--- configs: - config_name: preference_pairs data_files: - split: train path: data/preference_pairs/train.parquet - split: validation path: data/preference_pairs/validation.parquet - split: test path: data/preference_pairs/test.parquet default: true - config_name: swap_complexes data_files: - split: train path: data/swap_complexes/train.parquet - split: validation path: data/swap_complexes/validation.parquet - split: test path: data/swap_complexes/test.parquet - config_name: swap_relationships data_files: data/swap_relationships.parquet - config_name: natives data_files: data/natives.parquet - config_name: natives-filtered data_files: data/natives-filtered.parquet license: cc-by-4.0 task_categories: - other tags: - protein - complex - protein-protein-interaction - self-supervised - preference - swap - ortholog - alphafold size_categories: - 100K **Identity filter (v3.1):** L2/L3/L4 are restricted to ortholog pairs whose > **BLAST-style identity ≥ 0.80 on both chain A and chain B**. Below this the > "swap" is too divergent to be an informative dispreferred negative for > preference training. Unfiltered parquets remain available locally as > `*_unfiltered.parquet` for ablation studies. > > **Same-species & no-data filter (v3.4):** Rows where both interacting > partners come from the same organism (`taxid_1 == taxid_2`) are dropped — > they were leaking through from AFCDB genus-only labels (e.g. `"Homo"` vs > `"Homo sapiens"`, both taxid 9606) and don't represent a real cross-species > swap. Rows whose species pair has no TimeTree calibration (e.g. *Alouatta* > vs other primates) are also dropped. As a result, every shipped row has > `divergence_mya` non-null and strictly positive (0.29 – 1530 Mya). > > **Unified configs (v3.7):** All layers (`swap_relationships`, > `swap_complexes`, `preference_pairs`) bundle within-source AND > cross-source rows into single configs. Filter by the `source` column > ∈ {`nvda`, `conglab`, `cross`} on any layer. The `*_cross_source` > configs from earlier versions are retired. > > **Train/val/test splits (v3.6 — rescued):** L3 and L4 carry a `split` > column ∈ {`train`, `validation`, `test`}. Pipeline: > 1. mmseqs2 `easy-cluster` every unique chain across L3 (within + > cross-source) at **30% min-seq-id, 80% coverage** → 14,814 chains > collapse to 1,716 clusters. > 2. **Cluster co-occurrence merge:** if any L3 row has chains in two > different clusters, union those clusters. The 1,716 clusters collapse > to 545 connected components (largest = ~21% of rows). > 3. **Greedy 80/10/10 bin-pack by component size:** largest-first, place > each component into the split with the most remaining capacity. > Lands exactly at 80.00% / 10.00% / 10.00%. > 4. Every L3 row's 4 chains live in a single component → single split. > **No rows dropped** (was ~25% drop in v3.5). L4 inherits the parent > L3 row's split via `swap_complex_id`. > 5. **Leak-free guarantee:** every chain belongs to exactly one component, > every component belongs to exactly one split, so no chain appears in > multiple splits. Verified by `mmseqs2 easy-search` at -s 7.5: > 0 exact-string overlaps train↔val/test; residual borderline-PID hits > (mean ~33%, 95th 39–44%) reflect mmseqs2 cluster's heuristic boundary, > typical of paper-grade homology-reduced splits. > > The combined swap_complexes row split is exactly **80/10/10** > (58,008 / 7,251 / 7,251). Earlier *strict-drop* variants are reproducible > via `Phase 23` + a chosen seed; the rescue path is `Phase 36`. ## Five-layer structure | Layer | config_name | Rows | Purpose | |---|---|---:|---| | L1 | `natives` | 85,554 | All native heterodimers, no filter. | | L1ʹ | `natives-filtered` | 50,357 | Same, Opisthokonta + NCBI-genus-mappable. | | L2 | `swap_relationships` | 77,571 | FK-only pairs of natives sharing an ortholog group. Combined within + cross-source (filter via the `source` column). *(unsplit)* | | L3 | `swap_complexes` | 72,510 | **wjiaqi/evo-style** — 4 sequences inline per row + gene/species/identity/mya labels. Combined within + cross-source. Split: 58,008 / 7,251 / 7,251 (exact 80/10/10). | | **L4** | `preference_pairs` *(default)* | **290,040** | **Train-ready.** Per-row (preferred native, dispreferred chimera) with species+gene labels on every chain. Combined within + cross-source. Split: 232,032 / 29,004 / 29,004 (exact 80/10/10). | ```python from datasets import load_dataset prefs = load_dataset("wjiaqi/evo-final")["train"] # L4, combined (default) complexes = load_dataset("wjiaqi/evo-final", "swap_complexes")["train"] # L3, combined relations = load_dataset("wjiaqi/evo-final", "swap_relationships")["train"] # L2, combined # Filter by source via the `source` column ∈ {"nvda", "conglab", "cross"} on any layer. ``` L1, L2, L3, L4 are mathematically equivalent at the *information* level — L3 is `expand(L2) ⨝ natives`, L4 is `expand(L3, {AB,BA} × {sp1,sp2}) = 4 × L3`. Pick whichever layer matches your consumption pattern. ## Three sequence-identity metrics on every L2/L3/L4 row | metric | formula | use case | |---|---|---| | `seq_identity_*` (**primary**) | matches / **aligned columns (non-gap)** | BLAST-style "% identity"; what's in literature; aligned-region conservation | | `seq_identity_*_min_len` | matches / min(len_a, len_b) | core-domain identity; bounded by shorter chain | | `seq_identity_*_max_len` | matches / max(len_a, len_b) | conservative; penalises length mismatch | Computed via BLOSUM62 global alignment (Biopython `PairwiseAligner`, open gap −11, extend −1). For orthologs with different lengths (e.g. human UQCRH 91 aa vs yeast QCR6 135 aa), the three values differ — BLAST ≈ 44%, max-len ≈ 30%. Standard literature reports the BLAST number. ## Distance label B — TimeTree divergence time `divergence_mya` = species-pairwise Median Time from [TimeTree](http://timetree.org), molecular-clock calibrated. Examples: | species pair | mya | |---|---:| | Mouse ↔ Rat | 13 | | Human ↔ Mouse | 87 | | Human ↔ Zebrafish | 429 | | S. cerevisiae ↔ S. pombe | 543 | | Fly ↔ Worm | 727 | | **Human ↔ Yeast** | **1,275** | We rejected NCBI tree-edge distance as a label: it's cladistic, not chronological, and inverts biological ordering (NCBI: human-fly 58 > human-yeast 38; reality: 686 < 1,275 Mya). For conglab, where lineage only carries genus, we map each genus to its lowest-taxid species descendant in NCBI before querying TimeTree. ## L1 sources and species filter | Source | `natives` | `natives-filtered` | Provenance | |---|---:|---:|---| | `nvda` | 69,693 | 34,620 | [wjiaqi/evo-afcdb-final](https://huggingface.co/datasets/wjiaqi/evo-afcdb-final) — NVIDIA AFDB strict (ipSAE_min ≥ 0.6 & N_clash_backbone ≤ 10). | | `conglab` | 15,861 | 15,737 | [wjiaqi/evo-conglab-final](https://huggingface.co/datasets/wjiaqi/evo-conglab-final) — Cong-lab human-PPI orthologs filtered through Boltz (iptm ≥ 0.7 & complex_plddt ≥ 70). 124 rows dropped from filtered for no-NCBI-genus. | Filter: **Opisthokonta** (Metazoa + Fungi = clade ≤ human-yeast distance) AND a valid NCBI scientific-name genus so every row maps to a taxid for TimeTree lookup. ## Schemas ### L3 — `swap_complexes` (wjiaqi/evo-style; one row = one cross-species ortholog pair) | Column | Notes | |---|---| | `swap_complex_id`, `relationship_id`, `source`, `interaction_group_id` | identifiers; `source` ∈ {nvda, conglab, cross} | | `species_1`, `taxid_1`, `organism_acc_1` | sp1 organism metadata (organism_acc filled for conglab only) | | `protein_A_sp1_{uniprot,gene,seq,len}` | chain A of native 1 | | `protein_B_sp1_{uniprot,gene,seq,len}` | chain B of native 1 | | `species_2`, `taxid_2`, `organism_acc_2`, `protein_A_sp2_*`, `protein_B_sp2_*` | same for native 2 | | `role_key_A`, `role_key_B` | ortholog group ids — STRING `OG_` for nvda and cross, canonical human UniProt for conglab | | `seq_identity_A`, `seq_identity_A_min_len`, `seq_identity_A_max_len` | three identity metrics for chain A | | `seq_identity_B`, `seq_identity_B_min_len`, `seq_identity_B_max_len` | three identity metrics for chain B | | `divergence_mya` | TimeTree species divergence in Mya (NaN where TimeTree has no data) | | `split` | `train` / `val` / `test`; assigned by 30%-PID mmseqs2 cluster, 60/20/20 ratio, mixed-cluster rows → train (no leakage) | ### L4 — `preference_pairs` (training-ready) | Column | Notes | |---|---| | `preference_id`, `swap_complex_id`, `relationship_id`, `source` | identifiers + FKs | | `crossover` | `'AB'` = dispreferred is (sp1.A, sp2.B); `'BA'` = (sp2.A, sp1.B) | | `preferred_native_id`, `preferred_species`, `preferred_taxid`, `preferred_organism_acc` | which native is the positive | | `preferred_A_{uniprot,gene,seq,len}`, `preferred_B_*` | the native heterodimer | | `dispreferred_A_species`, `dispreferred_A_taxid`, `dispreferred_A_organism_acc` | species of the A chain of the swap chimera | | `dispreferred_A_{uniprot,gene,seq,len}` | the A chain of the swap | | `dispreferred_B_*` | same for B (different species — that's the swap!) | | `seq_identity_A`, `seq_identity_A_min_len`, `seq_identity_A_max_len`, `seq_identity_B`, … | identity metrics inherited from L3 | | `divergence_mya` | TimeTree distance between the two natives' species (inherited from L3) | | `split` | inherited from L3 via `swap_complex_id` | ### L2 — `swap_relationships` (FK-only, lightweight) | Column | Notes | |---|---| | `relationship_id`, `source`, `swap_kind` | identifiers; `swap_kind` always `'cross_species_ortholog'` | | `native1_id`, `native2_id` | FK → L1 | | `taxid_1`, `taxid_2`, `organism_1`, `organism_2`, `role_key_A`, `role_key_B`, `interaction_group_id` | minimal metadata for filtering / joining | | `seq_identity_A`, `seq_identity_A_min_len`, `seq_identity_A_max_len`, `seq_identity_B`, …, `divergence_mya` | labels (same values as L3/L4) | ### L1 — `natives` / `natives-filtered` Vocabulary of native heterodimers; identical schema, the `_filtered` variant just applies the Opisthokonta + NCBI-genus filter. | Column | Notes | |---|---| | `native_id`, `source` | `nvda-AF-…` or `conglab-PPI…@organism_acc` | | `taxid`, `organism`, `organism_key`, `lineage`, `clade`, `phylum` | taxonomy (NaN where unavailable) | | `A_{uniprot,gene,seq,len}`, `B_{uniprot,gene,seq,len}` | the two chains | | `nvda_ipSAE_min`, `nvda_ipTM`, `nvda_pDockQ`, `nvda_n_clash`, `nvda_complex_id` | NVIDIA AF-Multimer scores (NaN for conglab) | | `boltz_iptm`, `boltz_plddt`, `boltz_ptm`, `boltz_conf`, `ppi_id` | conglab Boltz scores (NaN for nvda) | ## Within-source vs cross-source Each layer has two variants: - **within-source** (no `_cross_source` suffix): pairs are NVIDIA × NVIDIA (via STRING ortholog groups) or conglab × conglab (via shared `ppi_id`). Both natives are from the same upstream dataset. - **cross-source** (`_cross_source`): pairs are NVIDIA × conglab. They exist because 337 STRING ortholog signatures appear in BOTH NVIDIA and conglab (= 30% of NVIDIA signatures, 40% of conglab signatures). For these signatures we get to pair an NVIDIA AF-multimer-verified native with a conglab Boltz-verified native of the same conserved interaction. Keep them separate so consumers can opt in to one, the other, or both. ## Construction summary ``` NVIDIA AFDB strict + conglab Boltz-filtered → L1 natives ↓ STRING orthologs over 30 Opisthokonta species → role_keys ↓ L2 swap_relationships (within-source: nvda×nvda via STRING, conglab×conglab via ppi_id; cross-source: signature overlap) ↓ L3 swap_complexes ( = L2 ⨝ natives, 4 sequences inline per row) ↓ L4 preference_pairs ( = expand L3 to 4 rows: {AB, BA} × {sp1 preferred, sp2 preferred} ) ``` Plus per-row: - BLOSUM62 global alignment → 3 identity metrics - TimeTree pairwise Median Time → divergence Mya Code: [w-jiaqi/EvoData](https://github.com/w-jiaqi/EvoData/tree/cursor/build-nvidia-afcdb-dataset-4a41) (scripts 01–20, plus `species_config.py`). ## 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). - TimeTree: Kumar et al., *Mol Biol Evol* (2022). - Cong-lab Boltz-filtered PPI set (forthcoming).