NeuralBind — Boltz-1x predicted protein–ligand complexes with ChEMBL affinities

Overview

NeuralBind pairs a predicted three-dimensional binding pose with a measured binding affinity for every entry. Protein–ligand complexes were predicted with Boltz-1x from ChEMBL target sequences and ligand SMILES, and each prediction is retained alongside the model’s own confidence estimates.

Affinity labels are pChEMBL values drawn from curated Ki and Kd measurements, spanning 0.85–11.40 (mean 6.99). The dataset is intended to complement existing resources for training and evaluating machine-learning scoring functions for binding affinity prediction, particularly where experimentally determined structures are unavailable.

Across the collection the median Boltz-1x confidence score is 0.77, with 88% of complexes scoring above 0.70 and 89% passing PoseBusters validity checks. Per-complex metrics are distributed with the dataset so that users can apply their own confidence thresholds.

What each entry contains

Every complex is stored in its own directory named Candidate_<index>_<target>_<molecule>_<seq> and includes:

Input FASTA — the target protein sequence and ligand SMILES submitted to Boltz-1x.
Predicted structure — the Boltz-1x complex model for the protein and its bound ligand.
Multiple sequence alignment — the MSA generated for the target sequence.
Confidence metrics — the overall Boltz-1x confidence score together with pLDDT, PAE, PDE, pTM and ipTM terms, reported globally and for the binding pocket, shell and ligand.

A companion table links every complex back to its ChEMBL provenance and affinity label:

Field	Description
`system_id`	Unique identifier of the complex; matches the directory name
`smiles`	Ligand structure in SMILES notation
`molecule_chembl_id`	ChEMBL identifier of the ligand
`target_chembl_id`	ChEMBL identifier of the protein target
`target_pref_name`	Preferred name of the target, e.g. Dopamine D1 receptor
`uniprot_id`	UniProt accession of the target
`fasta_seqs`	Amino acid sequence used for structure prediction
`assay_chembl_id`	ChEMBL identifier of the source assay
`standard_type`	Measurement type: `Ki` (164,162) or `Kd` (13,489)
`pchembl_value`	Binding affinity on a negative logarithmic scale

Download

Download neuralbind.tar.gz

gzip-compressed tar archive · approximately 440 GB — please ensure sufficient disk space and a stable connection before downloading.

wget https://neuralbind-dataset.bioch.ox.ac.uk/downloads/neuralbind.tar.gz
tar -xzf neuralbind.tar.gz

Contact

For questions or feedback, please contact Wei-Tse Hsu at wei-tse.hsu@bioch.ox.ac.uk.