THU-ATOM/DrugCLIP_data

---

license: cc-by-4.0

🧬 DrugCLIP data repository

This repository hosts benchmark datasets, pre-computed molecular embeddings, pretrained model weights, and supporting files used in the DrugCLIP project. It also includes data and models used for wet lab validation experiments.

---

📁 Repository Contents

1. DUD-E.zip

• Full dataset for the DUD-E benchmark.
• Includes ligand and target files for all targets.

---

2. LIT-PCBA.zip

• Full dataset for the LIT-PCBA benchmark.
• Includes ligand and target files for all targets.

---

3. encoded_mol_embs.zip

• Pre-encoded molecular embeddings from the ChemDiv compound library.
• Each .pkl file contains:
  • name_list: [hitid, SMILES]
  • embedding_list: list of 128-dimensional vectors
• Versions included:
  • 8-fold version of the full ChemDiv library
  • 6-fold version of the full ChemDiv library
  • 6-fold version of a filtered ChemDiv library

---

4. benchmark_weights.zip

Contains pretrained model weights for benchmark experiments on the DUD-E and LIT-PCBA datasets using various ligand and target filtering strategies.

🔬 DUD-E: Ligand Filtering Strategies

Filename	Description
dude_ecfp_90.pt	Trained by removing ligands with ECFP4 similarity > 0.9.
dude_ecfp_60.pt	Trained by removing ligands with ECFP4 similarity > 0.6.
dude_ecfp_30.pt	Trained by removing ligands with ECFP4 similarity > 0.3.
dude_scaffold.pt	Trained by removing ligands sharing scaffolds with test set.

🧬 DUD-E: Target Filtering Strategies

Filename	Description
dude_identity_90.pt	Removed targets with MMseqs2 identity > 0.9.
dude_identity_60.pt	Removed targets with MMseqs2 identity > 0.6.
dude_identity_30.pt	Removed targets with MMseqs2 identity > 0.3.
dude_identity_0.pt	Removed targets based on HMMER sequence identity.

🧪 LIT-PCBA: Target Filtering Strategy

Filename	Description
litpcba_identity_90.pt	Removed targets with MMseqs2 identity > 0.9.

---

5. model_weights.zip

Contains model weights trained specifically for wet lab experiments. These models were trained using:

• 6-fold data splits
• 8-fold data splits

Used to predict compounds validated in real-world assays for the following targets:

• 5HT2a
• NET
• Trip12

---

6. WetLab_PDBs_and_LMDBs

Target data used for wet lab validation experiments:

• LMDB files: For DrugCLIP screening

Includes data for:

• 5HT2a
• NET
• Trip12

---

7. benchmark_throughput

Files for reproducing throughput benchmark results.