This document provides a detailed description of all files produced by the dataset construction pipeline.

For the full end-to-end dataset construction process, see Dataset Pipeline Overview.

Outputs are organized into two categories:

1. SRC_DATA - Intermediate and source data files used during dataset construction.
2. SUPPORTED_DATA - Final curated files required for evaluation.

Each file’s format, key fields, and intended use are described below.

Table of Contents

• 1. SRC_DATA
  • 1.1. PDB Meta Info
  • 1.2. PDB Sequences
  • 1.3. RecentPDB Chain / Interface Index
  • 1.4. Entity-Type Counts (Low Homology Subset)
  • 1.5. Train-Test High-Similarity Matches
• 2. SUPPORTED_DATA
  • 2.1. Low Homology Chain / Interface Index
  • 2.2. Low Homology Cluster File
  • 2.3. Low-Homology → Train Homology
  • 2.4. Ligand Information File
  • 2.5. Dataset Statistics
    • a. Statistic Report
    • b. PDB ID Lists
    • c. Token Distribution Figures

---

1. SRC_DATA

(Intermediate products / files generated during dataset filtering and preparation)

src_data/

---

1.1. PDB Meta Info

File: src_data/pdb_meta_info.csv Contains metadata for all mmCIF entries, including structure quality, polymer composition, and filter status.

Columns

Column	Meaning	Type	Notes
entry_id	PDB ID	str	—
exptl_methods	Experimental methods	str	Multiple values separated by ";"
classification	Classification field from the RCSB PDB record	str	-
release_date	Public release date	YYYY-MM-DD	—
resolution	Resolution	int	Missing: -1
num_tokens	Token count (AF3-style tokenization)	int	Filtered-out values set to -1
no_standard_polymer	No Protein(L), DNA, RNA in the entry	bool	Filtered-out values default to False
max_chain_copies	Max number of chains within any entity	int	Filtered-out values: -1
lacking_resolved	No chain meets the resolved-residue requirement (>4 & >30%)	bool	Filtered-out values default to False
all_chains_unk	All polymer chains contain only unknown residues	bool	Filtered-out values default to False
all_chains_break	All polymer chains contain Cα-Cα breaks	bool	Filtered-out values default to False
pass_filter	Whether the entry passes all Step 1 filters	bool	Determines candidate RecentPDB set

---

1.2. PDB Sequences

File: src_data/pdb_seqs.csv Contains all entity sequences extracted from mmCIF files.

Columns

entry_id, entity_id, release_date, entity_type, seq, seq_length

Includes all polymer and non-polymer entity sequences.

---

1.3. RecentPDB Chain / Interface Index

File: src_data/RecentPDB_chain_interface.csv Lists all chains and interfaces in the evaluation date window after Step 1 filtering.

Columns

Column	Meaning	Type	Notes
id	Unified ID for complex / chain / interface	str	pdbid/ pdbid_chain1/ pdbid_chain1_chain2 (sorted chain IDs)
type	"chain" or "interface"	str	—
entry_id	PDB ID	str	—
entity_id_1	Entity ID of chain 1	str	—
entity_type_1	Entity type of chain 1	str	Protein / DNA / RNA / Ligand / etc.
entity_type_2	Entity type of chain 2	str	Only for interfaces
chain_id_1	label_asym_id of chain 1	str	Non-Asym Unit chains may contain ".", e.g., "A.1"
chain_id_2	label_asym_id of chain 2	str	Same note as above
auth_chain_id_1	auth_asym_id of chain 1	str	—
auth_chain_id_2	auth_asym_id of chain 2	str	—
seq_length_1	Length of chain 1 sequence	int	—
seq_length_2	Length of chain 2 sequence	int	—
resolved_seq_length_1	Number of resolved residues on chain 1	int	—
resolved_seq_length_2	Number of resolved residues on chain 2	int	—

---

1.4. Entity-Type Counts (Low Homology Subset)

File: src_data/RecentPDB_low_homology_entity_types_count.csv

Contains entity-type composition for entries in the Low Homology subset.

Columns

entry_id, polypeptide(L), ligand, polyribonucleotide, polydeoxyribonucleotide, polydeoxyribonucleotide/polyribonucleotide hybrid, polypeptide(D), is_protein_monomer, is_rna_monomer

Used to determine monomer / homomer labels.

---

1.5. Train-Test High-Similarity Matches

File: src_data/test_to_train_entity_homo.parquet

This Parquet file stores the similarity relationships between test-set entities (those released between after_date and before_date) and training-set entities (those released on or before after_date).

The file is written in columnar Parquet format using pyarrow with zstd compression.

Columns

Column	Meaning	Type	Notes
query_id	Test-set entity ID	string	Format: "<test_entry_id>_<entity_id>"
db_id	Training-set entity ID	string	Format: "<train_entry_id>_<entity_id>"
similarity	Similarity score	float	Protein/DNA/RNA: sequence identity; ligand: Tanimoto similarity
aligned_res_num	Number of aligned residues	float/int	Protein/DNA/RNA: aligned residue count; ligand: NaN

Notes:

• For proteins and nucleic acids, similarity and alignment come from MMseqs2 (or exact-identity matching for very short sequences).
• For ligands, similarity comes from Morgan fingerprint Tanimoto values; aligned_res_num is not applicable.
• This Parquet file is used in Step 3 for low-homology filtering.

---

2. SUPPORTED_DATA

(Final curated dataset files for evaluation)

supported_data/

---

2.1. Low Homology Chain / Interface Index

File: supported_data/RecentPDB_low_homology.csv

Final chain and interface records after all filters (including homology, ligand QC, short-chain rules, symmetry checks).

Columns

id, type, entry_id, entity_id_1, entity_id_2, entity_type_1, entity_type_2, chain_id_1, chain_id_2, auth_chain_id_1, auth_chain_id_2, seq_length_1, seq_length_2, resolved_seq_length_1, resolved_seq_length_2, subset

Subset Labels

A chain or interface can have multiple labels, separated by ";".

Example:

[antibody-protein];[antibody_HL-protein]

Supported Labels

Chain-level

• [antibody]

  • [antibody_HL]
  • [antibody_scFv]
  • [antibody_H]
  • [antibody_L]

• [protein-monomer]

Interface-level

• [antibody-antibody]

• [antibody-protein]

  • [antibody_HL-protein]
  • [antibody_scFv-protein]
  • [antibody_H-protein]
  • [antibody_L-protein]

• [peptide-interface]

  • [peptide-peptide]
  • [peptide-protein] (polymer length ≥ 25)
  • [peptide-dna]
  • [peptide-rna]

Antibody annotation source: SAbDab summary TSV

• Hchain = NaN → [antibody_L]
• Lchain = NaN → [antibody_H]
• Both present → [antibody_HL]

---

2.2. Low Homology Cluster File

File: supported_data/RecentPDB_low_homology_cluster_info.csv

Lists cluster IDs for all polymer and ligand entities in the low-homology dataset.

Columns

entry_id, label_entity_id, cluster_id, entity_type

• Polymer clusters come from MMseqs2 clustering
• Ligand clusters use CCD code directly

---

2.3. Low-Homology → Train Homology (Parquet, 0.3 Threshold)

File: supported_data/RecentPDB_low_homology_entity_homo.parquet

This Parquet file stores sequence homology relationships between entities in the Low Homology subset and all training-set entities (release date < after_date), using a relaxed sequence identity threshold of 0.3. It is primarily used for grouping evaluation samples by homology.

Columns

Column	Meaning	Type	Notes
query_id	Low-homology entity ID	string	Format: "<lowh_entry_id>_<entity_id>"
db_id	Training-set entity ID	string	Format: "<train_entry_id>_<entity_id>"
similarity	Sequence identity	float	Smith-Waterman or MMseqs-based identity (≥ 0.3)
aligned_res_num	Number of aligned residues	float/int	Number of query residues aligned to the training sequence

Notes:

• Only polymer entities (protein, DNA, RNA) are included; ligands are excluded.
• Multiple rows per query_id are possible because a test entity may align to multiple training entities.

---

2.4. Ligand Information File

File: supported_data/RecentPDB_low_homology_lig_info.csv This file lists the ligand chains associated with structures in the low-homology dataset.

Columns:

• entry_id - The PDB entry identifier.
• label_asym_id - The chain identifier of the ligand within the entry.

In the evaluation pipeline, this file is also used to identify ligand entities that require pocket-aligned RMSD computation and PoseBusters ligand validity checks. Only ligands listed in this file will be included in these ligand-specific quality assessments during benchmarking.

---

2.5. Dataset Statistics

Directory: supported_data/stat_data/

Contains all statistical summaries and plots.

a. Statistic Report

File: stat.txt

Includes:

• Step 1 filtering counts
• Complex / Chain / Interface counts
• Per-eval-type and per-subset statistics
• Cluster counts

b. PDB ID Lists

Directory: supported_data/stat_data/pdb_ids/

Contains:

• all - all low-homology PDB IDs
• lowh_polymer_only - IDs containing only polymer low-homology data
• Per-subset and per-eval-type lists for controlled evaluation sampling

c. Token Distribution Figures

Directory: supported_data/stat_data/figs/

Contains token-count histograms for:

• Complex-level tokens
• Chain/interface-level tokens
• Per-eval-type token distributions