ajiang2025/cath-4.3-backbone

CATH 4.3 protein backbones, cropped to 128 residues

Protein backbone coordinates prepared for generative modelling (CIS 6270 project, modality 2). Every chain is cropped to a fixed 128-residue window, unresolved residues carry an explicit mask, and coordinates are centred on the resolved alpha carbons.

from datasets import load_dataset
ds = load_dataset("ajiang2025/cath-4.3-backbone")

Split	Chains	Purpose
train	16,691	fit the model
validation	1,528	tune, select, early-stop
test	1,880	report once, at the end

What is in a row

Column	Type	Meaning
name	string	source chain, e.g. 12as.A
cath	list of string	CATH superfamily codes, e.g. ["3.30.930"]
seq	string	amino-acid sequence of the cropped window
coords	[L, 4, 3] float	backbone coordinates in Angstrom, atom order N, CA, C, O
mask	[L] bool	True where the residue is resolved
length	int	residues in the window (<= 128)
n_resolved	int	number of True entries in mask

import numpy as np
r = ds["train"][0]
coords = np.array(r["coords"])      # [L, 4, 3]
mask   = np.array(r["mask"])        # [L]
ca     = coords[mask][:, 1]         # resolved alpha carbons

Scope: chains, not domains

This holds roughly 20,000 chains, not CATH 4.3's ~151,000 classified domains. The splits come from the ESM inverse-folding data split (Ingraham et al., 2019), where chains are clustered at 40% sequence identity for redundancy reduction. That is the standard benchmark partition used by ProteinMPNN, GVP and ESM-IF, so results here are comparable to those papers. The full domain set is not on the Hub and would have to come from CATH's own FTP.

Processing

Derived from cctien/protein_backbone_cath_4.3 (CC-BY-4.0). Three changes, each for a concrete reason:

1. Unresolved residues. The source stores them as [None, None, None]. The arrays still line up with the sequence, so lengths look correct while the values are null — feeding them forward produces NaN losses with no error. Here they are zeros plus an explicit boolean mask. A residue is marked unresolved if any of its four backbone atoms is missing.
2. Cropping to 128 residues. Chains run to several hundred residues, and pairwise operations scale as N^2. Windows are chosen at random, preferring one that starts on a resolved residue. Set --crop 0 in the preparation script to keep full chains.
3. Centring. Each window is centred on the mean of its resolved CA atoms only. Including the zeroed-out unresolved positions in that mean would shift the chain by an amount depending on how much structure happens to be missing.

Chains with fewer than 32 resolved residues after cropping are dropped (8 of 20,110).

Resolved fraction after processing: train 96.4%, validation 96.4%, test 95.2%.

Sanity check

Consecutive alpha carbons in a real backbone sit about 3.8 A apart. In this dataset the median is 3.80 A, which is the quickest way to confirm a crop or mask has not scrambled the geometry:

import numpy as np
r = ds["validation"][0]
ca = np.array(r["coords"])[np.array(r["mask"])][:, 1]
print(np.median(np.linalg.norm(np.diff(ca, axis=0), axis=-1)))   # ~3.80

Attribution

• CATH: Sillitoe et al., CATH: increased structural coverage of functional space, Nucleic Acids Research 2021. https://www.cathdb.info/
• Split: Ingraham, Garg, Barzilay & Jaakkola, Generative Models for Graph-Based Protein Design, NeurIPS 2019; as distributed with ESM inverse folding.
• Immediate source: cctien/protein_backbone_cath_4.3, CC-BY-4.0.

Redistributed under CC-BY-4.0.