evo-final

A multi-source protein complex dataset for self-supervised preference training. This is v3 — Layer 1 only. L2 (swap relationships) and L3 (preference pairs) will be re-published after the STRING ortholog graph finishes building (currently a multi-hour Phase 1 SLURM job).

v2 of this dataset published L2 + L3 derived from wjiaqi/evo-afcdb-final's same-organism random swap shuffle. That swap construction was biologically uninformative (it mixes unrelated proteins from one organism), so v2's L2/L3 have been removed. The v3 rebuild requires every swap to be ortholog-aligned (swap human_A with yeast_A, not yeast_C).

Two configs

config_name	Rows	Description
natives (default)	85,554	All native heterodimers from both sources, no species filter.
natives-filtered	50,357	Same data, restricted to the Opisthokonta clade (Metazoa + Fungi) AND with a valid NCBI scientific-name genus — biological scope ≤ human–yeast taxonomic distance, every row addressable by the same NCBI tree-edge distance metric used at L2.

Load:

from datasets import load_dataset
all_nat   = load_dataset("wjiaqi/evo-final")["train"]                  # default = natives
filtered  = load_dataset("wjiaqi/evo-final", "natives-filtered")["train"]

Sources

Source	natives rows	natives-filtered rows	Provenance
nvda	69,693	34,620	wjiaqi/evo-afcdb-final → upstream nvidia_heterodimers_strict.parquet (NVIDIA AFDB strict subset, ipSAE_min ≥ 0.6 & N_clash_backbone ≤ 10).
conglab	15,861	15,737	wjiaqi/evo-conglab-final (Cong-lab human-PPI orthologs filtered through Boltz with iptm ≥ 0.7 & complex_plddt ≥ 70). 124 rows dropped from the filtered config: 1 row whose genus (Aeorestes) has no NCBI scientific-name match, and 123 rows whose lineage is literally na:na:na:Mammalia:Chordata (missing genus).
Total	85,554	50,357	unique natives

How the species filter works

• NVIDIA: each native's strain taxid is normalised to species-level (via a small STRAIN_TO_SPECIES map for yeast/fungi) and intersected with a 30-species whitelist of Opisthokonta organisms — the 22 NVIDIA-covered species plus 8 well-studied vertebrates/fungi (chimp, macaque, cow, dog, chicken, X. tropicalis, A. niger, A. nidulans) that densify the cross-species ortholog graph downstream. 34,620 of 69,693 strict natives survive (50.1%); the ~35k dropped are mostly soybean, Arabidopsis, rice/maize, kinetoplastid parasites, and bacteria.
• Cong-lab: each native's colon-separated taxonomy_a lineage is inspected; the terminal phylum is checked against an Opisthokonta whitelist. All 15,861 rows pass that whitelist (the dataset is human-PPI ortholog-expanded across vertebrates so every native is Chordata). 124 rows are then dropped because their genus has no scientific-name entry in the NCBI taxonomy — required so that L2 can give every conglab native an NCBI taxid for the tree-edge distance metric. Final conglab contribution to natives-filtered: 15,737.

The whitelist:

• Metazoa (animals): 36 phyla including Chordata, Arthropoda, Nematoda, Mollusca, Platyhelminthes, …
• Fungi: 11 phyla including Ascomycota, Basidiomycota, Mucoromycota, …

Schema (same in both configs)

Column	Type	Notes
native_id	str	Stable globally-unique key: nvda-AF-… or conglab-PPI…@organism_acc
source	str	'nvda' or 'conglab'
Taxonomy
taxid	int / NaN	NCBI species taxid (NVIDIA only; conglab uses opaque assembly ids)
organism	str	Latin binomial (NVIDIA) or genus (conglab)
organism_key	str / NaN	Conglab assembly id (e.g. HUMAN_QUERY, GCA_000146795.3)
lineage	str / NaN	Conglab colon-separated lineage (Genus:Family:Order:Class:Phylum)
clade	str / NaN	'Metazoa' or 'Fungi' (conglab rows; NVIDIA derivable from taxid)
phylum	str / NaN	Conglab last lineage segment
Chains
A_uniprot	str	UniProt accession. Conglab uses synthetic <canonical_uniprot>@<organism_acc> since per-organism accessions are not in the source — strip the suffix to recover the canonical id.
A_gene	str / NaN	Gene symbol (NVIDIA only)
A_seq, A_len	str, int	Full chain A sequence + length
B_uniprot, B_gene, B_seq, B_len	—	Same for chain B
NVIDIA confidence (NaN for conglab)
nvda_ipSAE_min	float	min(ipSAE_AB, ipSAE_BA) — bidirectional interface confidence lower bound (filter ≥ 0.6)
nvda_ipTM	float	Interface predicted TM-score
nvda_pDockQ	float	Global pDockQ
nvda_n_clash	int	Backbone clashes (filter ≤ 10)
nvda_complex_id	str	AFDB model entity id (e.g. AF-0000000211760987)
Conglab confidence (NaN for NVIDIA)
boltz_iptm	float	Boltz iptm
boltz_plddt	float	Boltz complex pLDDT
boltz_ptm	float	Boltz pTM
boltz_conf	float	Boltz confidence_score
ppi_id	str / NaN	Cong-lab interaction id — groups orthologs across organisms within the conglab source

Per-source score columns are mutually exclusive: an nvda row has all boltz_* null, and vice versa.

L2 / L3 roadmap

L2 (swap relationships) and L3 (preference pairs) require an ortholog group key (role_key) shared between NVIDIA and conglab natives. For NVIDIA this is being built from STRING orthologs over the same 30-species set (Phase 1 of the pipeline at w-jiaqi/EvoData), which is a multi-hour download job. For conglab, ppi_id already encodes ortholog grouping by construction.

When Phase 1 lands, this dataset will gain two more configs:

load_dataset("wjiaqi/evo-final", "swap_relationships")
load_dataset("wjiaqi/evo-final", "preference_pairs")

Both will use the same ortholog-aligned recipe across sources (no more same-organism random swaps).

Citation

• NVIDIA AFDB: Han Y, Tsenkov MI, Venanzi NAE, et al. bioRxiv (2026), doi:10.64898/2026.03.27.714458.
• STRING: Szklarczyk et al., Nucleic Acids Research (2023).
• UniProt: UniProt Consortium, Nucleic Acids Research (2025).
• Cong-lab Boltz-filtered PPI set (forthcoming).