Datasets:
File size: 7,163 Bytes
84a242a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 | # Legacy/internal README below
# **MolAS: Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking**
[](https://link.springer.com/article/10.1186/s13321-026-01168-8)
[](https://arxiv.org/abs/2512.02328)
[](https://zenodo.org/records/17760688)
[](LICENSE)
MolAS is a learning-based framework that predicts, for each protein–ligand pair, which docking or pose-generation method is most likely to succeed.
This repository contains the full training, evaluation, and analysis code used in the accompanying paper *Molecular Embedding–Based Algorithm Selection in Protein–Ligand Docking*.
---
## **Repository Layout**
```
molas/
│
├── main.py # Training / testing entry point (PyTorch Lightning)
├── main.sh # 5-fold and multi-GPU script (torchrun presets)
│
├── data/
│ ├── dataset.py # Dataset construction from CSVs and precomputed graphs
│ ├── data_interface.py. # LightningDataModule with K-Fold / split support
│ └── ...
│
├── model/ # MolAS architectures and loss functions
│ └── ...
│
├── embedding_generation/ # Scripts for generating ESM and ChemBERTa embeddings (for completeness)
│
├── stats_test.py # Fold-concatenated reporting and paired statistics
├── environment.yaml # Python dependencies
└── README.md
```
---
## **Installation**
Create a new environment and install PyTorch **separately**.
```bash
# 1) Create env
conda env create -f environment.yaml
conda activate as_dock
# 2) Install PyTorch (choose a CUDA build)
pip install --index-url https://download.pytorch.org/whl/cu124 \
torch torchvision torchaudio
# 3) Install remaining dependencies
pip install -r requirements.txt
```
---
## **Data Preparation**
Download the processed MolAS dataset and trained checkpoints from:
**Zenodo: https://zenodo.org/records/17760688**
Then update the `data_path` variables in `data/dataset.py`.
### **Expected directory structure**
```
{data_path}/
├── protein_graphs_esmc_600m/
│ └── pyg_graph_{pdbid}_{ligandid}_esmc_600m.pt
│
├── ligand_chemberta.pt
│
├── {benchmark}_pb_ratio.csv # PoseBusters validity (per-algorithm)
└── {benchmark}_rmsd.csv # RMSD targets (per-algorithm)
```
### **Active-learning matrices**
Active-learning runs use a separate reconstructed score matrix for each outer
cross-validation fold and AL seed:
```
AL/PORT-AS_cost_aware/
{method}/
fold{0..4}/
seed{al_seed}/
strict_score_matrices/
strict_score_{percentage:03d}pct.csv
```
For outer fold `f`, the matrix contains only rows with `test_fold != f`.
Acquisition and reconstruction are performed separately on each fold's training
pool. A globally reconstructed matrix cropped to the training rows afterward is
not equivalent and is not used by the current pipeline. The matching
`split_exports/...csv` file defines both row membership and row order.
The split lifecycle is:
```
outer-fold training rows
→ fold-specific active learning and reconstruction
→ internal MolAS train/validation split
→ evaluation on the untouched outer test fold using ground-truth labels
```
### Notes on checkpoints
Log folder of checkpoints (named `lightning_logs`) should be placed inside the repository root, **parallel** to this README.md.
---
## **Quick Start**
### **5-Fold and Multi-GPU Training**
Use the preset runner:
```bash
bash main.sh
```
Configure:
* `benchmark` (e.g., `posex_self_docking`, `posex_cross_docking`, `astex_posex`, `moad`, `posebusters`)
* `relaxation` (`both | true | false`)
* `incl_columns` (algorithm list; ensure count matches `--num_classes`)
* `devices` (e.g., `"0,1,2,3"` for torchrun)
* `test=1` to run test-only mode and export reports
* `version=0` to set the initial version of the 5-version bundle for testing
#### **Version Directory**
To reproduce the results using the Zenodo-released checkpoints, set `version` according to the target benchmark and post-processing mode. Each entry refers to a contiguous 5-checkpoint bundle (`version`–`version+4`).
| Benchmark | Post-process | `version` | Version bundle |
| ------------- | ------------------- | --------- | -------------- |
| MOAD-curated | — | 5 | 5–9 |
| PoseX + Astex | Mixed | 219 | 219–223 |
| PoseX-SD | Mixed | 5 | 5–9 |
| PoseX-CD | Mixed | 5 | 5–9 |
| PoseBusters | Mixed | 20 | 20–24 |
| PoseX + Astex | No | 209 | 209–213 |
| PoseX + Astex | Relaxation | 214 | 214–218 |
| PoseX-SD | No | 10 | 10–14 |
| PoseX-SD | Relaxation | 30 | 30–34 |
| PoseX-CD | No | 30 | 30–34 |
| PoseX-CD | Relaxation | 25 | 25–29 |
| PoseBusters | No | 10 | 10–14 |
| PoseBusters | Energy minimisation | 15 | 15–19 |
### **Single-GPU Example**
```bash
python main.py \
--benchmark posex_self_docking \
--batch_size 16 \
--lr 1e-3 \
--max_epochs 100 \
--devices 0 \
--model MolAS \
--num_node_features 1152 \
--num_classes 24 \
--incl_columns surfdock unimol diffdock diffdock_l gnina autodock glide alphafold3 moe dynamicbind rfaa fabind equibind boltz1x tankbind neuralplexer chai deepdock boltz ifd protenix diffdock_pocket DS interformer
```
### **Testing With a Checkpoint**
```bash
python main.py \
--benchmark posex_self_docking \
--devices 0 \
--model MolAS \
--num_node_features 1152 \
--num_classes 24 \
--incl_columns <algorithm list> \
--test \
--ckpt_path path/to/checkpoint.ckpt
```
---
## **Important Arguments**
* `--model`
`MolAS` (default) or multi-GNN variants (e.g., `MolAS_GCN_GAT_GINE`, `MolAS_EGNN_GAT_GINE`, `MolASGT`).
Multi-GNN variants require node features and adjacency consistent with provided graphs.
* `--incl_columns`
Space-separated algorithm names. Must match `--num_classes`.
* `--use_ndcg_loss`, `--ndcg_loss_weight`, `--ndcg_k`
Enable LambdaNDCG ranking loss.
* `--use_logistic_loss`, `--logistic_loss_weight`
Pairwise logistic ranking objective.
* `--k_folds`, `--fold_num`
Enable K-fold cross-validation.
* `--log_root`, `--log_name`, `--log_version`
Control Lightning logging layout.
---
## **Acknowledgements**
MolAS uses embeddings from **ESM-C** and **ChemBERTa**, and builds on PyTorch, PyTorch Lightning, and PyTorch Geometric.
|