遇见数据集

Mad2: all models and scores

收藏
Zenodo2026-10-23 更新2026-10-29 收录
官方服务:

资源简介:

This record contains all 11,000 original Artificial Intelligence Sampling (AIS) models of the 1KLQ-derived Mad2 monomer construct with 197 modeled residues and wild-type sequence at modeled positions, their 11,000 matched Boltz2 no-MSA template-only refolds, and associated scores. There are 1,000 original AIS models from each of eleven sampling methods: AF2/AFSample2, AF3/AFSample3, OpenDDE Exact, Boltz2 Exact, BioEmu v1.1, BioEmu2-preview, SPEACH_AF, AF-Cluster, AlphaFlow-PDB, AlphaFlow-MD and ESMFlow-PDB. The first four methods used 40% stochastic MSA column masking. Both BioEmu methods used native unmasked conditioning; the other five methods retained their own native sampling protocols. All generated AIS models are retained, including native physicality-filter failures, with passed, failed and unavailable flags kept distinct. Sampling counts are not equilibrium populations. Original AIS structures and their matched refolds are organized separately within this target record. Scores include native inference confidence where available, Boltz2 template-only refold pTM, per-residue refold pLDDT, experimental-reference TM similarity and original-AIS versus matched-refold TM similarity. Confidence values are not structural TM scores. Raw refold pLDDT values use their recorded 0–1 scale; any 0–100 display conversion is explicit. Missing native confidence is not replaced with refold confidence. Original/refold pairs are joined by model ID and both coordinate hashes. Experimental controls are PDB 1DUJ (open) and 1S2H (closed, R133A mutant), using the 175 shared resolved target positions 3–101 and 112–187. Controls, original deposited coordinates, comparison copies and source citations are separately identified and excluded from the 11,000 AIS-model and 11,000-refold counts. The 1S2H author-numbered R133A substitution maps to target position 125; unresolved residues and construct differences remain explicit. No residues are silently trimmed from deposited originals. Original/refold TM similarity instead uses the full 197-residue predicted construct, with bidirectional full US-align searches and average-length normalization; the reported score is the mean of both directions. The package preserves model/run/checkpoint provenance, physical-filter flags, sequence/residue mappings, scored-model identities and applicable upstream notices. PDB archive sources retain their CC0 attribution; other data and code retain their applicable rights, including AF3 output terms. A blanket license is not inferred from a software repository license. BioEmu2-preview is identified by the v0.1.0rc0 release family, the actual inference commit and checkpoint revision in provenance. The Adk and Mad2 resources are separate target-level Zenodo records. Each target record contains both the original AIS models and matched template-only refolds; these are separate file groups within the record. This research sponsored in part by NIH/NIGMS T32GM141865 and NIH/NIA R25AG088409. Thanks goes also to members of the Montelione Lab for useful insight and discussion: T.A Ramelot, Y.J. Huang, B.A. Shurina, and L.E. Spaman.Release and provenance scheraga-poster-v1.0.0; commit 27f2b5dd98732a9596652717165e36cd0b8e4115. Current eleven-method Mad2 cohort: 11,000 AIS predictions and 11,000 exact matched Boltz2 no-MSA template-only refolds. Construct sequence SHA-256 5e536a6215dbe1fb1b2f11628d6e182409edfa252fb3287058e81f192b5e8f26. All-model eleven-method analysis and intake records are retained. Original/refold pairing and 22,000 bidirectional two-target comparisons passed independent validation (44,000 directional results). The complete two-target score table SHA-256 is 269ffe3fba009d3c388fff9b81697b1aed41342c393cb6829bce531560710152; each target package must include only its 11,000 rows. Freeze the per-target membership and coordinate/score hashes before packaging. No new prediction or scientific regeneration is authorized by this config. This is the Mad2-alone monomer collection, not the earlier Mad2–peptide complex collection. Packaging preserves 11000 original models and 11000 refolds. 10000 native/analysis file aliases are recorded explicitly rather than archived twice; 1000 additional analysis copies are labeled derivatives. Portable per-target score/provenance exports preserve all numeric and model/hash identities. Browse the deposit Start with README.md or download index.html. files.tsv and manifest.json list every selected file, byte count and SHA-256. Citation files identify this version. inputs.tar.gz: Unscored PDB originals and explicitly labeled protein/CA comparison copies; excluded from AI-model totals.; Frozen target/settings/reference descriptors with portable locators; scientific values, residue masks and pointers preserved. outputs.tar.gz: Exactly 11000 native AIS coordinate files, 1000 per method, retaining physical-filter failures.; Exactly 11000 matched Boltz2 no-MSA template-only refold coordinate files.; 1000 physically distinct analysis-coordinate copies; 10000 byte-identical native/analysis aliases are represented by the archived native bytes and the complete alias table. These are not additional AIS predictions. provenance.tar.gz: Every native/analysis physical alias plus raw-to-derived hashes and explicit privacy export provenance.; Owner-selected CC BY 4.0 authored-content license, MIT source-code notice and explicit third-party/PDB exceptions. reports.tar.gz: 11000 matched refold confidence JSON summaries; native scores are preserved separately where available.; 11000 matched per-residue refold pLDDT NPZ files, preserving native array units.; Per-target 11000-row original/refold TM table, 22000-row reference-pair table, 11000-row reference-wide table and full confidence/identity joins; no recomputation.; Compact frozen projections, assignments, selections, matrices, masks and Bonsai outputs; reference rows remain explicitly separated. source.tar.gz: source at the exact release; submodule pins in manifest.json. AF3 output notices AFSample3/AF3 outputs and applicable derivatives retain the separate AlphaFold 3 output terms. The authored-content CC BY designation does not replace or override those terms. Official terms and the original binding notice are included in provenance.tar.gz under provenance/af3-output-terms/. No model weights are distributed.

提供机构:
Zenodo
创建时间:
2026-10-23
二维码
社区交流群
二维码
科研交流群
商业服务