Upload README.md with huggingface_hub

README.md

@@ -7,58 +7,63 @@ tags:
   - proteins
   - protein-protein-interaction
   - coevolution
-  - cross-species
-  - structure-prediction
-pretty_name: EvoBind loose same-species ortholog pool
+pretty_name: EvoBind moderately-filtered same-species PPI orthologs
 size_categories:
-  - 1M<n<10M
+  - 100K<n<1M
 ---
 
-# EvoBind — loose same-species ortholog pool
+# EvoBind — moderately-filtered same-species PPI orthologs
 
 ## What this is
 
 For each of the **17,849 high-confidence human PPI predictions** in
 `final_predictions_90.tsv` (90% expected precision) from
 [Cong Lab humanPPI](https://conglab.swmed.edu/humanPPI/humanPPI_download.html),
-this dataset extracts up to **N same-species ortholog pairs** from the Cong
-Lab `protein_omicMSAs` (multi-genome MSAs covering thousands of eukaryotic
-genomes — vertebrates, insects, fungi, plants).
-
-For each PPI `(A_human, B_human)` we emit:
+this dataset provides:
 
 * the **human pair** itself (`pair_type = human_pair`,
   `ortholog_rank = 0`), and
-* up to **`max_orthologs_per_ppi`** same-species pairs `(A_org, B_org)` for
-  every organism that appears in both `MSA(A)` and `MSA(B)`, sorted by
-  combined sequence identity to the human queries (descending), with
-  `pair_type = natural` and `ortholog_rank = 1..N`.
+* up to **10 same-species ortholog pairs** `(A_org, B_org)` taken from
+  the Cong Lab `protein_omicMSAs`, sorted by combined sequence identity
+  to the human queries (descending), with `pair_type = natural` and
+  `ortholog_rank = 1..N`.
 
-This is intentionally a **loose pool of candidate training pairs**. Quality is
-mixed: many rows will fail structure prediction (e.g. very distant fungal /
-insect orthologs with low identity, fragmentary genome assemblies, missing
-N-/C-termini). The intended workflow is to **predict structure and confidence
-for each natural pair downstream**, then use the confidence scores to filter
-this pool into a clean training set.
+Layout per PPI:
 
-No `swap` or `human_swap` rows are included — those will be added in a
-later revision once a structure-aware quality filter is in place.
+```
+PPI0000001_R00 (A_human, B_human)
+PPI0000001_R01 (A_org1, B_org1)   # closest co-evolved ortholog pair
+PPI0000001_R02 (A_org2, B_org2)
+... up to R10
+PPI0000002_R00 (A_human, B_human)
+...
+```
 
-## Filters used
+Total: **196,269 rows** across **17,849 PPIs** (17,849 `human_pair`
++ 178,420 `natural`).
 
-| filter | value | description |
-|---|---|---|
-| `min_identity` | 0.20 | drop orthologs whose identity to human is below this on either side |
-| `max_gap_fraction` | 0.50 | drop orthologs whose aligned gap fraction is above this on either side |
-| `max_orthologs_per_ppi` | 300 | top-N closest orthologs by combined identity (set to a non-trivial number to keep dataset tractable while preserving diversity) |
-| same-genus exclusion | NO | unlike the v2 release, same-genus orthologs (e.g. non-reference Homo assemblies) are **kept** in this pool. Filter them downstream if needed. |
-| taxonomic class restriction | NO | mammals, birds, reptiles, fish, insects, fungi etc. are all kept |
+## Filters
+
+| filter | value |
+|---|---|
+| same-genus exclusion | drop natural rows whose genus matches the query genus (`Homo`); avoids the assembly-artefact pollution from non-reference Homo assemblies |
+| taxonomic class restriction | Mammalia, Aves, Lepidosauria, Amphibia, Actinopteri, Chondrichthyes (vertebrates only) |
+| identity floor | `identity_to_human_{a,b} >= 0.50` on both sides |
+| max orthologs per PPI | 10 (closest by combined identity) |
+
+This is a "moderately loose" filter that keeps the dataset compact (~34 MB)
+while still providing meaningful taxonomic diversity beyond just primates.
+
+A separate, larger candidate pool (~5.36M rows × 53 columns, no taxonomic
+restriction, up to 300 orthologs per PPI, identity floor 0.20) is available
+on the [`cursor/loose-natural-pairs-aaab`](https://github.com/w-jiaqi/EvoBind/tree/cursor/loose-natural-pairs-aaab)
+branch's pipeline; this curated dataset is the published default.
 
 ## Schema
 
 | column | type | notes |
 |---|---|---|
-| `row_id` | string | `PPI0000123_R042` — globally unique row identifier |
+| `row_id` | string | `PPI0000123_R02` — globally unique row identifier |
 | `ppi_id` | string | `PPI0000123` — shared by all rows of one PPI |
 | `original_protein_a`, `original_protein_b` | string | Human UniProt accessions |
 | `gene_name_a`, `gene_name_b` | string | Gene symbols (from Cong predictions) |
@@ -72,7 +77,8 @@ later revision once a structure-aware quality filter is in place.
 | `length_a`, `length_b` | int | Sequence lengths |
 | `identity_to_human_a`, `identity_to_human_b` | float | Aligned-position identity to the human query in [0, 1] |
 | `pair_type` | string | `human_pair` or `natural` |
-| `ortholog_rank` | int | 0 for the human row; 1..N for surviving orthologs sorted by combined identity, descending |
+| `ortholog_rank` | int | 0 for the human row; 1..N for surviving orthologs in identity order (after filtering) |
+| `ortholog_rank_loose` | int | The rank this row had in the larger loose pool (preserved for traceability) |
 | `interaction_label` | int | Always 1 (positives only) |
 | `taxonomic_distance` | int | 0 = same genus … 5 = different phylum (always 0 for `natural` since both sides are the same organism) |
 | `source` | string | `conglab_humanppi_omicmsa_loose` |
@@ -80,6 +86,25 @@ later revision once a structure-aware quality filter is in place.
 | `split_ppi_disjoint` | string | All rows of a PPI go to the same split |
 | `split_protein_disjoint` | string | Human proteins in test/valid don't appear in train; cross-split rows fall back to train |
 
+## Class distribution (natural rows)
+
+```
+Mammalia        177,506
+Aves                294
+Amphibia            214
+Actinopteri         195
+Lepidosauria        185
+Chondrichthyes       26
+```
+
+## Splits (90/5/5)
+
+```
+split_random:           train=176,509   valid=9,906   test=9,854
+split_ppi_disjoint:     train=176,662   valid=9,812   test=9,795
+split_protein_disjoint: train=195,268   valid=  484   test=  517
+```
+
 ## Usage
 
 ```python
@@ -93,12 +118,6 @@ for ppi_id, group in itertools.groupby(ds, key=lambda r: r["ppi_id"]):
     rows = list(group)
     human, *orthologs = rows
     # ortholog_rank == 0 is the human pair; 1..N are same-species orthologs
-
-# Iterate only natural rows for a structure-prediction filter pass
-natural = ds.filter(lambda r: r["pair_type"] == "natural")
-for r in natural:
-    seq_a, seq_b = r["sequence_a"], r["sequence_b"]
-    # ... fold and score with your structure model, keep high-confidence
 ```
 
 ## Provenance & licensing
@@ -108,16 +127,12 @@ for r in natural:
 
 ## Known limitations
 
-- Quality is intentionally mixed; many rows are expected to be filtered out
-  by downstream structure prediction. The dataset is a **candidate pool**,
-  not a curated training set.
+- Only positive interactions; no negatives.
+- No cross-species swap rows (will be added once a structure-aware quality filter is in place).
 - Sequences come from each ortholog's MSA-aligned sequence with gaps
   stripped; this means the resulting "native" sequence may be missing
   residues where the MSA had gaps relative to human. Per-row completeness
   signals are **not** stored as columns yet.
-- `max_orthologs_per_ppi = 300` is a hard cap. Very deep MSAs (e.g.
-  immunoglobulin V-regions with thousands of orthologs) get truncated to
-  the closest 300 by combined sequence identity.
-- Same-genus assemblies (e.g. non-reference `Homo` assemblies like
-  `GCA_012933715.1`) are **kept** in this pool unlike the previous v2
-  release. Filter them downstream if you want strictly cross-species data.
+- `ortholog_rank_loose` references the original ranking before filtering;
+  gaps in this sequence indicate intermediate ranks that were dropped by
+  the moderate filters.