Add dataset card

README.md

@@ -95,4 +95,115 @@ configs:
     path: data/prs_rrs_human-*
   - split: orthogonal_assays
     path: data/orthogonal_assays-*
+license: cc-by-4.0
+language:
+- en
+tags:
+- protein
+- protein-protein-interaction
+- interactome
+- biology
+- alphafold
+pretty_name: CCSB AI Interactome Benchmark
 ---
+
+# CCSB-AI-Interactome
+
+Harmonized protein-protein interaction benchmark built from the Supplementary Data of:
+
+> Lambourne, L., Yadav, A., Wang, Y. et al. **Experimental assessment of AI-based interactome mapping.** *Nature Communications* **17**, 4894 (2026). https://doi.org/10.1038/s41467-026-70942-x
+
+The study screened a near-complete *S. cerevisiae* ORFeome (5,854 sequence-validated ORFs, 99.5% coverage) with an improved yeast two-hybrid assay to produce YeRI, then put proteome-scale AlphaFold/RoseTTAFold predictions through the same wet-lab pipeline. Its headline finding is that high-confidence AI predictions match experimental data in quality but recover far fewer strictly novel interactions in proteome-wide screening.
+
+Every table has been resolved to amino acid sequences and coerced to a single schema, so all splits share identical columns and can be iterated without special-casing.
+
+## Splits
+
+| Split | Rows | Positive | Negative | Unscorable | Organism | What it is |
+|---|---|---|---|---|---|---|
+| `afrf_y2h_yeast` | 597 | 17 | 559 | 21 | yeast | AlphaFold/RoseTTAFold-predicted pairs, each given a wet-lab Y2H verdict (Supp. Data 19) |
+| `afrf_y2h_human` | 4,046 | 402 | 2,916 | 728 | human | Zhang et al. AF/RF human predictions, each given a Y2H verdict (Supp. Data 22) |
+| `yeri` | 1,970 | 1,970 | 0 | 0 | yeast | The yeast reference interactome, with AlphaFold metrics and orthogonal assays joined (Supp. Data 16 + 24) |
+| `y2h_union_25` | 4,556 | 4,556 | 0 | 0 | yeast | Union of four systematic yeast Y2H maps (Supp. Data 17) |
+| `valbin_25` | 12,706 | 12,706 | 0 | 0 | yeast | Validated binary yeast PPIs across structural, literature, and Y2H evidence (Supp. Data 21) |
+| `prs_rrs_yeast` | 306 | 108 | 198 | 0 | yeast | scPRS-v2 positive and scRRS-v2 random reference sets, with every assay readout attached (Supp. Data 2, 3, 4, 23) |
+| `prs_rrs_human` | 729 | 342 | 387 | 0 | human | hsPRS-v2 / hsRRS-v2 plus the literature and random comparators tested alongside (Supp. Data 22) |
+| `orthogonal_assays` | 8,846 | 858 | 6,484 | 1,504 | yeast | Raw MAPPIT and GPCA assay-level results, one row per assayed configuration (Supp. Data 11) |
+
+## The interesting one
+
+`afrf_y2h_yeast` and `afrf_y2h_human` are pairs that a structure-based AI method called an interaction, each carrying an experimental verdict. Every negative is a structural near-miss rather than a random pair, which makes them a far harder discrimination target than the usual random-negative benchmarks, and a direct test of whether a sequence-only model recovers signal that structure-based virtual screening got wrong.
+
+## Schema
+
+All splits share these columns.
+
+| Column | Type | Meaning |
+|---|---|---|
+| `A`, `B` | string | UniProt accession |
+| `SeqA`, `SeqB` | string | Amino acid sequence |
+| `OrgA`, `OrgB` | string | Organism binomial |
+| `NameA`, `NameB` | string | Systematic ORF name (yeast) or source accession (human) |
+| `GeneA`, `GeneB` | string | Common gene name |
+| `labels` | int8 | `1` interacting, `0` non-interacting, `-1` not scorable |
+| `source_dataset` | string | Provenance set within the paper |
+| `evidence` | string | `;`-joined evidence flags, for example `I3D-exp-24;Lit-BM-24` |
+| `y2h_v1_result`, `y2h_v4_result` | string | Raw Y2H verdict, empty when not tested |
+| `mappit_result`, `gpca_result` | string | Orthogonal assay verdict, empty when not tested |
+| `score` | float32 | Assay score, or YeRI manual growth score |
+| `contact_probability`, `model_confidence`, `pDockQ`, `iPAE`, `CCC` | float32 | AlphaFold confidence metrics, `NaN` when unavailable |
+| `novel` | int8 | `1` if strictly novel at publication, `0` if not, `-1` unknown |
+
+### Label semantics differ by split
+
+This is the one thing to read before using `labels`.
+
+- `afrf_y2h_yeast`, `afrf_y2h_human`, `orthogonal_assays`: the label is the **assay outcome**. This is the adjudication being benchmarked.
+- `prs_rrs_yeast`, `prs_rrs_human`: the label is **reference-set membership** (PRS positive, RRS random). The assay outcome is in the result columns instead, since reference-set membership is the ground truth and the assay is what is being calibrated.
+- `yeri`, `y2h_union_25`, `valbin_25`: positive-only interaction lists, so every label is `1`. Supply your own negatives.
+
+`labels = -1` marks rows that cannot be scored: autoactivators, failed tests, and clones that failed sequence confirmation. They are retained rather than dropped so nothing is silently lost. Filter with `ds.filter(lambda x: x["labels"] >= 0)`.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+# The AI-adjudicated benchmark
+ds = load_dataset("GleghornLab/CCSB-AI-Interactome", split="afrf_y2h_human")
+ds = ds.filter(lambda x: x["labels"] >= 0)   # 3,318 scorable pairs
+
+# The yeast reference interactome
+yeri = load_dataset("GleghornLab/CCSB-AI-Interactome", split="yeri")
+```
+
+## Caveats
+
+- **Y2H negatives are not verified non-interactions.** The assay recovers roughly 20 to 30% of scPRS-v2 positives, so a `Negative` means "not detected in this assay" and not "these proteins do not interact." Treat `afrf_y2h_*` as a Y2H-verifiability benchmark, not a ground-truth interaction benchmark.
+- **Yeast is well represented in common PPI training corpora** (STRING, BioGRID). Any evaluation on the yeast splits needs a homology or cluster-level leakage control against the model's training set. The `novel` column on `yeri` marks the 1,446 rows that were strictly novel at publication and are the least likely to be memorized.
+- **`yeri` and `prs_rrs_human` contain both-orientation rows.** `yeri` has 1,970 rows over 1,910 unique unordered pairs, matching the paper's reported 1,910 PPIs; `prs_rrs_human` has 729 rows over 590 unique pairs. Deduplicate on a sorted `(NameA, NameB)` key if you need one row per pair.
+- **`mappit_result` and `gpca_result` are a pair-level collapse** of the assay-level table, scoring a pair positive if any tested configuration was positive. This does not exactly reproduce the per-assay counts in the paper's Figure 2a, which applies additional quality control that is not recoverable from the released tables. Use the `orthogonal_assays` split to recompute under a different rule.
+- **10 rows lack sequences.** Four human accessions referenced by the source tables are now inactive in UniProt (`A0A0C4DGZ8`, `O43930` and `Q8WV35` deleted, `P01562` demerged). Their rows are retained with empty `SeqA` / `SeqB`.
+
+## Provenance
+
+Sequences were resolved from UniProt reference proteomes `UP000002311` (*S. cerevisiae* S288C) and `UP000005640` (*H. sapiens*), with per-identifier lookups for entries outside those proteomes. Yeast ORF names were matched on the UniProt ordered locus name field.
+
+## License and citation
+
+Source Supplementary Data are distributed under CC BY 4.0 as part of the open access article, and this derived dataset carries the same license. Cite the original paper:
+
+```bibtex
+@article{lambourne2026interactome,
+  title   = {Experimental assessment of AI-based interactome mapping},
+  author  = {Lambourne, Luke and Yadav, Anupama and Wang, Yang and others},
+  journal = {Nature Communications},
+  volume  = {17},
+  number  = {1},
+  pages   = {4894},
+  year    = {2026},
+  doi     = {10.1038/s41467-026-70942-x}
+}
+```
+
+The yeast interaction maps are also browsable at the [Yeast Interactome Portal](https://yeast.interactome-atlas.org/).