Thal-Kak: Boltz-2, Chai-1, ESMFold and Protenix v1/v2 predictions on the FoldBench benchmark

Structure predictions from five biomolecular structure predictors, generated for FoldBench benchmarking with Thal-Kak. Each target carries 200 predictions — 40 seeds × 5 samples. 47.33 GB across five archives. Model Targets Archive Size Boltz-2 654 boltz2_foldbench.tar.gz 8.67 GB Chai-1 662 chai1_foldbench.tar.gz 9.83 GB ESMFold 602 esmfold2_foldbench.tar.gz 6.55 GB Protenix v1 679 protenix_v1_base_foldbench.tar.gz 11.14 GB Protenix v2 679 protenix_v2_foldbench.tar.gz 11.13 GB Target counts differ between models because not every target completed on every predictor; some exceeded available GPU memory. Targets are PDB biological assemblies, named <pdbid>-assembly1, drawn from FoldBench dataset. Archive layout. Each archive expands to one directory per target: <model>/<model>_results_<pdbid>-assembly1_foldbench/common/ containing <target>_seed_<N>_sample_<M>.pdb — 200 predictions per target, from 40 seeds (seed_1–seed_40) × 5 samples (sample_0–sample_4), with per-residue pLDDT in the B-factor column (0–100) — plus <target>_results_summary.csv. Some targets additionally appear with a _chain_ligand suffix, denoting the chain-and-ligand input variant. Confidence scores. <target>_results_summary.csv holds one row per prediction: target, option, model, seed-sample, ranking_score, mean_plddt, ptm, iptm. Licensing. This record is released under the Apache License 2.0. Citation. Please cite the Thal-Kak paper which will be available soon, together with the papers for whichever predictors you use — listed under References below. Code. https://github.com/CSSB-SNU/Thal-Kak