Boltz2-Notebook: BindingDB affinity benchmark data, scripts and supplementary materials

Data, code and supplementary materials for the manuscript “Boltz2-Notebook: An Interactive Google Colab Platform for Diffusion-Based Biomolecular Structure and Binding Affinity Prediction Using the Boltz-2 Model”. This record holds an independent benchmark of Boltz-2 binding-affinity prediction: 317 protein–ligand pairs curated from BindingDB (122 unique protein sequences, 277 unique ligands, 114 PDB IDs), each predicted in triplicate with Boltz-2 (951 predictions). Against experimental pIC50 the benchmark gives Pearson r = 0.609, Spearman ρ = 0.625, MAE = 0.968 and RMSE = 1.239 (N = 317); the included scripts reproduce these numbers. File: boltz2-bindingdb-affinity-benchmark.zip (single archive; start with its README.md). It contains: the curated 317 protein–ligand pairs (CSV) and the 317 Boltz-2 input YAML files; all predictions together with the experimental values (XLSX); the raw Boltz-2 outputs of all 951 predictions: predicted structures (mmCIF), confidence and affinity JSON files, and PAE / PDE / pLDDT arrays (about 3.5 GB unpacked), with SHA-256 checksums; the raw BindingDB “curated from articles” bulk export (94,219 data rows; a snapshot, because BindingDB changes over time) from which the benchmark was derived; the curation, YAML-generation, score-collection and analysis scripts; result figures and metrics, a step-by-step protocol with the exact Boltz-2 parameters, and ten example Boltz-2 input files. Not included: the multiple sequence alignments returned by the MSA server, Boltz-2’s preprocessed tensors and the run logs. The Colab notebooks and user documentation are maintained at https://github.com/AtharvaTilewale/boltz2-notebook and https://atharvatilewale.github.io/boltz2-notebook/. BindingDB-derived data are used under BindingDB’s Creative Commons Attribution terms. Please cite BindingDB (doi:10.1093/nar/gkl999; doi:10.1093/nar/gkae1075) and Boltz-2 (doi:10.1101/2025.06.14.659707).