Datasets:

wjulius
/

HumanBindome

@@ -75,30 +75,35 @@ The dataset provides predefined train, validation, and test partitions. Complexe
 | Split | Complexes | Share |
 |---|---:|---:|
-| train | 233,170 | 76.16% |
-| validation | 21,260 | 6.94% |
-| test | 19,777 | 6.46% |
-| deleaked | 31,939 | 10.43% |
 | total | 306,146 | 100.00% |
-| usable benchmark | 274,207 | 89.57% |
 The usable benchmark consists of the train, validation, and test partitions only.
 ## Split Construction
-Leakage was defined at the level of structural interface similarity, rather than by target identity, binder identity, sequence identity, or design metadata alone. This is important because designed complexes can share related binding epitopes or recurrent interface geometries even when their identifiers or sequences differ.
 The split was constructed using an interface-level procedure adapted from [PINDER](https://www.biorxiv.org/content/10.1101/2024.07.17.603980v4). Interface residues were defined as residues with at least one heavy backbone atom within 10 Å of the opposite chain. Target and binder chains were aligned all-vs-all with [Foldseek](https://www.nature.com/articles/s41587-023-01773-0). Alignments were retained for clustering only when the aligned region covered at least 50% of an interface residue set.
-The resulting interface-similarity graph was thresholded using the PINDER clustering threshold of 0.70, and weighted asynchronous label propagation was used to assign chain-interface cluster labels. Each complex was assigned to a paired-interface cluster defined by its target-side and binder-side interface cluster identifiers. Train, validation, and test assignment was performed at the paired-interface-cluster level.
 After holdout clusters were selected, PINDER’s depth-2 transitive deleaking logic was applied using the deleaking threshold of 0.55. Candidate training complexes connected to validation or test complexes through one or two interface-similarity graph edges were assigned to the `deleaked` partition and excluded from the usable benchmark.
 ## Redundancy Audit
 As an independent audit, we assessed whether ProteinMPNN sister variants derived from the same binder backbone on the same target domain were assigned to the same paired-interface cluster. ProteinMPNN grouping was not used as input to the clustering algorithm.
-Among 120,609 multi-member backbone groups, 120,082 were contained within a single paired-interface cluster, whereas 527 were split across multiple clusters. 99.56% of multi-member backbone groups were captured by the paired-interface clustering procedure.
 ## Intended Use

 | Split | Complexes | Share |
 |---|---:|---:|
+| train | 131,939 | 43.10% |
+| validation | 53,549 | 17.49% |
+| test | 52,137 | 17.03% |
+| deleaked | 68,521 | 22.38% |
 | total | 306,146 | 100.00% |
+| usable benchmark | 237,625 | 77.62% |
 The usable benchmark consists of the train, validation, and test partitions only.
 ## Split Construction
+Leakage was defined at the level of interface similarity, using both structural (Foldseek) and sequence (MMseqs) interface similarity during the subsequent deleaking procedure, rather than by target identity, binder identity or design metadata alone. This is important because designed complexes can share related binding epitopes or recurrent interface geometries even when their identifiers or sequences differ.
 The split was constructed using an interface-level procedure adapted from [PINDER](https://www.biorxiv.org/content/10.1101/2024.07.17.603980v4). Interface residues were defined as residues with at least one heavy backbone atom within 10 Å of the opposite chain. Target and binder chains were aligned all-vs-all with [Foldseek](https://www.nature.com/articles/s41587-023-01773-0). Alignments were retained for clustering only when the aligned region covered at least 50% of an interface residue set.
+The resulting Foldseek interface-similarity graph was thresholded using the PINDER clustering threshold of 0.70, and weighted asynchronous label propagation was used to assign chain-interface cluster labels. Each complex was assigned to a paired-interface cluster defined by its target-side and binder-side interface cluster identifiers. Train, validation, and test assignment was performed at the paired-interface-cluster level.
 After holdout clusters were selected, PINDER’s depth-2 transitive deleaking logic was applied using the deleaking threshold of 0.55. Candidate training complexes connected to validation or test complexes through one or two interface-similarity graph edges were assigned to the `deleaked` partition and excluded from the usable benchmark.
+After holdout clusters were selected, PINDER’s depth-2 transitive deleaking procedure was applied. Candidate training complexes connected to validation or test complexes through one or two edges in either the Foldseek interface-similarity graph or a separately constructed interface-filtered MMseqs sequence-similarity graph were assigned to the deleaked partition and excluded from the usable benchmark.
 ## Redundancy Audit
 As an independent audit, we assessed whether ProteinMPNN sister variants derived from the same binder backbone on the same target domain were assigned to the same paired-interface cluster. ProteinMPNN grouping was not used as input to the clustering algorithm.
+Among 120,609 backbones with ProteinMPNN sister variants, 120,082 were contained within a single paired-interface cluster, whereas 527 were split across multiple clusters. 99.56% of multi-member backbone groups were captured by the paired-interface clustering procedure.
+After deleaking, only 14 of the 85,037 multi-member backbone groups in the usable benchmark (0.016%) have sister variants split between `train` and the held-out (`validation` or `test`) split.
 ## Intended Use