#!/usr/bin/env bash # Download the public assets used by CausalBind into the repository root. # All datasets are third-party resources released by their original authors # (see DATA.md for sources and licenses); this script only fetches them. # # Usage: bash scripts/download_data.sh [all|pretrain|train|test] set -euo pipefail part="${1:-all}" REPO_DIR="$(cd "$(dirname "$0")/.." && pwd)" cd "${REPO_DIR}" fetch() { # fetch [ -s "$2" ] && { echo " exists: $2"; return; } echo " downloading: $2" wget -nv -O "$2" "$1" } download_pretrain() { echo "[1/3] Pretrained backbones -> pretrain/" mkdir -p pretrain # Uni-Mol molecule and pocket encoders (Uni-Mol v0.1 release) fetch https://github.com/deepmodeling/Uni-Mol/releases/download/v0.1/mol_pre_no_h_220816.pt pretrain/mol_pre_no_h_220816.pt fetch https://github.com/deepmodeling/Uni-Mol/releases/download/v0.1/pocket_pre_220816.pt pretrain/pocket_pre_220816.pt # ESM-2 (35M) protein language model from Hugging Face python - <<'EOF' from huggingface_hub import snapshot_download snapshot_download("facebook/esm2_t12_35M_UR50D", local_dir="pretrain/esm2_t12_35M_UR50D") EOF } download_train() { echo "[2/3] LigUnity training data (figshare 10.6084/m9.figshare.27966819) -> data/" mkdir -p data fetch https://ndownloader.figshare.com/files/50992173 data/valid_label_seq.json fetch https://ndownloader.figshare.com/files/50992176 data/valid_lig.lmdb fetch https://ndownloader.figshare.com/files/50992182 data/valid_prot.lmdb fetch https://ndownloader.figshare.com/files/50992179 data/train_label.zip fetch https://ndownloader.figshare.com/files/50992185 data/pocket_name2idx_train_blend.json fetch https://ndownloader.figshare.com/files/50992191 data/mol_smi2idx_train_blend.json fetch https://ndownloader.figshare.com/files/50992194 data/train_prot_all_blend.lmdb fetch https://ndownloader.figshare.com/files/50992728 data/train_lig_all_blend.lmdb fetch https://ndownloader.figshare.com/files/54239555 data/uniport80.clstr fetch https://ndownloader.figshare.com/files/54239558 data/uniport40.clstr (cd data && unzip -n -q train_label.zip) # train_label_blend_seq_full.json, train_label_pdbbind_seq.json } download_test() { echo "[3/3] Test benchmarks -> test_datasets/" mkdir -p test_datasets # Small benchmark index files (CASF validation, DUD-E / LIT-PCBA / FEP target lists) # from the LigUnity repository. for f in casf.lmdb casf_label_seq.json dude.json PCBA.json FEP.json dekois.json \ fep_assay_ids.json fep_repeat_ligands_can.json fep_similar_ligands_0d5.json; do fetch "https://raw.githubusercontent.com/IDEA-XL/LigUnity/main/test_datasets/${f}" "test_datasets/${f}" done # The FEP-overlap filters are also read from the training directory. mkdir -p data ln -sf ../test_datasets/fep_assay_ids.json data/fep_assays.json ln -sf ../test_datasets/fep_repeat_ligands_can.json data/fep_repeat_ligands_can.json ln -sf ../test_datasets/fep_similar_ligands_0d5.json data/fep_similar_ligands_0d5.json # DEKOIS 2.0 (LigUnity release, figshare 10.6084/m9.figshare.27967422), used for OOD evaluation fetch https://ndownloader.figshare.com/files/50994021 test_datasets/DEKOIS_2.0x.zip (cd test_datasets && unzip -n -q DEKOIS_2.0x.zip) # -> DEKOIS_2.0x// cat <<'EOF' DUD-E and LIT-PCBA are distributed by the LigUnity authors through Google Drive: https://drive.google.com/drive/folders/1zW1MGpgunynFxTKXC2Q4RgWxZmg6CInV Download dude.zip and pcba.zip into test_datasets/ (e.g. with `gdown --folder`), then run: cd test_datasets unzip -q pcba.zip # -> lit_pcba// unzip -q dude.zip # -> data/protein/DUD-E/raw/all// mv data/protein/DUD-E/raw/all DUD-E && rm -r data EOF } case "${part}" in pretrain) download_pretrain ;; train) download_train ;; test) download_test ;; all) download_pretrain; download_train; download_test ;; *) echo "Usage: bash scripts/download_data.sh [all|pretrain|train|test]"; exit 1 ;; esac