design latency. (Hugging Face model card, 'Measured design efficiency', paragraph beginning 'There are 200 original attempts per target'; benchmarks/METHODS.md, 'Common accounting and timing')
  • RFD4-Proteina frozen project benchmark, October 9 2026 supplement: conditioning differences: Conditioning differences Conditioning differs between some methods: PPI uses RFD4 backbone conditioning versus all-atom conditioning for several baselines; DNA compares unfixed RFD4 inputs with fixed-DNA baselines; Proteina-Complexa's small-molecule route uses fixed ligands. Only three DNA targets are included. Computational passers do not establish experimental activity. (Hugging Face model card, 'Measured design efficiency', paragraph beginning 'Conditioning differs between some methods')
  • RFD4-Proteina frozen project benchmark, October 9 2026 supplement: metric definition: Metric definition The card defines the table's metric as 'mean per-target structurally distinct computational passers per GPU-hour, including generation and common AF3 evaluation', higher being better, and adds that computational passers do not establish experimental activity. The provenance file gives the expression mean_over_targets(3600 * passing_structure_clusters / (generation_gpu_s + af3_gpu_s)). The methods supplement defines a passing cluster as a distinct Foldseek cluster among passing originals, computed within each model, protocol and target at TM threshold 0.5, and says these are algorithmic structural clusters, not experimental successes or globally distinct proteins. (Hugging Face model card, 'Measured design efficiency' metric line and the paragraph below the table; benchmarks/provenance.json, key metric; benchmarks/METHODS.md, final paragraph of 'Common AF3 scoring')
  • RFD4-Proteina frozen project benchmark, October 9 2026 supplement: pass criteria: Pass criteria Protein binding, 24 targets: AF3 binder pTM at least 80 on the native 0 to 100 scale, minimum interface PAE at most 1.5 A, and target-backbone-aligned binder backbone RMSD at most 2.5 A, with valid required target mapping and conditioning, with CA-only RMSD for one CA-only comparator. Small molecule, 77 targets: AF3 minimum interface PAE at most 1.5 A, chain-of-interest ipTM at least 80 out of 100, protein-backbone RMSD at most 1.5 A, and protein-backbone-aligned ligand RMSD at most 5 A (the sm_binder criterion, not the looser historical disco criterion). Atomic Motif Enzyme, 37 targets: protein-backbone RMSD at most 2 A, required motif all-atom RMSD at most 1.5 A, and AF3 minimum interface PAE below 1.5 A. DNA, 3 targets: AF3 minimum interface PAE at most 1.5 A and DNA-target-aligned binder backbone RMSD at most 8 A, aligning on all target DNA atoms in both the fixed and unfixed modes. The pass criteria are a common AlphaFold3 evaluation, stated not to be the model's own acceptance filter, and the packet reuses existing scores rather than launching new folding or rescoring. (benchmarks/METHODS.md, 'Common AF3 scoring' table, all five rows, and the paragraph above it)

Models, benchmarks and other records

Models (1)

Configurations (4)

Benchmarks (1)

Protocols (6)

Datasets (4)

Evaluations (8)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

28 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance
Archive member
artifacts/hf_benchmarks_METHODS.md.gz
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_member

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Archive sha256
874fbc30bc22db1b38acee7a244cdc59daddb2ff88600aabcc2209888d93783e
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Artifact SHA-256
46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Licence
BSD 3-Clause
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.licence

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Limitations
1 values
  • States that the executed schedule and checkpoint for each RFD4-Proteina cohort come from its own native config, which is not published in this packet, so no row in it is tied to a named checkpoint file.
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.limitations

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Media type
text/markdown
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.media_type

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Publication status
official_project_source
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.publication_status

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Retrieval url
https://huggingface.co/nvidia/RFD4-Proteina/raw/621b785abf9d61dd24f4b81ae4dbd1330728f030/benchmarks/METHODS.md
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieval_url

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Retrieved at
2026-11-15T23:10:00Z
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.retrieved_at

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

Review: artifact sha256
46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b
Source metadata
RFD4-Proteina benchmarks/METHODS.md at revision 621b785

Original source ↗

No field-specific location recorded

Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Retrieved: 2026-11-15T23:10:00Z

catalogued

No individual claim review recorded

Audit details

Field: attributes.review.artifact_sha256

Source artifact SHA-256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b

Hash scope: Hash scope not separately documented; inspect source record

Archive member: artifacts/hf_benchmarks_METHODS.md.gz

How to read the evidence

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

The methods supplement of the published benchmark packet, giving the timing accounting, the per-method recipes, the common AlphaFold3 pass criteria and the clustering rule.

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: rfd4-proteina-20261115-source-hf-benchmark-methods

areas
proteins-complexes; protein-structure; molecular-interactions
contexts
research
archive member
artifacts/hf_benchmarks_METHODS.md.gz
archive sha256
874fbc30bc22db1b38acee7a244cdc59daddb2ff88600aabcc2209888d93783e
artifact sha256
46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b
licence
BSD 3-Clause
limitations
States that the executed schedule and checkpoint for each RFD4-Proteina cohort come from its own native config, which is not published in this packet, so no row in it is tied to a named checkpoint file.
media type
text/markdown
publication status
official_project_source
retrieval url
https://huggingface.co/nvidia/RFD4-Proteina/raw/621b785abf9d61dd24f4b81ae4dbd1330728f030/benchmarks/METHODS.md
retrieved at
2026-11-15T23:10:00Z
review
artifact sha256: 46d1e181c495d4ecc9a3992a4e7f3cd8fae0b4904fb799ab5c1a883090ca837b; date: 2026-11-15; method: source-hash-verification; automated-source-review; note: Retrieved twice on 2026-11-15 with identical bytes; the hash is of the retrieved artifact, not of a text layer.; retrieval url: https://huggingface.co/nvidia/RFD4-Proteina/raw/621b785abf9d61dd24f4b81ae4dbd1330728f030/benchmarks/METHODS.md; reviewer: claude; reviewer note: Claude (Opus 5) extraction agent; no human review claimed. Extraction only: a different agent reviews this batch, and nothing here is reproduced.
source locator
Sections 'Common accounting and timing', 'Model recipes and adaptations', 'Common AF3 scoring' and 'Supplement selection and conditioning comparisons'
url
https://huggingface.co/nvidia/RFD4-Proteina/blob/621b785abf9d61dd24f4b81ae4dbd1330728f030/benchmarks/METHODS.md
version
Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
Related records

Suggest a correction